From 2d049ba1d462073284497bf93f5a254da9074ba4 Mon Sep 17 00:00:00 2001 From: YeonwooSung Date: Mon, 28 Sep 2026 15:15:47 +0900 Subject: [PATCH 1/7] feat(intrinsics): admit bf16x2 to FP8x2 pack conversions Admit the four selectionless bf16x2 pack records as inline PTX. They require PTX ISA 9.1 on sm_100f, sm_110f, and sm_120f. Runtime validation stays unexecuted because ptxas and libNVVM are not available here. Signed-off-by: YeonwooSung --- .../cuda-device/src/generated/async_copy.rs | 2 +- crates/cuda-device/src/generated/atomic.rs | 2 +- crates/cuda-device/src/generated/bf16.rs | 2 +- crates/cuda-device/src/generated/bf16x2.rs | 2 +- crates/cuda-device/src/generated/clc.rs | 2 +- .../src/generated/cluster_barrier.rs | 2 +- .../src/generated/cluster_memory.rs | 2 +- .../cuda-device/src/generated/cluster_sreg.rs | 2 +- crates/cuda-device/src/generated/convert.rs | 38 +- .../src/generated/counted_barrier.rs | 2 +- .../src/generated/debug_control.rs | 2 +- .../cuda-device/src/generated/debug_sreg.rs | 2 +- crates/cuda-device/src/generated/dotprod.rs | 2 +- crates/cuda-device/src/generated/f16.rs | 2 +- crates/cuda-device/src/generated/f16x2.rs | 2 +- crates/cuda-device/src/generated/f32x2.rs | 2 +- crates/cuda-device/src/generated/fence.rs | 2 +- crates/cuda-device/src/generated/float.rs | 2 +- .../src/generated/grid_dependency.rs | 2 +- crates/cuda-device/src/generated/grid_sreg.rs | 2 +- crates/cuda-device/src/generated/i16x2.rs | 2 +- crates/cuda-device/src/generated/int.rs | 2 +- crates/cuda-device/src/generated/ldmatrix.rs | 2 +- .../src/generated/mbarrier_basic.rs | 2 +- .../src/generated/mbarrier_extended.rs | 2 +- crates/cuda-device/src/generated/movmatrix.rs | 2 +- crates/cuda-device/src/generated/prmt.rs | 2 +- .../src/generated/register_control.rs | 2 +- .../cuda-device/src/generated/register_mma.rs | 2 +- .../cuda-device/src/generated/shared_sreg.rs | 2 +- .../cuda-device/src/generated/sparse_mma.rs | 2 +- crates/cuda-device/src/generated/sreg.rs | 2 +- crates/cuda-device/src/generated/stmatrix.rs | 2 +- crates/cuda-device/src/generated/tcgen05.rs | 2 +- crates/cuda-device/src/generated/tma.rs | 2 +- crates/cuda-device/src/generated/warp_sreg.rs | 2 +- .../src/generated/wgmma_control.rs | 2 +- .../src/model/contracts/scalar.rs | 6 +- .../cuda-intrinsics-gen/src/model/overlay.rs | 16 + .../cuda-intrinsics-gen/src/render/compat.rs | 3 +- .../src/render/dialect/scalar_packed.rs | 3 +- .../src/render/families/packed.rs | 33 +- .../src/render/lowering.rs | 3 +- .../src/render/tests/mma.rs | 2 +- .../src/render/tests/scalar_misc.rs | 2 +- .../src/resolve/evidence.rs | 12 + .../src/resolve/families/packed_conversion.rs | 189 +- .../src/resolve/overlay.rs | 17 +- .../src/resolve/tests/fixtures.rs | 23 +- .../src/resolve/tests/misc_families.rs | 2 + .../src/resolve/tests/overlay_schema.rs | 10 +- .../src/resolve/tests/scalar_packed.rs | 46 +- .../src/resolve/tests/sreg_threadfence.rs | 1 + .../src/resolve/tests/tcgen05.rs | 1 + .../src/resolve/tests/tma_clc_cluster.rs | 3 + .../cuda-intrinsics/src/generated/abi_v1.rs | 50 +- crates/cuda-intrinsics/src/generated/mod.rs | 30 +- .../active_mask.rs | 2 +- .../src/generated_intrinsic_targets/clc.rs | 2 +- .../cluster_barrier.rs | 2 +- .../cluster_memory.rs | 2 +- .../generated_intrinsic_targets/cp_async.rs | 2 +- .../debug_control.rs | 2 +- .../generated_intrinsic_targets/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/elect.rs | 2 +- .../execution_control.rs | 2 +- .../extended_minmax.rs | 2 +- .../integer_minmax.rs | 2 +- .../generated_intrinsic_targets/ldmatrix.rs | 2 +- .../mbarrier_basic.rs | 2 +- .../mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/mod.rs | 2 +- .../generated_intrinsic_targets/movmatrix.rs | 2 +- .../generated_intrinsic_targets/packed_alu.rs | 2 +- .../packed_atomic.rs | 2 +- .../packed_conversion.rs | 182 +- .../src/generated_intrinsic_targets/prmt.rs | 2 +- .../src/generated_intrinsic_targets/redux.rs | 2 +- .../register_mma.rs | 2 +- .../scalar_arithmetic.rs | 2 +- .../scalar_conversion.rs | 2 +- .../scalar_math.rs | 2 +- .../generated_intrinsic_targets/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/sreg.rs | 2 +- .../generated_intrinsic_targets/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/sync.rs | 2 +- .../generated_intrinsic_targets/tcgen05_cp.rs | 2 +- .../generated_intrinsic_targets/tcgen05_ld.rs | 2 +- .../tcgen05_mma.rs | 2 +- .../tcgen05_mma_ws.rs | 2 +- .../tcgen05_other.rs | 2 +- .../generated_intrinsic_targets/tcgen05_st.rs | 2 +- .../tests/active_mask.rs | 2 +- .../generated_intrinsic_targets/tests/clc.rs | 2 +- .../tests/cluster_barrier.rs | 2 +- .../tests/cluster_memory.rs | 2 +- .../tests/cp_async.rs | 2 +- .../tests/debug_control.rs | 2 +- .../tests/dotprod.rs | 2 +- .../tests/elect.rs | 2 +- .../tests/execution_control.rs | 2 +- .../tests/extended_minmax.rs | 2 +- .../tests/integer_minmax.rs | 2 +- .../tests/ldmatrix.rs | 2 +- .../tests/mbarrier_basic.rs | 2 +- .../tests/mbarrier_extended.rs | 2 +- .../generated_intrinsic_targets/tests/mod.rs | 2 +- .../tests/movmatrix.rs | 2 +- .../tests/packed_alu.rs | 2 +- .../tests/packed_atomic.rs | 2 +- .../tests/packed_conversion.rs | 224 ++- .../generated_intrinsic_targets/tests/prmt.rs | 2 +- .../tests/redux.rs | 2 +- .../tests/register_mma_float.rs | 2 +- .../tests/register_mma_s32.rs | 2 +- .../tests/scalar_arithmetic.rs | 2 +- .../tests/scalar_conversion.rs | 2 +- .../tests/scalar_math.rs | 2 +- .../tests/sparse_mma.rs | 2 +- .../generated_intrinsic_targets/tests/sreg.rs | 2 +- .../tests/stmatrix.rs | 2 +- .../generated_intrinsic_targets/tests/sync.rs | 2 +- .../tests/tcgen05_cp.rs | 2 +- .../tests/tcgen05_ld.rs | 2 +- .../tests/tcgen05_mma.rs | 2 +- .../tests/tcgen05_mma_ws.rs | 2 +- .../tests/tcgen05_other.rs | 2 +- .../tests/tcgen05_st.rs | 2 +- .../generated_intrinsic_targets/tests/tma.rs | 2 +- .../generated_intrinsic_targets/tests/vote.rs | 2 +- .../tests/warp_barrier.rs | 2 +- .../tests/warp_match.rs | 2 +- .../tests/warp_shuffle.rs | 2 +- .../tests/wgmma_control.rs | 2 +- .../src/generated_intrinsic_targets/tma.rs | 2 +- .../src/generated_intrinsic_targets/vote.rs | 2 +- .../warp_barrier.rs | 2 +- .../generated_intrinsic_targets/warp_match.rs | 2 +- .../warp_shuffle.rs | 2 +- .../wgmma_control.rs | 2 +- crates/cuda-target-spec/src/lib.rs | 6 +- .../src/ops/generated/active_mask.rs | 2 +- crates/dialect-nvvm/src/ops/generated/clc.rs | 2 +- .../src/ops/generated/cluster_barrier.rs | 2 +- .../src/ops/generated/cluster_memory.rs | 2 +- .../src/ops/generated/cp_async.rs | 2 +- .../src/ops/generated/debug_control.rs | 2 +- .../dialect-nvvm/src/ops/generated/dotprod.rs | 2 +- .../dialect-nvvm/src/ops/generated/elect.rs | 2 +- .../src/ops/generated/execution_control.rs | 2 +- .../src/ops/generated/extended_minmax.rs | 2 +- .../src/ops/generated/integer_minmax.rs | 2 +- .../src/ops/generated/ldmatrix.rs | 2 +- .../src/ops/generated/mbarrier_basic.rs | 2 +- .../src/ops/generated/mbarrier_extended.rs | 2 +- crates/dialect-nvvm/src/ops/generated/mod.rs | 2 +- .../src/ops/generated/movmatrix.rs | 2 +- .../src/ops/generated/packed_alu.rs | 2 +- .../src/ops/generated/packed_atomic.rs | 2 +- .../src/ops/generated/packed_conversion.rs | 202 +- crates/dialect-nvvm/src/ops/generated/prmt.rs | 2 +- .../dialect-nvvm/src/ops/generated/redux.rs | 2 +- .../src/ops/generated/register_mma.rs | 2 +- .../src/ops/generated/scalar_arithmetic.rs | 2 +- .../src/ops/generated/scalar_conversion.rs | 2 +- .../src/ops/generated/scalar_math.rs | 2 +- .../src/ops/generated/sparse_mma.rs | 2 +- crates/dialect-nvvm/src/ops/generated/sreg.rs | 2 +- .../src/ops/generated/stmatrix.rs | 2 +- crates/dialect-nvvm/src/ops/generated/sync.rs | 2 +- .../dialect-nvvm/src/ops/generated/tcgen05.rs | 2 +- crates/dialect-nvvm/src/ops/generated/tma.rs | 2 +- crates/dialect-nvvm/src/ops/generated/vote.rs | 2 +- .../src/ops/generated/warp_barrier.rs | 2 +- .../src/ops/generated/warp_match.rs | 2 +- .../src/ops/generated/warp_shuffle.rs | 2 +- .../src/ops/generated/wgmma_control.rs | 2 +- .../intrinsics/generated/active_mask.rs | 2 +- .../terminator/intrinsics/generated/clc.rs | 2 +- .../intrinsics/generated/cluster_barrier.rs | 2 +- .../intrinsics/generated/cluster_memory.rs | 2 +- .../intrinsics/generated/cp_async.rs | 2 +- .../intrinsics/generated/debug_control.rs | 2 +- .../intrinsics/generated/dotprod.rs | 2 +- .../terminator/intrinsics/generated/elect.rs | 2 +- .../intrinsics/generated/execution_control.rs | 2 +- .../intrinsics/generated/extended_minmax.rs | 2 +- .../intrinsics/generated/helpers.rs | 2 +- .../intrinsics/generated/integer_minmax.rs | 2 +- .../intrinsics/generated/ldmatrix.rs | 2 +- .../intrinsics/generated/mbarrier_basic.rs | 2 +- .../intrinsics/generated/mbarrier_extended.rs | 2 +- .../terminator/intrinsics/generated/mod.rs | 22 +- .../intrinsics/generated/movmatrix.rs | 2 +- .../intrinsics/generated/packed_alu.rs | 2 +- .../intrinsics/generated/packed_atomic.rs | 2 +- .../intrinsics/generated/packed_conversion.rs | 164 +- .../terminator/intrinsics/generated/prmt.rs | 2 +- .../terminator/intrinsics/generated/redux.rs | 2 +- .../intrinsics/generated/register_mma.rs | 2 +- .../intrinsics/generated/scalar_arithmetic.rs | 2 +- .../intrinsics/generated/scalar_conversion.rs | 2 +- .../intrinsics/generated/scalar_math.rs | 2 +- .../intrinsics/generated/sparse_mma.rs | 2 +- .../terminator/intrinsics/generated/sreg.rs | 2 +- .../intrinsics/generated/stmatrix.rs | 2 +- .../terminator/intrinsics/generated/sync.rs | 2 +- .../intrinsics/generated/tcgen05_cp.rs | 2 +- .../intrinsics/generated/tcgen05_ld.rs | 2 +- .../intrinsics/generated/tcgen05_mma.rs | 2 +- .../intrinsics/generated/tcgen05_other.rs | 2 +- .../intrinsics/generated/tcgen05_st.rs | 2 +- .../intrinsics/generated/tests/active_mask.rs | 2 +- .../intrinsics/generated/tests/clc.rs | 2 +- .../generated/tests/cluster_barrier.rs | 2 +- .../generated/tests/cluster_memory.rs | 2 +- .../intrinsics/generated/tests/cp_async.rs | 2 +- .../generated/tests/debug_control.rs | 2 +- .../intrinsics/generated/tests/dotprod.rs | 2 +- .../intrinsics/generated/tests/elect.rs | 2 +- .../generated/tests/execution_control.rs | 2 +- .../generated/tests/extended_minmax.rs | 2 +- .../generated/tests/integer_minmax.rs | 2 +- .../intrinsics/generated/tests/ldmatrix.rs | 2 +- .../generated/tests/mbarrier_basic.rs | 2 +- .../generated/tests/mbarrier_extended.rs | 2 +- .../intrinsics/generated/tests/mod.rs | 2 +- .../intrinsics/generated/tests/movmatrix.rs | 2 +- .../intrinsics/generated/tests/packed_alu.rs | 2 +- .../generated/tests/packed_atomic.rs | 2 +- .../generated/tests/packed_conversion.rs | 146 +- .../intrinsics/generated/tests/prmt.rs | 2 +- .../intrinsics/generated/tests/redux.rs | 2 +- .../generated/tests/register_mma.rs | 2 +- .../generated/tests/scalar_arithmetic.rs | 2 +- .../generated/tests/scalar_conversion.rs | 2 +- .../intrinsics/generated/tests/scalar_math.rs | 2 +- .../intrinsics/generated/tests/sparse_mma.rs | 2 +- .../intrinsics/generated/tests/sreg.rs | 2 +- .../intrinsics/generated/tests/stmatrix.rs | 2 +- .../intrinsics/generated/tests/sync.rs | 2 +- .../intrinsics/generated/tests/tcgen05.rs | 2 +- .../intrinsics/generated/tests/tma.rs | 2 +- .../intrinsics/generated/tests/vote.rs | 2 +- .../generated/tests/warp_barrier.rs | 2 +- .../intrinsics/generated/tests/warp_match.rs | 2 +- .../generated/tests/warp_shuffle.rs | 2 +- .../generated/tests/wgmma_control.rs | 2 +- .../terminator/intrinsics/generated/tma.rs | 2 +- .../terminator/intrinsics/generated/vote.rs | 2 +- .../intrinsics/generated/warp_barrier.rs | 2 +- .../intrinsics/generated/warp_match.rs | 2 +- .../intrinsics/generated/warp_shuffle.rs | 2 +- .../intrinsics/generated/wgmma_control.rs | 2 +- .../generated_intrinsics/active_mask.rs | 2 +- .../src/convert/generated_intrinsics/clc.rs | 2 +- .../generated_intrinsics/cluster_barrier.rs | 2 +- .../generated_intrinsics/cluster_memory.rs | 2 +- .../convert/generated_intrinsics/cp_async.rs | 2 +- .../generated_intrinsics/debug_control.rs | 2 +- .../convert/generated_intrinsics/dotprod.rs | 2 +- .../src/convert/generated_intrinsics/elect.rs | 2 +- .../generated_intrinsics/execution_control.rs | 2 +- .../generated_intrinsics/extended_minmax.rs | 2 +- .../generated_intrinsics/integer_minmax.rs | 2 +- .../convert/generated_intrinsics/ldmatrix.rs | 2 +- .../generated_intrinsics/mbarrier_basic.rs | 2 +- .../generated_intrinsics/mbarrier_extended.rs | 2 +- .../src/convert/generated_intrinsics/mod.rs | 2 +- .../convert/generated_intrinsics/movmatrix.rs | 2 +- .../generated_intrinsics/packed_alu.rs | 2 +- .../generated_intrinsics/packed_atomic.rs | 2 +- .../generated_intrinsics/packed_conversion.rs | 84 +- .../src/convert/generated_intrinsics/prmt.rs | 2 +- .../src/convert/generated_intrinsics/redux.rs | 2 +- .../generated_intrinsics/register_mma.rs | 2 +- .../generated_intrinsics/scalar_arithmetic.rs | 2 +- .../generated_intrinsics/scalar_conversion.rs | 2 +- .../generated_intrinsics/scalar_math.rs | 2 +- .../generated_intrinsics/sparse_mma.rs | 2 +- .../src/convert/generated_intrinsics/sreg.rs | 2 +- .../convert/generated_intrinsics/stmatrix.rs | 2 +- .../src/convert/generated_intrinsics/sync.rs | 2 +- .../convert/generated_intrinsics/tcgen05.rs | 2 +- .../src/convert/generated_intrinsics/tma.rs | 2 +- .../src/convert/generated_intrinsics/vote.rs | 2 +- .../generated_intrinsics/warp_barrier.rs | 2 +- .../generated_intrinsics/warp_match.rs | 2 +- .../generated_intrinsics/warp_shuffle.rs | 2 +- .../generated_intrinsics/wgmma_control.rs | 2 +- .../src/generated_intrinsics.rs | 18 +- intrinsics/abi-v1.toml | 28 + intrinsics/catalog.json | 1626 +++++++++++++---- ...-13.3.33-packed-conversion-fp8-bf16x2.json | 200 ++ ...16696adc-packed-conversion-fp8-bf16x2.json | 196 ++ intrinsics/generated-reference.md | 50 +- intrinsics/overlay.toml | 1 + .../overlay/packed_conversion_fp8_bf16x2.toml | 18 + intrinsics/probes/abs_bf16x2.ll | 2 +- intrinsics/probes/abs_f16x2.ll | 2 +- intrinsics/probes/active_mask.ll | 2 +- intrinsics/probes/add_bf16x2.ll | 2 +- intrinsics/probes/add_f16x2.ll | 2 +- intrinsics/probes/add_f32x2.ll | 2 +- intrinsics/probes/add_ftz_f32x2.ll | 2 +- intrinsics/probes/add_rm_f32.ll | 2 +- intrinsics/probes/add_rm_f64.ll | 2 +- intrinsics/probes/add_rm_ftz_f32.ll | 2 +- intrinsics/probes/add_rm_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rm_sat_f32.ll | 2 +- intrinsics/probes/add_rn_f32.ll | 2 +- intrinsics/probes/add_rn_f64.ll | 2 +- intrinsics/probes/add_rn_ftz_f32.ll | 2 +- intrinsics/probes/add_rn_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rn_sat_f32.ll | 2 +- intrinsics/probes/add_rp_f32.ll | 2 +- intrinsics/probes/add_rp_f64.ll | 2 +- intrinsics/probes/add_rp_ftz_f32.ll | 2 +- intrinsics/probes/add_rp_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rp_sat_f32.ll | 2 +- intrinsics/probes/add_rz_f32.ll | 2 +- intrinsics/probes/add_rz_f64.ll | 2 +- intrinsics/probes/add_rz_ftz_f32.ll | 2 +- intrinsics/probes/add_rz_ftz_sat_f32.ll | 2 +- intrinsics/probes/add_rz_sat_f32.ll | 2 +- intrinsics/probes/all_sync.ll | 2 +- intrinsics/probes/any_sync.ll | 2 +- intrinsics/probes/ballot_sync.ll | 2 +- intrinsics/probes/barrier_cluster_arrive.ll | 2 +- .../probes/barrier_cluster_arrive_aligned.ll | 2 +- .../probes/barrier_cluster_arrive_relaxed.ll | 2 +- .../barrier_cluster_arrive_relaxed_aligned.ll | 2 +- intrinsics/probes/barrier_cluster_wait.ll | 2 +- .../probes/barrier_cluster_wait_aligned.ll | 2 +- intrinsics/probes/barrier_cta_arrive.ll | 2 +- .../probes/barrier_cta_arrive_aligned.ll | 2 +- intrinsics/probes/barrier_cta_sync.ll | 2 +- intrinsics/probes/barrier_cta_sync_aligned.ll | 2 +- intrinsics/probes/block_dim_x.ll | 2 +- intrinsics/probes/block_dim_y.ll | 2 +- intrinsics/probes/block_dim_z.ll | 2 +- intrinsics/probes/block_idx_x.ll | 2 +- intrinsics/probes/block_idx_y.ll | 2 +- intrinsics/probes/block_idx_z.ll | 2 +- intrinsics/probes/breakpoint.ll | 2 +- .../probes/clc_query_get_first_ctaid_x.ll | 2 +- .../probes/clc_query_get_first_ctaid_y.ll | 2 +- .../probes/clc_query_get_first_ctaid_z.ll | 2 +- intrinsics/probes/clc_query_is_canceled.ll | 2 +- intrinsics/probes/clc_try_cancel.ll | 2 +- intrinsics/probes/clc_try_cancel_multicast.ll | 2 +- intrinsics/probes/clock.lib_nvvm.ll | 2 +- intrinsics/probes/clock.llvm_nvptx.ll | 2 +- intrinsics/probes/clock64.lib_nvvm.ll | 2 +- intrinsics/probes/clock64.llvm_nvptx.ll | 2 +- intrinsics/probes/cluster_block_count.ll | 2 +- intrinsics/probes/cluster_block_idx_x.ll | 2 +- intrinsics/probes/cluster_block_idx_y.ll | 2 +- intrinsics/probes/cluster_block_idx_z.ll | 2 +- intrinsics/probes/cluster_block_rank.ll | 2 +- intrinsics/probes/cluster_dim_x.ll | 2 +- intrinsics/probes/cluster_dim_y.ll | 2 +- intrinsics/probes/cluster_dim_z.ll | 2 +- intrinsics/probes/cluster_grid_dim_x.ll | 2 +- intrinsics/probes/cluster_grid_dim_y.ll | 2 +- intrinsics/probes/cluster_grid_dim_z.ll | 2 +- intrinsics/probes/cluster_idx_x.ll | 2 +- intrinsics/probes/cluster_idx_y.ll | 2 +- intrinsics/probes/cluster_idx_z.ll | 2 +- intrinsics/probes/cos_approx_f32.ll | 2 +- intrinsics/probes/cos_approx_ftz_f32.ll | 2 +- .../probes/cp_async_bulk_commit_group.ll | 2 +- .../cp_async_bulk_prefetch_tensor_1d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_1d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_2d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_2d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_3d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_3d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_4d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_4d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_5d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_5d_l2_cache_hint.ll | 2 +- ...sync_bulk_prefetch_tensor_gather4_2d_l2.ll | 2 +- ...refetch_tensor_gather4_2d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s.ll | 2 +- .../cp_async_bulk_tensor_2d_g2s_multicast.ll | 2 +- ..._async_bulk_tensor_2d_g2s_multicast_cg2.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_s2g.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_5d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_3d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_4d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_5d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_1d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_2d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_3d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_4d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_5d.ll | 2 +- intrinsics/probes/cp_async_bulk_wait_group.ll | 2 +- .../probes/cp_async_bulk_wait_group_read.ll | 2 +- intrinsics/probes/cp_async_ca_16.ll | 2 +- intrinsics/probes/cp_async_ca_4.ll | 2 +- intrinsics/probes/cp_async_ca_8.ll | 2 +- intrinsics/probes/cp_async_ca_zfill_16.ll | 2 +- intrinsics/probes/cp_async_ca_zfill_4.ll | 2 +- intrinsics/probes/cp_async_ca_zfill_8.ll | 2 +- intrinsics/probes/cp_async_cg_16.ll | 2 +- intrinsics/probes/cp_async_cg_zfill_16.ll | 2 +- intrinsics/probes/cp_async_commit_group.ll | 2 +- intrinsics/probes/cp_async_mbarrier_arrive.ll | 2 +- .../probes/cp_async_mbarrier_arrive_noinc.ll | 2 +- .../cp_async_mbarrier_arrive_noinc_shared.ll | 2 +- .../probes/cp_async_mbarrier_arrive_shared.ll | 2 +- intrinsics/probes/cp_async_wait_all.ll | 2 +- intrinsics/probes/cp_async_wait_group.ll | 2 +- intrinsics/probes/cvt_f16x2_f32.ll | 2 +- intrinsics/probes/cvt_f32x2_bf16x2.ll | 2 +- intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll | 2 +- intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll | 2 +- intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll | 2 +- intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll | 2 +- intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll | 2 +- intrinsics/probes/cvt_rn_relu_f16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rn_relu_tf32_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll | 12 + .../probes/cvt_rn_satfinite_e4m3x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll | 12 + .../probes/cvt_rn_satfinite_e5m2x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll | 12 + .../cvt_rn_satfinite_relu_e4m3x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll | 12 + .../cvt_rn_satfinite_relu_e5m2x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rn_tf32_f32.ll | 2 +- .../probes/cvt_rna_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rna_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rz_bf16x2_f32.ll | 2 +- intrinsics/probes/cvt_rz_f16x2_f32.ll | 2 +- .../probes/cvt_rz_relu_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rz_relu_tf32_f32.ll | 2 +- .../probes/cvt_rz_satfinite_tf32_f32.ll | 2 +- intrinsics/probes/cvt_rz_tf32_f32.ll | 2 +- intrinsics/probes/div_rm_f32.ll | 2 +- intrinsics/probes/div_rm_f64.ll | 2 +- intrinsics/probes/div_rm_ftz_f32.ll | 2 +- intrinsics/probes/div_rn_f32.ll | 2 +- intrinsics/probes/div_rn_f64.ll | 2 +- intrinsics/probes/div_rn_ftz_f32.ll | 2 +- intrinsics/probes/div_rp_f32.ll | 2 +- intrinsics/probes/div_rp_f64.ll | 2 +- intrinsics/probes/div_rp_ftz_f32.ll | 2 +- intrinsics/probes/div_rz_f32.ll | 2 +- intrinsics/probes/div_rz_f64.ll | 2 +- intrinsics/probes/div_rz_ftz_f32.ll | 2 +- intrinsics/probes/dp2a_s32.ll | 2 +- intrinsics/probes/dp2a_u32.ll | 2 +- intrinsics/probes/dp4a_s32.ll | 2 +- intrinsics/probes/dp4a_u32.ll | 2 +- intrinsics/probes/dsmem_read_u32.ll | 2 +- .../probes/dynamic_smem_size.lib_nvvm.ll | 2 +- .../probes/dynamic_smem_size.llvm_nvptx.ll | 2 +- intrinsics/probes/elect_sync.lib_nvvm.ll | 2 +- intrinsics/probes/elect_sync.llvm_nvptx.ll | 2 +- intrinsics/probes/envreg1.lib_nvvm.ll | 2 +- intrinsics/probes/envreg1.llvm_nvptx.ll | 2 +- intrinsics/probes/envreg2.lib_nvvm.ll | 2 +- intrinsics/probes/envreg2.llvm_nvptx.ll | 2 +- intrinsics/probes/ex2_approx_f16.ll | 2 +- intrinsics/probes/ex2_approx_f32.ll | 2 +- intrinsics/probes/ex2_approx_ftz_f32.ll | 2 +- .../fence_mbarrier_init_release_cluster.ll | 2 +- ..._generic_acquire_shared_cluster_cluster.ll | 2 +- ...sync_generic_release_shared_cta_cluster.ll | 2 +- .../probes/fence_proxy_async_shared_cta.ll | 2 +- ...proxy_tensormap_generic_acquire_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_cta.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_gpu.ll | 2 +- ..._proxy_tensormap_generic_acquire_system.ll | 2 +- ...proxy_tensormap_generic_release_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_release_cta.ll | 2 +- ...nce_proxy_tensormap_generic_release_gpu.ll | 2 +- ..._proxy_tensormap_generic_release_system.ll | 2 +- intrinsics/probes/fma_bf16x2.ll | 2 +- intrinsics/probes/fma_f16x2.ll | 2 +- intrinsics/probes/fma_f32x2.ll | 2 +- intrinsics/probes/fma_ftz_f16x2.ll | 2 +- intrinsics/probes/fma_ftz_f32x2.ll | 2 +- intrinsics/probes/fma_ftz_relu_f16x2.ll | 2 +- intrinsics/probes/fma_ftz_sat_f16x2.ll | 2 +- intrinsics/probes/fma_relu_bf16x2.ll | 2 +- intrinsics/probes/fma_relu_f16x2.ll | 2 +- intrinsics/probes/fma_rm_f32.ll | 2 +- intrinsics/probes/fma_rm_f64.ll | 2 +- intrinsics/probes/fma_rm_ftz_f32.ll | 2 +- intrinsics/probes/fma_rm_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rm_sat_f32.ll | 2 +- intrinsics/probes/fma_rn_f32.ll | 2 +- intrinsics/probes/fma_rn_f64.ll | 2 +- intrinsics/probes/fma_rn_ftz_f32.ll | 2 +- intrinsics/probes/fma_rn_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rn_sat_f32.ll | 2 +- intrinsics/probes/fma_rp_f32.ll | 2 +- intrinsics/probes/fma_rp_f64.ll | 2 +- intrinsics/probes/fma_rp_ftz_f32.ll | 2 +- intrinsics/probes/fma_rp_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rp_sat_f32.ll | 2 +- intrinsics/probes/fma_rz_f32.ll | 2 +- intrinsics/probes/fma_rz_f64.ll | 2 +- intrinsics/probes/fma_rz_ftz_f32.ll | 2 +- intrinsics/probes/fma_rz_ftz_sat_f32.ll | 2 +- intrinsics/probes/fma_rz_sat_f32.ll | 2 +- intrinsics/probes/fma_sat_f16x2.ll | 2 +- intrinsics/probes/globaltimer.lib_nvvm.ll | 2 +- intrinsics/probes/globaltimer.llvm_nvptx.ll | 2 +- .../grid_dependency_launch_dependents.ll | 2 +- intrinsics/probes/grid_dependency_wait.ll | 2 +- intrinsics/probes/grid_dim_x.ll | 2 +- intrinsics/probes/grid_dim_y.ll | 2 +- intrinsics/probes/grid_dim_z.ll | 2 +- intrinsics/probes/gridid.lib_nvvm.ll | 2 +- intrinsics/probes/gridid.llvm_nvptx.ll | 2 +- intrinsics/probes/lane_id.ll | 2 +- intrinsics/probes/lanemask_eq.ll | 2 +- intrinsics/probes/lanemask_ge.ll | 2 +- intrinsics/probes/lanemask_gt.ll | 2 +- intrinsics/probes/lanemask_le.ll | 2 +- intrinsics/probes/lanemask_lt.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll | 2 +- intrinsics/probes/ldmatrix_m8n8_x1_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x1_trans_b16.ll | 2 +- intrinsics/probes/ldmatrix_m8n8_x2_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x2_trans_b16.ll | 2 +- intrinsics/probes/ldmatrix_m8n8_x4_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x4_trans_b16.ll | 2 +- intrinsics/probes/lg2_approx_f32.ll | 2 +- intrinsics/probes/lg2_approx_ftz_f32.ll | 2 +- intrinsics/probes/map_shared_rank.ll | 2 +- intrinsics/probes/match_all_i64_sync.ll | 2 +- intrinsics/probes/match_all_sync.ll | 2 +- intrinsics/probes/match_any_i64_sync.ll | 2 +- intrinsics/probes/match_any_sync.ll | 2 +- intrinsics/probes/max_bf16.ll | 2 +- intrinsics/probes/max_bf16x2.ll | 2 +- intrinsics/probes/max_f16.ll | 2 +- intrinsics/probes/max_f16x2.ll | 2 +- intrinsics/probes/max_ftz_f16.ll | 2 +- intrinsics/probes/max_ftz_f16x2.ll | 2 +- intrinsics/probes/max_ftz_nan_f16.ll | 2 +- intrinsics/probes/max_ftz_nan_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/max_ftz_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/max_ftz_xorsign_abs_f32.ll | 2 +- intrinsics/probes/max_nan_bf16.ll | 2 +- intrinsics/probes/max_nan_bf16x2.ll | 2 +- intrinsics/probes/max_nan_f16.ll | 2 +- intrinsics/probes/max_nan_f16x2.ll | 2 +- intrinsics/probes/max_nan_xorsign_abs_bf16.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/max_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_nan_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/max_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/max_relu_s16x2.ll | 2 +- intrinsics/probes/max_relu_s32.ll | 2 +- intrinsics/probes/max_s16x2.ll | 2 +- intrinsics/probes/max_u16x2.ll | 2 +- intrinsics/probes/max_xorsign_abs_bf16.ll | 2 +- intrinsics/probes/max_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/max_xorsign_abs_f16.ll | 2 +- intrinsics/probes/max_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/max_xorsign_abs_f32.ll | 2 +- intrinsics/probes/mbarrier_arrive.ll | 2 +- intrinsics/probes/mbarrier_arrive_cluster.ll | 2 +- .../probes/mbarrier_arrive_expect_tx.ll | 2 +- .../mbarrier_arrive_expect_tx_cluster.ll | 2 +- .../probes/mbarrier_arrive_no_complete.ll | 2 +- intrinsics/probes/mbarrier_init.ll | 2 +- intrinsics/probes/mbarrier_inval.ll | 2 +- intrinsics/probes/mbarrier_test_wait.ll | 2 +- intrinsics/probes/mbarrier_try_wait.ll | 2 +- intrinsics/probes/mbarrier_try_wait_parity.ll | 2 +- .../mbarrier_try_wait_parity_cluster.ll | 2 +- intrinsics/probes/min_bf16.ll | 2 +- intrinsics/probes/min_bf16x2.ll | 2 +- intrinsics/probes/min_f16.ll | 2 +- intrinsics/probes/min_f16x2.ll | 2 +- intrinsics/probes/min_ftz_f16.ll | 2 +- intrinsics/probes/min_ftz_f16x2.ll | 2 +- intrinsics/probes/min_ftz_nan_f16.ll | 2 +- intrinsics/probes/min_ftz_nan_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/min_ftz_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/min_ftz_xorsign_abs_f32.ll | 2 +- intrinsics/probes/min_nan_bf16.ll | 2 +- intrinsics/probes/min_nan_bf16x2.ll | 2 +- intrinsics/probes/min_nan_f16.ll | 2 +- intrinsics/probes/min_nan_f16x2.ll | 2 +- intrinsics/probes/min_nan_xorsign_abs_bf16.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/min_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_nan_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/min_nan_xorsign_abs_f32.ll | 2 +- intrinsics/probes/min_relu_s16x2.ll | 2 +- intrinsics/probes/min_relu_s32.ll | 2 +- intrinsics/probes/min_s16x2.ll | 2 +- intrinsics/probes/min_u16x2.ll | 2 +- intrinsics/probes/min_xorsign_abs_bf16.ll | 2 +- intrinsics/probes/min_xorsign_abs_bf16x2.ll | 2 +- intrinsics/probes/min_xorsign_abs_f16.ll | 2 +- intrinsics/probes/min_xorsign_abs_f16x2.ll | 2 +- intrinsics/probes/min_xorsign_abs_f32.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_xor_popc.ll | 2 +- intrinsics/probes/mma_m16n8k16_f16_f16.ll | 2 +- intrinsics/probes/mma_m16n8k16_f32_bf16.ll | 2 +- intrinsics/probes/mma_m16n8k16_f32_f16.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_s8.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll | 2 +- .../mma_m16n8k16_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_u8.ll | 2 +- intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll | 2 +- .../mma_m16n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_satfinite.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_xor_popc.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s4.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll | 2 +- .../mma_m16n8k32_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s8.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u4.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll | 2 +- .../mma_m16n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u8.ll | 2 +- intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k4_f32_tf32.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_s4.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_u4.ll | 2 +- intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m16n8k8_f16_f16.ll | 2 +- intrinsics/probes/mma_m16n8k8_f32_bf16.ll | 2 +- intrinsics/probes/mma_m16n8k8_f32_f16.ll | 2 +- intrinsics/probes/mma_m16n8k8_f32_tf32.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_xor_popc.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_u8.ll | 2 +- intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_u4.ll | 2 +- intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_m8n8k4_f64.ll | 2 +- intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_satfinite.ll | 2 +- intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_s4.ll | 2 +- ...red_metadata_m16n8k128_s32_s4_satfinite.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_s4_u4.ll | 2 +- ..._metadata_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_u4.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_u4_s4.ll | 2 +- ..._metadata_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- ...red_metadata_m16n8k128_s32_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f32_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_tf32.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k32_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f32_f16.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_s8.ll | 2 +- ...ered_metadata_m16n8k32_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k32_s32_u8_satfinite.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s4.ll | 2 +- ...ered_metadata_m16n8k64_s32_s4_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s4_u4.ll | 2 +- ...d_metadata_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s8.ll | 2 +- ...ered_metadata_m16n8k64_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u4.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u4_s4.ll | 2 +- ...d_metadata_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k8_f32_tf32.ll | 2 +- intrinsics/probes/movmatrix_trans_b16.ll | 2 +- intrinsics/probes/mul_bf16x2.ll | 2 +- intrinsics/probes/mul_f16x2.ll | 2 +- intrinsics/probes/mul_f32x2.ll | 2 +- intrinsics/probes/mul_ftz_f32x2.ll | 2 +- intrinsics/probes/mul_rm_f32.ll | 2 +- intrinsics/probes/mul_rm_f64.ll | 2 +- intrinsics/probes/mul_rm_ftz_f32.ll | 2 +- intrinsics/probes/mul_rn_f32.ll | 2 +- intrinsics/probes/mul_rn_f64.ll | 2 +- intrinsics/probes/mul_rn_ftz_f32.ll | 2 +- intrinsics/probes/mul_rp_f32.ll | 2 +- intrinsics/probes/mul_rp_f64.ll | 2 +- intrinsics/probes/mul_rp_ftz_f32.ll | 2 +- intrinsics/probes/mul_rz_f32.ll | 2 +- intrinsics/probes/mul_rz_f64.ll | 2 +- intrinsics/probes/mul_rz_ftz_f32.ll | 2 +- intrinsics/probes/nanosleep.ll | 2 +- intrinsics/probes/neg_bf16x2.ll | 2 +- intrinsics/probes/neg_f16x2.ll | 2 +- intrinsics/probes/nsmid.lib_nvvm.ll | 2 +- intrinsics/probes/nsmid.llvm_nvptx.ll | 2 +- intrinsics/probes/nwarpid.lib_nvvm.ll | 2 +- intrinsics/probes/nwarpid.llvm_nvptx.ll | 2 +- intrinsics/probes/packed_atomic_add_bf16x2.ll | 2 +- intrinsics/probes/packed_atomic_add_f16x2.ll | 2 +- intrinsics/probes/pmevent.ll | 2 +- intrinsics/probes/prefetch_tma_descriptor.ll | 2 +- intrinsics/probes/prmt.ll | 2 +- intrinsics/probes/prmt_b4e.ll | 2 +- intrinsics/probes/prmt_ecl.ll | 2 +- intrinsics/probes/prmt_ecr.ll | 2 +- intrinsics/probes/prmt_f4e.ll | 2 +- intrinsics/probes/prmt_rc16.ll | 2 +- intrinsics/probes/prmt_rc8.ll | 2 +- intrinsics/probes/rcp_approx_ftz_f32.ll | 2 +- intrinsics/probes/rcp_approx_ftz_f64.ll | 2 +- intrinsics/probes/rcp_rm_f32.ll | 2 +- intrinsics/probes/rcp_rm_f64.ll | 2 +- intrinsics/probes/rcp_rm_ftz_f32.ll | 2 +- intrinsics/probes/rcp_rn_f32.ll | 2 +- intrinsics/probes/rcp_rn_f64.ll | 2 +- intrinsics/probes/rcp_rn_ftz_f32.ll | 2 +- intrinsics/probes/rcp_rp_f32.ll | 2 +- intrinsics/probes/rcp_rp_f64.ll | 2 +- intrinsics/probes/rcp_rp_ftz_f32.ll | 2 +- intrinsics/probes/rcp_rz_f32.ll | 2 +- intrinsics/probes/rcp_rz_f64.ll | 2 +- intrinsics/probes/rcp_rz_ftz_f32.ll | 2 +- intrinsics/probes/redux_sync_add.ll | 2 +- intrinsics/probes/redux_sync_and.ll | 2 +- intrinsics/probes/redux_sync_max_abs_f32.ll | 2 +- .../probes/redux_sync_max_abs_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_max_f32.ll | 2 +- intrinsics/probes/redux_sync_max_i32.ll | 2 +- intrinsics/probes/redux_sync_max_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_max_u32.ll | 2 +- intrinsics/probes/redux_sync_min_abs_f32.ll | 2 +- .../probes/redux_sync_min_abs_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_min_f32.ll | 2 +- intrinsics/probes/redux_sync_min_i32.ll | 2 +- intrinsics/probes/redux_sync_min_nan_f32.ll | 2 +- intrinsics/probes/redux_sync_min_u32.ll | 2 +- intrinsics/probes/redux_sync_or.ll | 2 +- intrinsics/probes/redux_sync_xor.ll | 2 +- intrinsics/probes/rsqrt_approx_f32.ll | 2 +- intrinsics/probes/rsqrt_approx_f64.ll | 2 +- intrinsics/probes/rsqrt_approx_ftz_f32.ll | 2 +- intrinsics/probes/setmaxnreg_dec.ll | 2 +- intrinsics/probes/setmaxnreg_inc.ll | 2 +- intrinsics/probes/shuffle_down_f32_sync.ll | 2 +- intrinsics/probes/shuffle_down_sync.ll | 2 +- intrinsics/probes/shuffle_down_u64_sync.ll | 2 +- intrinsics/probes/shuffle_f32_sync.ll | 2 +- intrinsics/probes/shuffle_sync.ll | 2 +- intrinsics/probes/shuffle_u64_sync.ll | 2 +- intrinsics/probes/shuffle_up_f32_sync.ll | 2 +- intrinsics/probes/shuffle_up_sync.ll | 2 +- intrinsics/probes/shuffle_up_u64_sync.ll | 2 +- intrinsics/probes/shuffle_xor_f32_sync.ll | 2 +- intrinsics/probes/shuffle_xor_sync.ll | 2 +- intrinsics/probes/shuffle_xor_u64_sync.ll | 2 +- intrinsics/probes/sin_approx_f32.ll | 2 +- intrinsics/probes/sin_approx_ftz_f32.ll | 2 +- intrinsics/probes/smid.lib_nvvm.ll | 2 +- intrinsics/probes/smid.llvm_nvptx.ll | 2 +- intrinsics/probes/sqrt_approx_f32.ll | 2 +- intrinsics/probes/sqrt_approx_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rm_f32.ll | 2 +- intrinsics/probes/sqrt_rm_f64.ll | 2 +- intrinsics/probes/sqrt_rm_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rn_f32.ll | 2 +- intrinsics/probes/sqrt_rn_f64.ll | 2 +- intrinsics/probes/sqrt_rn_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rp_f32.ll | 2 +- intrinsics/probes/sqrt_rp_f64.ll | 2 +- intrinsics/probes/sqrt_rp_ftz_f32.ll | 2 +- intrinsics/probes/sqrt_rz_f32.ll | 2 +- intrinsics/probes/sqrt_rz_f64.ll | 2 +- intrinsics/probes/sqrt_rz_ftz_f32.ll | 2 +- intrinsics/probes/stmatrix_m8n8_x2_b16.ll | 2 +- .../probes/stmatrix_m8n8_x2_trans_b16.ll | 2 +- intrinsics/probes/stmatrix_m8n8_x4_b16.ll | 2 +- .../probes/stmatrix_m8n8_x4_trans_b16.ll | 2 +- intrinsics/probes/sub_bf16x2.ll | 2 +- intrinsics/probes/sub_f16x2.ll | 2 +- intrinsics/probes/sub_f32x2.ll | 2 +- intrinsics/probes/sub_ftz_f32x2.ll | 2 +- intrinsics/probes/sync_mask.ll | 2 +- intrinsics/probes/sync_threads.ll | 2 +- intrinsics/probes/tanh_approx_f32.ll | 2 +- intrinsics/probes/tcgen05_alloc.ll | 2 +- intrinsics/probes/tcgen05_alloc_cg2.ll | 2 +- intrinsics/probes/tcgen05_commit.ll | 2 +- intrinsics/probes/tcgen05_commit_cg2.ll | 2 +- intrinsics/probes/tcgen05_commit_multicast.ll | 2 +- .../probes/tcgen05_commit_multicast_cg2.ll | 2 +- .../probes/tcgen05_commit_shared_cluster.ll | 2 +- .../tcgen05_commit_shared_cluster_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_128x128b.ll | 2 +- .../probes/tcgen05_cp_128x128b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x128b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x128b_b6x16_p32_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_128x128b_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b4x16_p64.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b6x16_p32.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_4x256b.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_4x256b_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_01_23_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_02_13_cg2.ll | 2 +- intrinsics/probes/tcgen05_cp_smem_to_tmem.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem_cg2.ll | 2 +- intrinsics/probes/tcgen05_dealloc.ll | 2 +- intrinsics/probes/tcgen05_dealloc_cg2.ll | 2 +- .../probes/tcgen05_fence_after_thread_sync.ll | 2 +- .../tcgen05_fence_before_thread_sync.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_raw.ll | 2 +- intrinsics/probes/tcgen05_ld_16x256b_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_pack16.ll | 2 +- intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll | 2 +- intrinsics/probes/tcgen05_load_wait.ll | 2 +- intrinsics/probes/tcgen05_mma_e2m1.ll | 2 +- intrinsics/probes/tcgen05_mma_e2m3.ll | 2 +- intrinsics/probes/tcgen05_mma_e3m2.ll | 2 +- intrinsics/probes/tcgen05_mma_e4m3.ll | 2 +- intrinsics/probes/tcgen05_mma_e5m2.ll | 2 +- intrinsics/probes/tcgen05_mma_f16.ll | 2 +- intrinsics/probes/tcgen05_mma_f16_cg2.ll | 2 +- intrinsics/probes/tcgen05_mma_shared.ll | 2 +- intrinsics/probes/tcgen05_mma_sp_shared.ll | 2 +- intrinsics/probes/tcgen05_mma_sp_tensor.ll | 2 +- .../probes/tcgen05_mma_sp_tensor_ashift.ll | 2 +- intrinsics/probes/tcgen05_mma_tensor.ll | 2 +- .../probes/tcgen05_mma_tensor_ashift.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_bf16.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e2m1.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e2m3.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e3m2.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e4m3.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_e5m2.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_f16.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_shared.ll | 2 +- .../tcgen05_mma_ws_shared_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_sp_shared.ll | 2 +- .../tcgen05_mma_ws_sp_shared_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll | 2 +- .../tcgen05_mma_ws_sp_tensor_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_tensor.ll | 2 +- .../tcgen05_mma_ws_tensor_zero_col_mask.ll | 2 +- intrinsics/probes/tcgen05_mma_ws_tf32.ll | 2 +- .../probes/tcgen05_relinquish_alloc_permit.ll | 2 +- .../tcgen05_relinquish_alloc_permit_cg2.ll | 2 +- intrinsics/probes/tcgen05_shift_down.ll | 2 +- intrinsics/probes/tcgen05_shift_down_cg2.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x128_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x16_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x32_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x64_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_unpack16.ll | 2 +- intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_unpack16.ll | 2 +- intrinsics/probes/tcgen05_store_wait.ll | 2 +- .../probes/tensormap_replace_box_dim.ll | 2 +- .../tensormap_replace_element_stride.ll | 2 +- .../probes/tensormap_replace_element_type.ll | 2 +- .../probes/tensormap_replace_fill_mode.ll | 2 +- .../tensormap_replace_global_address.ll | 2 +- .../probes/tensormap_replace_global_dim.ll | 2 +- .../probes/tensormap_replace_global_stride.ll | 2 +- .../tensormap_replace_interleave_layout.ll | 2 +- intrinsics/probes/tensormap_replace_rank.ll | 2 +- .../tensormap_replace_swizzle_atomicity.ll | 2 +- .../probes/tensormap_replace_swizzle_mode.ll | 2 +- intrinsics/probes/thread_idx_x.ll | 2 +- intrinsics/probes/thread_idx_y.ll | 2 +- intrinsics/probes/thread_idx_z.ll | 2 +- intrinsics/probes/threadfence.ll | 2 +- intrinsics/probes/threadfence_block.ll | 2 +- intrinsics/probes/threadfence_system.ll | 2 +- intrinsics/probes/total_smem_size.lib_nvvm.ll | 2 +- .../probes/total_smem_size.llvm_nvptx.ll | 2 +- intrinsics/probes/trap.ll | 2 +- intrinsics/probes/uni_sync.ll | 2 +- intrinsics/probes/warpid.lib_nvvm.ll | 2 +- intrinsics/probes/warpid.llvm_nvptx.ll | 2 +- intrinsics/probes/wgmma_commit_group.ll | 2 +- intrinsics/probes/wgmma_fence.ll | 2 +- intrinsics/probes/wgmma_wait_group.ll | 2 +- 1344 files changed, 4577 insertions(+), 1734 deletions(-) create mode 100644 intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json create mode 100644 intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json create mode 100644 intrinsics/overlay/packed_conversion_fp8_bf16x2.toml create mode 100644 intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll create mode 100644 intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll create mode 100644 intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll create mode 100644 intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll diff --git a/crates/cuda-device/src/generated/async_copy.rs b/crates/cuda-device/src/generated/async_copy.rs index e44c9a333c..af0208b5f0 100644 --- a/crates/cuda-device/src/generated/async_copy.rs +++ b/crates/cuda-device/src/generated/async_copy.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::async_copy` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/atomic.rs b/crates/cuda-device/src/generated/atomic.rs index 47a5053d24..cc45bc5931 100644 --- a/crates/cuda-device/src/generated/atomic.rs +++ b/crates/cuda-device/src/generated/atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::atomic` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/bf16.rs b/crates/cuda-device/src/generated/bf16.rs index fb8215c889..cbcb9110f8 100644 --- a/crates/cuda-device/src/generated/bf16.rs +++ b/crates/cuda-device/src/generated/bf16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::bf16` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/bf16x2.rs b/crates/cuda-device/src/generated/bf16x2.rs index b9f4be0b75..c3a7ba972d 100644 --- a/crates/cuda-device/src/generated/bf16x2.rs +++ b/crates/cuda-device/src/generated/bf16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::bf16x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/clc.rs b/crates/cuda-device/src/generated/clc.rs index 94c9cae176..3b806ef080 100644 --- a/crates/cuda-device/src/generated/clc.rs +++ b/crates/cuda-device/src/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::clc` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/cluster_barrier.rs b/crates/cuda-device/src/generated/cluster_barrier.rs index 9115aec1b5..b6ee7ae7c1 100644 --- a/crates/cuda-device/src/generated/cluster_barrier.rs +++ b/crates/cuda-device/src/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::cluster` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/cluster_memory.rs b/crates/cuda-device/src/generated/cluster_memory.rs index d665c6fc59..d70551cfdd 100644 --- a/crates/cuda-device/src/generated/cluster_memory.rs +++ b/crates/cuda-device/src/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::cluster` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/cluster_sreg.rs b/crates/cuda-device/src/generated/cluster_sreg.rs index 4e2ce888c0..c0f0867ba6 100644 --- a/crates/cuda-device/src/generated/cluster_sreg.rs +++ b/crates/cuda-device/src/generated/cluster_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // This file is included inside `cuda_device::cluster`. diff --git a/crates/cuda-device/src/generated/convert.rs b/crates/cuda-device/src/generated/convert.rs index f84862afdd..e4aaaf3ada 100644 --- a/crates/cuda-device/src/generated/convert.rs +++ b/crates/cuda-device/src/generated/convert.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::convert`. @@ -81,6 +81,15 @@ pub fn cvt_rn_relu_f16x2_f32(lo: f32, hi: f32) -> u32 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_e4m3x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_e4m3x2_f16x2(packed: u32) -> u16 { @@ -99,6 +108,15 @@ pub fn cvt_rn_satfinite_e4m3x2_f32(lo: f32, hi: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_e5m2x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_e5m2x2_f16x2(packed: u32) -> u16 { @@ -117,6 +135,15 @@ pub fn cvt_rn_satfinite_e5m2x2_f32(lo: f32, hi: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_relu_e4m3x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_relu_e4m3x2_f16x2(packed: u32) -> u16 { @@ -135,6 +162,15 @@ pub fn cvt_rn_satfinite_relu_e4m3x2_f32(lo: f32, hi: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_relu_e5m2x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_relu_e5m2x2_f16x2(packed: u32) -> u16 { diff --git a/crates/cuda-device/src/generated/counted_barrier.rs b/crates/cuda-device/src/generated/counted_barrier.rs index 877e00f15f..0deb657f01 100644 --- a/crates/cuda-device/src/generated/counted_barrier.rs +++ b/crates/cuda-device/src/generated/counted_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::barrier` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/debug_control.rs b/crates/cuda-device/src/generated/debug_control.rs index 6f52412e9a..3f6e9e8953 100644 --- a/crates/cuda-device/src/generated/debug_control.rs +++ b/crates/cuda-device/src/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::debug` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/debug_sreg.rs b/crates/cuda-device/src/generated/debug_sreg.rs index 2028a15d31..93d781a696 100644 --- a/crates/cuda-device/src/generated/debug_sreg.rs +++ b/crates/cuda-device/src/generated/debug_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::debug` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/dotprod.rs b/crates/cuda-device/src/generated/dotprod.rs index 2e1e16f829..b39b0454d7 100644 --- a/crates/cuda-device/src/generated/dotprod.rs +++ b/crates/cuda-device/src/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::dotprod` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/f16.rs b/crates/cuda-device/src/generated/f16.rs index e0308d047a..4c4ff7b1f8 100644 --- a/crates/cuda-device/src/generated/f16.rs +++ b/crates/cuda-device/src/generated/f16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::f16` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/f16x2.rs b/crates/cuda-device/src/generated/f16x2.rs index 00085fa0b6..4ff4cb4529 100644 --- a/crates/cuda-device/src/generated/f16x2.rs +++ b/crates/cuda-device/src/generated/f16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::f16x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/f32x2.rs b/crates/cuda-device/src/generated/f32x2.rs index c5de9a8539..79d9cb68b7 100644 --- a/crates/cuda-device/src/generated/f32x2.rs +++ b/crates/cuda-device/src/generated/f32x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::f32x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/fence.rs b/crates/cuda-device/src/generated/fence.rs index 685f7e4c38..dcd5a6e174 100644 --- a/crates/cuda-device/src/generated/fence.rs +++ b/crates/cuda-device/src/generated/fence.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // This file is included inside `cuda_device::fence` so existing paths stay stable. diff --git a/crates/cuda-device/src/generated/float.rs b/crates/cuda-device/src/generated/float.rs index fbe53bd878..d493596d11 100644 --- a/crates/cuda-device/src/generated/float.rs +++ b/crates/cuda-device/src/generated/float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::float`. diff --git a/crates/cuda-device/src/generated/grid_dependency.rs b/crates/cuda-device/src/generated/grid_dependency.rs index 94d7848ee7..29b2822fe1 100644 --- a/crates/cuda-device/src/generated/grid_dependency.rs +++ b/crates/cuda-device/src/generated/grid_dependency.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::grid` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/grid_sreg.rs b/crates/cuda-device/src/generated/grid_sreg.rs index ed3fa06aeb..84d8ee488e 100644 --- a/crates/cuda-device/src/generated/grid_sreg.rs +++ b/crates/cuda-device/src/generated/grid_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::grid` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/i16x2.rs b/crates/cuda-device/src/generated/i16x2.rs index 054da50cd7..5a07b04f01 100644 --- a/crates/cuda-device/src/generated/i16x2.rs +++ b/crates/cuda-device/src/generated/i16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::i16x2` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/int.rs b/crates/cuda-device/src/generated/int.rs index 87a798b163..af6d3395e0 100644 --- a/crates/cuda-device/src/generated/int.rs +++ b/crates/cuda-device/src/generated/int.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::int` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/ldmatrix.rs b/crates/cuda-device/src/generated/ldmatrix.rs index b288951019..109038cc75 100644 --- a/crates/cuda-device/src/generated/ldmatrix.rs +++ b/crates/cuda-device/src/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/mbarrier_basic.rs b/crates/cuda-device/src/generated/mbarrier_basic.rs index eb377051e6..86588700a4 100644 --- a/crates/cuda-device/src/generated/mbarrier_basic.rs +++ b/crates/cuda-device/src/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/mbarrier_extended.rs b/crates/cuda-device/src/generated/mbarrier_extended.rs index 7787d693cf..e6d86882ce 100644 --- a/crates/cuda-device/src/generated/mbarrier_extended.rs +++ b/crates/cuda-device/src/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/movmatrix.rs b/crates/cuda-device/src/generated/movmatrix.rs index 09ccf2ea4d..81bfe023f2 100644 --- a/crates/cuda-device/src/generated/movmatrix.rs +++ b/crates/cuda-device/src/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/prmt.rs b/crates/cuda-device/src/generated/prmt.rs index 30fb07902d..e0f5651f15 100644 --- a/crates/cuda-device/src/generated/prmt.rs +++ b/crates/cuda-device/src/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::prmt` to keep the public paths stable. diff --git a/crates/cuda-device/src/generated/register_control.rs b/crates/cuda-device/src/generated/register_control.rs index 91d3e7f239..e8b9f1297f 100644 --- a/crates/cuda-device/src/generated/register_control.rs +++ b/crates/cuda-device/src/generated/register_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::thread` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/register_mma.rs b/crates/cuda-device/src/generated/register_mma.rs index f623acbe07..6ccb2dc62a 100644 --- a/crates/cuda-device/src/generated/register_mma.rs +++ b/crates/cuda-device/src/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/shared_sreg.rs b/crates/cuda-device/src/generated/shared_sreg.rs index 35198aaa6f..661672f59f 100644 --- a/crates/cuda-device/src/generated/shared_sreg.rs +++ b/crates/cuda-device/src/generated/shared_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::shared` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/sparse_mma.rs b/crates/cuda-device/src/generated/sparse_mma.rs index a1e8bd0f63..542bc9bba7 100644 --- a/crates/cuda-device/src/generated/sparse_mma.rs +++ b/crates/cuda-device/src/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/sreg.rs b/crates/cuda-device/src/generated/sreg.rs index f972a0a046..6855af9d1a 100644 --- a/crates/cuda-device/src/generated/sreg.rs +++ b/crates/cuda-device/src/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // This file is included lexically inside `cuda_device::thread` so existing diff --git a/crates/cuda-device/src/generated/stmatrix.rs b/crates/cuda-device/src/generated/stmatrix.rs index 73f22a70d0..42c520e6f5 100644 --- a/crates/cuda-device/src/generated/stmatrix.rs +++ b/crates/cuda-device/src/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::tcgen05` to keep public paths stable. diff --git a/crates/cuda-device/src/generated/tcgen05.rs b/crates/cuda-device/src/generated/tcgen05.rs index cb1712552e..008e071141 100644 --- a/crates/cuda-device/src/generated/tcgen05.rs +++ b/crates/cuda-device/src/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::tcgen05` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/tma.rs b/crates/cuda-device/src/generated/tma.rs index 36156f0b8d..4e6fd16aa6 100644 --- a/crates/cuda-device/src/generated/tma.rs +++ b/crates/cuda-device/src/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::tma` to keep its public API stable. diff --git a/crates/cuda-device/src/generated/warp_sreg.rs b/crates/cuda-device/src/generated/warp_sreg.rs index 8019adef7b..f3c2407f56 100644 --- a/crates/cuda-device/src/generated/warp_sreg.rs +++ b/crates/cuda-device/src/generated/warp_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::warp` to keep existing paths stable. diff --git a/crates/cuda-device/src/generated/wgmma_control.rs b/crates/cuda-device/src/generated/wgmma_control.rs index fd1ea8426e..5f6024f87f 100644 --- a/crates/cuda-device/src/generated/wgmma_control.rs +++ b/crates/cuda-device/src/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. // Included inside `cuda_device::wgmma` to keep its public API stable. diff --git a/crates/cuda-intrinsics-gen/src/model/contracts/scalar.rs b/crates/cuda-intrinsics-gen/src/model/contracts/scalar.rs index 257a1a85e5..f3da03ecca 100644 --- a/crates/cuda-intrinsics-gen/src/model/contracts/scalar.rs +++ b/crates/cuda-intrinsics-gen/src/model/contracts/scalar.rs @@ -311,7 +311,7 @@ pub enum ExtendedMinMaxAdapter { /// Closed contract for converting between scalar and packed values. /// /// The source may be a scalar pair (`f32x2`, two operands) or an already-packed -/// 16-bit or 32-bit register (`f16x2`, `e4m3x2`, `e5m2x2`, one operand), so the +/// 16-bit or 32-bit register (`bf16x2`, `f16x2`, `e4m3x2`, `e5m2x2`, one operand), so the /// operand arity follows [`PackedConversionSourceFormat`] rather than being /// fixed at two. #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)] @@ -327,6 +327,7 @@ pub struct PackedConversion { #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)] #[serde(rename_all = "snake_case")] pub enum PackedConversionSourceFormat { + Bf16x2, E4m3x2, E5m2x2, F16x2, @@ -341,13 +342,14 @@ impl PackedConversionSourceFormat { pub fn operand_count(self) -> usize { match self { Self::F32x2 => 2, - Self::E4m3x2 | Self::E5m2x2 | Self::F16x2 => 1, + Self::Bf16x2 | Self::E4m3x2 | Self::E5m2x2 | Self::F16x2 => 1, } } /// PTX source-type token, used as the trailing `cvt` modifier. pub fn ptx_token(self) -> &'static str { match self { + Self::Bf16x2 => "bf16x2", Self::E4m3x2 => "e4m3x2", Self::E5m2x2 => "e5m2x2", Self::F16x2 => "f16x2", diff --git a/crates/cuda-intrinsics-gen/src/model/overlay.rs b/crates/cuda-intrinsics-gen/src/model/overlay.rs index 081f4b33b7..4f3e94cddb 100644 --- a/crates/cuda-intrinsics-gen/src/model/overlay.rs +++ b/crates/cuda-intrinsics-gen/src/model/overlay.rs @@ -81,6 +81,7 @@ pub struct OverlayShardFile { #[serde(default)] pub packed_conversion_fp8: Option, #[serde(default)] + pub packed_conversion_fp8_bf16x2: Option, pub packed_conversion_fp8_f16x2: Option, #[serde(default)] pub scalar_conversion: Option, @@ -505,6 +506,21 @@ pub enum PackedConversionFp8Format { E5m2x2, } +/// Compact admission for packing `bf16x2` into `e4m3x2` or `e5m2x2`. +/// +/// Pack only. The imported reverse conversions are block-scaled +/// (`scale_n2_ue8m0`) and are not part of this family. +#[derive(Debug, Clone, Deserialize)] +#[serde(deny_unknown_fields)] +pub struct PackedConversionFp8Bf16x2Admission { + pub llvm_evidence_profile: String, + pub libnvvm_evidence_profile: String, + pub runtime_validation: RuntimeValidation, + pub fp8_formats: Vec, + pub relu_variants: bool, + pub product_count: usize, +} + /// Which way an FP8/`f16x2` conversion runs. #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Deserialize)] #[serde(rename_all = "snake_case")] diff --git a/crates/cuda-intrinsics-gen/src/render/compat.rs b/crates/cuda-intrinsics-gen/src/render/compat.rs index ecc48feca7..7275698428 100644 --- a/crates/cuda-intrinsics-gen/src/render/compat.rs +++ b/crates/cuda-intrinsics-gen/src/render/compat.rs @@ -2092,7 +2092,8 @@ pub(super) fn render_compat_packed_conversion( // follows the record's own argument types rather than a fixed f32 pair. let parameter_names: Vec<&str> = match packed_conversion_source(record) { PackedConversionSourceFormat::F32x2 => vec![argument_names.0, argument_names.1], - PackedConversionSourceFormat::E4m3x2 + PackedConversionSourceFormat::Bf16x2 + | PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 | PackedConversionSourceFormat::F16x2 => vec!["packed"], }; diff --git a/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs b/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs index 045f7c5c83..f739235c41 100644 --- a/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs +++ b/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs @@ -383,7 +383,8 @@ pub(in crate::render) fn render_dialect_packed_conversion( ) .unwrap(); } - PackedConversionSourceFormat::E4m3x2 + PackedConversionSourceFormat::Bf16x2 + | PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 | PackedConversionSourceFormat::F16x2 => { let source_width = packed_conversion_source_width(record); diff --git a/crates/cuda-intrinsics-gen/src/render/families/packed.rs b/crates/cuda-intrinsics-gen/src/render/families/packed.rs index f4977f10af..f377c9a597 100644 --- a/crates/cuda-intrinsics-gen/src/render/families/packed.rs +++ b/crates/cuda-intrinsics-gen/src/render/families/packed.rs @@ -240,7 +240,9 @@ pub(in crate::render) fn packed_conversion_is_closed_recipe( let adapter_matches = conversion.adapter == match conversion.source_format { Src::F32x2 => PackedConversionAdapter::ReverseHighLowOperands, - Src::E4m3x2 | Src::E5m2x2 | Src::F16x2 => PackedConversionAdapter::Identity, + Src::Bf16x2 | Src::E4m3x2 | Src::E5m2x2 | Src::F16x2 => { + PackedConversionAdapter::Identity + } }; adapter_matches @@ -285,6 +287,20 @@ pub(in crate::render) fn packed_conversion_is_closed_recipe( Round::NearestEven, Sat::SatfiniteRelu ) + | (Src::Bf16x2, Dst::E4m3x2, Round::NearestEven, Sat::Satfinite) + | ( + Src::Bf16x2, + Dst::E4m3x2, + Round::NearestEven, + Sat::SatfiniteRelu + ) + | (Src::Bf16x2, Dst::E5m2x2, Round::NearestEven, Sat::Satfinite) + | ( + Src::Bf16x2, + Dst::E5m2x2, + Round::NearestEven, + Sat::SatfiniteRelu + ) | (Src::E4m3x2, Dst::F16x2, Round::NearestEven, Sat::None) | (Src::E4m3x2, Dst::F16x2, Round::NearestEven, Sat::Relu) | (Src::E5m2x2, Dst::F16x2, Round::NearestEven, Sat::None) @@ -305,7 +321,7 @@ pub(in crate::render) fn packed_conversion_source( /// Register width of a single packed source operand. pub(in crate::render) fn packed_conversion_source_width(record: &CatalogIntrinsic) -> u32 { match packed_conversion_source(record) { - PackedConversionSourceFormat::F16x2 => 32, + PackedConversionSourceFormat::Bf16x2 | PackedConversionSourceFormat::F16x2 => 32, PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 => 16, PackedConversionSourceFormat::F32x2 => { unreachable!("f32x2 conversions do not have a single packed source") @@ -319,7 +335,9 @@ pub(in crate::render) fn packed_conversion_rust_arguments( ) -> Vec<&'static str> { match packed_conversion_source(record) { PackedConversionSourceFormat::F32x2 => vec!["f32", "f32"], - PackedConversionSourceFormat::F16x2 => vec!["u32"], + PackedConversionSourceFormat::Bf16x2 | PackedConversionSourceFormat::F16x2 => { + vec!["u32"] + } PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 => vec!["u16"], } } @@ -330,7 +348,9 @@ pub(in crate::render) fn packed_conversion_dialect_operands( ) -> Vec<&'static str> { match packed_conversion_source(record) { PackedConversionSourceFormat::F32x2 => vec!["f32", "f32"], - PackedConversionSourceFormat::F16x2 => vec!["i32"], + PackedConversionSourceFormat::Bf16x2 | PackedConversionSourceFormat::F16x2 => { + vec!["i32"] + } PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 => vec!["i16"], } } @@ -372,7 +392,7 @@ pub(in crate::render) fn packed_conversion_constraint(record: &CatalogIntrinsic) ) { (16, PackedConversionSourceFormat::F32x2) => "=h,f,f", (32, PackedConversionSourceFormat::F32x2) => "=r,f,f", - (16, PackedConversionSourceFormat::F16x2) => "=h,r", + (16, PackedConversionSourceFormat::Bf16x2 | PackedConversionSourceFormat::F16x2) => "=h,r", (32, PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2) => "=r,h", _ => unreachable!("closed packed-conversion result width and source format"), } @@ -426,7 +446,8 @@ pub(in crate::render) fn packed_conversion_ptx_mnemonic(record: &CatalogIntrinsi conversion.adapter, match conversion.source_format { PackedConversionSourceFormat::F32x2 => PackedConversionAdapter::ReverseHighLowOperands, - PackedConversionSourceFormat::E4m3x2 + PackedConversionSourceFormat::Bf16x2 + | PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 | PackedConversionSourceFormat::F16x2 => PackedConversionAdapter::Identity, } diff --git a/crates/cuda-intrinsics-gen/src/render/lowering.rs b/crates/cuda-intrinsics-gen/src/render/lowering.rs index 36b73243f1..4f530a1627 100644 --- a/crates/cuda-intrinsics-gen/src/render/lowering.rs +++ b/crates/cuda-intrinsics-gen/src/render/lowering.rs @@ -1085,7 +1085,8 @@ fn packed_conversion_impls(catalog: &CatalogFile) -> String { ) .unwrap(); } - PackedConversionSourceFormat::E4m3x2 + PackedConversionSourceFormat::Bf16x2 + | PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 | PackedConversionSourceFormat::F16x2 => { debug_assert_eq!(conversion.adapter, PackedConversionAdapter::Identity); diff --git a/crates/cuda-intrinsics-gen/src/render/tests/mma.rs b/crates/cuda-intrinsics-gen/src/render/tests/mma.rs index 67123ff4e6..749c5eaf03 100644 --- a/crates/cuda-intrinsics-gen/src/render/tests/mma.rs +++ b/crates/cuda-intrinsics-gen/src/render/tests/mma.rs @@ -345,7 +345,7 @@ fn register_mma_rendering_preserves_apis_order_convergence_and_variants() { let repo_root = Path::new(env!("CARGO_MANIFEST_DIR")).join("../.."); let catalog = crate::resolve::resolve(&repo_root).unwrap(); validate_renderable(&catalog).unwrap(); - assert_eq!(catalog.intrinsics.len(), 1029); + assert_eq!(catalog.intrinsics.len(), 1033); let records: Vec<_> = register_mmas(&catalog).collect(); assert_eq!(records.len(), 154); let generated_records = records diff --git a/crates/cuda-intrinsics-gen/src/render/tests/scalar_misc.rs b/crates/cuda-intrinsics-gen/src/render/tests/scalar_misc.rs index b947253880..6c555c0456 100644 --- a/crates/cuda-intrinsics-gen/src/render/tests/scalar_misc.rs +++ b/crates/cuda-intrinsics-gen/src/render/tests/scalar_misc.rs @@ -108,7 +108,7 @@ fn packed_alu_and_conversion_render_exact_pure_inline_ptx_adapters() { let catalog = crate::resolve::resolve(&repo_root).unwrap(); validate_renderable(&catalog).unwrap(); assert_eq!(packed_alus(&catalog).count(), 30); - assert_eq!(packed_conversions(&catalog).count(), 18); + assert_eq!(packed_conversions(&catalog).count(), 22); let dialect = render_dialect_packed_alu(&catalog, "test-hash"); for op in [ diff --git a/crates/cuda-intrinsics-gen/src/resolve/evidence.rs b/crates/cuda-intrinsics-gen/src/resolve/evidence.rs index 30387a6c9a..694c2579fc 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/evidence.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/evidence.rs @@ -831,6 +831,18 @@ pub(super) fn validate_packed_conversion_backend_evidence( } match lowering.backend { IntrinsicBackend::LlvmNvptx => { + // `bf16x2` pack records have empty selections, so there is no typed + // NVVM route to assemble. Inline PTX is the admitted lowering. + if policy.packed_conversion.as_ref().is_some_and(|conversion| { + conversion.source_format == crate::model::PackedConversionSourceFormat::Bf16x2 + }) { + ensure!( + lowering.mechanism == BackendLoweringMechanism::InlinePtx, + "{} bf16x2 packed conversion must lower through inline PTX", + policy.id + ); + return Ok(()); + } validate_typed_llvm_evidence(policy, record)?; for stage in [ EvidenceStageKind::DeclarationCanonicalization, diff --git a/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs b/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs index b3c0a82a7d..d3e2220b8f 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs @@ -6,7 +6,8 @@ use crate::model::{ BackendLoweringMechanism, ImportedIntrinsic, IntrinsicBackend, IntrinsicSource, OverlayBackendLowering, OverlayIntrinsic, PackedConversionAdapter, - PackedConversionDestinationFormat, PackedConversionFp8Admission, PackedConversionFp8Direction, + PackedConversionDestinationFormat, PackedConversionFp8Admission, + PackedConversionFp8Bf16x2Admission, PackedConversionFp8Direction, PackedConversionFp8F16x2Admission, PackedConversionFp8Format, PackedConversionRounding, PackedConversionSaturation, PackedConversionSourceFormat, RuntimeValidation, }; @@ -35,12 +36,103 @@ pub(in crate::resolve) fn packed_conversion_recipe( ) -> Option { match conversion.source_format { PackedConversionSourceFormat::F32x2 => packed_conversion_recipe_f32x2(conversion), + PackedConversionSourceFormat::Bf16x2 => packed_conversion_recipe_fp8_bf16x2(conversion), PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 | PackedConversionSourceFormat::F16x2 => packed_conversion_recipe_fp8_f16x2(conversion), } } +/// Recipes for packing `bf16x2` into `e4m3x2` or `e5m2x2`. +/// +/// These imported records have empty selections, so the lowering is inline PTX. +/// PTX ISA 9.1 introduces them on the `sm_100f`, `sm_110f`, and `sm_120f` +/// families, not on the Ada `sm_89` floor used by `f16x2`. +pub(in crate::resolve) fn packed_conversion_recipe_fp8_bf16x2( + conversion: &crate::model::PackedConversion, +) -> Option { + match ( + conversion.source_format, + conversion.destination_format, + conversion.rounding, + conversion.saturation, + ) { + ( + PackedConversionSourceFormat::Bf16x2, + PackedConversionDestinationFormat::E4m3x2, + PackedConversionRounding::NearestEven, + PackedConversionSaturation::Satfinite, + ) => Some(PackedConversionRecipe { + id: "cvt_rn_satfinite_e4m3x2_bf16x2", + abi_id: "i1030", + operation_key: "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite", + rust_name: "cvt_rn_satfinite_e4m3x2_bf16x2", + compatibility_path: "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2", + dialect_op_type: "CvtRnSatfiniteE4m3x2Bf16x2Op", + dialect_op_name: "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2", + source_record: "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite", + llvm_symbol: "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", + llvm_result: "i16", + summary: "Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order.", + }), + ( + PackedConversionSourceFormat::Bf16x2, + PackedConversionDestinationFormat::E4m3x2, + PackedConversionRounding::NearestEven, + PackedConversionSaturation::SatfiniteRelu, + ) => Some(PackedConversionRecipe { + id: "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + abi_id: "i1031", + operation_key: "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite.relu", + rust_name: "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + compatibility_path: "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2", + dialect_op_type: "CvtRnSatfiniteReluE4m3x2Bf16x2Op", + dialect_op_name: "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2", + source_record: "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite", + llvm_symbol: "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", + llvm_result: "i16", + summary: "Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order.", + }), + ( + PackedConversionSourceFormat::Bf16x2, + PackedConversionDestinationFormat::E5m2x2, + PackedConversionRounding::NearestEven, + PackedConversionSaturation::Satfinite, + ) => Some(PackedConversionRecipe { + id: "cvt_rn_satfinite_e5m2x2_bf16x2", + abi_id: "i1032", + operation_key: "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite", + rust_name: "cvt_rn_satfinite_e5m2x2_bf16x2", + compatibility_path: "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2", + dialect_op_type: "CvtRnSatfiniteE5m2x2Bf16x2Op", + dialect_op_name: "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2", + source_record: "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite", + llvm_symbol: "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", + llvm_result: "i16", + summary: "Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order.", + }), + ( + PackedConversionSourceFormat::Bf16x2, + PackedConversionDestinationFormat::E5m2x2, + PackedConversionRounding::NearestEven, + PackedConversionSaturation::SatfiniteRelu, + ) => Some(PackedConversionRecipe { + id: "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + abi_id: "i1033", + operation_key: "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite.relu", + rust_name: "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + compatibility_path: "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2", + dialect_op_type: "CvtRnSatfiniteReluE5m2x2Bf16x2Op", + dialect_op_name: "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2", + source_record: "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite", + llvm_symbol: "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", + llvm_result: "i16", + summary: "Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order.", + }), + _ => None, + } +} + /// Recipes for the packed FP8 conversions whose other side is `f16x2`. /// /// Keyed on source as well as destination: unpacking to `f16x2` shares its @@ -434,6 +526,11 @@ pub(in crate::resolve) fn packed_conversion_source_types( vec!["f32".into(), "f32".into()], vec!["f32".into(), "f32".into()], ), + PackedConversionSourceFormat::Bf16x2 => ( + vec!["u32".into()], + vec!["i32".into()], + vec!["v2bf16".into()], + ), PackedConversionSourceFormat::F16x2 => { (vec!["u32".into()], vec!["i32".into()], vec!["v2f16".into()]) } @@ -454,18 +551,23 @@ pub(in crate::resolve) fn packed_conversion_result_width( pub(in crate::resolve) fn packed_conversion_floor( conversion: &crate::model::PackedConversion, -) -> (&'static str, &'static str) { +) -> (&'static str, Option<&'static str>, &'static str) { + // `bf16x2` to packed FP8 is a PTX ISA 9.1 family-specific conversion. + // It is not available on the Ada `sm_89` floor used by `f16x2`. + if conversion.source_format == PackedConversionSourceFormat::Bf16x2 { + return ("9.1", None, "sm_100f|sm_110f|sm_120f"); + } // FP8 on either side carries the Ada floor, including when FP8 is the // source and the destination is the older `f16x2`. if packed_conversion_uses_fp8(conversion) { - return ("8.1", "sm_89"); + return ("8.1", Some("sm_89"), "all"); } match conversion.destination_format { PackedConversionDestinationFormat::Bf16x2 | PackedConversionDestinationFormat::F16x2 => { - ("7.0", "sm_80") + ("7.0", Some("sm_80"), "all") } PackedConversionDestinationFormat::E4m3x2 | PackedConversionDestinationFormat::E5m2x2 => { - ("8.1", "sm_89") + ("8.1", Some("sm_89"), "all") } } } @@ -563,6 +665,64 @@ pub(in crate::resolve) fn expand_packed_conversion_fp8_admission( Ok(records) } +pub(in crate::resolve) fn expand_packed_conversion_fp8_bf16x2_admission( + admission: &PackedConversionFp8Bf16x2Admission, +) -> Result> { + ensure!( + admission.runtime_validation == RuntimeValidation::Unexecuted, + "FP8 bf16x2 conversion runtime validation may be marked executed only with GPU evidence" + ); + ensure!( + admission.fp8_formats + == [ + PackedConversionFp8Format::E4m3x2, + PackedConversionFp8Format::E5m2x2, + ], + "compact FP8 bf16x2 conversion admission must list the canonical two formats" + ); + ensure!( + admission.relu_variants, + "compact FP8 bf16x2 conversion admission must admit the ReLU variants" + ); + ensure!( + admission.product_count + == admission + .fp8_formats + .len() + .checked_mul(2) + .context("compact FP8 bf16x2 conversion product count overflow")? + && admission.product_count == 4, + "compact FP8 bf16x2 conversion product_count must be exactly 4" + ); + + let mut records = Vec::with_capacity(admission.product_count); + for &fp8_format in &admission.fp8_formats { + for relu in [false, true] { + let conversion = crate::model::PackedConversion { + source_format: PackedConversionSourceFormat::Bf16x2, + destination_format: match fp8_format { + PackedConversionFp8Format::E4m3x2 => PackedConversionDestinationFormat::E4m3x2, + PackedConversionFp8Format::E5m2x2 => PackedConversionDestinationFormat::E5m2x2, + }, + rounding: PackedConversionRounding::NearestEven, + saturation: if relu { + PackedConversionSaturation::SatfiniteRelu + } else { + PackedConversionSaturation::Satfinite + }, + adapter: PackedConversionAdapter::Identity, + }; + records.push(packed_conversion_overlay_record( + conversion, + &admission.llvm_evidence_profile, + &admission.libnvvm_evidence_profile, + )?); + } + } + ensure!(records.len() == admission.product_count); + Ok(records) +} + pub(in crate::resolve) fn expand_packed_conversion_fp8_f16x2_admission( admission: &PackedConversionFp8F16x2Admission, ) -> Result> { @@ -669,7 +829,7 @@ pub(in crate::resolve) fn packed_conversion_overlay_record( let result_width = packed_conversion_result_width(&conversion); let rust_result = format!("u{result_width}"); let dialect_result = format!("i{result_width}"); - let (minimum_ptx, minimum_sm) = packed_conversion_floor(&conversion); + let (minimum_ptx, minimum_sm, targets) = packed_conversion_floor(&conversion); let (rust_arguments, dialect_operands, llvm_arguments) = packed_conversion_source_types(&conversion); // `cvt` writes one destination and reads every source operand. @@ -707,9 +867,9 @@ pub(in crate::resolve) fn packed_conversion_overlay_record( convergent: false, execution_scope: "thread".into(), minimum_ptx: minimum_ptx.into(), - minimum_sm: Some(minimum_sm.into()), + minimum_sm: minimum_sm.map(str::to_owned), ptx_result: rust_result, - targets: "all".into(), + targets: targets.into(), ptx_isa_version: "9.3".into(), ptx_isa_section: "9.7.9.22 Data Movement and Conversion Instructions: cvt".into(), ptx_isa_url: "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt".into(), @@ -725,7 +885,7 @@ pub(in crate::resolve) fn packed_conversion_overlay_record( evidence_profile: evidence_profile.into(), targets: None, minimum_ptx: Some(minimum_ptx.into()), - minimum_sm: Some(minimum_sm.into()), + minimum_sm: minimum_sm.map(str::to_owned), }) .collect(), packed_atomic: None, @@ -790,7 +950,8 @@ pub(in crate::resolve) fn validate_packed_conversion_policy( // packed operand is forwarded unchanged. let expected_adapter = match conversion.source_format { PackedConversionSourceFormat::F32x2 => PackedConversionAdapter::ReverseHighLowOperands, - PackedConversionSourceFormat::E4m3x2 + PackedConversionSourceFormat::Bf16x2 + | PackedConversionSourceFormat::E4m3x2 | PackedConversionSourceFormat::E5m2x2 | PackedConversionSourceFormat::F16x2 => PackedConversionAdapter::Identity, }; @@ -808,7 +969,7 @@ pub(in crate::resolve) fn validate_packed_conversion_policy( let result_width = packed_conversion_result_width(conversion); let rust_result = format!("u{result_width}"); let dialect_result = format!("i{result_width}"); - let (minimum_ptx, minimum_sm) = packed_conversion_floor(conversion); + let (minimum_ptx, minimum_sm, targets) = packed_conversion_floor(conversion); let (expected_rust_arguments, expected_dialect_operands, expected_llvm_arguments) = packed_conversion_source_types(conversion); ensure!( @@ -864,9 +1025,9 @@ pub(in crate::resolve) fn validate_packed_conversion_policy( && !policy.convergent && policy.execution_scope == "thread" && policy.minimum_ptx == minimum_ptx - && policy.minimum_sm.as_deref() == Some(minimum_sm) + && policy.minimum_sm.as_deref() == minimum_sm && policy.ptx_result == rust_result - && policy.targets == "all" + && policy.targets == targets && policy.ptx_isa_version == "9.3" && policy.ptx_isa_section == "9.7.9.22 Data Movement and Conversion Instructions: cvt" && policy.ptx_isa_url @@ -916,7 +1077,7 @@ pub(in crate::resolve) fn validate_packed_conversion_policy( for lowering in &policy.backend_lowerings { ensure!( lowering.minimum_ptx.as_deref() == Some(minimum_ptx) - && lowering.minimum_sm.as_deref() == Some(minimum_sm) + && lowering.minimum_sm.as_deref() == minimum_sm && !lowering.evidence_profile.trim().is_empty(), "{} backend {:?} does not carry its exact packed-conversion floor", policy.id, diff --git a/crates/cuda-intrinsics-gen/src/resolve/overlay.rs b/crates/cuda-intrinsics-gen/src/resolve/overlay.rs index 03a02dbea6..4b6206f094 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/overlay.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/overlay.rs @@ -19,7 +19,7 @@ use super::guards::*; pub(super) const OVERLAY_SCHEMA: u32 = 44; pub(super) const MINIMUM_OVERLAY_SHARD_SCHEMA: u32 = 26; -pub(super) const OVERLAY_SHARD_SCHEMA: u32 = 64; +pub(super) const OVERLAY_SHARD_SCHEMA: u32 = 65; pub(super) const REGISTER_MMA_F8F6F4_SHARD_SCHEMA: u32 = 46; pub(super) const REGISTER_MMA_F8F6F4_F16_SHARD_SCHEMA: u32 = 47; pub(super) const REGISTER_MMA_MXF8F6F4_SHARD_SCHEMA: u32 = 60; @@ -32,6 +32,7 @@ pub(super) const SPARSE_MMA_FP8_F32_SHARD_SCHEMA: u32 = 64; pub(super) const PRMT_SHARD_SCHEMA: u32 = 28; pub(super) const PACKED_CONVERSION_FP8_SHARD_SCHEMA: u32 = 29; pub(super) const PACKED_CONVERSION_FP8_F16X2_SHARD_SCHEMA: u32 = 59; +pub(super) const PACKED_CONVERSION_FP8_BF16X2_SHARD_SCHEMA: u32 = 65; pub(super) const CLUSTER_SREG_SHARD_SCHEMA: u32 = 30; pub(super) const CLUSTER_BARRIER_SHARD_SCHEMA: u32 = 31; pub(super) const SPECIAL_REGISTER_SHARD_SCHEMA: u32 = 32; @@ -114,6 +115,7 @@ pub(super) fn read_overlay( let sparse_mma_ampere_float_admission = shard.sparse_mma_ordered_ampere_float.take(); let prmt_admission = shard.prmt.take(); let packed_conversion_fp8_admission = shard.packed_conversion_fp8.take(); + let packed_conversion_fp8_bf16x2_admission = shard.packed_conversion_fp8_bf16x2.take(); let packed_conversion_fp8_f16x2_admission = shard.packed_conversion_fp8_f16x2.take(); let scalar_conversion_admission = shard.scalar_conversion.take(); let scalar_arithmetic_admission = shard.scalar_arithmetic.take(); @@ -254,6 +256,13 @@ pub(super) fn read_overlay( ); shard.intrinsics = expand_packed_conversion_fp8_admission(&admission)?; } + if let Some(admission) = packed_conversion_fp8_bf16x2_admission { + ensure!( + shard.family == "packed_conversion" && shard.intrinsics.is_empty(), + "compact FP8 bf16x2 conversion admission must be the only content of a packed_conversion shard" + ); + shard.intrinsics = expand_packed_conversion_fp8_bf16x2_admission(&admission)?; + } if let Some(admission) = packed_conversion_fp8_f16x2_admission { ensure!( shard.family == "packed_conversion" && shard.intrinsics.is_empty(), @@ -482,6 +491,12 @@ pub(super) fn validate_overlay_shard_schema_with_max( "compact FP8 f16x2 conversion admission requires overlay shard schema {}", PACKED_CONVERSION_FP8_F16X2_SHARD_SCHEMA ); + ensure!( + shard.packed_conversion_fp8_bf16x2.is_none() + || shard.schema >= PACKED_CONVERSION_FP8_BF16X2_SHARD_SCHEMA, + "compact FP8 bf16x2 conversion admission requires overlay shard schema {}", + PACKED_CONVERSION_FP8_BF16X2_SHARD_SCHEMA + ); ensure!( shard.scalar_conversion.is_none() || shard.schema >= SCALAR_CONVERSION_SHARD_SCHEMA, "compact scalar-conversion admission requires overlay shard schema {}", diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs index 082bfbf15e..d78a265d52 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs @@ -12,7 +12,8 @@ use crate::model::{ MaskEncoding, MbarrierExtendedAdmission, MbarrierExtendedOperation, MovmatrixAdapter, MovmatrixParticipation, OverlayBackendLowering, OverlayFile, OverlayIntrinsic, PackedAluAdapter, PackedAluFormat, PackedAluOperation, PackedConversionAdapter, - PackedConversionDestinationFormat, PackedConversionFp8Admission, PackedConversionFp8Direction, + PackedConversionDestinationFormat, PackedConversionFp8Admission, + PackedConversionFp8Bf16x2Admission, PackedConversionFp8Direction, PackedConversionFp8F16x2Admission, PackedConversionFp8Format, PackedConversionRounding, PackedConversionSaturation, PackedConversionSourceFormat, PreSm70MemberMaskRule, PrmtAdmission, PrmtMode, RegisterMmaAccumulator, RegisterMmaAmpereFloatAdmission, RegisterMmaF8F6F4Admission, @@ -725,9 +726,9 @@ pub(super) fn packed_conversion_policy( record.memory = "none".into(); record.convergent = false; record.execution_scope = "thread".into(); - let (minimum_ptx, minimum_sm) = packed_conversion_floor(&conversion); + let (minimum_ptx, minimum_sm, _targets) = packed_conversion_floor(&conversion); record.minimum_ptx = minimum_ptx.into(); - record.minimum_sm = Some(minimum_sm.into()); + record.minimum_sm = minimum_sm.map(str::to_owned); record.ptx_result = format!("u{result_width}"); record.ptx_isa_section = "9.7.9.22 Data Movement and Conversion Instructions: cvt".into(); record.ptx_isa_url = "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt".into(); @@ -740,7 +741,7 @@ pub(super) fn packed_conversion_policy( evidence_profile: "test".into(), targets: None, minimum_ptx: Some(minimum_ptx.into()), - minimum_sm: Some(minimum_sm.into()), + minimum_sm: minimum_sm.map(str::to_owned), }) .collect(); let modifiers = packed_conversion_ptx_modifiers(&conversion); @@ -1410,6 +1411,20 @@ pub(super) fn test_fp8_conversion_admission() -> PackedConversionFp8Admission { } } +pub(super) fn test_fp8_bf16x2_conversion_admission() -> PackedConversionFp8Bf16x2Admission { + PackedConversionFp8Bf16x2Admission { + llvm_evidence_profile: "llvm-test".into(), + libnvvm_evidence_profile: "libnvvm-test".into(), + runtime_validation: RuntimeValidation::Unexecuted, + fp8_formats: vec![ + PackedConversionFp8Format::E4m3x2, + PackedConversionFp8Format::E5m2x2, + ], + relu_variants: true, + product_count: 4, + } +} + pub(super) fn test_fp8_f16x2_conversion_admission() -> PackedConversionFp8F16x2Admission { PackedConversionFp8F16x2Admission { llvm_evidence_profile: "llvm-test".into(), diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/misc_families.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/misc_families.rs index 8f50c158f2..00bea0aa6f 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/misc_families.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/misc_families.rs @@ -191,6 +191,7 @@ fn debug_control_compact_schema_is_reserved_for_aggregation() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, @@ -460,6 +461,7 @@ fn wgmma_control_compact_schema_is_reserved_for_aggregation() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs index 44707537ae..d10dccace9 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs @@ -48,8 +48,8 @@ fn overlay_manifest_loads_sorted_family_shards() { let (overlay, hash) = read_overlay(&repo_root, &repo_root.join("intrinsics/overlay.toml")).unwrap(); assert_eq!(overlay.schema, OVERLAY_SCHEMA); - assert_eq!(overlay.shards.len(), 66); - assert_eq!(overlay.intrinsics.len(), 1029); + assert_eq!(overlay.shards.len(), 67); + assert_eq!(overlay.intrinsics.len(), 1033); assert_eq!( overlay .intrinsics @@ -104,7 +104,7 @@ fn overlay_manifest_loads_sorted_family_shards() { .iter() .filter(|record| record.family == "packed_conversion") .count(), - 18 + 22 ); assert_eq!( overlay @@ -268,6 +268,7 @@ fn overlay_shard_schema_range_is_composable_and_new_fields_fail_closed() { sparse_mma_ordered_ampere_float: None, prmt, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, @@ -409,6 +410,7 @@ fn overlay_shard_schema_range_is_composable_and_new_fields_fail_closed() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: Some(test_fp8_conversion_admission()), + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, @@ -455,6 +457,7 @@ fn overlay_shard_schema_range_is_composable_and_new_fields_fail_closed() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, @@ -502,6 +505,7 @@ fn overlay_shard_schema_range_is_composable_and_new_fields_fail_closed() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs index 4d7b0dc448..ac4cfbae1b 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs @@ -171,6 +171,50 @@ fn compact_fp8_f16x2_conversion_axes_require_the_exact_closed_product() { assert!(expand_packed_conversion_fp8_f16x2_admission(&executed).is_err()); } +#[test] +fn compact_fp8_bf16x2_pack_admits_rn_satfinite_forms() { + let records = + expand_packed_conversion_fp8_bf16x2_admission(&test_fp8_bf16x2_conversion_admission()) + .unwrap(); + assert_eq!(records.len(), 4); + assert_eq!(records[0].id, "cvt_rn_satfinite_e4m3x2_bf16x2"); + assert_eq!(records[0].rust_name, "cvt_rn_satfinite_e4m3x2_bf16x2"); + assert_eq!( + records[0].expected_ptx.modifiers, + ["rn", "satfinite", "e4m3x2", "bf16x2"] + ); + assert_eq!(records[1].rust_name, "cvt_rn_satfinite_relu_e4m3x2_bf16x2"); + assert_eq!( + records[1].expected_ptx.modifiers, + ["rn", "satfinite", "relu", "e4m3x2", "bf16x2"] + ); + assert_eq!(records[2].rust_name, "cvt_rn_satfinite_e5m2x2_bf16x2"); + assert_eq!( + records[2].expected_ptx.modifiers, + ["rn", "satfinite", "e5m2x2", "bf16x2"] + ); + assert_eq!(records[3].rust_name, "cvt_rn_satfinite_relu_e5m2x2_bf16x2"); + assert_eq!( + records[3].expected_ptx.modifiers, + ["rn", "satfinite", "relu", "e5m2x2", "bf16x2"] + ); + assert!(records.iter().all(|record| { + record.rust_arguments == ["u32"] + && record.rust_result == "u16" + && record.llvm_arguments == ["v2bf16"] + && record.llvm_results == ["i16"] + && record.expected_ptx.mnemonic == "cvt" + && record.expected_ptx.operands.len() == 2 + && record.minimum_ptx == "9.1" + && record.minimum_sm.is_none() + && record.targets == "sm_100f|sm_110f|sm_120f" + && record + .source_record + .as_deref() + .is_some_and(|source| source.starts_with("int_nvvm_bf16x2_to_")) + })); +} + #[test] fn packed_atomic_closed_semantics_reject_every_unreviewed_mutation() { let valid = packed_policy("packed_atomic_add_f16x2"); @@ -424,7 +468,7 @@ fn pinned_packed_conversion_records_match_the_closed_recipes() { .iter() .filter(|record| record.family == "packed_conversion") .collect(); - assert_eq!(packed.len(), 18); + assert_eq!(packed.len(), 22); for policy in packed { let source = resolve_policy_source(policy).unwrap(); let declaration = policy diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/sreg_threadfence.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/sreg_threadfence.rs index 593cd62229..0262e4dc90 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/sreg_threadfence.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/sreg_threadfence.rs @@ -116,6 +116,7 @@ fn special_register_admission_is_closed_and_schema_gated() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/tcgen05.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/tcgen05.rs index 91aae6815a..48c6f8c1e3 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/tcgen05.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/tcgen05.rs @@ -1188,6 +1188,7 @@ fn tcgen05_compact_schema_is_reserved_for_aggregation() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, diff --git a/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs b/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs index ceb416cb46..ae2bada23a 100644 --- a/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs +++ b/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs @@ -116,6 +116,7 @@ fn clc_compact_schema_is_reserved_for_aggregation() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, @@ -349,6 +350,7 @@ fn tma_compact_schema_is_reserved_for_aggregation() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, @@ -426,6 +428,7 @@ fn cluster_memory_compact_schema_is_reserved_and_fail_closed() { sparse_mma_ordered_ampere_float: None, prmt: None, packed_conversion_fp8: None, + packed_conversion_fp8_bf16x2: None, packed_conversion_fp8_f16x2: None, scalar_conversion: None, scalar_arithmetic: None, diff --git a/crates/cuda-intrinsics/src/generated/abi_v1.rs b/crates/cuda-intrinsics/src/generated/abi_v1.rs index ce0289c0dd..f1a4298994 100644 --- a/crates/cuda-intrinsics/src/generated/abi_v1.rs +++ b/crates/cuda-intrinsics/src/generated/abi_v1.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Raw ABI functions recognized by cuda-oxide. @@ -3002,6 +3002,18 @@ pub fn i0371(_arg0: f32) -> u32 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_e4m3x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_satfinite`; expects PTX `cvt.rn.satfinite.e4m3x2.bf16x2 , ;`. +/// Available on `sm_100f or sm_110f or sm_120f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1030(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_e4m3x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn` from `int_nvvm_f16x2_to_e4m3x2_rn`; expects PTX `cvt.rn.satfinite.e4m3x2.f16x2 , ;`. @@ -3026,6 +3038,18 @@ pub fn i0259(_arg0: f32, _arg1: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_e5m2x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_satfinite`; expects PTX `cvt.rn.satfinite.e5m2x2.bf16x2 , ;`. +/// Available on `sm_100f or sm_110f or sm_120f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1032(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_e5m2x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn` from `int_nvvm_f16x2_to_e5m2x2_rn`; expects PTX `cvt.rn.satfinite.e5m2x2.f16x2 , ;`. @@ -3050,6 +3074,18 @@ pub fn i0261(_arg0: f32, _arg1: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_relu_e4m3x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite`; expects PTX `cvt.rn.satfinite.relu.e4m3x2.bf16x2 , ;`. +/// Available on `sm_100f or sm_110f or sm_120f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1031(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_relu_e4m3x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn.relu` from `int_nvvm_f16x2_to_e4m3x2_rn_relu`; expects PTX `cvt.rn.satfinite.relu.e4m3x2.f16x2 , ;`. @@ -3074,6 +3110,18 @@ pub fn i0260(_arg0: f32, _arg1: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_relu_e5m2x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite`; expects PTX `cvt.rn.satfinite.relu.e5m2x2.bf16x2 , ;`. +/// Available on `sm_100f or sm_110f or sm_120f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1033(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_relu_e5m2x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn.relu` from `int_nvvm_f16x2_to_e5m2x2_rn_relu`; expects PTX `cvt.rn.satfinite.relu.e5m2x2.f16x2 , ;`. diff --git a/crates/cuda-intrinsics/src/generated/mod.rs b/crates/cuda-intrinsics/src/generated/mod.rs index 3ab1989e8d..e9d7108860 100644 --- a/crates/cuda-intrinsics/src/generated/mod.rs +++ b/crates/cuda-intrinsics/src/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. #[doc(hidden)] @@ -143,6 +143,10 @@ pub mod convert { pub use crate::__cuda_oxide_intrinsic_abi_v1::i0827 as cvt_rn_relu_f16x2_e4m3x2; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0828 as cvt_rn_f16x2_e5m2x2; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0829 as cvt_rn_relu_f16x2_e5m2x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1030 as cvt_rn_satfinite_e4m3x2_bf16x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1031 as cvt_rn_satfinite_relu_e4m3x2_bf16x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1032 as cvt_rn_satfinite_e5m2x2_bf16x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1033 as cvt_rn_satfinite_relu_e5m2x2_bf16x2; } /// Generated `debug` intrinsic source API. @@ -2465,6 +2469,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_e4m3x2_bf16x2_reexports_abi_i1030() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e4m3x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1030; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_e4m3x2_f16x2_reexports_abi_i0822() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e4m3x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0822; @@ -2477,6 +2487,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_e5m2x2_bf16x2_reexports_abi_i1032() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e5m2x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1032; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_e5m2x2_f16x2_reexports_abi_i0824() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e5m2x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0824; @@ -2489,6 +2505,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_relu_e4m3x2_bf16x2_reexports_abi_i1031() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1031; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_relu_e4m3x2_f16x2_reexports_abi_i0823() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0823; @@ -2501,6 +2523,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_relu_e5m2x2_bf16x2_reexports_abi_i1033() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1033; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_relu_e5m2x2_f16x2_reexports_abi_i0825() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0825; diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs index f0cc44758a..11be3bf49f 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `active_mask` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs index 9a62dd9586..3d59888475 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `clc` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs index aa284eb54e..92016954be 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `cluster_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs index a3e2c44a3b..66f4e9a331 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `cluster_memory` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs index 8955f8cc8c..4b3819b184 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `cp_async` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs index b50a23afae..e4ca290041 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `debug_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs index f73fef865f..5fc647b5d8 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `dotprod` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs index 0ada366dc9..af2b892116 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `elect` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs index 52c9ffdd8d..eac85ec386 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `execution_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs index 735115d8a9..2e4e664df9 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `extended_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs index dde0b8cefd..58cb04a54d 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `integer_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs index 4f3de70be4..aad3e72fd1 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `ldmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs index 31cc8f6be6..9077c198d3 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `mbarrier_basic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs index 6c9be5ac8a..ac4530c192 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `mbarrier_extended` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs index c6aa47c031..1bdac0426f 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated target requirements and separately imported LLVM/selection facts. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs index ea446d4825..e6bc4349ed 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `movmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs index 74aa24dacd..3e3564d910 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `packed_alu` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs index c7e5547922..84a1d75719 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `packed_atomic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs index 1fd06948b5..dfd9c0cc99 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `packed_conversion` intrinsics. @@ -718,4 +718,184 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ result_range: None, }), }, + GeneratedIntrinsicTarget { + marker: "v1:i1030", + id: "cvt_rn_satfinite_e4m3x2_bf16x2", + abi_id: "i1030", + dialect_op: "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1031", + id: "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + abi_id: "i1031", + dialect_op: "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1032", + id: "cvt_rn_satfinite_e5m2x2_bf16x2", + abi_id: "i1032", + dialect_op: "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1033", + id: "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + abi_id: "i1033", + dialect_op: "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, ]; diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs index 4c4a2f8287..0cc10d341b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `prmt` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs index bdceaba496..f4a2f3ffd0 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `redux` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs index 2ccb0c0db1..6f07fa5f37 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `register_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs index 102a59e264..1301d504f9 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `scalar_arithmetic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs index 66bef39204..865de04706 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `scalar_conversion` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs index c0b78a50bf..b9144add3b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `scalar_math` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs index 712fbc5a18..32c5541427 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `sparse_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs index 8ac059d9cc..c9d6ebbeaf 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `sreg` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs index 556a7e788e..de493af2a9 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `stmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs index c58ca91aab..50123a2b5c 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `sync` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs index e667da42b1..8b7d03950a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_cp` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs index f9a55b392c..9265db6037 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_ld` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs index ab96fe16aa..c2e6bd4232 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs index edd04c7507..619f0520cf 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_mma_ws` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs index fa7086dcc6..cf420823f6 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_other` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs index a00308f306..73f2c5ef3f 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tcgen05_st` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs index 2fd40c18ed..ae35bfefa9 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `active_mask` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs index 6076281fa9..5604b23d43 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `clc` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs index cc0926521e..007dbae082 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `cluster_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs index 4635787a76..30154e620d 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `cluster_memory` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs index 4c2a20c02a..287178b79b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `cp_async` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs index 123f15d932..9ee713df49 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `debug_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs index 8081db240d..315daefccc 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `dotprod` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs index 169baef8cf..9c8b62291c 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `elect` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs index 78ab2245bb..b7315c12b8 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `execution_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs index faf89ff5ba..baf0e7250f 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `extended_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs index 053f68fc88..b21a740646 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `integer_minmax` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs index 675e1ca2cf..e1473eb479 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `ldmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs index 5e4809b884..2abdfb7ad3 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `mbarrier_basic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs index b25414b0a5..11b2388bf2 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `mbarrier_extended` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs index f34109e6e6..00e841e353 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests, grouped per family shard. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs index 66b12afa5e..b95c5c0c70 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `movmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs index 75ff1a863c..44d1197b08 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `packed_alu` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs index d8ceed3973..8677ad0fe5 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `packed_atomic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs index 6d7fad75c4..851e02619b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `packed_conversion` intrinsics. @@ -802,3 +802,225 @@ fn cvt_rn_relu_f16x2_e5m2x2_target_matches_the_catalog() { &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] ); } + +#[test] +fn cvt_rn_satfinite_e4m3x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1030").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_e4m3x2_bf16x2"); + assert_eq!(target.abi_id, "i1030"); + assert_eq!(target.dialect_op, "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2"); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} + +#[test] +fn cvt_rn_satfinite_relu_e4m3x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1031").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_relu_e4m3x2_bf16x2"); + assert_eq!(target.abi_id, "i1031"); + assert_eq!( + target.dialect_op, + "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} + +#[test] +fn cvt_rn_satfinite_e5m2x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1032").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_e5m2x2_bf16x2"); + assert_eq!(target.abi_id, "i1032"); + assert_eq!(target.dialect_op, "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2"); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} + +#[test] +fn cvt_rn_satfinite_relu_e5m2x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1033").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_relu_e5m2x2_bf16x2"); + assert_eq!(target.abi_id, "i1033"); + assert_eq!( + target.dialect_op, + "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::FamilyTarget(120), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs index 214502fbda..33867bfe35 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `prmt` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs index c4857e8d56..05ccc820e8 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `redux` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs index 48b805f004..43e7722cec 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `register_mma_float` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs index 5216dc659d..c4b28640cd 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `register_mma_s32` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs index ca62c3c7de..c7fbb39693 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `scalar_arithmetic` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs index 98af462f88..3978262f8a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `scalar_conversion` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs index 570b8b0232..bc85a4b2e2 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `scalar_math` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs index bad07267a2..330ae6da5a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `sparse_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs index b800b7b44c..929b9f768e 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `sreg` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs index aa3027531b..626cdfcf3e 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `stmatrix` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs index fc5ddff329..6a5249929b 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `sync` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs index 3dbb77ebd1..dc0ba8c9a7 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_cp` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs index b01dcd7de4..8c206fd8ed 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_ld` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs index 90092d1fb8..ae699fdd35 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_mma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs index 589e374de8..4949f5b746 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_mma_ws` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs index e2ea3a9ddc..cfa7e0a5a7 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_other` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs index a324705f5e..8c4baf0adf 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tcgen05_st` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs index cf0d62e8d4..9f98211de6 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `tma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs index 34ab9c2e50..8906b0d5dc 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `vote` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs index e0c0274b88..3e950c0d44 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `warp_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs index 0a7b559a3b..4b2a536eda 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `warp_match` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs index 17608cd3ac..468801ab0e 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `warp_shuffle` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs index 2fe8dd720f..9a33877a3e 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target tests: `wgmma_control` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs index c0323beb8d..ffb194ad0a 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `tma` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs index 73d122c8b5..f54142df36 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `vote` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs index b6829d7a5a..d2aa0b1ffe 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `warp_barrier` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs index 8517fd196b..5925ca3517 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `warp_match` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs index 20cb4a2df6..1856a4ecc4 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `warp_shuffle` intrinsics. diff --git a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs index e06d161614..4fc50858bc 100644 --- a/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs +++ b/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated intrinsic target records: `wgmma_control` intrinsics. diff --git a/crates/cuda-target-spec/src/lib.rs b/crates/cuda-target-spec/src/lib.rs index a1ef76a680..62b2f65d99 100644 --- a/crates/cuda-target-spec/src/lib.rs +++ b/crates/cuda-target-spec/src/lib.rs @@ -481,7 +481,7 @@ pub fn recorded_ptx_floor(arch: &CudaArch) -> Result "+ptx87", 88 => "+ptx88", 90 => "+ptx90", + 91 => "+ptx91", _ => unreachable!(), } } @@ -670,7 +671,8 @@ mod tests { assert_eq!(spelling_feature(74), None); assert_eq!(PtxSpelling::from_spelling(74), None); assert_eq!(PtxSpelling::round_up(74).map(PtxSpelling::get), Some(78)); - assert_eq!(PtxSpelling::round_up(91), None); + assert_eq!(PtxSpelling::round_up(91).map(PtxSpelling::get), Some(91)); + assert_eq!(PtxSpelling::round_up(92), None); } #[test] diff --git a/crates/dialect-nvvm/src/ops/generated/active_mask.rs b/crates/dialect-nvvm/src/ops/generated/active_mask.rs index 449ea2c5e7..c9179f16f9 100644 --- a/crates/dialect-nvvm/src/ops/generated/active_mask.rs +++ b/crates/dialect-nvvm/src/ops/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for the generated active warp mask. diff --git a/crates/dialect-nvvm/src/ops/generated/clc.rs b/crates/dialect-nvvm/src/ops/generated/clc.rs index 1c52191590..563d98acc2 100644 --- a/crates/dialect-nvvm/src/ops/generated/clc.rs +++ b/crates/dialect-nvvm/src/ops/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated Cluster Launch Control operations. diff --git a/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs b/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs index fe423c1171..37c03cf4c7 100644 --- a/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs +++ b/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated operations for the closed cluster-barrier family. diff --git a/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs b/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs index 488275d4d4..2602079ed1 100644 --- a/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs +++ b/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for cluster address mapping and remote shared reads. diff --git a/crates/dialect-nvvm/src/ops/generated/cp_async.rs b/crates/dialect-nvvm/src/ops/generated/cp_async.rs index bfb3bed78c..d94828b598 100644 --- a/crates/dialect-nvvm/src/ops/generated/cp_async.rs +++ b/crates/dialect-nvvm/src/ops/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for classic `cp.async` instructions. diff --git a/crates/dialect-nvvm/src/ops/generated/debug_control.rs b/crates/dialect-nvvm/src/ops/generated/debug_control.rs index d82e88c6aa..eec957b535 100644 --- a/crates/dialect-nvvm/src/ops/generated/debug_control.rs +++ b/crates/dialect-nvvm/src/ops/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated PTX debug controls. diff --git a/crates/dialect-nvvm/src/ops/generated/dotprod.rs b/crates/dialect-nvvm/src/ops/generated/dotprod.rs index ae2ab10cf8..4b1630daa4 100644 --- a/crates/dialect-nvvm/src/ops/generated/dotprod.rs +++ b/crates/dialect-nvvm/src/ops/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated packed integer dot products. diff --git a/crates/dialect-nvvm/src/ops/generated/elect.rs b/crates/dialect-nvvm/src/ops/generated/elect.rs index 666b1b6c3f..886ce9441b 100644 --- a/crates/dialect-nvvm/src/ops/generated/elect.rs +++ b/crates/dialect-nvvm/src/ops/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated warp leader-election operation. diff --git a/crates/dialect-nvvm/src/ops/generated/execution_control.rs b/crates/dialect-nvvm/src/ops/generated/execution_control.rs index ec29e13a32..7f55c03b31 100644 --- a/crates/dialect-nvvm/src/ops/generated/execution_control.rs +++ b/crates/dialect-nvvm/src/ops/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated counted-barrier, grid-dependency, and register-control operations. diff --git a/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs b/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs index 693b61f9c3..bad7f0a34b 100644 --- a/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs +++ b/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for extended floating-point min/max. diff --git a/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs b/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs index 93e4b44632..c5476d57d5 100644 --- a/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs +++ b/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated extended integer min/max. diff --git a/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs b/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs index fac1180302..50cc00f18a 100644 --- a/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs +++ b/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation and compatibility carriers for generated `ldmatrix`. diff --git a/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs b/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs index 38578d4237..8b1274120b 100644 --- a/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs +++ b/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the basic shared-memory mbarrier lifecycle. diff --git a/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs b/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs index ec97c54ac3..b3ba101af1 100644 --- a/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs +++ b/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for extended mbarrier and async-proxy instructions. diff --git a/crates/dialect-nvvm/src/ops/generated/mod.rs b/crates/dialect-nvvm/src/ops/generated/mod.rs index 3c2463da32..dd165611be 100644 --- a/crates/dialect-nvvm/src/ops/generated/mod.rs +++ b/crates/dialect-nvvm/src/ops/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. mod active_mask; diff --git a/crates/dialect-nvvm/src/ops/generated/movmatrix.rs b/crates/dialect-nvvm/src/ops/generated/movmatrix.rs index e7a3136723..dc79a6d0ca 100644 --- a/crates/dialect-nvvm/src/ops/generated/movmatrix.rs +++ b/crates/dialect-nvvm/src/ops/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated in-register matrix transpose. diff --git a/crates/dialect-nvvm/src/ops/generated/packed_alu.rs b/crates/dialect-nvvm/src/ops/generated/packed_alu.rs index 30c46117b3..752c297681 100644 --- a/crates/dialect-nvvm/src/ops/generated/packed_alu.rs +++ b/crates/dialect-nvvm/src/ops/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated packed floating-point arithmetic. diff --git a/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs b/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs index b7ac0ce129..1449007739 100644 --- a/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs +++ b/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for the closed generated packed-atomic family. diff --git a/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs b/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs index d992cb6310..380acafe85 100644 --- a/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs +++ b/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated packed conversion. @@ -432,6 +432,55 @@ impl Verify for CvtRnReluF16x2F32Op { } } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteE4m3x2Bf16x2Op; + +impl CvtRnSatfiniteE4m3x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteE4m3x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -531,6 +580,55 @@ impl Verify for CvtRnSatfiniteE4m3x2F32Op { } } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteE5m2x2Bf16x2Op; + +impl CvtRnSatfiniteE5m2x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteE5m2x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -630,6 +728,55 @@ impl Verify for CvtRnSatfiniteE5m2x2F32Op { } } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteReluE4m3x2Bf16x2Op; + +impl CvtRnSatfiniteReluE4m3x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteReluE4m3x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -729,6 +876,55 @@ impl Verify for CvtRnSatfiniteReluE4m3x2F32Op { } } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteReluE5m2x2Bf16x2Op; + +impl CvtRnSatfiniteReluE5m2x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteReluE5m2x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -937,12 +1133,16 @@ pub(super) fn register(ctx: &mut Context) { CvtRnReluF16x2E4m3x2Op::register(ctx); CvtRnReluF16x2E5m2x2Op::register(ctx); CvtRnReluF16x2F32Op::register(ctx); + CvtRnSatfiniteE4m3x2Bf16x2Op::register(ctx); CvtRnSatfiniteE4m3x2F16x2Op::register(ctx); CvtRnSatfiniteE4m3x2F32Op::register(ctx); + CvtRnSatfiniteE5m2x2Bf16x2Op::register(ctx); CvtRnSatfiniteE5m2x2F16x2Op::register(ctx); CvtRnSatfiniteE5m2x2F32Op::register(ctx); + CvtRnSatfiniteReluE4m3x2Bf16x2Op::register(ctx); CvtRnSatfiniteReluE4m3x2F16x2Op::register(ctx); CvtRnSatfiniteReluE4m3x2F32Op::register(ctx); + CvtRnSatfiniteReluE5m2x2Bf16x2Op::register(ctx); CvtRnSatfiniteReluE5m2x2F16x2Op::register(ctx); CvtRnSatfiniteReluE5m2x2F32Op::register(ctx); CvtRzBf16x2F32Op::register(ctx); diff --git a/crates/dialect-nvvm/src/ops/generated/prmt.rs b/crates/dialect-nvvm/src/ops/generated/prmt.rs index ea9a88d3ac..84215c4da5 100644 --- a/crates/dialect-nvvm/src/ops/generated/prmt.rs +++ b/crates/dialect-nvvm/src/ops/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for the closed generated `prmt` family. diff --git a/crates/dialect-nvvm/src/ops/generated/redux.rs b/crates/dialect-nvvm/src/ops/generated/redux.rs index 2236ae8a42..78f07f7d55 100644 --- a/crates/dialect-nvvm/src/ops/generated/redux.rs +++ b/crates/dialect-nvvm/src/ops/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural NVVM operations for the closed generated `redux.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/register_mma.rs b/crates/dialect-nvvm/src/ops/generated/register_mma.rs index 552a0cb3c3..15fa5f6fa1 100644 --- a/crates/dialect-nvvm/src/ops/generated/register_mma.rs +++ b/crates/dialect-nvvm/src/ops/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated register-only `mma.sync` variants. diff --git a/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs b/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs index 726a2256e7..5e5b12dd84 100644 --- a/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs +++ b/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated scalar arithmetic. diff --git a/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs b/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs index 1d2c361ae5..38bf4b6ab1 100644 --- a/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs +++ b/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated `f32` to TF32 conversions. diff --git a/crates/dialect-nvvm/src/ops/generated/scalar_math.rs b/crates/dialect-nvvm/src/ops/generated/scalar_math.rs index 424fd246dc..7948255b0c 100644 --- a/crates/dialect-nvvm/src/ops/generated/scalar_math.rs +++ b/crates/dialect-nvvm/src/ops/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated scalar math. diff --git a/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs b/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs index 578e3a3e4e..c5c7122039 100644 --- a/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs +++ b/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! One structural operation for generated sparse MMA variants. diff --git a/crates/dialect-nvvm/src/ops/generated/sreg.rs b/crates/dialect-nvvm/src/ops/generated/sreg.rs index ed4f1c62d1..11e65803d5 100644 --- a/crates/dialect-nvvm/src/ops/generated/sreg.rs +++ b/crates/dialect-nvvm/src/ops/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural NVVM operations for generated special-register reads. diff --git a/crates/dialect-nvvm/src/ops/generated/stmatrix.rs b/crates/dialect-nvvm/src/ops/generated/stmatrix.rs index c3e587573f..438602c954 100644 --- a/crates/dialect-nvvm/src/ops/generated/stmatrix.rs +++ b/crates/dialect-nvvm/src/ops/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the four generated `stmatrix` stores. diff --git a/crates/dialect-nvvm/src/ops/generated/sync.rs b/crates/dialect-nvvm/src/ops/generated/sync.rs index 02d2db9500..1c0af542b3 100644 --- a/crates/dialect-nvvm/src/ops/generated/sync.rs +++ b/crates/dialect-nvvm/src/ops/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated CTA synchronization. diff --git a/crates/dialect-nvvm/src/ops/generated/tcgen05.rs b/crates/dialect-nvvm/src/ops/generated/tcgen05.rs index 990baf92b6..0081bfbab8 100644 --- a/crates/dialect-nvvm/src/ops/generated/tcgen05.rs +++ b/crates/dialect-nvvm/src/ops/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated Tensor Core Generation 5 operations. diff --git a/crates/dialect-nvvm/src/ops/generated/tma.rs b/crates/dialect-nvvm/src/ops/generated/tma.rs index b6457435ee..781a34016c 100644 --- a/crates/dialect-nvvm/src/ops/generated/tma.rs +++ b/crates/dialect-nvvm/src/ops/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated Tensor Memory Accelerator operations. diff --git a/crates/dialect-nvvm/src/ops/generated/vote.rs b/crates/dialect-nvvm/src/ops/generated/vote.rs index def9f11d20..f287ceea9c 100644 --- a/crates/dialect-nvvm/src/ops/generated/vote.rs +++ b/crates/dialect-nvvm/src/ops/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the generated `vote.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs b/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs index 32aab1739b..4ae64bff09 100644 --- a/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs +++ b/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operation for generated warp synchronization. diff --git a/crates/dialect-nvvm/src/ops/generated/warp_match.rs b/crates/dialect-nvvm/src/ops/generated/warp_match.rs index c0c01f5ee1..fa9ca4b02f 100644 --- a/crates/dialect-nvvm/src/ops/generated/warp_match.rs +++ b/crates/dialect-nvvm/src/ops/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the generated `match.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs b/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs index 767391efc9..8f7eea91f8 100644 --- a/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs +++ b/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for the generated `shfl.sync` family. diff --git a/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs b/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs index 9872b0e4b9..ae1573a1ab 100644 --- a/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs +++ b/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Structural operations for generated WGMMA controls. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs index 7363584ce2..73d314373d 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `active_mask` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs index 54643ce57c..a7ecea053b 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `clc` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs index ea162f2edc..e37e973478 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `cluster_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs index 695d3c8f6c..1cdac28a5d 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `cluster_memory` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs index 017ed0c92e..17e60bdc85 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `cp_async` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs index 2bd7ca2bae..08cf94ad40 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `debug_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs index 14363fb24f..b99895db32 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `dotprod` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs index 7d585a7a52..9c4d09e9b3 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `elect` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs index 697372cc9e..3390f342f1 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `execution_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs index 07e0c4f542..ee5548ce2f 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `extended_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs index 7e76c72bb5..c64255b1f0 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Shared glue for the generated intrinsic dispatch shards. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs index 62a9d968a9..6e5affdda6 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `integer_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs index cc0df48deb..35f27ecfd2 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `ldmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs index 6a719bb8a1..7f363652f4 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `mbarrier_basic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs index 48657a9535..7ed943c3ab 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `mbarrier_extended` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs index c1370c9621..86942bb51d 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility path dispatch for CUDA intrinsics. @@ -517,18 +517,26 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::convert::cvt_rn_relu_satfinite_tf32_f32" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" | "cuda_device::convert::cvt_rn_relu_tf32_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" @@ -2403,12 +2411,16 @@ pub fn is_raw_generated_intrinsic_path(name: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0083" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0373" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0372" @@ -3635,18 +3647,26 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { | "cuda_device::convert::cvt_rn_relu_satfinite_tf32_f32" => Some("v1:i0373"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" | "cuda_device::convert::cvt_rn_relu_tf32_f32" => Some("v1:i0371"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" => Some("v1:i1030"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" => Some("v1:i0822"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" => Some("v1:i0259"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" => Some("v1:i1032"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" => Some("v1:i0824"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" => Some("v1:i0261"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" => Some("v1:i1031"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" => Some("v1:i0823"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" => Some("v1:i0260"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" => Some("v1:i1033"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" => Some("v1:i0825"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs index 271dc99c59..bb79fbd7e1 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `movmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs index 5f22170a73..04f3e99753 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `packed_alu` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs index 95ae3a7713..8aa3fe7de4 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `packed_atomic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs index 811e9681b3..f3c0fe926a 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `packed_conversion` intrinsics. @@ -15,8 +15,10 @@ use crate::translator::{rvalue, terminator::helpers, values::ValueMap}; use dialect_nvvm::ops::{ CvtF16x2F32Op, CvtF32x2Bf16x2Op, CvtRnF16x2E4m3x2Op, CvtRnF16x2E5m2x2Op, CvtRnReluBf16x2F32Op, CvtRnReluF16x2E4m3x2Op, CvtRnReluF16x2E5m2x2Op, CvtRnReluF16x2F32Op, - CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, CvtRnSatfiniteE5m2x2F16x2Op, - CvtRnSatfiniteE5m2x2F32Op, CvtRnSatfiniteReluE4m3x2F16x2Op, CvtRnSatfiniteReluE4m3x2F32Op, + CvtRnSatfiniteE4m3x2Bf16x2Op, CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, + CvtRnSatfiniteE5m2x2Bf16x2Op, CvtRnSatfiniteE5m2x2F16x2Op, CvtRnSatfiniteE5m2x2F32Op, + CvtRnSatfiniteReluE4m3x2Bf16x2Op, CvtRnSatfiniteReluE4m3x2F16x2Op, + CvtRnSatfiniteReluE4m3x2F32Op, CvtRnSatfiniteReluE5m2x2Bf16x2Op, CvtRnSatfiniteReluE5m2x2F16x2Op, CvtRnSatfiniteReluE5m2x2F32Op, CvtRzBf16x2F32Op, CvtRzF16x2F32Op, }; @@ -389,6 +391,45 @@ pub(super) fn try_dispatch( "cvt_rn_relu_f16x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteE4m3x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1030"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_e4m3x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; @@ -476,6 +517,45 @@ pub(super) fn try_dispatch( "cvt_rn_satfinite_e4m3x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteE5m2x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1032"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_e5m2x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; @@ -563,6 +643,45 @@ pub(super) fn try_dispatch( "cvt_rn_satfinite_e5m2x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteReluE4m3x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1031"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_relu_e4m3x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; @@ -650,6 +769,45 @@ pub(super) fn try_dispatch( "cvt_rn_satfinite_relu_e4m3x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteReluE5m2x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1033"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_relu_e5m2x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs index 47f3018ae3..f8d721ca8b 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `prmt` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs index acb9d2d2be..5c83144e68 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `redux` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs index 2d333957b1..2611f18356 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `register_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs index 2593523055..79ae16eabe 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `scalar_arithmetic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs index e91e9a70f9..0e5492bc30 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `scalar_conversion` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs index e33d6cec27..a809bbce3f 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `scalar_math` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs index 9ad19f8a62..8628801788 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `sparse_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs index 0817a6c2da..c2f045130d 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `sreg` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs index 6789f53f39..48f38ae92d 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `stmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs index e0d6949b1a..55401de0a7 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `sync` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs index 1f4f183f86..08002a8419 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_cp` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs index ee83464604..9896220458 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_ld` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs index 3c4cff1501..663e51fd17 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs index 02cf1c9317..4b1295dba5 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_other` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs index e7123a1cbb..c6de2a43dd 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tcgen05_st` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs index 299cdaee53..b300ff04a8 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `active_mask` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs index 04e44ae65b..904e52b741 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `clc` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs index 522e70ad78..ccadc03948 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `cluster_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs index b1525803a9..ad63ebc33c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `cluster_memory` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs index fab4c9c642..503578602e 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `cp_async` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs index 365e539146..c74b305720 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `debug_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs index 57bca785f5..d3a8fd0e0e 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `dotprod` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs index 63a533fa45..b9af80bb0a 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `elect` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs index 7b3ca56b96..b4dd31afaa 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `execution_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs index 01a5ccca63..e73d272b55 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `extended_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs index 9e25c77525..863ba8b1ed 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `integer_minmax` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs index 2cdb7b9a67..e673b9fc63 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `ldmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs index 569197d5fa..5c4f411082 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `mbarrier_basic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs index 0670447d85..e17753421d 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `mbarrier_extended` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs index cc78b6207b..c98ecfda38 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests, grouped per intrinsic family. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs index d52350e08f..4f69090670 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `movmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs index 453444a7f5..ea908c49c3 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `packed_alu` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs index c6c75b16f3..ed04238bcb 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `packed_atomic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs index 04c8091720..5320e77d53 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `packed_conversion` intrinsics. @@ -300,6 +300,42 @@ fn cvt_rn_relu_f16x2_f32_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cvt_rn_satfinite_e4m3x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030"), + Some("v1:i1030") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2"), + Some("v1:i1030") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1030" + )); +} +#[test] fn cvt_rn_satfinite_e4m3x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" @@ -372,6 +408,42 @@ fn cvt_rn_satfinite_e4m3x2_f32_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cvt_rn_satfinite_e5m2x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032"), + Some("v1:i1032") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2"), + Some("v1:i1032") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1032" + )); +} +#[test] fn cvt_rn_satfinite_e5m2x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" @@ -444,6 +516,42 @@ fn cvt_rn_satfinite_e5m2x2_f32_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cvt_rn_satfinite_relu_e4m3x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031"), + Some("v1:i1031") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2"), + Some("v1:i1031") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1031" + )); +} +#[test] fn cvt_rn_satfinite_relu_e4m3x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" @@ -516,6 +624,42 @@ fn cvt_rn_satfinite_relu_e4m3x2_f32_uses_only_canonical_or_compatibility_defpath )); } #[test] +fn cvt_rn_satfinite_relu_e5m2x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033"), + Some("v1:i1033") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2"), + Some("v1:i1033") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1033" + )); +} +#[test] fn cvt_rn_satfinite_relu_e5m2x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs index df75e02029..b96a6780fc 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `prmt` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs index 3963fd8c5e..f4e85b5461 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `redux` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs index 2d65aaa95e..48717a2510 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `register_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs index a228c4b86d..e3a4e6405d 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `scalar_arithmetic` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs index d0ad70c41c..ff83945495 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `scalar_conversion` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs index 90726a69e7..c588be33cc 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `scalar_math` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs index 6cf699ea72..45a403645c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `sparse_mma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs index 7ba1abc20c..ddc0dd1772 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `sreg` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs index 8c2e34bd8f..519770e956 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `stmatrix` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs index 0e2925c790..a7ec38d88c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `sync` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs index 71a2fcd9bd..6034203e0c 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `tcgen05` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs index 023347d56c..7f96d083c7 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `tma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs index bf76ce6cd9..890c4ba866 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `vote` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs index 481095a476..fe3ee0c38b 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `warp_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs index ed61420ea6..d6872b3037 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `warp_match` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs index 6698b2e851..b288e670f2 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `warp_shuffle` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs index f587d63acd..c7183598ee 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated dispatch-path tests: `wgmma_control` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs index 7d87b7e96a..deae86b3d3 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `tma` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs index 016a3ae61b..ea34959c8f 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `vote` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs index ea2efec42c..9d814e21b7 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `warp_barrier` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs index 7da0eb23dd..6f94e6c9d1 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `warp_match` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs index f445adee0d..4aed294a09 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `warp_shuffle` intrinsics. diff --git a/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs b/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs index 966b12094d..1033d63eb3 100644 --- a/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs +++ b/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated raw/compatibility dispatch arms: `wgmma_control` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs b/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs index 95b3f0daf5..9be51405e7 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `active_mask` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/clc.rs b/crates/mir-lower/src/convert/generated_intrinsics/clc.rs index fc7b13775f..0d40e5f6fc 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/clc.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `clc` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs b/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs index 79e59a186a..7a62d80d07 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `cluster_barrier` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs b/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs index 809bdd2fe6..9e9fa3c7bc 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `cluster_memory` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs b/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs index 30ec31ca6c..cdddfc983d 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `cp_async` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs b/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs index cc37b0bcac..75b335fc46 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `debug_control` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs b/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs index 5f1c71aa9e..ca122482a2 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `dotprod` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/elect.rs b/crates/mir-lower/src/convert/generated_intrinsics/elect.rs index 49d0485ba0..9006656ac8 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/elect.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `elect` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs b/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs index eb33f917ed..86132f4e38 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `execution_control` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs b/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs index 64976d1c34..f95e6b8684 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `extended_minmax` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs b/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs index dbfe47cc8c..e8a312bbb8 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `integer_minmax` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs b/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs index 3c029721fe..0c3ef9f8f0 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `ldmatrix` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs index 760b526608..0209b992f5 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `mbarrier_basic` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs index 6a2e00d2da..3f0a2e66a1 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `mbarrier_extended` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/mod.rs b/crates/mir-lower/src/convert/generated_intrinsics/mod.rs index fdaa4bf21c..98522f695b 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/mod.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces for admitted CUDA intrinsic families. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs b/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs index ce5d980b17..aa7fa53482 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `movmatrix` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs b/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs index 3d9d39e5ca..21bcf70d4b 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `packed_alu` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs b/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs index 8b10729fce..a4c8a39c2c 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `packed_atomic` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs b/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs index 7f9e1b49fb..20165c24ae 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `packed_conversion` intrinsics. @@ -16,8 +16,10 @@ use crate::convert::intrinsics::{ use dialect_nvvm::ops::{ CvtF16x2F32Op, CvtF32x2Bf16x2Op, CvtRnF16x2E4m3x2Op, CvtRnF16x2E5m2x2Op, CvtRnReluBf16x2F32Op, CvtRnReluF16x2E4m3x2Op, CvtRnReluF16x2E5m2x2Op, CvtRnReluF16x2F32Op, - CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, CvtRnSatfiniteE5m2x2F16x2Op, - CvtRnSatfiniteE5m2x2F32Op, CvtRnSatfiniteReluE4m3x2F16x2Op, CvtRnSatfiniteReluE4m3x2F32Op, + CvtRnSatfiniteE4m3x2Bf16x2Op, CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, + CvtRnSatfiniteE5m2x2Bf16x2Op, CvtRnSatfiniteE5m2x2F16x2Op, CvtRnSatfiniteE5m2x2F32Op, + CvtRnSatfiniteReluE4m3x2Bf16x2Op, CvtRnSatfiniteReluE4m3x2F16x2Op, + CvtRnSatfiniteReluE4m3x2F32Op, CvtRnSatfiniteReluE5m2x2Bf16x2Op, CvtRnSatfiniteReluE5m2x2F16x2Op, CvtRnSatfiniteReluE5m2x2F32Op, CvtRzBf16x2F32Op, CvtRzF16x2F32Op, }; @@ -179,6 +181,25 @@ impl MirToLlvmConversion for CvtRnReluF16x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteE4m3x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.e4m3x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteE4m3x2F16x2Op { fn convert( @@ -217,6 +238,25 @@ impl MirToLlvmConversion for CvtRnSatfiniteE4m3x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteE5m2x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.e5m2x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteE5m2x2F16x2Op { fn convert( @@ -255,6 +295,25 @@ impl MirToLlvmConversion for CvtRnSatfiniteE5m2x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteReluE4m3x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.relu.e4m3x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteReluE4m3x2F16x2Op { fn convert( @@ -293,6 +352,25 @@ impl MirToLlvmConversion for CvtRnSatfiniteReluE4m3x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteReluE5m2x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.relu.e5m2x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteReluE5m2x2F16x2Op { fn convert( diff --git a/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs b/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs index 7fcacbd0b9..97eecc001e 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `prmt` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/redux.rs b/crates/mir-lower/src/convert/generated_intrinsics/redux.rs index a776f60f9d..531e1180b3 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/redux.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `redux` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs b/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs index 3258cf9bd4..c9b29759e0 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `register_mma` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs b/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs index eb2f26f21f..2fc4ace5d9 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `scalar_arithmetic` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs b/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs index 0d5779dab4..57a31146f8 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `scalar_conversion` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs b/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs index 329b2242f7..00ea3592ab 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `scalar_math` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs b/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs index 6d6b104715..45736f692d 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `sparse_mma` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs b/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs index e8c341fbba..0e3feb2a2f 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `sreg` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs b/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs index 639c29ce43..c609e5f5eb 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `stmatrix` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/sync.rs b/crates/mir-lower/src/convert/generated_intrinsics/sync.rs index 0385565067..a5ead3c720 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/sync.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `sync` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs b/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs index 6819daab9d..eebaec11d4 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `tcgen05` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/tma.rs b/crates/mir-lower/src/convert/generated_intrinsics/tma.rs index 5ba6cbc944..aaa75041f6 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/tma.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `tma` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/vote.rs b/crates/mir-lower/src/convert/generated_intrinsics/vote.rs index f385fe8109..b19f250bb2 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/vote.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `vote` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs b/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs index 6d49318480..981bcf5fe8 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `warp_barrier` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs b/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs index fd4aa60aa9..32525269ec 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `warp_match` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs b/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs index d6822754f5..045085f1bf 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `warp_shuffle` intrinsics. diff --git a/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs b/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs index c60e551e5a..c3c0364a43 100644 --- a/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs +++ b/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated conversion interfaces: `wgmma_control` intrinsics. diff --git a/crates/rustc-codegen-cuda/src/generated_intrinsics.rs b/crates/rustc-codegen-cuda/src/generated_intrinsics.rs index 2c0b79d2e7..f1c3e00d4a 100644 --- a/crates/rustc-codegen-cuda/src/generated_intrinsics.rs +++ b/crates/rustc-codegen-cuda/src/generated_intrinsics.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. // LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. //! Generated collector predicates for intrinsic placeholder functions. @@ -201,12 +201,16 @@ pub(crate) const GENERATED_INTRINSIC_CANONICAL_PATHS: &[&str] = &[ "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0083", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0373", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0372", @@ -1233,12 +1237,16 @@ pub(crate) const GENERATED_INTRINSIC_PUBLIC_PATHS: &[&str] = &[ "cuda_intrinsics::convert::cvt_rn_relu_f16x2_f32", "cuda_intrinsics::convert::cvt_rn_relu_satfinite_tf32_f32", "cuda_intrinsics::convert::cvt_rn_relu_tf32_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f32", "cuda_intrinsics::convert::cvt_rn_satfinite_tf32_f32", @@ -2271,12 +2279,16 @@ pub(crate) fn is_generated_intrinsic_canonical_path(path: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0083" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0373" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0372" @@ -3310,12 +3322,16 @@ pub(crate) fn is_generated_intrinsic_compatibility_path(path: &str) -> bool { | "cuda_device::convert::cvt_rn_relu_f16x2_f32" | "cuda_device::convert::cvt_rn_relu_satfinite_tf32_f32" | "cuda_device::convert::cvt_rn_relu_tf32_f32" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f32" | "cuda_device::convert::cvt_rn_satfinite_tf32_f32" diff --git a/intrinsics/abi-v1.toml b/intrinsics/abi-v1.toml index 04cefc916a..c82c2cfa98 100644 --- a/intrinsics/abi-v1.toml +++ b/intrinsics/abi-v1.toml @@ -7173,3 +7173,31 @@ status = "active" catalog_id = "mma_sp_m16n8k64_f32_e5m2_e5m2_f32" operation_key = "matrix.mma.sp.m16n8k64.row.col.f32.e5m2.e5m2.f32.not_applicable.standard_metadata" raw_rust_signature = { safe = false, arguments = ["[f32; 4]", "[u32; 4]", "[u32; 4]", "u32", "u32"], result = "[f32; 4]" } + +[[entry]] +abi_id = "i1030" +status = "active" +catalog_id = "cvt_rn_satfinite_e4m3x2_bf16x2" +operation_key = "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite" +raw_rust_signature = { safe = true, arguments = ["u32"], result = "u16" } + +[[entry]] +abi_id = "i1031" +status = "active" +catalog_id = "cvt_rn_satfinite_relu_e4m3x2_bf16x2" +operation_key = "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite.relu" +raw_rust_signature = { safe = true, arguments = ["u32"], result = "u16" } + +[[entry]] +abi_id = "i1032" +status = "active" +catalog_id = "cvt_rn_satfinite_e5m2x2_bf16x2" +operation_key = "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite" +raw_rust_signature = { safe = true, arguments = ["u32"], result = "u16" } + +[[entry]] +abi_id = "i1033" +status = "active" +catalog_id = "cvt_rn_satfinite_relu_e5m2x2_bf16x2" +operation_key = "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite.relu" +raw_rust_signature = { safe = true, arguments = ["u32"], result = "u16" } diff --git a/intrinsics/catalog.json b/intrinsics/catalog.json index 1066e57eaf..6aced9912d 100644 --- a/intrinsics/catalog.json +++ b/intrinsics/catalog.json @@ -11,8 +11,8 @@ }, "inputs": { "imported_sha256": "3f54245b231a5b6e19f3225710c3865b7e1a106242cce873250ef5fdc32e532e", - "overlay_sha256": "ecc7d0d7d70db737303abb6271666c67e3757a2727ba2191946a6d4e66e9f717", - "abi_ledger_sha256": "fc378e74ef6da50df31de62c3fafcb3264dfb414034e5852147359fcad4184eb", + "overlay_sha256": "0cfa29280838cc6d492d29636907dfc0d36818a20f02c8eb58bc1250b47e0e62", + "abi_ledger_sha256": "0e8cec7fdd52ad28ff4365bea512101aef06f29ad7781dcde6b867ad93a0a5bc", "evidence_sha256": [ "cuda-12.0-libnvvm-12.0.140-ex2-f16.json:7102ea7d567ff2d2eb75a0384148d2e5eb6e605c26a5b98c71371fa3c46adef6", "cuda-12.9-libnvvm-12.9.86-f32x2.json:bc5bb505677989affa9ac1044da772b4eb349700dc65e16f09fffffa1f5c493e", @@ -28,6 +28,7 @@ "cuda-13.3-libnvvm-13.3.33-mbarrier-basic.json:dc5ffaa725aa8bd0201e17c1ec9375e70aa33c164bedd1aaf989653d2ed92bea", "cuda-13.3-libnvvm-13.3.33-mbarrier-extended.json:3e01efcd940547e514dcc750aa5fb83fd59cc53f8d2b888788044bcefc1152cd", "cuda-13.3-libnvvm-13.3.33-movmatrix.json:4063d72012f78cc980a2d8d7a2f7f888e1ebc5bee41ccf4969ab6f5077cf2d87", + "cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json:267354583947999f73259b2e17e8e9a5247c8c180d67b17a9a50b6875cc171a2", "cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-f16x2.json:4e1c360c16059f862409d3e6b73182ca0fcebf8b499068845aa0f49d7facc7eb", "cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8.json:167358662f1512c6ca7dba1080140834a22f5e122ae5e8db87fe8e19ff05ad9b", "cuda-13.3-libnvvm-13.3.33-prmt.json:e06d8b99575304ab780b58f665babfb7d24dfd2f57e1a9b5eac420208022520c", @@ -88,6 +89,7 @@ "rust-llvm-23.1.0-16696adc-integer-minmax.json:bf56449724c0ed5325586222a6e03bc2f53cad25ba50ac894fd6e75e86296e48", "rust-llvm-23.1.0-16696adc-mbarrier-basic.json:2663376dd611ba3c38bfdd00948ec45c6a4ad004e972f9392097eec7d19c8292", "rust-llvm-23.1.0-16696adc-mbarrier-extended.json:c53bd6fcbf47833fc102e7a6c94ca53c858ffc7e09429841bf4385b42526eb02", + "rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json:c9db101f443efb333aae997adb0f9463c2e1fb8d9f0a2354d5efdccfe3dd0c8b", "rust-llvm-23.1.0-16696adc-packed-conversion-fp8-f16x2.json:ff5c15aa45ad7177f83a669e293faaf44cf7302d5fc2b0f4163245652933d9c1", "rust-llvm-23.1.0-16696adc-packed-conversion-fp8.json:5d14d86b3ee6e17bcf6b07647c686649e97a27cc06dd45d8c31307b4705c4564", "rust-llvm-23.1.0-16696adc-prmt.json:69847316bd21a159daed299392114df5d9596faa93d5a2dfee9e12003f7b20cc", @@ -45202,18 +45204,18 @@ "summary": "Converts one f32 value with rn and returns raw TF32 bits." }, { - "id": "cvt_rn_satfinite_e4m3x2_f16x2", - "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e4m3x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_f16x2_to_e4m3x2_rn" + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite" }, "selections": [], "rust": { - "abi_id": "i0822", + "abi_id": "i1030", "module": "convert", - "name": "cvt_rn_satfinite_e4m3x2_f16x2", + "name": "cvt_rn_satfinite_e4m3x2_bf16x2", "arguments": [ "u32" ], @@ -45221,15 +45223,15 @@ "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE4m3x2F16x2Op", - "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f16x2", + "op_type": "CvtRnSatfiniteE4m3x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2", "operands": [ "i32" ], @@ -45238,9 +45240,9 @@ ] }, "llvm": { - "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn", + "symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", "arguments": [ - "v2f16" + "v2bf16" ], "results": [ "i16" @@ -45262,18 +45264,26 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 } ] }, "ptx_result": "u16", - "targets": "all", + "targets": "sm_100f|sm_110f|sm_120f", "ptx_isa_version": "9.3", "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" @@ -45282,10 +45292,10 @@ "profile": "rust-llvm-23.1.0-16696adc", "version": "LLVM version 23.1.0-rust-1.100.0-nightly", "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", - "status": "validated", + "status": "lowered", "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_89", - "ptx_feature": "+ptx81" + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91" }, "backend_lowerings": [ { @@ -45293,13 +45303,21 @@ "mechanism": "inline_ptx", "evidence_profile": "rust-llvm-23.1.0-16696adc", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 } ] } @@ -45308,82 +45326,22 @@ "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", "artifact_path": null, "build_id_prefix": null, - "status": "validated", + "status": "lowered", "stages": [ { "targets": [ - "sm_89", - "ptx81" - ], - "representation": "canonical LLVM declarations for all eight f16x2 and packed-FP8 conversions", - "stage": "declaration_canonicalization", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452).", - "artifact_kind": null, - "tool_path": null, - "tool_version": null, - "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx81" - ], - "representation": "typed LLVM intrinsic calls on a single packed source operand", - "stage": "backend_codegen", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,815-byte fixture (SHA-256 76eab1d5b8e485d46480392bfbcd8414b8c84ddd9f82cbf642da68a84452ce57) to 4,830-byte PTX (SHA-256 323876d5eb482939160e6718240f5358876c7abf50de4f000f64cee27f64c449) containing all eight exact conversions. PTX 8.0 and sm_87 were rejected with 'Cannot select: intrinsic %llvm.nvvm.f16x2.to.e4m3x2.rn'.", - "artifact_kind": null, - "tool_path": null, - "tool_version": null, - "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx81" - ], - "representation": "typed LLVM output assembled at the exact floor", - "stage": "ptx_assembly", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95) retaining the E4M3/E5M2 base and ReLU forms in both directions. sm_87 was rejected.", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" - }, - { - "targets": [ - "sm_89", - "ptx81" + "sm_100f", + "ptx91" ], - "representation": "pure inline PTX with a single packed source operand", + "representation": "pure inline PTX with a single packed bf16x2 source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,366-byte inline fixture (SHA-256 51a75a316c1334d60ef82ea936bab8ac02dfe57562890c494c017110471f0fa7) to 5,102-byte PTX (SHA-256 a45759f6aa9ead00769d3ec786eb46407d3152295ce0fb57743fa0244c2a7bd6). Every form used the reviewed $0, $1 order with =h,r constraints when packing and =r,h constraints when unpacking.", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", "artifact_kind": null, "tool_path": null, "tool_version": null, "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx81" - ], - "representation": "inline PTX assembled at the exact floor", - "stage": "ptx_assembly", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95), byte-identical to the typed-NVVM cubin, confirming both routes emit the same machine code.", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" } ] }, @@ -45392,13 +45350,21 @@ "mechanism": "inline_ptx", "evidence_profile": "cuda-13.3-libnvvm-13.3.33", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 } ] } @@ -45407,58 +45373,28 @@ "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", "build_id_prefix": "2330c0", - "status": "validated", + "status": "lowered", "stages": [ { "targets": [ - "compute_89", - "ptx93" + "sm_100f", + "ptx91" ], - "representation": "pure inline PTX with a single packed source operand", + "representation": "pure inline PTX with a single packed bf16x2 source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", "artifact_kind": null, "tool_path": null, "tool_version": null, "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx93" - ], - "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", - "stage": "ptx_assembly", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" - }, - { - "targets": [ - "sm_89", - "ptx93" - ], - "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", - "stage": "device_link", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "CUDA 13.3 nvJitLink 13.3", - "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f16x2", + "source_format": "bf16x2", "destination_format": "e4m3x2", "rounding": "nearest_even", "saturation": "satfinite", @@ -45471,7 +45407,7 @@ "rn", "satfinite", "e4m3x2", - "f16x2" + "bf16x2" ], "operands": [ { @@ -45482,51 +45418,48 @@ } ] }, - "summary": "Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order." + "summary": "Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order." }, { - "id": "cvt_rn_satfinite_e4m3x2_f32", - "operation_key": "packed.convert.f32x2.e4m3x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e4m3x2_f16x2", + "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_ff_to_e4m3x2_rn" + "source_record": "int_nvvm_f16x2_to_e4m3x2_rn" }, "selections": [], "rust": { - "abi_id": "i0259", + "abi_id": "i0822", "module": "convert", - "name": "cvt_rn_satfinite_e4m3x2_f32", + "name": "cvt_rn_satfinite_e4m3x2_f16x2", "arguments": [ - "f32", - "f32" + "u32" ], "result": "u16", "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE4m3x2F32Op", - "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f32", + "op_type": "CvtRnSatfiniteE4m3x2F16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f16x2", "operands": [ - "f32", - "f32" + "i32" ], "results": [ "i16" ] }, "llvm": { - "symbol": "llvm.nvvm.ff.to.e4m3x2.rn", + "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn", "arguments": [ - "f32", - "f32" + "v2f16" ], "results": [ "i16" @@ -45576,7 +45509,7 @@ "backend_lowerings": [ { "backend": "llvm_nvptx", - "mechanism": "typed_nvvm", + "mechanism": "inline_ptx", "evidence_profile": "rust-llvm-23.1.0-16696adc", "target": { "minimum_ptx": "8.1", @@ -45601,11 +45534,11 @@ "sm_89", "ptx81" ], - "representation": "canonical LLVM declarations for all four packed-FP8 conversions", + "representation": "canonical LLVM declarations for all eight f16x2 and packed-FP8 conversions", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Pinned LLVM canonicalized the four i16(float, float) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 1,724-byte IR (SHA-256 d6a1f164920582f756e24a01e4ac1ccd87042b45fb7202bcabfc27dad23e095a).", + "detail": "Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452).", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45616,11 +45549,11 @@ "sm_89", "ptx81" ], - "representation": "typed LLVM intrinsic calls in PTX high, low source order", + "representation": "typed LLVM intrinsic calls on a single packed source operand", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,128-byte fixture (SHA-256 c34fcffc96185a5a3d746fb99435955858780171b842b88b1939e94469ca0fe7) to 977-byte PTX (SHA-256 8da714495549c455c6a77d168e5d193463037164645fa90f8cc505dccbc1adb1) containing all four exact conversions. PTX 8.0 and sm_87 were rejected.", + "detail": "Pinned LLVM lowered the 1,815-byte fixture (SHA-256 76eab1d5b8e485d46480392bfbcd8414b8c84ddd9f82cbf642da68a84452ce57) to 4,830-byte PTX (SHA-256 323876d5eb482939160e6718240f5358876c7abf50de4f000f64cee27f64c449) containing all eight exact conversions. PTX 8.0 and sm_87 were rejected with 'Cannot select: intrinsic %llvm.nvvm.f16x2.to.e4m3x2.rn'.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45635,7 +45568,7 @@ "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 8758c825ff05674628381bbf19cfe4ba99f5077971ce6066e863c292fb42b82a) retaining E4M3/E5M2 base and ReLU forms.", + "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95) retaining the E4M3/E5M2 base and ReLU forms in both directions. sm_87 was rejected.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -45646,11 +45579,11 @@ "sm_89", "ptx81" ], - "representation": "pure inline PTX with u16 output and public low, high inputs", + "representation": "pure inline PTX with a single packed source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,022-byte inline fixture (SHA-256 5919ef9b476cd61b2e276cd2726e6db9cf443ac6dced739366284450a50087fa) to 1,305-byte PTX (SHA-256 ac6134577dc1f9083d0373e5883bd028f2b7d2cd27891c600598725eb10c1701). Every form used the reviewed $0, $2, $1 order and =h,f,f constraints.", + "detail": "Pinned LLVM lowered the 1,366-byte inline fixture (SHA-256 51a75a316c1334d60ef82ea936bab8ac02dfe57562890c494c017110471f0fa7) to 5,102-byte PTX (SHA-256 a45759f6aa9ead00769d3ec786eb46407d3152295ce0fb57743fa0244c2a7bd6). Every form used the reviewed $0, $1 order with =h,r constraints when packing and =r,h constraints when unpacking.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45665,7 +45598,7 @@ "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 f238c85c61f29667b999a4bc0fadcbfafa85b980558687e1b6ba6e1f3217482d). PTX 8.0 and sm_87 failed with the expected floor diagnostics.", + "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95), byte-identical to the typed-NVVM cubin, confirming both routes emit the same machine code.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -45690,7 +45623,7 @@ } }, "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", + "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", "build_id_prefix": "2330c0", "status": "validated", @@ -45700,11 +45633,11 @@ "compute_89", "ptx93" ], - "representation": "pure inline PTX with u16 output and public low, high inputs", + "representation": "pure inline PTX with a single packed source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", + "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45715,11 +45648,11 @@ "sm_89", "ptx93" ], - "representation": "libNVVM-emitted packed-FP8 PTX", + "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", + "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -45730,89 +45663,89 @@ "sm_89", "ptx93" ], - "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", + "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", + "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "V13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3", "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f32x2", + "source_format": "f16x2", "destination_format": "e4m3x2", "rounding": "nearest_even", "saturation": "satfinite", - "adapter": "reverse_high_low_operands" + "adapter": "identity" }, - "lowering": "generated_packed_conversion_backend", + "lowering": "generated_packed_conversion_inline_ptx", "expected_ptx": { "mnemonic": "cvt", "modifiers": [ "rn", "satfinite", "e4m3x2", - "f32" + "f16x2" ], "operands": [ { "kind": "register" }, - { - "kind": "register" - }, { "kind": "register" } ] }, - "summary": "Converts two f32 values to packed e4m3x2 with nearest-even finite saturation and the first argument in the low byte." + "summary": "Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order." }, { - "id": "cvt_rn_satfinite_e5m2x2_f16x2", - "operation_key": "packed.convert.f16x2.e5m2x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e4m3x2_f32", + "operation_key": "packed.convert.f32x2.e4m3x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_f16x2_to_e5m2x2_rn" + "source_record": "int_nvvm_ff_to_e4m3x2_rn" }, "selections": [], "rust": { - "abi_id": "i0824", + "abi_id": "i0259", "module": "convert", - "name": "cvt_rn_satfinite_e5m2x2_f16x2", + "name": "cvt_rn_satfinite_e4m3x2_f32", "arguments": [ - "u32" + "f32", + "f32" ], "result": "u16", "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE5m2x2F16x2Op", - "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f16x2", + "op_type": "CvtRnSatfiniteE4m3x2F32Op", + "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f32", "operands": [ - "i32" + "f32", + "f32" ], "results": [ "i16" ] }, "llvm": { - "symbol": "llvm.nvvm.f16x2.to.e5m2x2.rn", + "symbol": "llvm.nvvm.ff.to.e4m3x2.rn", "arguments": [ - "v2f16" + "f32", + "f32" ], "results": [ "i16" @@ -45862,7 +45795,7 @@ "backend_lowerings": [ { "backend": "llvm_nvptx", - "mechanism": "inline_ptx", + "mechanism": "typed_nvvm", "evidence_profile": "rust-llvm-23.1.0-16696adc", "target": { "minimum_ptx": "8.1", @@ -45887,11 +45820,11 @@ "sm_89", "ptx81" ], - "representation": "canonical LLVM declarations for all eight f16x2 and packed-FP8 conversions", + "representation": "canonical LLVM declarations for all four packed-FP8 conversions", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452).", + "detail": "Pinned LLVM canonicalized the four i16(float, float) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 1,724-byte IR (SHA-256 d6a1f164920582f756e24a01e4ac1ccd87042b45fb7202bcabfc27dad23e095a).", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45902,11 +45835,11 @@ "sm_89", "ptx81" ], - "representation": "typed LLVM intrinsic calls on a single packed source operand", + "representation": "typed LLVM intrinsic calls in PTX high, low source order", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,815-byte fixture (SHA-256 76eab1d5b8e485d46480392bfbcd8414b8c84ddd9f82cbf642da68a84452ce57) to 4,830-byte PTX (SHA-256 323876d5eb482939160e6718240f5358876c7abf50de4f000f64cee27f64c449) containing all eight exact conversions. PTX 8.0 and sm_87 were rejected with 'Cannot select: intrinsic %llvm.nvvm.f16x2.to.e4m3x2.rn'.", + "detail": "Pinned LLVM lowered the 1,128-byte fixture (SHA-256 c34fcffc96185a5a3d746fb99435955858780171b842b88b1939e94469ca0fe7) to 977-byte PTX (SHA-256 8da714495549c455c6a77d168e5d193463037164645fa90f8cc505dccbc1adb1) containing all four exact conversions. PTX 8.0 and sm_87 were rejected.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45921,7 +45854,7 @@ "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95) retaining the E4M3/E5M2 base and ReLU forms in both directions. sm_87 was rejected.", + "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 8758c825ff05674628381bbf19cfe4ba99f5077971ce6066e863c292fb42b82a) retaining E4M3/E5M2 base and ReLU forms.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -45932,11 +45865,11 @@ "sm_89", "ptx81" ], - "representation": "pure inline PTX with a single packed source operand", + "representation": "pure inline PTX with u16 output and public low, high inputs", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,366-byte inline fixture (SHA-256 51a75a316c1334d60ef82ea936bab8ac02dfe57562890c494c017110471f0fa7) to 5,102-byte PTX (SHA-256 a45759f6aa9ead00769d3ec786eb46407d3152295ce0fb57743fa0244c2a7bd6). Every form used the reviewed $0, $1 order with =h,r constraints when packing and =r,h constraints when unpacking.", + "detail": "Pinned LLVM lowered the 1,022-byte inline fixture (SHA-256 5919ef9b476cd61b2e276cd2726e6db9cf443ac6dced739366284450a50087fa) to 1,305-byte PTX (SHA-256 ac6134577dc1f9083d0373e5883bd028f2b7d2cd27891c600598725eb10c1701). Every form used the reviewed $0, $2, $1 order and =h,f,f constraints.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45951,7 +45884,7 @@ "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95), byte-identical to the typed-NVVM cubin, confirming both routes emit the same machine code.", + "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 f238c85c61f29667b999a4bc0fadcbfafa85b980558687e1b6ba6e1f3217482d). PTX 8.0 and sm_87 failed with the expected floor diagnostics.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -45976,7 +45909,7 @@ } }, "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", "build_id_prefix": "2330c0", "status": "validated", @@ -45986,11 +45919,11 @@ "compute_89", "ptx93" ], - "representation": "pure inline PTX with a single packed source operand", + "representation": "pure inline PTX with u16 output and public low, high inputs", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", + "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -46001,11 +45934,11 @@ "sm_89", "ptx93" ], - "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", + "representation": "libNVVM-emitted packed-FP8 PTX", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", + "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -46016,89 +45949,89 @@ "sm_89", "ptx93" ], - "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", + "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", + "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "CUDA 13.3 nvJitLink 13.3", + "tool_version": "V13.3.33", "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f16x2", - "destination_format": "e5m2x2", + "source_format": "f32x2", + "destination_format": "e4m3x2", "rounding": "nearest_even", "saturation": "satfinite", - "adapter": "identity" + "adapter": "reverse_high_low_operands" }, - "lowering": "generated_packed_conversion_inline_ptx", + "lowering": "generated_packed_conversion_backend", "expected_ptx": { "mnemonic": "cvt", "modifiers": [ "rn", "satfinite", - "e5m2x2", - "f16x2" + "e4m3x2", + "f32" ], "operands": [ { "kind": "register" }, + { + "kind": "register" + }, { "kind": "register" } ] }, - "summary": "Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order." + "summary": "Converts two f32 values to packed e4m3x2 with nearest-even finite saturation and the first argument in the low byte." }, { - "id": "cvt_rn_satfinite_e5m2x2_f32", - "operation_key": "packed.convert.f32x2.e5m2x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e5m2x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_ff_to_e5m2x2_rn" + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite" }, "selections": [], "rust": { - "abi_id": "i0261", + "abi_id": "i1032", "module": "convert", - "name": "cvt_rn_satfinite_e5m2x2_f32", + "name": "cvt_rn_satfinite_e5m2x2_bf16x2", "arguments": [ - "f32", - "f32" + "u32" ], "result": "u16", "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE5m2x2F32Op", - "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f32", + "op_type": "CvtRnSatfiniteE5m2x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2", "operands": [ - "f32", - "f32" + "i32" ], "results": [ "i16" ] }, "llvm": { - "symbol": "llvm.nvvm.ff.to.e5m2x2.rn", + "symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", "arguments": [ - "f32", - "f32" + "v2bf16" ], "results": [ "i16" @@ -46120,18 +46053,26 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 } ] }, "ptx_result": "u16", - "targets": "all", + "targets": "sm_100f|sm_110f|sm_120f", "ptx_isa_version": "9.3", "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" @@ -46140,24 +46081,32 @@ "profile": "rust-llvm-23.1.0-16696adc", "version": "LLVM version 23.1.0-rust-1.100.0-nightly", "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", - "status": "validated", + "status": "lowered", "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_89", - "ptx_feature": "+ptx81" + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91" }, "backend_lowerings": [ { "backend": "llvm_nvptx", - "mechanism": "typed_nvvm", + "mechanism": "inline_ptx", "evidence_profile": "rust-llvm-23.1.0-16696adc", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 } ] } @@ -46166,82 +46115,22 @@ "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", "artifact_path": null, "build_id_prefix": null, - "status": "validated", + "status": "lowered", "stages": [ { "targets": [ - "sm_89", - "ptx81" - ], - "representation": "canonical LLVM declarations for all four packed-FP8 conversions", - "stage": "declaration_canonicalization", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "Pinned LLVM canonicalized the four i16(float, float) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 1,724-byte IR (SHA-256 d6a1f164920582f756e24a01e4ac1ccd87042b45fb7202bcabfc27dad23e095a).", - "artifact_kind": null, - "tool_path": null, - "tool_version": null, - "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx81" - ], - "representation": "typed LLVM intrinsic calls in PTX high, low source order", - "stage": "backend_codegen", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,128-byte fixture (SHA-256 c34fcffc96185a5a3d746fb99435955858780171b842b88b1939e94469ca0fe7) to 977-byte PTX (SHA-256 8da714495549c455c6a77d168e5d193463037164645fa90f8cc505dccbc1adb1) containing all four exact conversions. PTX 8.0 and sm_87 were rejected.", - "artifact_kind": null, - "tool_path": null, - "tool_version": null, - "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx81" - ], - "representation": "typed LLVM output assembled at the exact floor", - "stage": "ptx_assembly", - "mechanism": "typed_nvvm", - "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 8758c825ff05674628381bbf19cfe4ba99f5077971ce6066e863c292fb42b82a) retaining E4M3/E5M2 base and ReLU forms.", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" - }, - { - "targets": [ - "sm_89", - "ptx81" + "sm_100f", + "ptx91" ], - "representation": "pure inline PTX with u16 output and public low, high inputs", + "representation": "pure inline PTX with a single packed bf16x2 source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Pinned LLVM lowered the 1,022-byte inline fixture (SHA-256 5919ef9b476cd61b2e276cd2726e6db9cf443ac6dced739366284450a50087fa) to 1,305-byte PTX (SHA-256 ac6134577dc1f9083d0373e5883bd028f2b7d2cd27891c600598725eb10c1701). Every form used the reviewed $0, $2, $1 order and =h,f,f constraints.", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", "artifact_kind": null, "tool_path": null, "tool_version": null, "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx81" - ], - "representation": "inline PTX assembled at the exact floor", - "stage": "ptx_assembly", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 f238c85c61f29667b999a4bc0fadcbfafa85b980558687e1b6ba6e1f3217482d). PTX 8.0 and sm_87 failed with the expected floor diagnostics.", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" } ] }, @@ -46250,114 +46139,89 @@ "mechanism": "inline_ptx", "evidence_profile": "cuda-13.3-libnvvm-13.3.33", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 } ] } }, "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", + "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", "build_id_prefix": "2330c0", - "status": "validated", + "status": "lowered", "stages": [ { "targets": [ - "compute_89", - "ptx93" + "sm_100f", + "ptx91" ], - "representation": "pure inline PTX with u16 output and public low, high inputs", + "representation": "pure inline PTX with a single packed bf16x2 source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", "artifact_kind": null, "tool_path": null, "tool_version": null, "tool_sha256": null - }, - { - "targets": [ - "sm_89", - "ptx93" - ], - "representation": "libNVVM-emitted packed-FP8 PTX", - "stage": "ptx_assembly", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" - }, - { - "targets": [ - "sm_89", - "ptx93" - ], - "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", - "stage": "device_link", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", - "artifact_kind": "cubin", - "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "V13.3.33", - "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f32x2", + "source_format": "bf16x2", "destination_format": "e5m2x2", "rounding": "nearest_even", "saturation": "satfinite", - "adapter": "reverse_high_low_operands" + "adapter": "identity" }, - "lowering": "generated_packed_conversion_backend", + "lowering": "generated_packed_conversion_inline_ptx", "expected_ptx": { "mnemonic": "cvt", "modifiers": [ "rn", "satfinite", "e5m2x2", - "f32" + "bf16x2" ], "operands": [ { "kind": "register" }, - { - "kind": "register" - }, { "kind": "register" } ] }, - "summary": "Converts two f32 values to packed e5m2x2 with nearest-even finite saturation and the first argument in the low byte." + "summary": "Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order." }, { - "id": "cvt_rn_satfinite_relu_e4m3x2_f16x2", - "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite.relu", + "id": "cvt_rn_satfinite_e5m2x2_f16x2", + "operation_key": "packed.convert.f16x2.e5m2x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_f16x2_to_e4m3x2_rn_relu" + "source_record": "int_nvvm_f16x2_to_e5m2x2_rn" }, "selections": [], "rust": { - "abi_id": "i0823", + "abi_id": "i0824", "module": "convert", - "name": "cvt_rn_satfinite_relu_e4m3x2_f16x2", + "name": "cvt_rn_satfinite_e5m2x2_f16x2", "arguments": [ "u32" ], @@ -46365,15 +46229,15 @@ "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteReluE4m3x2F16x2Op", - "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_f16x2", + "op_type": "CvtRnSatfiniteE5m2x2F16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f16x2", "operands": [ "i32" ], @@ -46382,7 +46246,7 @@ ] }, "llvm": { - "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn.relu", + "symbol": "llvm.nvvm.f16x2.to.e5m2x2.rn", "arguments": [ "v2f16" ], @@ -46603,9 +46467,9 @@ ], "packed_conversion": { "source_format": "f16x2", - "destination_format": "e4m3x2", + "destination_format": "e5m2x2", "rounding": "nearest_even", - "saturation": "satfinite_relu", + "saturation": "satfinite", "adapter": "identity" }, "lowering": "generated_packed_conversion_inline_ptx", @@ -46614,8 +46478,7 @@ "modifiers": [ "rn", "satfinite", - "relu", - "e4m3x2", + "e5m2x2", "f16x2" ], "operands": [ @@ -46627,21 +46490,21 @@ } ] }, - "summary": "Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order." + "summary": "Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order." }, { - "id": "cvt_rn_satfinite_relu_e4m3x2_f32", - "operation_key": "packed.convert.f32x2.e4m3x2.nearest_even.satfinite.relu", + "id": "cvt_rn_satfinite_e5m2x2_f32", + "operation_key": "packed.convert.f32x2.e5m2x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_ff_to_e4m3x2_rn_relu" + "source_record": "int_nvvm_ff_to_e5m2x2_rn" }, "selections": [], "rust": { - "abi_id": "i0260", + "abi_id": "i0261", "module": "convert", - "name": "cvt_rn_satfinite_relu_e4m3x2_f32", + "name": "cvt_rn_satfinite_e5m2x2_f32", "arguments": [ "f32", "f32" @@ -46650,15 +46513,15 @@ "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" ] }, "dialect": { - "op_type": "CvtRnSatfiniteReluE4m3x2F32Op", - "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_f32", + "op_type": "CvtRnSatfiniteE5m2x2F32Op", + "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f32", "operands": [ "f32", "f32" @@ -46668,7 +46531,798 @@ ] }, "llvm": { - "symbol": "llvm.nvvm.ff.to.e4m3x2.rn.relu", + "symbol": "llvm.nvvm.ff.to.e5m2x2.rn", + "arguments": [ + "f32", + "f32" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + }, + "ptx_result": "u16", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_89", + "ptx_feature": "+ptx81" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.0-16696adc", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "canonical LLVM declarations for all four packed-FP8 conversions", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Pinned LLVM canonicalized the four i16(float, float) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 1,724-byte IR (SHA-256 d6a1f164920582f756e24a01e4ac1ccd87042b45fb7202bcabfc27dad23e095a).", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "typed LLVM intrinsic calls in PTX high, low source order", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Pinned LLVM lowered the 1,128-byte fixture (SHA-256 c34fcffc96185a5a3d746fb99435955858780171b842b88b1939e94469ca0fe7) to 977-byte PTX (SHA-256 8da714495549c455c6a77d168e5d193463037164645fa90f8cc505dccbc1adb1) containing all four exact conversions. PTX 8.0 and sm_87 were rejected.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "typed LLVM output assembled at the exact floor", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 8758c825ff05674628381bbf19cfe4ba99f5077971ce6066e863c292fb42b82a) retaining E4M3/E5M2 base and ReLU forms.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "pure inline PTX with u16 output and public low, high inputs", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Pinned LLVM lowered the 1,022-byte inline fixture (SHA-256 5919ef9b476cd61b2e276cd2726e6db9cf443ac6dced739366284450a50087fa) to 1,305-byte PTX (SHA-256 ac6134577dc1f9083d0373e5883bd028f2b7d2cd27891c600598725eb10c1701). Every form used the reviewed $0, $2, $1 order and =h,f,f constraints.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "inline PTX assembled at the exact floor", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 3,368-byte cubin (SHA-256 f238c85c61f29667b999a4bc0fadcbfafa85b980558687e1b6ba6e1f3217482d). PTX 8.0 and sm_87 failed with the expected floor diagnostics.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", + "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "2330c0", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_89", + "ptx93" + ], + "representation": "pure inline PTX with u16 output and public low, high inputs", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM-emitted packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "V13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ] + } + ], + "packed_conversion": { + "source_format": "f32x2", + "destination_format": "e5m2x2", + "rounding": "nearest_even", + "saturation": "satfinite", + "adapter": "reverse_high_low_operands" + }, + "lowering": "generated_packed_conversion_backend", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e5m2x2", + "f32" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts two f32 values to packed e5m2x2 with nearest-even finite saturation and the first argument in the low byte." + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite.relu", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite" + }, + "selections": [], + "rust": { + "abi_id": "i1031", + "module": "convert", + "name": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "arguments": [ + "u32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteReluE4m3x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "operands": [ + "i32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", + "arguments": [ + "v2bf16" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 + } + ] + }, + "ptx_result": "u16", + "targets": "sm_100f|sm_110f|sm_120f", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "lowered", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "inline_ptx", + "evidence_profile": "rust-llvm-23.1.0-16696adc", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", + "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "2330c0", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + } + ] + } + ], + "packed_conversion": { + "source_format": "bf16x2", + "destination_format": "e4m3x2", + "rounding": "nearest_even", + "saturation": "satfinite_relu", + "adapter": "identity" + }, + "lowering": "generated_packed_conversion_inline_ptx", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order." + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_f16x2", + "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite.relu", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_f16x2_to_e4m3x2_rn_relu" + }, + "selections": [], + "rust": { + "abi_id": "i0823", + "module": "convert", + "name": "cvt_rn_satfinite_relu_e4m3x2_f16x2", + "arguments": [ + "u32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteReluE4m3x2F16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_f16x2", + "operands": [ + "i32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn.relu", + "arguments": [ + "v2f16" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + }, + "ptx_result": "u16", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_89", + "ptx_feature": "+ptx81" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "inline_ptx", + "evidence_profile": "rust-llvm-23.1.0-16696adc", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "canonical LLVM declarations for all eight f16x2 and packed-FP8 conversions", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452).", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "typed LLVM intrinsic calls on a single packed source operand", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Pinned LLVM lowered the 1,815-byte fixture (SHA-256 76eab1d5b8e485d46480392bfbcd8414b8c84ddd9f82cbf642da68a84452ce57) to 4,830-byte PTX (SHA-256 323876d5eb482939160e6718240f5358876c7abf50de4f000f64cee27f64c449) containing all eight exact conversions. PTX 8.0 and sm_87 were rejected with 'Cannot select: intrinsic %llvm.nvvm.f16x2.to.e4m3x2.rn'.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "typed LLVM output assembled at the exact floor", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95) retaining the E4M3/E5M2 base and ReLU forms in both directions. sm_87 was rejected.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Pinned LLVM lowered the 1,366-byte inline fixture (SHA-256 51a75a316c1334d60ef82ea936bab8ac02dfe57562890c494c017110471f0fa7) to 5,102-byte PTX (SHA-256 a45759f6aa9ead00769d3ec786eb46407d3152295ce0fb57743fa0244c2a7bd6). Every form used the reviewed $0, $1 order with =h,r constraints when packing and =r,h constraints when unpacking.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "inline PTX assembled at the exact floor", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas produced a 1,464-byte cubin (SHA-256 7d86e0e31bcf5c2c446f9426a20db5359dfb53f0bd1bffcc9032aa890f9e4c95), byte-identical to the typed-NVVM cubin, confirming both routes emit the same machine code.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", + "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "2330c0", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_89", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ] + } + ], + "packed_conversion": { + "source_format": "f16x2", + "destination_format": "e4m3x2", + "rounding": "nearest_even", + "saturation": "satfinite_relu", + "adapter": "identity" + }, + "lowering": "generated_packed_conversion_inline_ptx", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e4m3x2", + "f16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order." + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_f32", + "operation_key": "packed.convert.f32x2.e4m3x2.nearest_even.satfinite.relu", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_ff_to_e4m3x2_rn_relu" + }, + "selections": [], + "rust": { + "abi_id": "i0260", + "module": "convert", + "name": "cvt_rn_satfinite_relu_e4m3x2_f32", + "arguments": [ + "f32", + "f32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteReluE4m3x2F32Op", + "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_f32", + "operands": [ + "f32", + "f32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.ff.to.e4m3x2.rn.relu", "arguments": [ "f32", "f32" @@ -46919,6 +47573,224 @@ }, "summary": "Converts two f32 values to packed e4m3x2 with nearest-even finite saturation, ReLU, and the first argument in the low byte." }, + { + "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite.relu", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite" + }, + "selections": [], + "rust": { + "abi_id": "i1033", + "module": "convert", + "name": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "arguments": [ + "u32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteReluE5m2x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "operands": [ + "i32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", + "arguments": [ + "v2bf16" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 + } + ] + }, + "ptx_result": "u16", + "targets": "sm_100f|sm_110f|sm_120f", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.0-16696adc", + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "status": "lowered", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "inline_ptx", + "evidence_profile": "rust-llvm-23.1.0-16696adc", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 + } + ] + } + }, + "version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "artifact_path": null, + "build_id_prefix": null, + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 120 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", + "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "2330c0", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + } + ] + } + ], + "packed_conversion": { + "source_format": "bf16x2", + "destination_format": "e5m2x2", + "rounding": "nearest_even", + "saturation": "satfinite_relu", + "adapter": "identity" + }, + "lowering": "generated_packed_conversion_inline_ptx", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order." + }, { "id": "cvt_rn_satfinite_relu_e5m2x2_f16x2", "operation_key": "packed.convert.f16x2.e5m2x2.nearest_even.satfinite.relu", diff --git a/intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json b/intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json new file mode 100644 index 0000000000..033c904a10 --- /dev/null +++ b/intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json @@ -0,0 +1,200 @@ +{ + "schema": 6, + "backend_profile": "cuda-13.3-libnvvm-13.3.33", + "backend_kind": "lib_nvvm", + "llvm_revision": "not-applicable-libnvvm-ir-2.0", + "backend_version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", + "backend_sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "2330c0", + "nvvm_ir_version": "2.0", + "debug_ir_version": "3.2", + "records": [ + { + "id": "cvt_rn_satfinite_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + } + ] +} diff --git a/intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json b/intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json new file mode 100644 index 0000000000..aac4bc8454 --- /dev/null +++ b/intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json @@ -0,0 +1,196 @@ +{ + "schema": 6, + "backend_profile": "rust-llvm-23.1.0-16696adc", + "backend_kind": "llvm_nvptx", + "llvm_revision": "16696adcd119e6ba9cc175207d984d7021211acb", + "backend_version": "LLVM version 23.1.0-rust-1.100.0-nightly", + "backend_sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", + "records": [ + { + "id": "cvt_rn_satfinite_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_100f", + "ptx_feature": "+ptx91", + "status": "lowered", + "stages": [ + { + "targets": [ + "sm_100f", + "ptx91" + ], + "representation": "pure inline PTX with a single packed bf16x2 source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + } + ] +} diff --git a/intrinsics/generated-reference.md b/intrinsics/generated-reference.md index 71c68648ee..2fb83db646 100644 --- a/intrinsics/generated-reference.md +++ b/intrinsics/generated-reference.md @@ -2,7 +2,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0 @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. --> @@ -199,12 +199,16 @@ This table is generated from the resolved catalog. Evidence stages below disting | `cuda_intrinsics::convert::cvt_rn_relu_f16x2_f32` | `nvvm.cvt_rn_relu_f16x2_f32` | LLVM `llvm.nvvm.ff2f16x2.rn.relu` from `int_nvvm_ff2f16x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 7.0 on sm_80+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.relu.f16x2.f32 , , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_relu_satfinite_tf32_f32` | `nvvm.scalar_conversion` | LLVM `llvm.nvvm.f2tf32.rn.relu.satfinite` from `int_nvvm_f2tf32_rn_relu_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 8.6 on sm_100+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.relu.satfinite.tf32.f32 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_relu_tf32_f32` | `nvvm.scalar_conversion` | LLVM `llvm.nvvm.f2tf32.rn.relu` from `int_nvvm_f2tf32_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 7.8 on sm_90+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.relu.tf32.f32 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2` | `nvvm.cvt_rn_satfinite_e4m3x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100f or sm_110f or sm_120f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | lowered on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.e4m3x2.bf16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2` | `nvvm.cvt_rn_satfinite_e4m3x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn` from `int_nvvm_f16x2_to_e4m3x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.e4m3x2.f16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32` | `nvvm.cvt_rn_satfinite_e4m3x2_f32` | LLVM `llvm.nvvm.ff.to.e4m3x2.rn` from `int_nvvm_ff_to_e4m3x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.e4m3x2.f32 , , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2` | `nvvm.cvt_rn_satfinite_e5m2x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100f or sm_110f or sm_120f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | lowered on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.e5m2x2.bf16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2` | `nvvm.cvt_rn_satfinite_e5m2x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn` from `int_nvvm_f16x2_to_e5m2x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.e5m2x2.f16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32` | `nvvm.cvt_rn_satfinite_e5m2x2_f32` | LLVM `llvm.nvvm.ff.to.e5m2x2.rn` from `int_nvvm_ff_to_e5m2x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.e5m2x2.f32 , , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` | `nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100f or sm_110f or sm_120f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | lowered on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.relu.e4m3x2.bf16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2` | `nvvm.cvt_rn_satfinite_relu_e4m3x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn.relu` from `int_nvvm_f16x2_to_e4m3x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.relu.e4m3x2.f16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32` | `nvvm.cvt_rn_satfinite_relu_e4m3x2_f32` | LLVM `llvm.nvvm.ff.to.e4m3x2.rn.relu` from `int_nvvm_ff_to_e4m3x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.relu.e4m3x2.f32 , , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` | `nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100f or sm_110f or sm_120f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | lowered on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.relu.e5m2x2.bf16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2` | `nvvm.cvt_rn_satfinite_relu_e5m2x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn.relu` from `int_nvvm_f16x2_to_e5m2x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.relu.e5m2x2.f16x2 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f32` | `nvvm.cvt_rn_satfinite_relu_e5m2x2_f32` | LLVM `llvm.nvvm.ff.to.e5m2x2.rn.relu` from `int_nvvm_ff_to_e5m2x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.relu.e5m2x2.f32 , , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_tf32_f32` | `nvvm.scalar_conversion` | LLVM `llvm.nvvm.f2tf32.rn.satfinite` from `int_nvvm_f2tf32_rn_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 8.6 on sm_100+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.0-16696adc`; expects `cvt.rn.satfinite.tf32.f32 , ;` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) | @@ -1796,6 +1800,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371` - public source path: `cuda_intrinsics::convert::cvt_rn_relu_tf32_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_relu_tf32_f32` +- `cvt_rn_satfinite_e4m3x2_bf16x2` (`i1030`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2` - `cvt_rn_satfinite_e4m3x2_f16x2` (`i0822`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2` @@ -1804,6 +1812,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32` +- `cvt_rn_satfinite_e5m2x2_bf16x2` (`i1032`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2` - `cvt_rn_satfinite_e5m2x2_f16x2` (`i0824`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2` @@ -1812,6 +1824,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32` +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2` (`i1031`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` - `cvt_rn_satfinite_relu_e4m3x2_f16x2` (`i0823`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2` @@ -1820,6 +1836,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32` +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2` (`i1033`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` - `cvt_rn_satfinite_relu_e5m2x2_f16x2` (`i0825`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2` @@ -5527,12 +5547,16 @@ This table is generated from the resolved catalog. Evidence stages below disting - `cvt_rn_relu_f16x2_e4m3x2` converts two `f32` inputs to packed `f16x2` using nearest-even rounding with ReLU. It lowers to pure `cvt.rn.relu.f16x2.e4m3x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_relu_f16x2_e5m2x2` converts two `f32` inputs to packed `f16x2` using nearest-even rounding with ReLU. It lowers to pure `cvt.rn.relu.f16x2.e5m2x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_relu_f16x2_f32` converts two `f32` inputs to packed `f16x2` using nearest-even rounding with ReLU. It lowers to pure `cvt.rn.relu.f16x2.f32` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. +- `cvt_rn_satfinite_e4m3x2_bf16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e4m3x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e4m3x2_f16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e4m3x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e4m3x2_f32` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e4m3x2.rn` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.e4m3x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. +- `cvt_rn_satfinite_e5m2x2_bf16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e5m2x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e5m2x2_f16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e5m2x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e5m2x2_f32` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e5m2x2.rn` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.e5m2x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e4m3x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e4m3x2_f16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e4m3x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e4m3x2_f32` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation and ReLU. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e4m3x2.rn.relu` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.relu.e4m3x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e5m2x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e5m2x2_f16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e5m2x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e5m2x2_f32` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation and ReLU. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e5m2x2.rn.relu` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.relu.e5m2x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. - `cvt_rz_bf16x2_f32` converts two `f32` inputs to packed `bf16x2` using toward-zero rounding without saturation. It lowers to pure `cvt.rz.bf16x2.f32` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. @@ -5905,12 +5929,16 @@ This table is generated from the resolved catalog. Evidence stages below disting - `cvt_rn_relu_f16x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_relu_satfinite_tf32_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records `CVT_to_tf32_rn_relu_satf`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 100` - `cvt_rn_relu_tf32_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records `CVT_to_tf32_rn_relu`; selection predicates `Subtarget->getPTXVersion() >= 78`, `Subtarget->getSmVersion() >= 90` +- `cvt_rn_satfinite_e4m3x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e4m3x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e4m3x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none +- `cvt_rn_satfinite_e5m2x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e5m2x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e5m2x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e4m3x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e4m3x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e5m2x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e5m2x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_tf32_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records `CVT_to_tf32_rn_satf`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 100` @@ -8126,6 +8154,11 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the fresh 1,203-byte NVVM IR fixture (SHA-256 4b5b61cbde895ad4cca3777fb86c33c06e96d35b70d9fc589d3e0126201dd222) to 956-byte PTX 9.3 (SHA-256 aca3fe16c949e7614f988c423f06912fd05a431babafe2def5af07b03abaaea9) and 2,040-byte LTOIR (SHA-256 d0affd25b909632cee088d988a7778bde56d065edbd5c1324c3e1a8cf957bb23), retaining all four exact instructions and =r,f constraints. - PTX assembly on `sm_90, ptx78`: `succeeded` — CUDA 13.3 ptxas accepted the 469-byte exact sm_90/PTX 7.8 fixture (SHA-256 7455770182c5f46203de2170777b7bce100e445a41562bbc34945256325fa93a) and produced a 3,704-byte cubin (SHA-256 33f8941a912c4580054123f3ce55e599143c13cdfa161f0b3de1c55158a1780f). It also assembled the libNVVM-emitted PTX to a 3,704-byte cubin (SHA-256 1302e6a89ef1629ad728d8459ca785a8dc0a8901231869bbb56c0435e78922bb). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.33; build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink 13.3 consumed the fresh LTOIR and produced a 3,704-byte sm_90 cubin (SHA-256 7758fa36cb80e9d0f3bd35c3084f710b4dd47870ffb3e9c4f4f1f378a2beefab). Tool `/usr/local/cuda-13.3/targets/x86_64-linux/lib/libnvJitLink.so.13.3.33` reports `nvJitLinkVersion 13.3; CUDA 13.3 V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_e4m3x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.33` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33`, SHA-256 `b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. - `cvt_rn_satfinite_e4m3x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452). @@ -8148,6 +8181,11 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_89, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order. - PTX assembly on `sm_89, ptx93`: `succeeded` — CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_89, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_e5m2x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.33` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33`, SHA-256 `b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. - `cvt_rn_satfinite_e5m2x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452). @@ -8170,6 +8208,11 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_89, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order. - PTX assembly on `sm_89, ptx93`: `succeeded` — CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_89, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.33` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33`, SHA-256 `b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. - `cvt_rn_satfinite_relu_e4m3x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452). @@ -8192,6 +8235,11 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_89, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order. - PTX assembly on `sm_89, ptx93`: `succeeded` — CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_89, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.33` at PTX 9.1 / sm_100f or sm_110f or sm_120f: status `lowered` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33`, SHA-256 `b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0`) + - backend codegen on `sm_100f, ptx91`: `succeeded` — Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted. - `cvt_rn_satfinite_relu_e5m2x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.0-16696adc` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.0-rust-1.100.0-nightly`, SHA-256 `2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Pinned LLVM canonicalized the four i16(<2 x half>) and four <2 x half>(i16) declarations to nocallback, nofree, nosync, nounwind, speculatable, willreturn, and memory(none) in 2,785-byte IR (SHA-256 0f0b893b3fdc9d13bbd03894017d82295121adbf15942d5fae5c0d2d7a9f1452). diff --git a/intrinsics/overlay.toml b/intrinsics/overlay.toml index 1a4e179e5e..967349f617 100644 --- a/intrinsics/overlay.toml +++ b/intrinsics/overlay.toml @@ -26,6 +26,7 @@ shards = [ "overlay/packed_atomic.toml", "overlay/packed_conversion.toml", "overlay/packed_conversion_fp8.toml", + "overlay/packed_conversion_fp8_bf16x2.toml", "overlay/packed_conversion_fp8_f16x2.toml", "overlay/prmt.toml", "overlay/redux.toml", diff --git a/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml b/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml new file mode 100644 index 0000000000..bd52006d40 --- /dev/null +++ b/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml @@ -0,0 +1,18 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 +# +# Reviewed bf16x2-to-packed-FP8 conversions. +# +# Pack only. The source arrives already packed in one register. The imported +# reverse conversions carry a block scale and are not admitted here. + +schema = 65 +family = "packed_conversion" + +[packed_conversion_fp8_bf16x2] +llvm_evidence_profile = "rust-llvm-23.1.0-16696adc" +libnvvm_evidence_profile = "cuda-13.3-libnvvm-13.3.33" +runtime_validation = "unexecuted" +fp8_formats = ["e4m3x2", "e5m2x2"] +relu_variants = true +product_count = 4 diff --git a/intrinsics/probes/abs_bf16x2.ll b/intrinsics/probes/abs_bf16x2.ll index c759d6494a..38bcbef41a 100644 --- a/intrinsics/probes/abs_bf16x2.ll +++ b/intrinsics/probes/abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/abs_f16x2.ll b/intrinsics/probes/abs_f16x2.ll index f79af7cbb5..110c8a9bb2 100644 --- a/intrinsics/probes/abs_f16x2.ll +++ b/intrinsics/probes/abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/active_mask.ll b/intrinsics/probes/active_mask.ll index 3c522a8933..ad7285aec7 100644 --- a/intrinsics/probes/active_mask.ll +++ b/intrinsics/probes/active_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_bf16x2.ll b/intrinsics/probes/add_bf16x2.ll index c3ded6e68a..ad2c4b0756 100644 --- a/intrinsics/probes/add_bf16x2.ll +++ b/intrinsics/probes/add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_f16x2.ll b/intrinsics/probes/add_f16x2.ll index 899682ab2f..863654118b 100644 --- a/intrinsics/probes/add_f16x2.ll +++ b/intrinsics/probes/add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_f32x2.ll b/intrinsics/probes/add_f32x2.ll index 5642fd65d9..56c0c302bb 100644 --- a/intrinsics/probes/add_f32x2.ll +++ b/intrinsics/probes/add_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_ftz_f32x2.ll b/intrinsics/probes/add_ftz_f32x2.ll index 2cf4f23b8c..7a938dc45a 100644 --- a/intrinsics/probes/add_ftz_f32x2.ll +++ b/intrinsics/probes/add_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_f32.ll b/intrinsics/probes/add_rm_f32.ll index 22090a5d47..2d47dada03 100644 --- a/intrinsics/probes/add_rm_f32.ll +++ b/intrinsics/probes/add_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_f64.ll b/intrinsics/probes/add_rm_f64.ll index c142dc8d38..00f05d9f5c 100644 --- a/intrinsics/probes/add_rm_f64.ll +++ b/intrinsics/probes/add_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_ftz_f32.ll b/intrinsics/probes/add_rm_ftz_f32.ll index 07a1b1b8a9..a40d1eb1d9 100644 --- a/intrinsics/probes/add_rm_ftz_f32.ll +++ b/intrinsics/probes/add_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_ftz_sat_f32.ll b/intrinsics/probes/add_rm_ftz_sat_f32.ll index f42e2f9a98..23e94c2fe9 100644 --- a/intrinsics/probes/add_rm_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rm_sat_f32.ll b/intrinsics/probes/add_rm_sat_f32.ll index 25e4199acd..950927549a 100644 --- a/intrinsics/probes/add_rm_sat_f32.ll +++ b/intrinsics/probes/add_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_f32.ll b/intrinsics/probes/add_rn_f32.ll index a462c088b9..172f6fc709 100644 --- a/intrinsics/probes/add_rn_f32.ll +++ b/intrinsics/probes/add_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_f64.ll b/intrinsics/probes/add_rn_f64.ll index 3c07852ba2..251ef81942 100644 --- a/intrinsics/probes/add_rn_f64.ll +++ b/intrinsics/probes/add_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_ftz_f32.ll b/intrinsics/probes/add_rn_ftz_f32.ll index 65b6b22de1..23a37dcb3b 100644 --- a/intrinsics/probes/add_rn_ftz_f32.ll +++ b/intrinsics/probes/add_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_ftz_sat_f32.ll b/intrinsics/probes/add_rn_ftz_sat_f32.ll index bb5f6b1bfa..b20c12fe1e 100644 --- a/intrinsics/probes/add_rn_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rn_sat_f32.ll b/intrinsics/probes/add_rn_sat_f32.ll index 0b0a8722d0..e76959270a 100644 --- a/intrinsics/probes/add_rn_sat_f32.ll +++ b/intrinsics/probes/add_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_f32.ll b/intrinsics/probes/add_rp_f32.ll index e9831836d4..dbf01023ae 100644 --- a/intrinsics/probes/add_rp_f32.ll +++ b/intrinsics/probes/add_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_f64.ll b/intrinsics/probes/add_rp_f64.ll index a0e14b6d2d..cc08d8668b 100644 --- a/intrinsics/probes/add_rp_f64.ll +++ b/intrinsics/probes/add_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_ftz_f32.ll b/intrinsics/probes/add_rp_ftz_f32.ll index 4ef74410fc..323b15c026 100644 --- a/intrinsics/probes/add_rp_ftz_f32.ll +++ b/intrinsics/probes/add_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_ftz_sat_f32.ll b/intrinsics/probes/add_rp_ftz_sat_f32.ll index 37afea0e75..a918733e16 100644 --- a/intrinsics/probes/add_rp_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rp_sat_f32.ll b/intrinsics/probes/add_rp_sat_f32.ll index ac90be34ac..add89e15e1 100644 --- a/intrinsics/probes/add_rp_sat_f32.ll +++ b/intrinsics/probes/add_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_f32.ll b/intrinsics/probes/add_rz_f32.ll index dcd877842b..23f4c0acc5 100644 --- a/intrinsics/probes/add_rz_f32.ll +++ b/intrinsics/probes/add_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_f64.ll b/intrinsics/probes/add_rz_f64.ll index 26ca838cfe..b1317977c8 100644 --- a/intrinsics/probes/add_rz_f64.ll +++ b/intrinsics/probes/add_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_ftz_f32.ll b/intrinsics/probes/add_rz_ftz_f32.ll index e4b909f948..550edad960 100644 --- a/intrinsics/probes/add_rz_ftz_f32.ll +++ b/intrinsics/probes/add_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_ftz_sat_f32.ll b/intrinsics/probes/add_rz_ftz_sat_f32.ll index 01e08bb065..1e8be5ddd8 100644 --- a/intrinsics/probes/add_rz_ftz_sat_f32.ll +++ b/intrinsics/probes/add_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/add_rz_sat_f32.ll b/intrinsics/probes/add_rz_sat_f32.ll index 350187eb9a..5c759afef2 100644 --- a/intrinsics/probes/add_rz_sat_f32.ll +++ b/intrinsics/probes/add_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/all_sync.ll b/intrinsics/probes/all_sync.ll index 54df3f0fc5..e697f1293c 100644 --- a/intrinsics/probes/all_sync.ll +++ b/intrinsics/probes/all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/any_sync.ll b/intrinsics/probes/any_sync.ll index c62e870fbb..cd5714c7b7 100644 --- a/intrinsics/probes/any_sync.ll +++ b/intrinsics/probes/any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ballot_sync.ll b/intrinsics/probes/ballot_sync.ll index 1c0a90549d..35f543c791 100644 --- a/intrinsics/probes/ballot_sync.ll +++ b/intrinsics/probes/ballot_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive.ll b/intrinsics/probes/barrier_cluster_arrive.ll index 158357505b..25510c60cd 100644 --- a/intrinsics/probes/barrier_cluster_arrive.ll +++ b/intrinsics/probes/barrier_cluster_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive_aligned.ll b/intrinsics/probes/barrier_cluster_arrive_aligned.ll index 901fffaf32..9096b00344 100644 --- a/intrinsics/probes/barrier_cluster_arrive_aligned.ll +++ b/intrinsics/probes/barrier_cluster_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive_relaxed.ll b/intrinsics/probes/barrier_cluster_arrive_relaxed.ll index e85e359968..cfcf159714 100644 --- a/intrinsics/probes/barrier_cluster_arrive_relaxed.ll +++ b/intrinsics/probes/barrier_cluster_arrive_relaxed.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll b/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll index 9d983a9cd9..2f02157c12 100644 --- a/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll +++ b/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_wait.ll b/intrinsics/probes/barrier_cluster_wait.ll index b7bfb2f9c2..2724450d0f 100644 --- a/intrinsics/probes/barrier_cluster_wait.ll +++ b/intrinsics/probes/barrier_cluster_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cluster_wait_aligned.ll b/intrinsics/probes/barrier_cluster_wait_aligned.ll index 1ce1ef8a24..4d01d8dd7a 100644 --- a/intrinsics/probes/barrier_cluster_wait_aligned.ll +++ b/intrinsics/probes/barrier_cluster_wait_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_arrive.ll b/intrinsics/probes/barrier_cta_arrive.ll index 8c37c624cb..6e82642590 100644 --- a/intrinsics/probes/barrier_cta_arrive.ll +++ b/intrinsics/probes/barrier_cta_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_arrive_aligned.ll b/intrinsics/probes/barrier_cta_arrive_aligned.ll index 5e9555a1b5..aa5568d46c 100644 --- a/intrinsics/probes/barrier_cta_arrive_aligned.ll +++ b/intrinsics/probes/barrier_cta_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_sync.ll b/intrinsics/probes/barrier_cta_sync.ll index 026e7984d5..84c46278e4 100644 --- a/intrinsics/probes/barrier_cta_sync.ll +++ b/intrinsics/probes/barrier_cta_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/barrier_cta_sync_aligned.ll b/intrinsics/probes/barrier_cta_sync_aligned.ll index 6309af77f5..96598dddd9 100644 --- a/intrinsics/probes/barrier_cta_sync_aligned.ll +++ b/intrinsics/probes/barrier_cta_sync_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_dim_x.ll b/intrinsics/probes/block_dim_x.ll index eec87d7cb2..7ea7154cd4 100644 --- a/intrinsics/probes/block_dim_x.ll +++ b/intrinsics/probes/block_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_dim_y.ll b/intrinsics/probes/block_dim_y.ll index 78c06ddb31..a7704f9587 100644 --- a/intrinsics/probes/block_dim_y.ll +++ b/intrinsics/probes/block_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_dim_z.ll b/intrinsics/probes/block_dim_z.ll index fc09e16013..785be4ae9f 100644 --- a/intrinsics/probes/block_dim_z.ll +++ b/intrinsics/probes/block_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_idx_x.ll b/intrinsics/probes/block_idx_x.ll index 443bbc24f2..f97cdaf07d 100644 --- a/intrinsics/probes/block_idx_x.ll +++ b/intrinsics/probes/block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_idx_y.ll b/intrinsics/probes/block_idx_y.ll index a7099664d3..2e1fecf34a 100644 --- a/intrinsics/probes/block_idx_y.ll +++ b/intrinsics/probes/block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/block_idx_z.ll b/intrinsics/probes/block_idx_z.ll index 9953c01142..a826526e80 100644 --- a/intrinsics/probes/block_idx_z.ll +++ b/intrinsics/probes/block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/breakpoint.ll b/intrinsics/probes/breakpoint.ll index 032084ab2f..b1da6d8e50 100644 --- a/intrinsics/probes/breakpoint.ll +++ b/intrinsics/probes/breakpoint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_get_first_ctaid_x.ll b/intrinsics/probes/clc_query_get_first_ctaid_x.ll index 3dae251d11..f7a6e35aa5 100644 --- a/intrinsics/probes/clc_query_get_first_ctaid_x.ll +++ b/intrinsics/probes/clc_query_get_first_ctaid_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_get_first_ctaid_y.ll b/intrinsics/probes/clc_query_get_first_ctaid_y.ll index 77a2459449..adcc954d64 100644 --- a/intrinsics/probes/clc_query_get_first_ctaid_y.ll +++ b/intrinsics/probes/clc_query_get_first_ctaid_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_get_first_ctaid_z.ll b/intrinsics/probes/clc_query_get_first_ctaid_z.ll index 36566c4490..e27b6b8222 100644 --- a/intrinsics/probes/clc_query_get_first_ctaid_z.ll +++ b/intrinsics/probes/clc_query_get_first_ctaid_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_query_is_canceled.ll b/intrinsics/probes/clc_query_is_canceled.ll index 9f0ada4b39..8c2196b7a6 100644 --- a/intrinsics/probes/clc_query_is_canceled.ll +++ b/intrinsics/probes/clc_query_is_canceled.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_try_cancel.ll b/intrinsics/probes/clc_try_cancel.ll index f89bfecf25..28b72db7ad 100644 --- a/intrinsics/probes/clc_try_cancel.ll +++ b/intrinsics/probes/clc_try_cancel.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clc_try_cancel_multicast.ll b/intrinsics/probes/clc_try_cancel_multicast.ll index fa96d77c86..ed622fa2e6 100644 --- a/intrinsics/probes/clc_try_cancel_multicast.ll +++ b/intrinsics/probes/clc_try_cancel_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock.lib_nvvm.ll b/intrinsics/probes/clock.lib_nvvm.ll index 8ac5ec9072..0c7e571806 100644 --- a/intrinsics/probes/clock.lib_nvvm.ll +++ b/intrinsics/probes/clock.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock.llvm_nvptx.ll b/intrinsics/probes/clock.llvm_nvptx.ll index 1673854a38..6bd6b693ee 100644 --- a/intrinsics/probes/clock.llvm_nvptx.ll +++ b/intrinsics/probes/clock.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock64.lib_nvvm.ll b/intrinsics/probes/clock64.lib_nvvm.ll index 726cf4829a..d3df1c659c 100644 --- a/intrinsics/probes/clock64.lib_nvvm.ll +++ b/intrinsics/probes/clock64.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/clock64.llvm_nvptx.ll b/intrinsics/probes/clock64.llvm_nvptx.ll index 77d94474c2..3b81eb97cd 100644 --- a/intrinsics/probes/clock64.llvm_nvptx.ll +++ b/intrinsics/probes/clock64.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_count.ll b/intrinsics/probes/cluster_block_count.ll index c77af0cb8b..221cced29f 100644 --- a/intrinsics/probes/cluster_block_count.ll +++ b/intrinsics/probes/cluster_block_count.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_idx_x.ll b/intrinsics/probes/cluster_block_idx_x.ll index f3a4ddd821..4768687207 100644 --- a/intrinsics/probes/cluster_block_idx_x.ll +++ b/intrinsics/probes/cluster_block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_idx_y.ll b/intrinsics/probes/cluster_block_idx_y.ll index 7f21d0a12e..ffb926a085 100644 --- a/intrinsics/probes/cluster_block_idx_y.ll +++ b/intrinsics/probes/cluster_block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_idx_z.ll b/intrinsics/probes/cluster_block_idx_z.ll index 91ad3fd6cd..95f0193bc8 100644 --- a/intrinsics/probes/cluster_block_idx_z.ll +++ b/intrinsics/probes/cluster_block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_block_rank.ll b/intrinsics/probes/cluster_block_rank.ll index cfb6a0ea65..c2d8f3edef 100644 --- a/intrinsics/probes/cluster_block_rank.ll +++ b/intrinsics/probes/cluster_block_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_dim_x.ll b/intrinsics/probes/cluster_dim_x.ll index 38705818f8..4b19d7d9d4 100644 --- a/intrinsics/probes/cluster_dim_x.ll +++ b/intrinsics/probes/cluster_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_dim_y.ll b/intrinsics/probes/cluster_dim_y.ll index 45853409df..dd8e7fe5a1 100644 --- a/intrinsics/probes/cluster_dim_y.ll +++ b/intrinsics/probes/cluster_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_dim_z.ll b/intrinsics/probes/cluster_dim_z.ll index 9b9e9ab058..2ac06fcefe 100644 --- a/intrinsics/probes/cluster_dim_z.ll +++ b/intrinsics/probes/cluster_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_grid_dim_x.ll b/intrinsics/probes/cluster_grid_dim_x.ll index 8b304b0807..600bbbd76f 100644 --- a/intrinsics/probes/cluster_grid_dim_x.ll +++ b/intrinsics/probes/cluster_grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_grid_dim_y.ll b/intrinsics/probes/cluster_grid_dim_y.ll index 5146b69270..67a4362500 100644 --- a/intrinsics/probes/cluster_grid_dim_y.ll +++ b/intrinsics/probes/cluster_grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_grid_dim_z.ll b/intrinsics/probes/cluster_grid_dim_z.ll index ad3e521abf..d7c5285ef1 100644 --- a/intrinsics/probes/cluster_grid_dim_z.ll +++ b/intrinsics/probes/cluster_grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_idx_x.ll b/intrinsics/probes/cluster_idx_x.ll index 28a37d350f..86125ae121 100644 --- a/intrinsics/probes/cluster_idx_x.ll +++ b/intrinsics/probes/cluster_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_idx_y.ll b/intrinsics/probes/cluster_idx_y.ll index b7265dc311..1835baaa48 100644 --- a/intrinsics/probes/cluster_idx_y.ll +++ b/intrinsics/probes/cluster_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cluster_idx_z.ll b/intrinsics/probes/cluster_idx_z.ll index 2432fc25b3..14dacd1f67 100644 --- a/intrinsics/probes/cluster_idx_z.ll +++ b/intrinsics/probes/cluster_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cos_approx_f32.ll b/intrinsics/probes/cos_approx_f32.ll index 463c30ec0e..22043aeab6 100644 --- a/intrinsics/probes/cos_approx_f32.ll +++ b/intrinsics/probes/cos_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cos_approx_ftz_f32.ll b/intrinsics/probes/cos_approx_ftz_f32.ll index 0cc4f4c0fe..1fb257556d 100644 --- a/intrinsics/probes/cos_approx_ftz_f32.ll +++ b/intrinsics/probes/cos_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_commit_group.ll b/intrinsics/probes/cp_async_bulk_commit_group.ll index 57e937b295..0f3f38ad6f 100644 --- a/intrinsics/probes/cp_async_bulk_commit_group.ll +++ b/intrinsics/probes/cp_async_bulk_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll index 0e945a5908..61e6b9f618 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll index 80d5f1d28a..daac1edda9 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll index 2e92cb76af..99c1801308 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll index c2c54d7f6b..200f9f635b 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll index 05aa3cf7ea..e60d96fd59 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll index 78afd6cecf..b702932994 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll index bf5b50969a..703e94312f 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll index a6a48ff86d..3da030132d 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll index b6d31f8453..727ac1ebdf 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll index e9f51ee313..92f2660e63 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll index 979a66ff62..c733a03308 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll index 388124aedd..22187e1f88 100644 --- a/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll +++ b/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll index cf4946e20f..c9cd1d04a9 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll index 1f759f4e24..46837ab8bf 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll index 805fb69ca3..8774da2d45 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll index 37ed46fd0d..a583ab1864 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll index 3c74138fa2..d1b3776dcb 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll index 3a3aaf72c5..ebd2b2b6f2 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll index 2bddcc5431..863251ee81 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll index d49e54e46d..913838e3d7 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll index 171dcd3f52..688517df84 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll index 4a92de939e..70c0109dbb 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll b/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll index fdd9db39f5..06e81e4d49 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll b/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll index 57ba0b709b..8c0f6036d8 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll index 73c0167e00..80df780f32 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll index c535f257e9..2cc2bbaad0 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll index 10a45a3fd0..96a6e55dbb 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll index 3264bac5b8..bd1456819f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll index c31a4e0abe..bbb425f509 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll index c670cf96c5..c0de41d1a0 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll index b9c9b2f995..1dbc19b519 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll index f97b3728ad..1870070f0a 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll index 361cd7a4a9..c488746019 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll index cfcb0f8958..6174b4a628 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll index dce921c5da..53b008dac8 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll index 8411827e33..6e71b43779 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll index 66a8bc4819..427106007f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll index ac83efb656..3a5e848684 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll index 2abbcf4bb5..aa7daaa2b6 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll index b9e69dd89c..4f20ecea30 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll index 7d459c971e..94fe9f2f6a 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll index d721c627c3..1290484c67 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll index 289fd56aac..3f8339246d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll index f716e7af93..391f38c326 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll index 9434682d63..ae2667cc8f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll index d9248a8d04..bf406d09e5 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll index a287e08da9..4945be7895 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll index a901151478..8d33ae322e 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll index 78f91df966..b97d7221cc 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll index 9ab6cbe102..a26fb4388d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll index 9a94956802..97b3936d21 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll index e88074227e..81936abe8f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll index 40c298537e..1f9c81d9ac 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll index 1817063a88..bf0598ce14 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll index d8a420acc2..f28f676883 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll index 3c6f129025..541468e98f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll index 0873d0284c..c04f74e06a 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll index a56a952adf..3b6b6437f4 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll index b53d95e547..9ac0538b2d 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll index 0f8ece64b8..6ac14b4e2c 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll index eb2657563d..721386cb0a 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll index 8fb7afa4f3..eebd3b02c9 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll index 73391f8c73..e4c0e42ddc 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll index f6786f5755..7765d45b21 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll index 136abe2ccd..d7f633e4a8 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll index c022a43347..db5be957b9 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll index a5a9cef492..e8ecfe4a81 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll index 2e2b3a86bc..3839d9ac0e 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll index b2800ffc98..85579739e9 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll index 0a667ceaf5..0fb1afd61f 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll index 14711c751c..7c4a7631b7 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll index 0f28a56460..314f7b1139 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll index b163c49caa..3a81196f30 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll index 375b37477d..6ed36ed88c 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll index 64cd8a78b1..a15bb68946 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll index d1471d795e..e0f177766c 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll index f5e8cc6daf..6a58caab47 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll index 0dfcb9d37c..666b45a0a4 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll index d703ae7903..a3677fb795 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll index 6f7c1d709b..67f55dae9b 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll index dc8eb106e1..576652c2d6 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll index 9c41388b19..70cf575c27 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll index 60a9c2278a..2533877a9b 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll index 1fa891fd77..6acd60e77e 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll index 5474511ebd..0b934c9af1 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll index b8b0fae3a1..8728e7b313 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll index 840ca4b4dd..1cabaa11a5 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll index 406c297013..d16edb067e 100644 --- a/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll +++ b/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_wait_group.ll b/intrinsics/probes/cp_async_bulk_wait_group.ll index 9b7fcc408a..341aa2d663 100644 --- a/intrinsics/probes/cp_async_bulk_wait_group.ll +++ b/intrinsics/probes/cp_async_bulk_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_bulk_wait_group_read.ll b/intrinsics/probes/cp_async_bulk_wait_group_read.ll index 003374d612..427a0f37d9 100644 --- a/intrinsics/probes/cp_async_bulk_wait_group_read.ll +++ b/intrinsics/probes/cp_async_bulk_wait_group_read.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_16.ll b/intrinsics/probes/cp_async_ca_16.ll index 6b878036c1..9755feda00 100644 --- a/intrinsics/probes/cp_async_ca_16.ll +++ b/intrinsics/probes/cp_async_ca_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_4.ll b/intrinsics/probes/cp_async_ca_4.ll index d95bfc6ae6..9dde8f1603 100644 --- a/intrinsics/probes/cp_async_ca_4.ll +++ b/intrinsics/probes/cp_async_ca_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_8.ll b/intrinsics/probes/cp_async_ca_8.ll index 74320807a5..dcee19906e 100644 --- a/intrinsics/probes/cp_async_ca_8.ll +++ b/intrinsics/probes/cp_async_ca_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_zfill_16.ll b/intrinsics/probes/cp_async_ca_zfill_16.ll index 4d7bd1fe1a..5824e90f17 100644 --- a/intrinsics/probes/cp_async_ca_zfill_16.ll +++ b/intrinsics/probes/cp_async_ca_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_zfill_4.ll b/intrinsics/probes/cp_async_ca_zfill_4.ll index 3639ac6a73..dd56f48e07 100644 --- a/intrinsics/probes/cp_async_ca_zfill_4.ll +++ b/intrinsics/probes/cp_async_ca_zfill_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_ca_zfill_8.ll b/intrinsics/probes/cp_async_ca_zfill_8.ll index 0bd93f1922..05ff55b1ba 100644 --- a/intrinsics/probes/cp_async_ca_zfill_8.ll +++ b/intrinsics/probes/cp_async_ca_zfill_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_cg_16.ll b/intrinsics/probes/cp_async_cg_16.ll index 1e716fa3d8..307bf07006 100644 --- a/intrinsics/probes/cp_async_cg_16.ll +++ b/intrinsics/probes/cp_async_cg_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_cg_zfill_16.ll b/intrinsics/probes/cp_async_cg_zfill_16.ll index 31a9f08401..6479100e1d 100644 --- a/intrinsics/probes/cp_async_cg_zfill_16.ll +++ b/intrinsics/probes/cp_async_cg_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_commit_group.ll b/intrinsics/probes/cp_async_commit_group.ll index 170b62c9e1..1dd2713c16 100644 --- a/intrinsics/probes/cp_async_commit_group.ll +++ b/intrinsics/probes/cp_async_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive.ll b/intrinsics/probes/cp_async_mbarrier_arrive.ll index a2275af196..3422e8ba50 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll b/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll index 23b21f6087..e1c8f4769d 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll b/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll index 0060317dd7..5f11411a34 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll b/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll index e2a4eca4a0..222a4bcef2 100644 --- a/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll +++ b/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_wait_all.ll b/intrinsics/probes/cp_async_wait_all.ll index 9e65e90dd1..77d91f82cf 100644 --- a/intrinsics/probes/cp_async_wait_all.ll +++ b/intrinsics/probes/cp_async_wait_all.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cp_async_wait_group.ll b/intrinsics/probes/cp_async_wait_group.ll index d47231be2d..c76c5a4171 100644 --- a/intrinsics/probes/cp_async_wait_group.ll +++ b/intrinsics/probes/cp_async_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_f16x2_f32.ll b/intrinsics/probes/cvt_f16x2_f32.ll index 0674c78705..db16c0b6a2 100644 --- a/intrinsics/probes/cvt_f16x2_f32.ll +++ b/intrinsics/probes/cvt_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_f32x2_bf16x2.ll b/intrinsics/probes/cvt_f32x2_bf16x2.ll index 70276d8a43..86e7d0d53f 100644 --- a/intrinsics/probes/cvt_f32x2_bf16x2.ll +++ b/intrinsics/probes/cvt_f32x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll b/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll index 63e4f7b437..171aad6280 100644 --- a/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll +++ b/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll b/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll index 30185b23b0..b700342c52 100644 --- a/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll +++ b/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll b/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll index bf5a1ce5c2..8ad2b49519 100644 --- a/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll b/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll index 82db013f53..7926287fcd 100644 --- a/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll +++ b/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll b/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll index fed9d0a3f7..5be270e5f7 100644 --- a/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll +++ b/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll b/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll index 7807de0859..14d4fa600f 100644 --- a/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll index 03dde0e7eb..7b978033f0 100644 --- a/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_relu_tf32_f32.ll b/intrinsics/probes/cvt_rn_relu_tf32_f32.ll index ec79a85d5b..77338eeda9 100644 --- a/intrinsics/probes/cvt_rn_relu_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll new file mode 100644 index 0000000000..04556ff2e8 --- /dev/null +++ b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll @@ -0,0 +1,12 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +define i16 @probe_cvt_rn_satfinite_e4m3x2_bf16x2(i32 %packed) { + %result = call i16 asm "cvt.rn.satfinite.e4m3x2.bf16x2 $0, $1;", "=h,r"(i32 %packed) + ret i16 %result +} diff --git a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index 001698b45e..d3b957af76 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll index d77da8e78a..a5bee97b1c 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll new file mode 100644 index 0000000000..9dc6f41d3a --- /dev/null +++ b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll @@ -0,0 +1,12 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +define i16 @probe_cvt_rn_satfinite_e5m2x2_bf16x2(i32 %packed) { + %result = call i16 asm "cvt.rn.satfinite.e5m2x2.bf16x2 $0, $1;", "=h,r"(i32 %packed) + ret i16 %result +} diff --git a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index a771cfdd27..f2919d1d21 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll index d40e5056b0..398ef3ecf2 100644 --- a/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll new file mode 100644 index 0000000000..e6bf6bb677 --- /dev/null +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll @@ -0,0 +1,12 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +define i16 @probe_cvt_rn_satfinite_relu_e4m3x2_bf16x2(i32 %packed) { + %result = call i16 asm "cvt.rn.satfinite.relu.e4m3x2.bf16x2 $0, $1;", "=h,r"(i32 %packed) + ret i16 %result +} diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll index 38d5553e4f..05f87742c1 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll index f9078aba54..7aff158535 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll new file mode 100644 index 0000000000..a882c1163f --- /dev/null +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll @@ -0,0 +1,12 @@ +; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +; SPDX-License-Identifier: Apache-2.0 +; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. +; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. + +target triple = "nvptx64-nvidia-cuda" + +define i16 @probe_cvt_rn_satfinite_relu_e5m2x2_bf16x2(i32 %packed) { + %result = call i16 asm "cvt.rn.satfinite.relu.e5m2x2.bf16x2 $0, $1;", "=h,r"(i32 %packed) + ret i16 %result +} diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll index 65f3756f9c..f945e71e9c 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll index 0353b4ee4e..4205983727 100644 --- a/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll index a345d26bf9..86ef31fecf 100644 --- a/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rn_tf32_f32.ll b/intrinsics/probes/cvt_rn_tf32_f32.ll index 7298293841..56621599a7 100644 --- a/intrinsics/probes/cvt_rn_tf32_f32.ll +++ b/intrinsics/probes/cvt_rn_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll index b14cf39b8c..86c20f971d 100644 --- a/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rna_tf32_f32.ll b/intrinsics/probes/cvt_rna_tf32_f32.ll index ce19a950e7..52b03a244e 100644 --- a/intrinsics/probes/cvt_rna_tf32_f32.ll +++ b/intrinsics/probes/cvt_rna_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_bf16x2_f32.ll b/intrinsics/probes/cvt_rz_bf16x2_f32.ll index 685083190f..4fe372d26f 100644 --- a/intrinsics/probes/cvt_rz_bf16x2_f32.ll +++ b/intrinsics/probes/cvt_rz_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_f16x2_f32.ll b/intrinsics/probes/cvt_rz_f16x2_f32.ll index a4fb07c467..847203823f 100644 --- a/intrinsics/probes/cvt_rz_f16x2_f32.ll +++ b/intrinsics/probes/cvt_rz_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll index 95213c1296..b31513a1fc 100644 --- a/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_relu_tf32_f32.ll b/intrinsics/probes/cvt_rz_relu_tf32_f32.ll index 94080ccd34..735d036934 100644 --- a/intrinsics/probes/cvt_rz_relu_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll b/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll index 09cf6b5803..424c6e5fc7 100644 --- a/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/cvt_rz_tf32_f32.ll b/intrinsics/probes/cvt_rz_tf32_f32.ll index a3a8616f32..a0e5a01896 100644 --- a/intrinsics/probes/cvt_rz_tf32_f32.ll +++ b/intrinsics/probes/cvt_rz_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rm_f32.ll b/intrinsics/probes/div_rm_f32.ll index 0f61c9350e..a501e28952 100644 --- a/intrinsics/probes/div_rm_f32.ll +++ b/intrinsics/probes/div_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rm_f64.ll b/intrinsics/probes/div_rm_f64.ll index a190589ea7..c2b4f4a280 100644 --- a/intrinsics/probes/div_rm_f64.ll +++ b/intrinsics/probes/div_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rm_ftz_f32.ll b/intrinsics/probes/div_rm_ftz_f32.ll index 232079de93..29b53f39ec 100644 --- a/intrinsics/probes/div_rm_ftz_f32.ll +++ b/intrinsics/probes/div_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rn_f32.ll b/intrinsics/probes/div_rn_f32.ll index 714ec0e096..173baa7ad1 100644 --- a/intrinsics/probes/div_rn_f32.ll +++ b/intrinsics/probes/div_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rn_f64.ll b/intrinsics/probes/div_rn_f64.ll index 4d0717b270..3a4f87dd8f 100644 --- a/intrinsics/probes/div_rn_f64.ll +++ b/intrinsics/probes/div_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rn_ftz_f32.ll b/intrinsics/probes/div_rn_ftz_f32.ll index b93da31ab7..484f9ea57e 100644 --- a/intrinsics/probes/div_rn_ftz_f32.ll +++ b/intrinsics/probes/div_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rp_f32.ll b/intrinsics/probes/div_rp_f32.ll index c02cc35d92..e318f2cb85 100644 --- a/intrinsics/probes/div_rp_f32.ll +++ b/intrinsics/probes/div_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rp_f64.ll b/intrinsics/probes/div_rp_f64.ll index a886b936c6..c0a70eac11 100644 --- a/intrinsics/probes/div_rp_f64.ll +++ b/intrinsics/probes/div_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rp_ftz_f32.ll b/intrinsics/probes/div_rp_ftz_f32.ll index 2505944120..0b20c9f1dd 100644 --- a/intrinsics/probes/div_rp_ftz_f32.ll +++ b/intrinsics/probes/div_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rz_f32.ll b/intrinsics/probes/div_rz_f32.ll index 586c0d30b9..9292b454e0 100644 --- a/intrinsics/probes/div_rz_f32.ll +++ b/intrinsics/probes/div_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rz_f64.ll b/intrinsics/probes/div_rz_f64.ll index afa12610f9..996f0d77f7 100644 --- a/intrinsics/probes/div_rz_f64.ll +++ b/intrinsics/probes/div_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/div_rz_ftz_f32.ll b/intrinsics/probes/div_rz_ftz_f32.ll index 6f4bb9b7c5..9e01c1c2d6 100644 --- a/intrinsics/probes/div_rz_ftz_f32.ll +++ b/intrinsics/probes/div_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp2a_s32.ll b/intrinsics/probes/dp2a_s32.ll index 3bec0f096c..8b198b8a07 100644 --- a/intrinsics/probes/dp2a_s32.ll +++ b/intrinsics/probes/dp2a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp2a_u32.ll b/intrinsics/probes/dp2a_u32.ll index 620f3cb82b..07cb913a45 100644 --- a/intrinsics/probes/dp2a_u32.ll +++ b/intrinsics/probes/dp2a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp4a_s32.ll b/intrinsics/probes/dp4a_s32.ll index 59e3002b8a..2526290253 100644 --- a/intrinsics/probes/dp4a_s32.ll +++ b/intrinsics/probes/dp4a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dp4a_u32.ll b/intrinsics/probes/dp4a_u32.ll index 28e9ba0263..3c53aa9654 100644 --- a/intrinsics/probes/dp4a_u32.ll +++ b/intrinsics/probes/dp4a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dsmem_read_u32.ll b/intrinsics/probes/dsmem_read_u32.ll index 1af51dc3da..24588110d0 100644 --- a/intrinsics/probes/dsmem_read_u32.ll +++ b/intrinsics/probes/dsmem_read_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll b/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll index fd2a06a984..53a93d4ed0 100644 --- a/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll +++ b/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll b/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll index dd1a391113..7ec59f7205 100644 --- a/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll +++ b/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/elect_sync.lib_nvvm.ll b/intrinsics/probes/elect_sync.lib_nvvm.ll index 62e82e1c05..4b54e19761 100644 --- a/intrinsics/probes/elect_sync.lib_nvvm.ll +++ b/intrinsics/probes/elect_sync.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/elect_sync.llvm_nvptx.ll b/intrinsics/probes/elect_sync.llvm_nvptx.ll index 159f7f611e..8406ddf832 100644 --- a/intrinsics/probes/elect_sync.llvm_nvptx.ll +++ b/intrinsics/probes/elect_sync.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg1.lib_nvvm.ll b/intrinsics/probes/envreg1.lib_nvvm.ll index ab2c0faa35..ce763a4a3f 100644 --- a/intrinsics/probes/envreg1.lib_nvvm.ll +++ b/intrinsics/probes/envreg1.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg1.llvm_nvptx.ll b/intrinsics/probes/envreg1.llvm_nvptx.ll index 5a34008805..070a7558bb 100644 --- a/intrinsics/probes/envreg1.llvm_nvptx.ll +++ b/intrinsics/probes/envreg1.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg2.lib_nvvm.ll b/intrinsics/probes/envreg2.lib_nvvm.ll index a51f98d74c..38ec8e58a9 100644 --- a/intrinsics/probes/envreg2.lib_nvvm.ll +++ b/intrinsics/probes/envreg2.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/envreg2.llvm_nvptx.ll b/intrinsics/probes/envreg2.llvm_nvptx.ll index 63942849b8..c0e46074c2 100644 --- a/intrinsics/probes/envreg2.llvm_nvptx.ll +++ b/intrinsics/probes/envreg2.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ex2_approx_f16.ll b/intrinsics/probes/ex2_approx_f16.ll index e72dcc6a2c..0e0fabdb9e 100644 --- a/intrinsics/probes/ex2_approx_f16.ll +++ b/intrinsics/probes/ex2_approx_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ex2_approx_f32.ll b/intrinsics/probes/ex2_approx_f32.ll index d9ca0c7e96..3a6772690e 100644 --- a/intrinsics/probes/ex2_approx_f32.ll +++ b/intrinsics/probes/ex2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ex2_approx_ftz_f32.ll b/intrinsics/probes/ex2_approx_ftz_f32.ll index 5b3c4a6c53..f6a12040c5 100644 --- a/intrinsics/probes/ex2_approx_ftz_f32.ll +++ b/intrinsics/probes/ex2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_mbarrier_init_release_cluster.ll b/intrinsics/probes/fence_mbarrier_init_release_cluster.ll index 709ea6786b..10340f37b2 100644 --- a/intrinsics/probes/fence_mbarrier_init_release_cluster.ll +++ b/intrinsics/probes/fence_mbarrier_init_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll b/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll index d37dcb85c4..358c64d0b9 100644 --- a/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll +++ b/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll b/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll index 0420e795fb..22c33fea34 100644 --- a/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll +++ b/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_async_shared_cta.ll b/intrinsics/probes/fence_proxy_async_shared_cta.ll index 756ab152ef..bed3360638 100644 --- a/intrinsics/probes/fence_proxy_async_shared_cta.ll +++ b/intrinsics/probes/fence_proxy_async_shared_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll index 9795cc8a82..415f54380f 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll index af28e9032a..53893614cd 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll index e01e367782..183ec3b6ca 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll index 99925c3ec9..9e176bc61c 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll index a01a1e6b9a..ac7348e279 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll index 435e4e97c8..4da4d8d12d 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll index 15ab9ca536..4a8bfc3560 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll b/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll index 61f2340f44..38659d9a4c 100644 --- a/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll +++ b/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_bf16x2.ll b/intrinsics/probes/fma_bf16x2.ll index d03cb35b49..ba8db560a2 100644 --- a/intrinsics/probes/fma_bf16x2.ll +++ b/intrinsics/probes/fma_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_f16x2.ll b/intrinsics/probes/fma_f16x2.ll index 11570cb7dc..fa8594e3d9 100644 --- a/intrinsics/probes/fma_f16x2.ll +++ b/intrinsics/probes/fma_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_f32x2.ll b/intrinsics/probes/fma_f32x2.ll index 1441f8cb69..d59016c811 100644 --- a/intrinsics/probes/fma_f32x2.ll +++ b/intrinsics/probes/fma_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_f16x2.ll b/intrinsics/probes/fma_ftz_f16x2.ll index 56787d210a..869e7cf32d 100644 --- a/intrinsics/probes/fma_ftz_f16x2.ll +++ b/intrinsics/probes/fma_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_f32x2.ll b/intrinsics/probes/fma_ftz_f32x2.ll index aea071a677..e168193b1d 100644 --- a/intrinsics/probes/fma_ftz_f32x2.ll +++ b/intrinsics/probes/fma_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_relu_f16x2.ll b/intrinsics/probes/fma_ftz_relu_f16x2.ll index f95f66a01f..e30a83a40b 100644 --- a/intrinsics/probes/fma_ftz_relu_f16x2.ll +++ b/intrinsics/probes/fma_ftz_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_ftz_sat_f16x2.ll b/intrinsics/probes/fma_ftz_sat_f16x2.ll index 3b91ebee55..f98c9a86a5 100644 --- a/intrinsics/probes/fma_ftz_sat_f16x2.ll +++ b/intrinsics/probes/fma_ftz_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_relu_bf16x2.ll b/intrinsics/probes/fma_relu_bf16x2.ll index b9782e147a..6aea542489 100644 --- a/intrinsics/probes/fma_relu_bf16x2.ll +++ b/intrinsics/probes/fma_relu_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_relu_f16x2.ll b/intrinsics/probes/fma_relu_f16x2.ll index 5ce2a6ce6a..fa8a117542 100644 --- a/intrinsics/probes/fma_relu_f16x2.ll +++ b/intrinsics/probes/fma_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_f32.ll b/intrinsics/probes/fma_rm_f32.ll index f0c4c72614..bbada45196 100644 --- a/intrinsics/probes/fma_rm_f32.ll +++ b/intrinsics/probes/fma_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_f64.ll b/intrinsics/probes/fma_rm_f64.ll index cfc8075316..eecd1cbf31 100644 --- a/intrinsics/probes/fma_rm_f64.ll +++ b/intrinsics/probes/fma_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_ftz_f32.ll b/intrinsics/probes/fma_rm_ftz_f32.ll index 6324e54129..c6e6816e47 100644 --- a/intrinsics/probes/fma_rm_ftz_f32.ll +++ b/intrinsics/probes/fma_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_ftz_sat_f32.ll b/intrinsics/probes/fma_rm_ftz_sat_f32.ll index f4bf2b34c0..ee088dce7e 100644 --- a/intrinsics/probes/fma_rm_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rm_sat_f32.ll b/intrinsics/probes/fma_rm_sat_f32.ll index 7adc765483..0399cb1387 100644 --- a/intrinsics/probes/fma_rm_sat_f32.ll +++ b/intrinsics/probes/fma_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_f32.ll b/intrinsics/probes/fma_rn_f32.ll index 880d1833af..bf1958099b 100644 --- a/intrinsics/probes/fma_rn_f32.ll +++ b/intrinsics/probes/fma_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_f64.ll b/intrinsics/probes/fma_rn_f64.ll index 7a21f94476..82d86a2a28 100644 --- a/intrinsics/probes/fma_rn_f64.ll +++ b/intrinsics/probes/fma_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_ftz_f32.ll b/intrinsics/probes/fma_rn_ftz_f32.ll index 5701f6a574..a34ffa4022 100644 --- a/intrinsics/probes/fma_rn_ftz_f32.ll +++ b/intrinsics/probes/fma_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_ftz_sat_f32.ll b/intrinsics/probes/fma_rn_ftz_sat_f32.ll index c4eb49e0fe..5986c5c6d7 100644 --- a/intrinsics/probes/fma_rn_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rn_sat_f32.ll b/intrinsics/probes/fma_rn_sat_f32.ll index 9ecae992d5..fe5764f52b 100644 --- a/intrinsics/probes/fma_rn_sat_f32.ll +++ b/intrinsics/probes/fma_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_f32.ll b/intrinsics/probes/fma_rp_f32.ll index 5c7e2e252b..190f21f6e3 100644 --- a/intrinsics/probes/fma_rp_f32.ll +++ b/intrinsics/probes/fma_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_f64.ll b/intrinsics/probes/fma_rp_f64.ll index a69a5ffb03..5871d0aa7c 100644 --- a/intrinsics/probes/fma_rp_f64.ll +++ b/intrinsics/probes/fma_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_ftz_f32.ll b/intrinsics/probes/fma_rp_ftz_f32.ll index 59c486b424..d932acb75a 100644 --- a/intrinsics/probes/fma_rp_ftz_f32.ll +++ b/intrinsics/probes/fma_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_ftz_sat_f32.ll b/intrinsics/probes/fma_rp_ftz_sat_f32.ll index eb3e059038..9443c8931d 100644 --- a/intrinsics/probes/fma_rp_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rp_sat_f32.ll b/intrinsics/probes/fma_rp_sat_f32.ll index cd7e49669a..8c72199e18 100644 --- a/intrinsics/probes/fma_rp_sat_f32.ll +++ b/intrinsics/probes/fma_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_f32.ll b/intrinsics/probes/fma_rz_f32.ll index fad52e7945..525becd08b 100644 --- a/intrinsics/probes/fma_rz_f32.ll +++ b/intrinsics/probes/fma_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_f64.ll b/intrinsics/probes/fma_rz_f64.ll index b26219dbb9..21b4b359af 100644 --- a/intrinsics/probes/fma_rz_f64.ll +++ b/intrinsics/probes/fma_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_ftz_f32.ll b/intrinsics/probes/fma_rz_ftz_f32.ll index 0f7d190873..0db661af1f 100644 --- a/intrinsics/probes/fma_rz_ftz_f32.ll +++ b/intrinsics/probes/fma_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_ftz_sat_f32.ll b/intrinsics/probes/fma_rz_ftz_sat_f32.ll index 774bdb33fc..d70f028811 100644 --- a/intrinsics/probes/fma_rz_ftz_sat_f32.ll +++ b/intrinsics/probes/fma_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_rz_sat_f32.ll b/intrinsics/probes/fma_rz_sat_f32.ll index e8b2396b07..bdb920fb24 100644 --- a/intrinsics/probes/fma_rz_sat_f32.ll +++ b/intrinsics/probes/fma_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/fma_sat_f16x2.ll b/intrinsics/probes/fma_sat_f16x2.ll index c69f4d0b0d..84d8c890f8 100644 --- a/intrinsics/probes/fma_sat_f16x2.ll +++ b/intrinsics/probes/fma_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/globaltimer.lib_nvvm.ll b/intrinsics/probes/globaltimer.lib_nvvm.ll index bd0606ec80..ad02429eeb 100644 --- a/intrinsics/probes/globaltimer.lib_nvvm.ll +++ b/intrinsics/probes/globaltimer.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/globaltimer.llvm_nvptx.ll b/intrinsics/probes/globaltimer.llvm_nvptx.ll index f8833f42c6..f1f1f9c464 100644 --- a/intrinsics/probes/globaltimer.llvm_nvptx.ll +++ b/intrinsics/probes/globaltimer.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dependency_launch_dependents.ll b/intrinsics/probes/grid_dependency_launch_dependents.ll index 8793515f3e..9a9c89027d 100644 --- a/intrinsics/probes/grid_dependency_launch_dependents.ll +++ b/intrinsics/probes/grid_dependency_launch_dependents.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dependency_wait.ll b/intrinsics/probes/grid_dependency_wait.ll index 3dcc697fda..0873cd95c7 100644 --- a/intrinsics/probes/grid_dependency_wait.ll +++ b/intrinsics/probes/grid_dependency_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dim_x.ll b/intrinsics/probes/grid_dim_x.ll index 6160b58479..b12f311981 100644 --- a/intrinsics/probes/grid_dim_x.ll +++ b/intrinsics/probes/grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dim_y.ll b/intrinsics/probes/grid_dim_y.ll index e568eef09c..ba50650e81 100644 --- a/intrinsics/probes/grid_dim_y.ll +++ b/intrinsics/probes/grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/grid_dim_z.ll b/intrinsics/probes/grid_dim_z.ll index e3ab0c2c78..def5bcd4e9 100644 --- a/intrinsics/probes/grid_dim_z.ll +++ b/intrinsics/probes/grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/gridid.lib_nvvm.ll b/intrinsics/probes/gridid.lib_nvvm.ll index 91bf8bae59..0ba74ee194 100644 --- a/intrinsics/probes/gridid.lib_nvvm.ll +++ b/intrinsics/probes/gridid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/gridid.llvm_nvptx.ll b/intrinsics/probes/gridid.llvm_nvptx.ll index 4d3af9c030..b22fa3af20 100644 --- a/intrinsics/probes/gridid.llvm_nvptx.ll +++ b/intrinsics/probes/gridid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lane_id.ll b/intrinsics/probes/lane_id.ll index 4375748889..5959c7b0b6 100644 --- a/intrinsics/probes/lane_id.ll +++ b/intrinsics/probes/lane_id.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_eq.ll b/intrinsics/probes/lanemask_eq.ll index 3f0c5becda..42a95b5f5d 100644 --- a/intrinsics/probes/lanemask_eq.ll +++ b/intrinsics/probes/lanemask_eq.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_ge.ll b/intrinsics/probes/lanemask_ge.ll index dcaf2c81f7..9dc1308337 100644 --- a/intrinsics/probes/lanemask_ge.ll +++ b/intrinsics/probes/lanemask_ge.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_gt.ll b/intrinsics/probes/lanemask_gt.ll index 4d5ea2baab..3d7e570b94 100644 --- a/intrinsics/probes/lanemask_gt.ll +++ b/intrinsics/probes/lanemask_gt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_le.ll b/intrinsics/probes/lanemask_le.ll index 42a538ad53..878cdd9b7e 100644 --- a/intrinsics/probes/lanemask_le.ll +++ b/intrinsics/probes/lanemask_le.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lanemask_lt.ll b/intrinsics/probes/lanemask_lt.ll index 29445102d1..7cb0e4f6bf 100644 --- a/intrinsics/probes/lanemask_lt.ll +++ b/intrinsics/probes/lanemask_lt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll index 48aae7f31f..0fea35ae9c 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll index 335ea7f964..d03d4b2cff 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll index 46110462c5..15d9dcc9b7 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll index e54bd8ddb6..7eed70cbdd 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll index 186e16a9f4..66f2908c25 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll index 48a46dd253..caf47da318 100644 --- a/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll index a14091597c..ddb75b0eee 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll index 300fe10d75..f36709465b 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll index 4d08f3a647..05bf71bbfa 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll index 04357788fb..5047947ac8 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll index 47a24b0408..473cc25d7b 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll index 16b6279264..c2d36574fe 100644 --- a/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll +++ b/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll index 6de5101df0..464a9b4df8 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll index 2e62e6f8bc..9ad178e0dc 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll index ff9314f5c5..d447680e0f 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll index c8a8192730..9cb4b8d25b 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll index ef545a1dec..148ef81521 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll b/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll index 8feaa576cf..f32ea5568b 100644 --- a/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll +++ b/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lg2_approx_f32.ll b/intrinsics/probes/lg2_approx_f32.ll index f162fddc72..c9eb2c57b0 100644 --- a/intrinsics/probes/lg2_approx_f32.ll +++ b/intrinsics/probes/lg2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/lg2_approx_ftz_f32.ll b/intrinsics/probes/lg2_approx_ftz_f32.ll index fe37b75dd4..7237b41d13 100644 --- a/intrinsics/probes/lg2_approx_ftz_f32.ll +++ b/intrinsics/probes/lg2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/map_shared_rank.ll b/intrinsics/probes/map_shared_rank.ll index 3e92103cf9..f1068e8247 100644 --- a/intrinsics/probes/map_shared_rank.ll +++ b/intrinsics/probes/map_shared_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_all_i64_sync.ll b/intrinsics/probes/match_all_i64_sync.ll index 5c4808bd9a..3145c5df90 100644 --- a/intrinsics/probes/match_all_i64_sync.ll +++ b/intrinsics/probes/match_all_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_all_sync.ll b/intrinsics/probes/match_all_sync.ll index 10c5ee442b..78ea768a92 100644 --- a/intrinsics/probes/match_all_sync.ll +++ b/intrinsics/probes/match_all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_any_i64_sync.ll b/intrinsics/probes/match_any_i64_sync.ll index 90b70fd545..735be5fb3c 100644 --- a/intrinsics/probes/match_any_i64_sync.ll +++ b/intrinsics/probes/match_any_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/match_any_sync.ll b/intrinsics/probes/match_any_sync.ll index 9288ae011d..d96c5eaff4 100644 --- a/intrinsics/probes/match_any_sync.ll +++ b/intrinsics/probes/match_any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_bf16.ll b/intrinsics/probes/max_bf16.ll index 42e2be579f..f0928e8e07 100644 --- a/intrinsics/probes/max_bf16.ll +++ b/intrinsics/probes/max_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_bf16x2.ll b/intrinsics/probes/max_bf16x2.ll index ce7f8125b1..fca4dcc023 100644 --- a/intrinsics/probes/max_bf16x2.ll +++ b/intrinsics/probes/max_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_f16.ll b/intrinsics/probes/max_f16.ll index e208b199e1..ad3f333432 100644 --- a/intrinsics/probes/max_f16.ll +++ b/intrinsics/probes/max_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_f16x2.ll b/intrinsics/probes/max_f16x2.ll index 8789a5d3ec..d8d0e1cf48 100644 --- a/intrinsics/probes/max_f16x2.ll +++ b/intrinsics/probes/max_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_f16.ll b/intrinsics/probes/max_ftz_f16.ll index 458003c675..34f09e53cc 100644 --- a/intrinsics/probes/max_ftz_f16.ll +++ b/intrinsics/probes/max_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_f16x2.ll b/intrinsics/probes/max_ftz_f16x2.ll index 1bd819eaa1..5610d03d3e 100644 --- a/intrinsics/probes/max_ftz_f16x2.ll +++ b/intrinsics/probes/max_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_f16.ll b/intrinsics/probes/max_ftz_nan_f16.ll index f3ed7912a0..46619270fa 100644 --- a/intrinsics/probes/max_ftz_nan_f16.ll +++ b/intrinsics/probes/max_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_f16x2.ll b/intrinsics/probes/max_ftz_nan_f16x2.ll index cfc8fec1dd..4a05ff90d0 100644 --- a/intrinsics/probes/max_ftz_nan_f16x2.ll +++ b/intrinsics/probes/max_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll index dcc8bf5ce5..c04bcd30b6 100644 --- a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll index e1a1ac6d9c..843afa7baa 100644 --- a/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll b/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll index 5aab79c28f..3e75ac50bc 100644 --- a/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_xorsign_abs_f16.ll b/intrinsics/probes/max_ftz_xorsign_abs_f16.ll index 06a882d0cd..ad05e5ebb1 100644 --- a/intrinsics/probes/max_ftz_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll b/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll index 5cee3d6a52..985a8c34fd 100644 --- a/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_ftz_xorsign_abs_f32.ll b/intrinsics/probes/max_ftz_xorsign_abs_f32.ll index f6fed3626a..4aa3a55fe0 100644 --- a/intrinsics/probes/max_ftz_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_bf16.ll b/intrinsics/probes/max_nan_bf16.ll index 20f90efb9c..49918f25a2 100644 --- a/intrinsics/probes/max_nan_bf16.ll +++ b/intrinsics/probes/max_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_bf16x2.ll b/intrinsics/probes/max_nan_bf16x2.ll index 4938b05160..c1cb964d79 100644 --- a/intrinsics/probes/max_nan_bf16x2.ll +++ b/intrinsics/probes/max_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_f16.ll b/intrinsics/probes/max_nan_f16.ll index b48856a401..a345d4d540 100644 --- a/intrinsics/probes/max_nan_f16.ll +++ b/intrinsics/probes/max_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_f16x2.ll b/intrinsics/probes/max_nan_f16x2.ll index f7840ca365..97170ce273 100644 --- a/intrinsics/probes/max_nan_f16x2.ll +++ b/intrinsics/probes/max_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_bf16.ll b/intrinsics/probes/max_nan_xorsign_abs_bf16.ll index f234e4ba90..d19a17890f 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_bf16.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll b/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll index 0e1a0e22ed..1420004fb3 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_f16.ll b/intrinsics/probes/max_nan_xorsign_abs_f16.ll index 7f977d3f5c..e71f4ea8c6 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll index ad91a7b8cd..afb1067c5e 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_nan_xorsign_abs_f32.ll b/intrinsics/probes/max_nan_xorsign_abs_f32.ll index 505dc8d7c1..a2d24c6131 100644 --- a/intrinsics/probes/max_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_relu_s16x2.ll b/intrinsics/probes/max_relu_s16x2.ll index 25a6d49c6b..76ee6db04f 100644 --- a/intrinsics/probes/max_relu_s16x2.ll +++ b/intrinsics/probes/max_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_relu_s32.ll b/intrinsics/probes/max_relu_s32.ll index d364154157..c5bb092765 100644 --- a/intrinsics/probes/max_relu_s32.ll +++ b/intrinsics/probes/max_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_s16x2.ll b/intrinsics/probes/max_s16x2.ll index 1860d6088d..d10e60d834 100644 --- a/intrinsics/probes/max_s16x2.ll +++ b/intrinsics/probes/max_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_u16x2.ll b/intrinsics/probes/max_u16x2.ll index d2a27f7c79..1cbaf11d65 100644 --- a/intrinsics/probes/max_u16x2.ll +++ b/intrinsics/probes/max_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_bf16.ll b/intrinsics/probes/max_xorsign_abs_bf16.ll index 174a9ad97b..8efbb7c81e 100644 --- a/intrinsics/probes/max_xorsign_abs_bf16.ll +++ b/intrinsics/probes/max_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_bf16x2.ll b/intrinsics/probes/max_xorsign_abs_bf16x2.ll index 6c9213a7a4..fa1cf748ad 100644 --- a/intrinsics/probes/max_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/max_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_f16.ll b/intrinsics/probes/max_xorsign_abs_f16.ll index 12c7994b29..b0b6d680ff 100644 --- a/intrinsics/probes/max_xorsign_abs_f16.ll +++ b/intrinsics/probes/max_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_f16x2.ll b/intrinsics/probes/max_xorsign_abs_f16x2.ll index b7ca356ab1..71915edf44 100644 --- a/intrinsics/probes/max_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/max_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/max_xorsign_abs_f32.ll b/intrinsics/probes/max_xorsign_abs_f32.ll index e0be80518f..009bed5975 100644 --- a/intrinsics/probes/max_xorsign_abs_f32.ll +++ b/intrinsics/probes/max_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive.ll b/intrinsics/probes/mbarrier_arrive.ll index b592698137..a4f02543b9 100644 --- a/intrinsics/probes/mbarrier_arrive.ll +++ b/intrinsics/probes/mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_cluster.ll b/intrinsics/probes/mbarrier_arrive_cluster.ll index 93bb09121d..5bd27ef720 100644 --- a/intrinsics/probes/mbarrier_arrive_cluster.ll +++ b/intrinsics/probes/mbarrier_arrive_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_expect_tx.ll b/intrinsics/probes/mbarrier_arrive_expect_tx.ll index c36d6aad5c..abe49f73c6 100644 --- a/intrinsics/probes/mbarrier_arrive_expect_tx.ll +++ b/intrinsics/probes/mbarrier_arrive_expect_tx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll b/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll index ad001c0ebb..ea2bad33ad 100644 --- a/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll +++ b/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_arrive_no_complete.ll b/intrinsics/probes/mbarrier_arrive_no_complete.ll index fc17a2816f..a26f7cb605 100644 --- a/intrinsics/probes/mbarrier_arrive_no_complete.ll +++ b/intrinsics/probes/mbarrier_arrive_no_complete.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_init.ll b/intrinsics/probes/mbarrier_init.ll index ab6038402e..b7c4bb808a 100644 --- a/intrinsics/probes/mbarrier_init.ll +++ b/intrinsics/probes/mbarrier_init.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_inval.ll b/intrinsics/probes/mbarrier_inval.ll index 21d816dafa..2476830d8e 100644 --- a/intrinsics/probes/mbarrier_inval.ll +++ b/intrinsics/probes/mbarrier_inval.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_test_wait.ll b/intrinsics/probes/mbarrier_test_wait.ll index e16b8dc05c..5b43dcab16 100644 --- a/intrinsics/probes/mbarrier_test_wait.ll +++ b/intrinsics/probes/mbarrier_test_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_try_wait.ll b/intrinsics/probes/mbarrier_try_wait.ll index 9ea2000fdc..f04aa31d77 100644 --- a/intrinsics/probes/mbarrier_try_wait.ll +++ b/intrinsics/probes/mbarrier_try_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_try_wait_parity.ll b/intrinsics/probes/mbarrier_try_wait_parity.ll index ddfe48da54..afc4a5fde8 100644 --- a/intrinsics/probes/mbarrier_try_wait_parity.ll +++ b/intrinsics/probes/mbarrier_try_wait_parity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll b/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll index 609bfd5de7..e31e40d142 100644 --- a/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll +++ b/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_bf16.ll b/intrinsics/probes/min_bf16.ll index 685c8beee2..e51be3be95 100644 --- a/intrinsics/probes/min_bf16.ll +++ b/intrinsics/probes/min_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_bf16x2.ll b/intrinsics/probes/min_bf16x2.ll index 98d2a848a7..cb256d1dd7 100644 --- a/intrinsics/probes/min_bf16x2.ll +++ b/intrinsics/probes/min_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_f16.ll b/intrinsics/probes/min_f16.ll index db6d4cf788..b4057aa840 100644 --- a/intrinsics/probes/min_f16.ll +++ b/intrinsics/probes/min_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_f16x2.ll b/intrinsics/probes/min_f16x2.ll index fd6aad9c0a..3a7769c192 100644 --- a/intrinsics/probes/min_f16x2.ll +++ b/intrinsics/probes/min_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_f16.ll b/intrinsics/probes/min_ftz_f16.ll index d7f60d85a2..ae0d739075 100644 --- a/intrinsics/probes/min_ftz_f16.ll +++ b/intrinsics/probes/min_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_f16x2.ll b/intrinsics/probes/min_ftz_f16x2.ll index 2c0522f034..0afb2af2f5 100644 --- a/intrinsics/probes/min_ftz_f16x2.ll +++ b/intrinsics/probes/min_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_f16.ll b/intrinsics/probes/min_ftz_nan_f16.ll index 40383f5720..95b1c37302 100644 --- a/intrinsics/probes/min_ftz_nan_f16.ll +++ b/intrinsics/probes/min_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_f16x2.ll b/intrinsics/probes/min_ftz_nan_f16x2.ll index 9d8379e621..bee24a0a32 100644 --- a/intrinsics/probes/min_ftz_nan_f16x2.ll +++ b/intrinsics/probes/min_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll index b97e669b2c..01c9517bc6 100644 --- a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll index 984663ea51..cf2406018b 100644 --- a/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll b/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll index 5b1816dc56..94c9cc8b99 100644 --- a/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_xorsign_abs_f16.ll b/intrinsics/probes/min_ftz_xorsign_abs_f16.ll index c6cd152c15..bf7f0e5791 100644 --- a/intrinsics/probes/min_ftz_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll b/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll index 535f2934a3..faee75f973 100644 --- a/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_ftz_xorsign_abs_f32.ll b/intrinsics/probes/min_ftz_xorsign_abs_f32.ll index d4b65452e4..f43aea61ba 100644 --- a/intrinsics/probes/min_ftz_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_bf16.ll b/intrinsics/probes/min_nan_bf16.ll index 800e9c770b..bfa669bd4c 100644 --- a/intrinsics/probes/min_nan_bf16.ll +++ b/intrinsics/probes/min_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_bf16x2.ll b/intrinsics/probes/min_nan_bf16x2.ll index bfc7d311a0..4a91c96413 100644 --- a/intrinsics/probes/min_nan_bf16x2.ll +++ b/intrinsics/probes/min_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_f16.ll b/intrinsics/probes/min_nan_f16.ll index a591854fb3..cd893e858a 100644 --- a/intrinsics/probes/min_nan_f16.ll +++ b/intrinsics/probes/min_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_f16x2.ll b/intrinsics/probes/min_nan_f16x2.ll index 4761a5aef5..ff42939daf 100644 --- a/intrinsics/probes/min_nan_f16x2.ll +++ b/intrinsics/probes/min_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_bf16.ll b/intrinsics/probes/min_nan_xorsign_abs_bf16.ll index 92b345ae9e..4ec17a7777 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_bf16.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll b/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll index 8d77182d42..dee2c8104f 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_f16.ll b/intrinsics/probes/min_nan_xorsign_abs_f16.ll index 42ae49f2f1..bb507ad427 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll b/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll index 4f671739a8..3f8e8d76e6 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_nan_xorsign_abs_f32.ll b/intrinsics/probes/min_nan_xorsign_abs_f32.ll index 86ecde55d6..19a4c4b3b5 100644 --- a/intrinsics/probes/min_nan_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_relu_s16x2.ll b/intrinsics/probes/min_relu_s16x2.ll index 114ba0ea2b..22ed75a31c 100644 --- a/intrinsics/probes/min_relu_s16x2.ll +++ b/intrinsics/probes/min_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_relu_s32.ll b/intrinsics/probes/min_relu_s32.ll index 72ec4e9862..82dddf1bb3 100644 --- a/intrinsics/probes/min_relu_s32.ll +++ b/intrinsics/probes/min_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_s16x2.ll b/intrinsics/probes/min_s16x2.ll index 024ac9a79d..30f0d122b5 100644 --- a/intrinsics/probes/min_s16x2.ll +++ b/intrinsics/probes/min_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_u16x2.ll b/intrinsics/probes/min_u16x2.ll index 17efe05f7d..26c7388371 100644 --- a/intrinsics/probes/min_u16x2.ll +++ b/intrinsics/probes/min_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_bf16.ll b/intrinsics/probes/min_xorsign_abs_bf16.ll index b2c98a1396..7933dab6de 100644 --- a/intrinsics/probes/min_xorsign_abs_bf16.ll +++ b/intrinsics/probes/min_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_bf16x2.ll b/intrinsics/probes/min_xorsign_abs_bf16x2.ll index 5c5eda21cc..5f38b8c695 100644 --- a/intrinsics/probes/min_xorsign_abs_bf16x2.ll +++ b/intrinsics/probes/min_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_f16.ll b/intrinsics/probes/min_xorsign_abs_f16.ll index fad3ee263c..16783bfeea 100644 --- a/intrinsics/probes/min_xorsign_abs_f16.ll +++ b/intrinsics/probes/min_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_f16x2.ll b/intrinsics/probes/min_xorsign_abs_f16x2.ll index 83e2d083f1..4916a41772 100644 --- a/intrinsics/probes/min_xorsign_abs_f16x2.ll +++ b/intrinsics/probes/min_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/min_xorsign_abs_f32.ll b/intrinsics/probes/min_xorsign_abs_f32.ll index 0600e57238..e80ff3466d 100644 --- a/intrinsics/probes/min_xorsign_abs_f32.ll +++ b/intrinsics/probes/min_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll b/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll index d963d02bf2..f1a4923117 100644 --- a/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll +++ b/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll b/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll index ff9b31661d..434c1faddc 100644 --- a/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll +++ b/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_f16_f16.ll b/intrinsics/probes/mma_m16n8k16_f16_f16.ll index 61856d680a..68251e2dd5 100644 --- a/intrinsics/probes/mma_m16n8k16_f16_f16.ll +++ b/intrinsics/probes/mma_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_f32_bf16.ll b/intrinsics/probes/mma_m16n8k16_f32_bf16.ll index c7143f7d5c..c7f82122c0 100644 --- a/intrinsics/probes/mma_m16n8k16_f32_bf16.ll +++ b/intrinsics/probes/mma_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_f32_f16.ll b/intrinsics/probes/mma_m16n8k16_f32_f16.ll index 7c10c34523..48aece6f65 100644 --- a/intrinsics/probes/mma_m16n8k16_f32_f16.ll +++ b/intrinsics/probes/mma_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll index 68c5760734..d9cfccd295 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll index 3787a8eaf1..976192e05f 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll index 7bbb1fb8f0..eef183fb3e 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll index e93e734e66..aa2a8c669f 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll index e4996d459a..e57b06060e 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll index 0bcfef72ab..281cd37f2a 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll index 83b9d915f4..f29be114d6 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll index 9bbd1478a9..9624058aed 100644 --- a/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8.ll b/intrinsics/probes/mma_m16n8k16_s32_s8.ll index 9f41f357ca..9b6e6cd7ef 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll index 0d8a8d7e62..35c844572d 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll b/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll index 26470d0556..1fcf37940c 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll index 6fa872212a..584d54e136 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8.ll b/intrinsics/probes/mma_m16n8k16_s32_u8.ll index 1b7a2ab311..080435577f 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll b/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll index e1d99bcc83..8535891f8d 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll index 561ca82bce..a4ffe6db9f 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll index b06c35609f..1d02ba0255 100644 --- a/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll b/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll index 89a369f4e3..eb19a7790a 100644 --- a/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll +++ b/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll b/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll index 715d24a899..eee6188b09 100644 --- a/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll +++ b/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll index e46e29201b..43a1be0f00 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll index b02d641a4e..4051c45bb2 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll index 671c967568..855344055e 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll index 16314882c7..c8b53cec5b 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll index 78fc634a97..adac62ed3d 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll index 60a7744103..a9310ec29e 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll index 3b21922940..7a0fd36b0c 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll index dcd9c3f56a..e83d0c9228 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll index 8a15719960..0ee274927d 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll index 975bc9ed11..7190a4d562 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll index caf1cf19bd..2e79fb7d62 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll index 97a7ad5840..30fb795c42 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll index 04bc4acf53..ab3ecd3ca1 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll index 72f90eed95..b3c1792e75 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll index 202ca3ce08..cb41369bad 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll index 2038d886f5..5e3cff1321 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll index defcac001d..d0988291e8 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll index cd73de8eb0..a668a1406d 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll index 4431e4d8a5..f6e0b3452e 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll index aac88c2777..ccf6ba3d21 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll index 56d53ae963..b65d957f5a 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll index 94f34125c8..7218339e25 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll index cdd7e57398..64eb06820d 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll index 1ad6123d34..4bd69cbccf 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll index a23db44a2c..077d64ff52 100644 --- a/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll index d13ede0396..7fdd8b99c5 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll index af6ae69ee0..446168ecaf 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll index b9fb90a7cf..1adbfc233a 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll index b65a5e1edf..455291f0e4 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll index d8ae5eb98d..70b7e2f80a 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll index f76a320e85..b7daddbf82 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll index e186591fc6..301a4ff239 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll index 06fb18bdf6..a3e49afa2c 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll index 6330da3521..b8f1be88da 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll index 90ac832473..fa5d078963 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll index 89a311470b..7bf4538054 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll index fc423843d7..2565fd241a 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll index 62adbab13e..724d091714 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll index 3c95f08757..142312e178 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll index 595955a2bf..b9df091e71 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll index fe76080066..4fbfcd15a2 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll index 355be7c4d3..eaec383dd9 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll index 0ed0557252..54fa7a2d21 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll index 726e27d404..ddaa95175d 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll index a20d6e3566..fe90fc6907 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll index a05abbbcfe..80af64c4a9 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll index e4227f1110..43bf03ec65 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll index f3670b096d..2edab48172 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll index f26c6f95fe..c6674b82b0 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll index e97e238d87..0b423d0fc0 100644 --- a/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll index 02f46cc6bd..cfe36d1809 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll index 0704365d94..be5340f062 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll index 7fece7f172..426374e276 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll index 76a03fbe61..b4eb959176 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll index c17965b9e1..95384628cb 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll index ec293d8e6c..143d6ea2e5 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll index fcfe25612f..b349498344 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll index d5222ca264..3b86f7c7b7 100644 --- a/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll index fcf989dcdb..538e5ca376 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll index 497b1c61a8..ed6d6b35be 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll index 201c6b4011..2ca8e0c9ba 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll index a781318992..c90b149501 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll index 2dcbc0291f..d839c01b81 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll index 9ab7304639..c5626adaf7 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll index 6d42709e72..27df48a15e 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll index 07fe418d71..2efc4754bf 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll index 99db7845c6..7d419980ab 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll index f0162e7b5b..1aaea44094 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll index 8bc029fb6d..ebb9b489a9 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll index c814062020..5172f60aee 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll index 48ea460094..044958ecd3 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll index 7177de07ab..2a3161f7a7 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll index 2e3f5b56c2..87853930bb 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll index 1baca9d509..57e126a596 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll index e373189c4d..e51cf6ed8a 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll index bb9b4d535d..f233224211 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll index 9f8922664f..b654f9a4d2 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll index 83225dc35d..d1610e16e6 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll index b895e7eaf8..e63630744d 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll index 87a1a68e0a..26f91c0733 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll index 6e971e4550..54fe20885f 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll index de42fd30e4..9de2e3fad6 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll index 4f37af1c1e..bc81f671e2 100644 --- a/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll +++ b/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4.ll b/intrinsics/probes/mma_m16n8k32_s32_s4.ll index d98377829a..ee702587cc 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll index 1547583cf9..111fe0cd5d 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll b/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll index 887e048030..5a8ace9d74 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll index 8f9a007242..5df1397817 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8.ll b/intrinsics/probes/mma_m16n8k32_s32_s8.ll index 4870c39f1b..f485e7ce7f 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll index cfe37b5c89..0248e2f26a 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll b/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll index 3925250b55..b6931333fd 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll index 428e6c42e2..0e8f4fcf27 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4.ll b/intrinsics/probes/mma_m16n8k32_s32_u4.ll index eedaf28c42..ecf9a004ff 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll b/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll index f5ecd80ef9..41004245a0 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll index 9336a6d1f4..ec21d7ee61 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll index 4dae4a3d74..145dcb2109 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8.ll b/intrinsics/probes/mma_m16n8k32_s32_u8.ll index d0cd3714f7..f45726d8c8 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll b/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll index f3ca3a052d..f83596d8cf 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll index 7af581e38b..2e4f005051 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll b/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll index 35145ad43d..52ef0bc86c 100644 --- a/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k4_f32_tf32.ll b/intrinsics/probes/mma_m16n8k4_f32_tf32.ll index 1bee661ae6..f9b235213b 100644 --- a/intrinsics/probes/mma_m16n8k4_f32_tf32.ll +++ b/intrinsics/probes/mma_m16n8k4_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4.ll b/intrinsics/probes/mma_m16n8k64_s32_s4.ll index d0d6079f1f..75c848c75e 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll index b22126cd85..2277594edd 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll b/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll index 39a6330ed7..77c76d9f28 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll index e098daa90a..72f9df6401 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4.ll b/intrinsics/probes/mma_m16n8k64_s32_u4.ll index f408f9bcf9..f73c5fddff 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll b/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll index c233aa5d2f..229cb2e229 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll index bfec98968c..4c47adb526 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll b/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll index a5e0eb5719..e9898d1387 100644 --- a/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f16_f16.ll b/intrinsics/probes/mma_m16n8k8_f16_f16.ll index d71f481be1..fb9ebf3e27 100644 --- a/intrinsics/probes/mma_m16n8k8_f16_f16.ll +++ b/intrinsics/probes/mma_m16n8k8_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f32_bf16.ll b/intrinsics/probes/mma_m16n8k8_f32_bf16.ll index 0b66faadfd..c8aa2ba030 100644 --- a/intrinsics/probes/mma_m16n8k8_f32_bf16.ll +++ b/intrinsics/probes/mma_m16n8k8_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f32_f16.ll b/intrinsics/probes/mma_m16n8k8_f32_f16.ll index 2770e8712f..3ff4e1623c 100644 --- a/intrinsics/probes/mma_m16n8k8_f32_f16.ll +++ b/intrinsics/probes/mma_m16n8k8_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m16n8k8_f32_tf32.ll b/intrinsics/probes/mma_m16n8k8_f32_tf32.ll index 4c3c6fccff..e1275a858f 100644 --- a/intrinsics/probes/mma_m16n8k8_f32_tf32.ll +++ b/intrinsics/probes/mma_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll b/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll index e66d1165dd..bd420a69b7 100644 --- a/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll +++ b/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll b/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll index 7a5cb727fe..ddfae45cfd 100644 --- a/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll +++ b/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8.ll b/intrinsics/probes/mma_m8n8k16_s32_s8.ll index d6d1b5fa40..bc6eaba68c 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll index feffa7e53e..66e132ff9d 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll b/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll index 63488fee56..0cd22f2aa7 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll index 72786dd844..facf12c493 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8.ll b/intrinsics/probes/mma_m8n8k16_s32_u8.ll index 2a2d346a55..e856384d55 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll b/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll index 2c10dac69b..0c302b49dd 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll index fa86ab8e8c..95bc813083 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll b/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll index 6c6359518c..6caa892444 100644 --- a/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4.ll b/intrinsics/probes/mma_m8n8k32_s32_s4.ll index ab00bf2115..c62170b65a 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll index e7d49c1bef..740ecbdedf 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll b/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll index 33d8a273c5..08121e469d 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll index 01ce04f226..fc5f0ea2ec 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4.ll b/intrinsics/probes/mma_m8n8k32_s32_u4.ll index ced502c6b4..b2b2d55342 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll b/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll index 895b9dfa45..3adc7648c0 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll index 9aa33a996e..7b51c95974 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll b/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll index bf67485621..6fcdf4830f 100644 --- a/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_m8n8k4_f64.ll b/intrinsics/probes/mma_m8n8k4_f64.ll index dbf88178e5..cdb4bcddbc 100644 --- a/intrinsics/probes/mma_m8n8k4_f64.ll +++ b/intrinsics/probes/mma_m8n8k4_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll index d7f31a77d6..25aa589326 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll index 41d695ffe7..a45c928a75 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll index b6e0ba0453..6d050a1e17 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll index dd0f05e125..b3a8bbfd14 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll index 537f839550..c52c3dcca2 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll index e5b354e0f4..7473c42a1e 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll index a126f1005f..6622b69695 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll index 3e8feac629..a3ba444030 100644 --- a/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll index d5c6f74a7f..1d21dec156 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll index 2f4d672a76..442bebb366 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll index a7a9ea84d3..eefaa0d5a8 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll index 7f8ea390a9..e7e34e1d86 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll index 9be435913b..939627be27 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll index db618bc374..915a891269 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll index 2916d72356..3cc16f1b3d 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll index 71f8ad3386..64b90e3f88 100644 --- a/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll index 0850246521..38656cfd61 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll index 099f62af1d..4901868c20 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll index c3f9e73766..69b1b97079 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll index 2c966709c8..f2f6484cc6 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll index 964db6ffe5..f07124a7bd 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll index e9b7c2712a..af843d056f 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll index 08e9934829..d2b1c8d841 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll index 5aafcbe2c4..24e347f500 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll index b30f068331..6ea6d5a2b3 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll index 9da665295b..f5b32e7400 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll index 71bb98fcab..6f09533dc9 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll index d0636ec068..9f18e807dc 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll index a4d698320b..3a739ea7fa 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll index e7dee2e3e9..09d677d882 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll index 483be82f9a..3cdce052b2 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll index a13a90508c..f09d33a2ec 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll index 5227f4c146..fd4ca09f8c 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll index 31788fae16..1878b88aea 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll index 1f6165d13c..1b6c36958b 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll index 1bf74ed3fb..8c0c1bd954 100644 --- a/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll index d2c54f1295..5b70c967e9 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll index e0bcd14872..b9c6ba5ff2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll index 4ae9f848cb..47f3e30614 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll index 5b5bbe3323..aaa47974a0 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll index 28870da031..854c172025 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll index 3495c366ea..d7359ee9fa 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll index 6b0a05e818..faf204822f 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll index 25707ae659..21e3dab6f1 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll index 31d5a18c92..19769e1d63 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll index f9f7ffa76d..1fe25f1874 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll index 5059228b34..09c9d8c81d 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll index 12ce6f2a9a..abd3d458e9 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll index bff33f80c7..114425f187 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll index 04ba90da8b..310944187b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll index fadc501d15..3a6a577e17 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll index 1ce1ad0e48..c62442165b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll index c86c1ff5bd..ca32ef0372 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll index 439a394053..f704edee51 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll index c99c3e53e8..cfa40ef1b5 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll index f72f045c70..425f4aa1ef 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll index cf8d239357..177c0e7eac 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll index 4b2f0007bd..a549ff9994 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll index 65679c06d1..28bcfd8bb6 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll index 63e92f2a0e..2b3d3586d2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll index f1857a7789..50b6e59949 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll index 667fec5619..9bc386c3ba 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll index 6a51ca130e..68a4ad35e0 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll index 2996ed8bfa..1d33fcd8e7 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll index 148beeaffe..77cee99645 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll index 624fe97fe9..afda5988bb 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll index 77dde641da..c54945a0ac 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll index 1efd307b51..515b415ef2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll index ebf30bb152..37a0d4652a 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll index be386d3014..3b4c87c90a 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll index 8d28298011..979168fe5f 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll index 0a70815a88..b60c878698 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll index 12d3322d7d..5d0ded5c80 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll index 5972bab465..201c4b76a1 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll index 05bedd4ca7..1a09e7ca54 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll index 3c852f3458..7c4c2c2473 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll index ad7aca5d6b..71cc6e4e23 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll index 5651f7057e..dfe33c4f39 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll index 24b662e07c..23b0143a5c 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll index b59f4b8ae5..0d562d2081 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll index fc7f857d6f..b60f0efb19 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll index 2900a5a529..ecb2c8d951 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll index 3aea58be20..208a2ba382 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll index ee13db537e..05d508ff08 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll index 417b7daf67..cd1a69d300 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll index 15d50059d0..bfa7ec5c57 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll index ce60de9013..2dd83e7e1d 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll index 383c88c27d..adb2fd4932 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll index 4a78e54b7c..73450b5f9b 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll index 7df3e1f8ec..1b1bee9823 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll index 7d0036c102..f96ae6c8dd 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll index 1858a99fbe..18a3081efb 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll index 2eb2690023..c06a7e5900 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll index dabdf1f912..02569941bb 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll index db87bbc8d0..a7eda20f9a 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll index 6a2747a9f2..4252ee99e7 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll index a520bf42fd..3f2a83bdc6 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll index d91345bf6d..be31bd6175 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll index 6a8a68f043..832987550f 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll index badc172fb8..8cd75454ec 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll index d3ee678420..d5f0ec35d5 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll index 3612d40921..45c1d285d5 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll index 30048c40d0..9c2057c777 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll index 9d3a909982..8096bf8955 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll index 4c81d8736c..d9a877d4e5 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll index b921dbc6b4..1020e643d9 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll index 6cdbed3935..464b30e821 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll index a638854bab..ec162bdfda 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll index 8b7d7805ca..adc8843f27 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll index 1e3f034f45..4ab241b125 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll index 1ebbafd176..cd43ad70b2 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll index 447bdcf885..d8f85c93cc 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll index 39880d919a..07c4b88abc 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll index 29e49dfb6a..67b9bfd4d4 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll index e2ca8eaa43..0ab75b5214 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll index 24e55f62fb..eeea57554c 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll index 7af0218094..e2f273b4d8 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll index 0deadc777a..00d1e696f8 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll index b444e489a5..ba3964b339 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll index ee14d0a4f8..5392eb25ee 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll index c79a4f8547..ea05d6ec86 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll index 63ce3fa1c1..9cd02d0310 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll index f67d04342e..1e5d30f366 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll index 7450c5e86b..17d55083ed 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll index c56e48c5f5..f67f4b58ac 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll index fbbfec4ad9..f8d07af125 100644 --- a/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll +++ b/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/movmatrix_trans_b16.ll b/intrinsics/probes/movmatrix_trans_b16.ll index eb61aefd8c..de1be8e9a6 100644 --- a/intrinsics/probes/movmatrix_trans_b16.ll +++ b/intrinsics/probes/movmatrix_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_bf16x2.ll b/intrinsics/probes/mul_bf16x2.ll index c1aa9ab7fc..415434f0ec 100644 --- a/intrinsics/probes/mul_bf16x2.ll +++ b/intrinsics/probes/mul_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_f16x2.ll b/intrinsics/probes/mul_f16x2.ll index 54d9358d3f..a15c8b31bb 100644 --- a/intrinsics/probes/mul_f16x2.ll +++ b/intrinsics/probes/mul_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_f32x2.ll b/intrinsics/probes/mul_f32x2.ll index 6ba7f1a2cf..18ab882ad6 100644 --- a/intrinsics/probes/mul_f32x2.ll +++ b/intrinsics/probes/mul_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_ftz_f32x2.ll b/intrinsics/probes/mul_ftz_f32x2.ll index 92527a58e0..65e917029e 100644 --- a/intrinsics/probes/mul_ftz_f32x2.ll +++ b/intrinsics/probes/mul_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rm_f32.ll b/intrinsics/probes/mul_rm_f32.ll index 2f85919620..ee4f66b76e 100644 --- a/intrinsics/probes/mul_rm_f32.ll +++ b/intrinsics/probes/mul_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rm_f64.ll b/intrinsics/probes/mul_rm_f64.ll index d30395c092..5336510646 100644 --- a/intrinsics/probes/mul_rm_f64.ll +++ b/intrinsics/probes/mul_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rm_ftz_f32.ll b/intrinsics/probes/mul_rm_ftz_f32.ll index 1f81ed3f58..8dfbf95205 100644 --- a/intrinsics/probes/mul_rm_ftz_f32.ll +++ b/intrinsics/probes/mul_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rn_f32.ll b/intrinsics/probes/mul_rn_f32.ll index 10edac5a56..400dc07b77 100644 --- a/intrinsics/probes/mul_rn_f32.ll +++ b/intrinsics/probes/mul_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rn_f64.ll b/intrinsics/probes/mul_rn_f64.ll index 699ba1b364..18cd99901e 100644 --- a/intrinsics/probes/mul_rn_f64.ll +++ b/intrinsics/probes/mul_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rn_ftz_f32.ll b/intrinsics/probes/mul_rn_ftz_f32.ll index 14b830cd6c..d3a64a43ef 100644 --- a/intrinsics/probes/mul_rn_ftz_f32.ll +++ b/intrinsics/probes/mul_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rp_f32.ll b/intrinsics/probes/mul_rp_f32.ll index 6d4c8626fa..1d4df1b835 100644 --- a/intrinsics/probes/mul_rp_f32.ll +++ b/intrinsics/probes/mul_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rp_f64.ll b/intrinsics/probes/mul_rp_f64.ll index 01943608a3..db542dcf4a 100644 --- a/intrinsics/probes/mul_rp_f64.ll +++ b/intrinsics/probes/mul_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rp_ftz_f32.ll b/intrinsics/probes/mul_rp_ftz_f32.ll index 88058a3757..e5c4cfb9a3 100644 --- a/intrinsics/probes/mul_rp_ftz_f32.ll +++ b/intrinsics/probes/mul_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rz_f32.ll b/intrinsics/probes/mul_rz_f32.ll index 7094191e33..59b20f6df9 100644 --- a/intrinsics/probes/mul_rz_f32.ll +++ b/intrinsics/probes/mul_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rz_f64.ll b/intrinsics/probes/mul_rz_f64.ll index e4c450ed82..6d6c87f522 100644 --- a/intrinsics/probes/mul_rz_f64.ll +++ b/intrinsics/probes/mul_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/mul_rz_ftz_f32.ll b/intrinsics/probes/mul_rz_ftz_f32.ll index 80f9921d36..9921b03130 100644 --- a/intrinsics/probes/mul_rz_ftz_f32.ll +++ b/intrinsics/probes/mul_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nanosleep.ll b/intrinsics/probes/nanosleep.ll index 0e1169bca3..3d8946524c 100644 --- a/intrinsics/probes/nanosleep.ll +++ b/intrinsics/probes/nanosleep.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/neg_bf16x2.ll b/intrinsics/probes/neg_bf16x2.ll index c1b3631c78..2ba9788377 100644 --- a/intrinsics/probes/neg_bf16x2.ll +++ b/intrinsics/probes/neg_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/neg_f16x2.ll b/intrinsics/probes/neg_f16x2.ll index f80d43a2cc..fa0a028c79 100644 --- a/intrinsics/probes/neg_f16x2.ll +++ b/intrinsics/probes/neg_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nsmid.lib_nvvm.ll b/intrinsics/probes/nsmid.lib_nvvm.ll index 7721ebafad..726dcf2f53 100644 --- a/intrinsics/probes/nsmid.lib_nvvm.ll +++ b/intrinsics/probes/nsmid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nsmid.llvm_nvptx.ll b/intrinsics/probes/nsmid.llvm_nvptx.ll index 4e45521e75..f1719d7412 100644 --- a/intrinsics/probes/nsmid.llvm_nvptx.ll +++ b/intrinsics/probes/nsmid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nwarpid.lib_nvvm.ll b/intrinsics/probes/nwarpid.lib_nvvm.ll index 3e3bb08f14..1a84319df0 100644 --- a/intrinsics/probes/nwarpid.lib_nvvm.ll +++ b/intrinsics/probes/nwarpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/nwarpid.llvm_nvptx.ll b/intrinsics/probes/nwarpid.llvm_nvptx.ll index 05c381ad9c..f3f36e6560 100644 --- a/intrinsics/probes/nwarpid.llvm_nvptx.ll +++ b/intrinsics/probes/nwarpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/packed_atomic_add_bf16x2.ll b/intrinsics/probes/packed_atomic_add_bf16x2.ll index 8ce88380d6..afe810e64d 100644 --- a/intrinsics/probes/packed_atomic_add_bf16x2.ll +++ b/intrinsics/probes/packed_atomic_add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/packed_atomic_add_f16x2.ll b/intrinsics/probes/packed_atomic_add_f16x2.ll index d45b9003c9..44484e926d 100644 --- a/intrinsics/probes/packed_atomic_add_f16x2.ll +++ b/intrinsics/probes/packed_atomic_add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/pmevent.ll b/intrinsics/probes/pmevent.ll index 661b4b1ae9..db6a08a71f 100644 --- a/intrinsics/probes/pmevent.ll +++ b/intrinsics/probes/pmevent.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prefetch_tma_descriptor.ll b/intrinsics/probes/prefetch_tma_descriptor.ll index 06c1ccb4b4..4cdfcf7c1c 100644 --- a/intrinsics/probes/prefetch_tma_descriptor.ll +++ b/intrinsics/probes/prefetch_tma_descriptor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt.ll b/intrinsics/probes/prmt.ll index 8d363b04c1..b495f21de4 100644 --- a/intrinsics/probes/prmt.ll +++ b/intrinsics/probes/prmt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_b4e.ll b/intrinsics/probes/prmt_b4e.ll index fdd4f96867..bf1324533a 100644 --- a/intrinsics/probes/prmt_b4e.ll +++ b/intrinsics/probes/prmt_b4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_ecl.ll b/intrinsics/probes/prmt_ecl.ll index d9f3640d85..3a5be5e115 100644 --- a/intrinsics/probes/prmt_ecl.ll +++ b/intrinsics/probes/prmt_ecl.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_ecr.ll b/intrinsics/probes/prmt_ecr.ll index 8f22bef597..c515d5be6c 100644 --- a/intrinsics/probes/prmt_ecr.ll +++ b/intrinsics/probes/prmt_ecr.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_f4e.ll b/intrinsics/probes/prmt_f4e.ll index 45e26de539..c8abf69abf 100644 --- a/intrinsics/probes/prmt_f4e.ll +++ b/intrinsics/probes/prmt_f4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_rc16.ll b/intrinsics/probes/prmt_rc16.ll index 16b565c1d1..e606c61f8a 100644 --- a/intrinsics/probes/prmt_rc16.ll +++ b/intrinsics/probes/prmt_rc16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/prmt_rc8.ll b/intrinsics/probes/prmt_rc8.ll index 1fde8d5fe8..e2a4d3c6b3 100644 --- a/intrinsics/probes/prmt_rc8.ll +++ b/intrinsics/probes/prmt_rc8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_approx_ftz_f32.ll b/intrinsics/probes/rcp_approx_ftz_f32.ll index fef0dba0bd..1ceeaf8e17 100644 --- a/intrinsics/probes/rcp_approx_ftz_f32.ll +++ b/intrinsics/probes/rcp_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_approx_ftz_f64.ll b/intrinsics/probes/rcp_approx_ftz_f64.ll index 105ef4a2e5..4c476bb65e 100644 --- a/intrinsics/probes/rcp_approx_ftz_f64.ll +++ b/intrinsics/probes/rcp_approx_ftz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rm_f32.ll b/intrinsics/probes/rcp_rm_f32.ll index b441204b53..00f7acffd5 100644 --- a/intrinsics/probes/rcp_rm_f32.ll +++ b/intrinsics/probes/rcp_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rm_f64.ll b/intrinsics/probes/rcp_rm_f64.ll index cf4b693faa..a7835cbdc2 100644 --- a/intrinsics/probes/rcp_rm_f64.ll +++ b/intrinsics/probes/rcp_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rm_ftz_f32.ll b/intrinsics/probes/rcp_rm_ftz_f32.ll index d389565526..5b8fb78ee5 100644 --- a/intrinsics/probes/rcp_rm_ftz_f32.ll +++ b/intrinsics/probes/rcp_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rn_f32.ll b/intrinsics/probes/rcp_rn_f32.ll index d6a387a560..26d5db2be1 100644 --- a/intrinsics/probes/rcp_rn_f32.ll +++ b/intrinsics/probes/rcp_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rn_f64.ll b/intrinsics/probes/rcp_rn_f64.ll index da02ad2519..57abdb9de8 100644 --- a/intrinsics/probes/rcp_rn_f64.ll +++ b/intrinsics/probes/rcp_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rn_ftz_f32.ll b/intrinsics/probes/rcp_rn_ftz_f32.ll index 4d25aa11fe..216ce9934c 100644 --- a/intrinsics/probes/rcp_rn_ftz_f32.ll +++ b/intrinsics/probes/rcp_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rp_f32.ll b/intrinsics/probes/rcp_rp_f32.ll index bebb01b8b7..b84f8e3f3b 100644 --- a/intrinsics/probes/rcp_rp_f32.ll +++ b/intrinsics/probes/rcp_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rp_f64.ll b/intrinsics/probes/rcp_rp_f64.ll index 61ff35afaf..02afb81d7a 100644 --- a/intrinsics/probes/rcp_rp_f64.ll +++ b/intrinsics/probes/rcp_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rp_ftz_f32.ll b/intrinsics/probes/rcp_rp_ftz_f32.ll index 0263af7acd..3405a2cabe 100644 --- a/intrinsics/probes/rcp_rp_ftz_f32.ll +++ b/intrinsics/probes/rcp_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rz_f32.ll b/intrinsics/probes/rcp_rz_f32.ll index 923614eafa..17bc05536d 100644 --- a/intrinsics/probes/rcp_rz_f32.ll +++ b/intrinsics/probes/rcp_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rz_f64.ll b/intrinsics/probes/rcp_rz_f64.ll index 607567956e..fe2957f939 100644 --- a/intrinsics/probes/rcp_rz_f64.ll +++ b/intrinsics/probes/rcp_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rcp_rz_ftz_f32.ll b/intrinsics/probes/rcp_rz_ftz_f32.ll index 52af228886..8b92f7c0b1 100644 --- a/intrinsics/probes/rcp_rz_ftz_f32.ll +++ b/intrinsics/probes/rcp_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_add.ll b/intrinsics/probes/redux_sync_add.ll index 6fa409fedf..6bfff266eb 100644 --- a/intrinsics/probes/redux_sync_add.ll +++ b/intrinsics/probes/redux_sync_add.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_and.ll b/intrinsics/probes/redux_sync_and.ll index f636eaa949..c8695fc356 100644 --- a/intrinsics/probes/redux_sync_and.ll +++ b/intrinsics/probes/redux_sync_and.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_abs_f32.ll b/intrinsics/probes/redux_sync_max_abs_f32.ll index ed3d2b2b6b..b2f02dc8b0 100644 --- a/intrinsics/probes/redux_sync_max_abs_f32.ll +++ b/intrinsics/probes/redux_sync_max_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_abs_nan_f32.ll b/intrinsics/probes/redux_sync_max_abs_nan_f32.ll index 86c46f8c1b..c0b30ee9cd 100644 --- a/intrinsics/probes/redux_sync_max_abs_nan_f32.ll +++ b/intrinsics/probes/redux_sync_max_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_f32.ll b/intrinsics/probes/redux_sync_max_f32.ll index 7df4271426..aa5b7a9cfe 100644 --- a/intrinsics/probes/redux_sync_max_f32.ll +++ b/intrinsics/probes/redux_sync_max_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_i32.ll b/intrinsics/probes/redux_sync_max_i32.ll index ffe9bdf596..652c751471 100644 --- a/intrinsics/probes/redux_sync_max_i32.ll +++ b/intrinsics/probes/redux_sync_max_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_nan_f32.ll b/intrinsics/probes/redux_sync_max_nan_f32.ll index 1e8d732997..0b18049906 100644 --- a/intrinsics/probes/redux_sync_max_nan_f32.ll +++ b/intrinsics/probes/redux_sync_max_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_max_u32.ll b/intrinsics/probes/redux_sync_max_u32.ll index 94c0e36224..20e858609f 100644 --- a/intrinsics/probes/redux_sync_max_u32.ll +++ b/intrinsics/probes/redux_sync_max_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_abs_f32.ll b/intrinsics/probes/redux_sync_min_abs_f32.ll index acd58ca578..46e1a126b1 100644 --- a/intrinsics/probes/redux_sync_min_abs_f32.ll +++ b/intrinsics/probes/redux_sync_min_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_abs_nan_f32.ll b/intrinsics/probes/redux_sync_min_abs_nan_f32.ll index 024477177d..e7d1266e07 100644 --- a/intrinsics/probes/redux_sync_min_abs_nan_f32.ll +++ b/intrinsics/probes/redux_sync_min_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_f32.ll b/intrinsics/probes/redux_sync_min_f32.ll index dadb7ee5ba..31d3c6538c 100644 --- a/intrinsics/probes/redux_sync_min_f32.ll +++ b/intrinsics/probes/redux_sync_min_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_i32.ll b/intrinsics/probes/redux_sync_min_i32.ll index 1b68195374..272264a988 100644 --- a/intrinsics/probes/redux_sync_min_i32.ll +++ b/intrinsics/probes/redux_sync_min_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_nan_f32.ll b/intrinsics/probes/redux_sync_min_nan_f32.ll index 7acfed2132..6cd09f4d8c 100644 --- a/intrinsics/probes/redux_sync_min_nan_f32.ll +++ b/intrinsics/probes/redux_sync_min_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_min_u32.ll b/intrinsics/probes/redux_sync_min_u32.ll index 7ff8b78ae2..5ef296c4a6 100644 --- a/intrinsics/probes/redux_sync_min_u32.ll +++ b/intrinsics/probes/redux_sync_min_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_or.ll b/intrinsics/probes/redux_sync_or.ll index e89a02af0a..144ab93cfd 100644 --- a/intrinsics/probes/redux_sync_or.ll +++ b/intrinsics/probes/redux_sync_or.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/redux_sync_xor.ll b/intrinsics/probes/redux_sync_xor.ll index 84b947e298..e2cce3bd1c 100644 --- a/intrinsics/probes/redux_sync_xor.ll +++ b/intrinsics/probes/redux_sync_xor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rsqrt_approx_f32.ll b/intrinsics/probes/rsqrt_approx_f32.ll index de5e025417..d0b8f1e6b2 100644 --- a/intrinsics/probes/rsqrt_approx_f32.ll +++ b/intrinsics/probes/rsqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rsqrt_approx_f64.ll b/intrinsics/probes/rsqrt_approx_f64.ll index c381bb0c12..72645c2249 100644 --- a/intrinsics/probes/rsqrt_approx_f64.ll +++ b/intrinsics/probes/rsqrt_approx_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/rsqrt_approx_ftz_f32.ll b/intrinsics/probes/rsqrt_approx_ftz_f32.ll index d798d5a19f..e1ec9f0b8d 100644 --- a/intrinsics/probes/rsqrt_approx_ftz_f32.ll +++ b/intrinsics/probes/rsqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/setmaxnreg_dec.ll b/intrinsics/probes/setmaxnreg_dec.ll index af37b39561..a01d6120c9 100644 --- a/intrinsics/probes/setmaxnreg_dec.ll +++ b/intrinsics/probes/setmaxnreg_dec.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/setmaxnreg_inc.ll b/intrinsics/probes/setmaxnreg_inc.ll index 6a5a12f145..873403b585 100644 --- a/intrinsics/probes/setmaxnreg_inc.ll +++ b/intrinsics/probes/setmaxnreg_inc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_down_f32_sync.ll b/intrinsics/probes/shuffle_down_f32_sync.ll index f0cc07b049..eab673b7ee 100644 --- a/intrinsics/probes/shuffle_down_f32_sync.ll +++ b/intrinsics/probes/shuffle_down_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_down_sync.ll b/intrinsics/probes/shuffle_down_sync.ll index 0c190bb994..bd4ab233c1 100644 --- a/intrinsics/probes/shuffle_down_sync.ll +++ b/intrinsics/probes/shuffle_down_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_down_u64_sync.ll b/intrinsics/probes/shuffle_down_u64_sync.ll index 72e66c3ce4..84a108f10b 100644 --- a/intrinsics/probes/shuffle_down_u64_sync.ll +++ b/intrinsics/probes/shuffle_down_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_f32_sync.ll b/intrinsics/probes/shuffle_f32_sync.ll index 4e49b2feb7..dc4555a4ae 100644 --- a/intrinsics/probes/shuffle_f32_sync.ll +++ b/intrinsics/probes/shuffle_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_sync.ll b/intrinsics/probes/shuffle_sync.ll index 42ecc23be7..bdc2887e93 100644 --- a/intrinsics/probes/shuffle_sync.ll +++ b/intrinsics/probes/shuffle_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_u64_sync.ll b/intrinsics/probes/shuffle_u64_sync.ll index fe100ff054..10301a017f 100644 --- a/intrinsics/probes/shuffle_u64_sync.ll +++ b/intrinsics/probes/shuffle_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_up_f32_sync.ll b/intrinsics/probes/shuffle_up_f32_sync.ll index 38729f4ba6..65f3cbba26 100644 --- a/intrinsics/probes/shuffle_up_f32_sync.ll +++ b/intrinsics/probes/shuffle_up_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_up_sync.ll b/intrinsics/probes/shuffle_up_sync.ll index 9e09f6f83a..aa3d1e4519 100644 --- a/intrinsics/probes/shuffle_up_sync.ll +++ b/intrinsics/probes/shuffle_up_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_up_u64_sync.ll b/intrinsics/probes/shuffle_up_u64_sync.ll index 601928804b..6d0b4ea652 100644 --- a/intrinsics/probes/shuffle_up_u64_sync.ll +++ b/intrinsics/probes/shuffle_up_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_xor_f32_sync.ll b/intrinsics/probes/shuffle_xor_f32_sync.ll index 95da54f491..428793f0e1 100644 --- a/intrinsics/probes/shuffle_xor_f32_sync.ll +++ b/intrinsics/probes/shuffle_xor_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_xor_sync.ll b/intrinsics/probes/shuffle_xor_sync.ll index 4efa1e64ca..55fe8d89ae 100644 --- a/intrinsics/probes/shuffle_xor_sync.ll +++ b/intrinsics/probes/shuffle_xor_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/shuffle_xor_u64_sync.ll b/intrinsics/probes/shuffle_xor_u64_sync.ll index e6862cc95c..63710cfa3f 100644 --- a/intrinsics/probes/shuffle_xor_u64_sync.ll +++ b/intrinsics/probes/shuffle_xor_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sin_approx_f32.ll b/intrinsics/probes/sin_approx_f32.ll index cb89352f47..e174a28510 100644 --- a/intrinsics/probes/sin_approx_f32.ll +++ b/intrinsics/probes/sin_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sin_approx_ftz_f32.ll b/intrinsics/probes/sin_approx_ftz_f32.ll index b1fd971f5c..140135358b 100644 --- a/intrinsics/probes/sin_approx_ftz_f32.ll +++ b/intrinsics/probes/sin_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/smid.lib_nvvm.ll b/intrinsics/probes/smid.lib_nvvm.ll index a73644934d..ac99553d50 100644 --- a/intrinsics/probes/smid.lib_nvvm.ll +++ b/intrinsics/probes/smid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/smid.llvm_nvptx.ll b/intrinsics/probes/smid.llvm_nvptx.ll index eb1d182e2d..5ecd53ab59 100644 --- a/intrinsics/probes/smid.llvm_nvptx.ll +++ b/intrinsics/probes/smid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_approx_f32.ll b/intrinsics/probes/sqrt_approx_f32.ll index 72ff6b20c3..b775392711 100644 --- a/intrinsics/probes/sqrt_approx_f32.ll +++ b/intrinsics/probes/sqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_approx_ftz_f32.ll b/intrinsics/probes/sqrt_approx_ftz_f32.ll index 280aab1b93..7661b45531 100644 --- a/intrinsics/probes/sqrt_approx_ftz_f32.ll +++ b/intrinsics/probes/sqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rm_f32.ll b/intrinsics/probes/sqrt_rm_f32.ll index 8f985c9cd7..7e3f82ef06 100644 --- a/intrinsics/probes/sqrt_rm_f32.ll +++ b/intrinsics/probes/sqrt_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rm_f64.ll b/intrinsics/probes/sqrt_rm_f64.ll index a83ca9bc74..2cb5c081c0 100644 --- a/intrinsics/probes/sqrt_rm_f64.ll +++ b/intrinsics/probes/sqrt_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rm_ftz_f32.ll b/intrinsics/probes/sqrt_rm_ftz_f32.ll index 15e453699e..da609adc58 100644 --- a/intrinsics/probes/sqrt_rm_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rn_f32.ll b/intrinsics/probes/sqrt_rn_f32.ll index 31fee94478..41ff9308b7 100644 --- a/intrinsics/probes/sqrt_rn_f32.ll +++ b/intrinsics/probes/sqrt_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rn_f64.ll b/intrinsics/probes/sqrt_rn_f64.ll index e9a3f1ffbe..dec5449401 100644 --- a/intrinsics/probes/sqrt_rn_f64.ll +++ b/intrinsics/probes/sqrt_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rn_ftz_f32.ll b/intrinsics/probes/sqrt_rn_ftz_f32.ll index 3777543af4..c4748db2ba 100644 --- a/intrinsics/probes/sqrt_rn_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rp_f32.ll b/intrinsics/probes/sqrt_rp_f32.ll index 82037c4860..6ded37bae7 100644 --- a/intrinsics/probes/sqrt_rp_f32.ll +++ b/intrinsics/probes/sqrt_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rp_f64.ll b/intrinsics/probes/sqrt_rp_f64.ll index 9604d98682..38e9430e0b 100644 --- a/intrinsics/probes/sqrt_rp_f64.ll +++ b/intrinsics/probes/sqrt_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rp_ftz_f32.ll b/intrinsics/probes/sqrt_rp_ftz_f32.ll index 75ac3ae913..a32e5c433b 100644 --- a/intrinsics/probes/sqrt_rp_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rz_f32.ll b/intrinsics/probes/sqrt_rz_f32.ll index 367661169b..86a2602cf3 100644 --- a/intrinsics/probes/sqrt_rz_f32.ll +++ b/intrinsics/probes/sqrt_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rz_f64.ll b/intrinsics/probes/sqrt_rz_f64.ll index b1b24314cc..089d82fff3 100644 --- a/intrinsics/probes/sqrt_rz_f64.ll +++ b/intrinsics/probes/sqrt_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sqrt_rz_ftz_f32.ll b/intrinsics/probes/sqrt_rz_ftz_f32.ll index 01548379a0..4fcffbc1c9 100644 --- a/intrinsics/probes/sqrt_rz_ftz_f32.ll +++ b/intrinsics/probes/sqrt_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x2_b16.ll b/intrinsics/probes/stmatrix_m8n8_x2_b16.ll index 0d8744f163..cd3f3c379e 100644 --- a/intrinsics/probes/stmatrix_m8n8_x2_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll b/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll index cd059aa51f..28793e230b 100644 --- a/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x4_b16.ll b/intrinsics/probes/stmatrix_m8n8_x4_b16.ll index 972f13e200..ec256bfd4d 100644 --- a/intrinsics/probes/stmatrix_m8n8_x4_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll b/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll index 79566adecb..2cd595af90 100644 --- a/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll +++ b/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_bf16x2.ll b/intrinsics/probes/sub_bf16x2.ll index 02007ab138..459a8ef7e7 100644 --- a/intrinsics/probes/sub_bf16x2.ll +++ b/intrinsics/probes/sub_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_f16x2.ll b/intrinsics/probes/sub_f16x2.ll index ee084d0d9d..8de8b43bdd 100644 --- a/intrinsics/probes/sub_f16x2.ll +++ b/intrinsics/probes/sub_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_f32x2.ll b/intrinsics/probes/sub_f32x2.ll index c4dbe9a3d3..95b0becbbe 100644 --- a/intrinsics/probes/sub_f32x2.ll +++ b/intrinsics/probes/sub_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sub_ftz_f32x2.ll b/intrinsics/probes/sub_ftz_f32x2.ll index 9043eb52e7..9311c93cbc 100644 --- a/intrinsics/probes/sub_ftz_f32x2.ll +++ b/intrinsics/probes/sub_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sync_mask.ll b/intrinsics/probes/sync_mask.ll index ee16870a76..e145eb65f7 100644 --- a/intrinsics/probes/sync_mask.ll +++ b/intrinsics/probes/sync_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/sync_threads.ll b/intrinsics/probes/sync_threads.ll index a8130be45f..2a51012e39 100644 --- a/intrinsics/probes/sync_threads.ll +++ b/intrinsics/probes/sync_threads.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tanh_approx_f32.ll b/intrinsics/probes/tanh_approx_f32.ll index e3e34ae1df..f464c981b5 100644 --- a/intrinsics/probes/tanh_approx_f32.ll +++ b/intrinsics/probes/tanh_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_alloc.ll b/intrinsics/probes/tcgen05_alloc.ll index 6acfe45a1c..04626704bd 100644 --- a/intrinsics/probes/tcgen05_alloc.ll +++ b/intrinsics/probes/tcgen05_alloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_alloc_cg2.ll b/intrinsics/probes/tcgen05_alloc_cg2.ll index dc0cbca220..bbab4cc6a8 100644 --- a/intrinsics/probes/tcgen05_alloc_cg2.ll +++ b/intrinsics/probes/tcgen05_alloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit.ll b/intrinsics/probes/tcgen05_commit.ll index beae2ad8f3..14fd657648 100644 --- a/intrinsics/probes/tcgen05_commit.ll +++ b/intrinsics/probes/tcgen05_commit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_cg2.ll b/intrinsics/probes/tcgen05_commit_cg2.ll index 023e507d30..2a32f51928 100644 --- a/intrinsics/probes/tcgen05_commit_cg2.ll +++ b/intrinsics/probes/tcgen05_commit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_multicast.ll b/intrinsics/probes/tcgen05_commit_multicast.ll index 699a249e90..1205651f9b 100644 --- a/intrinsics/probes/tcgen05_commit_multicast.ll +++ b/intrinsics/probes/tcgen05_commit_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_multicast_cg2.ll b/intrinsics/probes/tcgen05_commit_multicast_cg2.ll index 4772d30167..bd4a7f8901 100644 --- a/intrinsics/probes/tcgen05_commit_multicast_cg2.ll +++ b/intrinsics/probes/tcgen05_commit_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_shared_cluster.ll b/intrinsics/probes/tcgen05_commit_shared_cluster.ll index acb1289973..b28f46d6ad 100644 --- a/intrinsics/probes/tcgen05_commit_shared_cluster.ll +++ b/intrinsics/probes/tcgen05_commit_shared_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll b/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll index a1433b90ed..a9ecdd2a48 100644 --- a/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll +++ b/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b.ll b/intrinsics/probes/tcgen05_cp_128x128b.ll index 874d863547..e57c657f86 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll index f4e3048836..c16c2fab4e 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll index 718a763a12..bffad6c71e 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll index 2d06edfdf3..baf94c3969 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll index 035c14ab55..1b61765a3c 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll b/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll index 57f39a88a9..4c8723be98 100644 --- a/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll index be74fe61f7..d0a98c8c68 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll index c011f0964b..27a92a5fb0 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll index d129867b93..a981053620 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll index 91b0ad59d5..412e09bc6c 100644 --- a/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll index 388a3a637d..667f2014f0 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll index e3d340db5d..6ee8a7942f 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll index 544639637e..29c6662ea1 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll index b88e27ade7..42da026122 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll index 0f61ebfa29..dbbe8b62ec 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll index 75716157c3..b1e9ef6b70 100644 --- a/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b.ll b/intrinsics/probes/tcgen05_cp_4x256b.ll index d0477d394b..7b32499835 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll index bc8bcb758c..1e1f233b5f 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll index b5008cc128..dafdf51991 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll index 2404cfa452..d854b87661 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll index 6346a9b18c..c825a89616 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll b/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll index 0823b74275..cda5e7673c 100644 --- a/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll index 851638d6ee..67ee318a50 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll index 6154c24fd5..26375a2392 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll index 9c81cd3854..48441f0bcd 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll index e6232c72ef..cc24a6295f 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll index 04eb00bd45..795d210c83 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll index 913e0c9653..f3fcb308f0 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll index cd6b4548c8..1fbd5549ad 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll index ca2f614a30..aca6e573bf 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll index ea163ebf31..b3051ff000 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll index 900231f06d..c792dfa543 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll index ee2c9aa47a..7103b9e62d 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll index 4b53f40f9a..f1889defe1 100644 --- a/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll b/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll index f2927ff61e..41f6624f1a 100644 --- a/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll +++ b/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll b/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll index 2c6d574821..da67006e71 100644 --- a/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll +++ b/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_dealloc.ll b/intrinsics/probes/tcgen05_dealloc.ll index 7bae53ecbc..654e7c9d44 100644 --- a/intrinsics/probes/tcgen05_dealloc.ll +++ b/intrinsics/probes/tcgen05_dealloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_dealloc_cg2.ll b/intrinsics/probes/tcgen05_dealloc_cg2.ll index 9c5c87deee..3d2911f033 100644 --- a/intrinsics/probes/tcgen05_dealloc_cg2.ll +++ b/intrinsics/probes/tcgen05_dealloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_fence_after_thread_sync.ll b/intrinsics/probes/tcgen05_fence_after_thread_sync.ll index 4d99f51ecf..4ac265279a 100644 --- a/intrinsics/probes/tcgen05_fence_after_thread_sync.ll +++ b/intrinsics/probes/tcgen05_fence_after_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_fence_before_thread_sync.ll b/intrinsics/probes/tcgen05_fence_before_thread_sync.ll index 2e85368df9..e4d8fb18d6 100644 --- a/intrinsics/probes/tcgen05_fence_before_thread_sync.ll +++ b/intrinsics/probes/tcgen05_fence_before_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll index 66eea7805c..782f6e7210 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll index ebcd6d5324..683c6fd04d 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll index 122e4ea3af..ea85b7bfd7 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll index ed09506f35..264902b6e9 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll index 2c16e905b2..a6873e90b6 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll index cb6e7fafc3..b158fc207a 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll index 38733d57e4..69e9fc8311 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll index 65fd5cc7b5..2fc42aee34 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll index 0676b6a0f6..e4e447819d 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll index e40101f924..c308cc9788 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll index dedb0a0aa8..f4d15159b9 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll index cc8df14d6f..53834f4967 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll index 3fd33fb9fb..afd627e03b 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll index ed89859528..296eb369ef 100644 --- a/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_pure.ll b/intrinsics/probes/tcgen05_ld_16x256b_pure.ll index a54c69118f..228656e966 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_pure.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll index ea56b1fcc8..d18b404fa2 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll index 9a97fcdb32..bb81253e88 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll index e0edac159d..d1816b3e5d 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll index e900efdf2c..72931df51e 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll index 77aadce51d..00a9532924 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll index 8c31fbdd5f..17c7d1bac5 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll index 9e62b6aa48..52b794b7b8 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll index 62e4a4c350..09acfb9636 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll index 26d54c0e9b..694aba4aea 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll index 89f8e40fb3..31a1e24d57 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll index 6a77ff0120..221ce5c4fc 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll b/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll index d50558f4de..23f9c2fc0e 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll index 40dbcdecff..a6fc9a3fff 100644 --- a/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll index 4ffa096247..35e57274a6 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll index 80e486edc9..f77be88622 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll index 3bc438421d..8a8e24db01 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll index b525a53143..e78e6d15f2 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll index ea91a76523..81b7b95264 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll index 00da06b059..13149359ea 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll index 4c8c9985a6..812ac18b1a 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll index 4b3daade89..c61ddd8ff4 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll index d11833737e..9961bba395 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll index 0c7fcf9f74..16ada894fe 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll index 03733b0c51..28c5118fdd 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll index 43a58a1463..211a432bf9 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll index cbbf017da2..2c1472ec56 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll index d9a4c1e9b7..cd4779f668 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll index dd2114bcc3..725c6971d0 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll index be835e1829..494e4c91e2 100644 --- a/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll index 1f9edb56a4..e0cde7339d 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll index f18bbd320d..cdc9a42f29 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll index e29c844fda..5998903cd9 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll index 624bc96f51..98bae0b8b6 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll index bda45b63ea..71a55a7302 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll index 8f179c43fb..e8dea6881b 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll index 0c0110ad7c..28d568f4ad 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll index 8b9844f114..1758f21396 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll index 5f44b02772..e28f0709b9 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll index d2416f0b00..354752229b 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll index 93e718a265..4407c5b7ac 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll index f1b7f0bfdb..f8ad8c369e 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll index 911d27c9de..d9bc761ac0 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll index b2da613b08..2fe1c19176 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll index f803c14f8c..98fa739f4b 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll index 41b56f833c..9184a0f5cd 100644 --- a/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll index cf894651a3..2641fcfc99 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll index eb6d10e4ee..0674bfab16 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll index 1de6de1955..0bd50594b8 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll index 2a64bffd5b..bb1d69cb30 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll index 2d5f8d1a89..e0fc82fbe9 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll index c00badb7f2..6bb8157234 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll index baaf6a82bd..bb108f5828 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll index 73e2290d58..3f72bb96b4 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll index a32e916295..a64c7879fb 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll index ea7f80426f..cbae20a4d2 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll index eb7afc20a7..9ae1a1bcbd 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll index 114b18dc87..09557160e6 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll index 78cfc8971e..d0cd8e8c2f 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll index 12f42ca913..cb3d4394ed 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll b/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll index 8695c8b553..85ea449c40 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll b/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll index 690bb364d6..620d351755 100644 --- a/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_load_wait.ll b/intrinsics/probes/tcgen05_load_wait.ll index 2bdc2cd00e..38f71d3c7a 100644 --- a/intrinsics/probes/tcgen05_load_wait.ll +++ b/intrinsics/probes/tcgen05_load_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e2m1.ll b/intrinsics/probes/tcgen05_mma_e2m1.ll index 3d86f3ec30..3a48362320 100644 --- a/intrinsics/probes/tcgen05_mma_e2m1.ll +++ b/intrinsics/probes/tcgen05_mma_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e2m3.ll b/intrinsics/probes/tcgen05_mma_e2m3.ll index 35aa6aad96..d3013d838a 100644 --- a/intrinsics/probes/tcgen05_mma_e2m3.ll +++ b/intrinsics/probes/tcgen05_mma_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e3m2.ll b/intrinsics/probes/tcgen05_mma_e3m2.ll index efddea9aba..6bd650a5c6 100644 --- a/intrinsics/probes/tcgen05_mma_e3m2.ll +++ b/intrinsics/probes/tcgen05_mma_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e4m3.ll b/intrinsics/probes/tcgen05_mma_e4m3.ll index a0d6aea99c..7c7c9930f6 100644 --- a/intrinsics/probes/tcgen05_mma_e4m3.ll +++ b/intrinsics/probes/tcgen05_mma_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_e5m2.ll b/intrinsics/probes/tcgen05_mma_e5m2.ll index 4b442a42ea..3ebab7e3f5 100644 --- a/intrinsics/probes/tcgen05_mma_e5m2.ll +++ b/intrinsics/probes/tcgen05_mma_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_f16.ll b/intrinsics/probes/tcgen05_mma_f16.ll index ea9cbec878..0fa9544c11 100644 --- a/intrinsics/probes/tcgen05_mma_f16.ll +++ b/intrinsics/probes/tcgen05_mma_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_f16_cg2.ll b/intrinsics/probes/tcgen05_mma_f16_cg2.ll index 2ff1642e13..75b56cf042 100644 --- a/intrinsics/probes/tcgen05_mma_f16_cg2.ll +++ b/intrinsics/probes/tcgen05_mma_f16_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_shared.ll b/intrinsics/probes/tcgen05_mma_shared.ll index bd1a07c0b6..20aecda5b1 100644 --- a/intrinsics/probes/tcgen05_mma_shared.ll +++ b/intrinsics/probes/tcgen05_mma_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_sp_shared.ll b/intrinsics/probes/tcgen05_mma_sp_shared.ll index 8e9bf92012..b04039688c 100644 --- a/intrinsics/probes/tcgen05_mma_sp_shared.ll +++ b/intrinsics/probes/tcgen05_mma_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_sp_tensor.ll b/intrinsics/probes/tcgen05_mma_sp_tensor.ll index 6cf4f5d3d3..e64fa383ff 100644 --- a/intrinsics/probes/tcgen05_mma_sp_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll b/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll index 7c7e82d3aa..f578ea7324 100644 --- a/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll +++ b/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_tensor.ll b/intrinsics/probes/tcgen05_mma_tensor.ll index 16f6ee0987..4657a4f473 100644 --- a/intrinsics/probes/tcgen05_mma_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_tensor_ashift.ll b/intrinsics/probes/tcgen05_mma_tensor_ashift.ll index 7c81061cd3..baab94d23a 100644 --- a/intrinsics/probes/tcgen05_mma_tensor_ashift.ll +++ b/intrinsics/probes/tcgen05_mma_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_bf16.ll b/intrinsics/probes/tcgen05_mma_ws_bf16.ll index 7bdec871ab..5a91d1ac38 100644 --- a/intrinsics/probes/tcgen05_mma_ws_bf16.ll +++ b/intrinsics/probes/tcgen05_mma_ws_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e2m1.ll b/intrinsics/probes/tcgen05_mma_ws_e2m1.ll index 1926738c37..bfc11defb7 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e2m1.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e2m3.ll b/intrinsics/probes/tcgen05_mma_ws_e2m3.ll index 6c8f8d352b..a85aea31c4 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e2m3.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e3m2.ll b/intrinsics/probes/tcgen05_mma_ws_e3m2.ll index edc85afaa4..6b97341b14 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e3m2.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e4m3.ll b/intrinsics/probes/tcgen05_mma_ws_e4m3.ll index d9b5d3ed36..991a111878 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e4m3.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_e5m2.ll b/intrinsics/probes/tcgen05_mma_ws_e5m2.ll index 1e8ae5c0fc..39982f6558 100644 --- a/intrinsics/probes/tcgen05_mma_ws_e5m2.ll +++ b/intrinsics/probes/tcgen05_mma_ws_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_f16.ll b/intrinsics/probes/tcgen05_mma_ws_f16.ll index f1e5272d20..65d4b4f6e5 100644 --- a/intrinsics/probes/tcgen05_mma_ws_f16.ll +++ b/intrinsics/probes/tcgen05_mma_ws_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_shared.ll b/intrinsics/probes/tcgen05_mma_ws_shared.ll index 06651cf6d4..219c873bf8 100644 --- a/intrinsics/probes/tcgen05_mma_ws_shared.ll +++ b/intrinsics/probes/tcgen05_mma_ws_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll index 12b9e3e785..8606c4b984 100644 --- a/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll b/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll index 6ac3fab437..6565d79e56 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll index bdfd06214a..daa6a6b8db 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll b/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll index fda46dbb86..b0d6d83b0a 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll index e477013ea4..6db1eab1d7 100644 --- a/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_tensor.ll b/intrinsics/probes/tcgen05_mma_ws_tensor.ll index 8c0f0911a5..19f4dc8a9c 100644 --- a/intrinsics/probes/tcgen05_mma_ws_tensor.ll +++ b/intrinsics/probes/tcgen05_mma_ws_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll b/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll index d0a1478f1e..ef33da446f 100644 --- a/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll +++ b/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_mma_ws_tf32.ll b/intrinsics/probes/tcgen05_mma_ws_tf32.ll index 1ef72f1863..bd9bd0324b 100644 --- a/intrinsics/probes/tcgen05_mma_ws_tf32.ll +++ b/intrinsics/probes/tcgen05_mma_ws_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll b/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll index e4d3b3cb80..c5f412d03a 100644 --- a/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll +++ b/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll b/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll index b018f24a45..54df62df76 100644 --- a/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll +++ b/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_shift_down.ll b/intrinsics/probes/tcgen05_shift_down.ll index a21d3b771b..3f312779b6 100644 --- a/intrinsics/probes/tcgen05_shift_down.ll +++ b/intrinsics/probes/tcgen05_shift_down.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_shift_down_cg2.ll b/intrinsics/probes/tcgen05_shift_down_cg2.ll index 10cb9cc4ae..988e978600 100644 --- a/intrinsics/probes/tcgen05_shift_down_cg2.ll +++ b/intrinsics/probes/tcgen05_shift_down_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll index 7c429f8205..a1239d179b 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll index a4d28ec63b..8ba276231d 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll index 04cea28f4e..a6b32589c8 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll index 89567ebd34..d2890806db 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll index 6dc84f6ecb..56cc19ea60 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll index 3aa981e65b..a5619e465e 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll index 2ab4eca27c..b10b12909b 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll index e8f73ec97e..b064c9613b 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll index 6991738971..8c0161ab04 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll index 95f5fc3501..cbad4620e0 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll index b405eb8847..74095567de 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll index 9214e24793..23dc15eb8b 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll index d6824be5c9..0928a7cfdd 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll index 9a193796f4..c78f4e73a3 100644 --- a/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll index fda79237c9..9945cca421 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll index 8f9d7122b3..ef7ec0a3ba 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll index fdf8be91d4..aecab84459 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll index 7e9d57b1d4..c05bbe253e 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll index 51f68b428f..3aa2a080eb 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll index 2a22da8dfb..774a773036 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll index 78d13c1601..dce2344a5e 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll index 4324fd6448..eb023ca814 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll index a02f312eac..e774602a19 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll index 445d595681..e944c77b46 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll index 179fe1d8ce..41c07fa667 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll index 7f2e27bb63..e3ff2960fe 100644 --- a/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll index 94ba07bf66..8c72f75f34 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll index 318fa8dfef..56ec6c6db4 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll index bfc9cad2b2..67dadf096c 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll index 68278293aa..9658b81113 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll index dd0ebe55da..8fc737e7f2 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll index c173bd426c..5bbbb314af 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll index 64704b9451..011b82c370 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll index 6c21d82c47..5b331b7724 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll index c33a5a2262..08bc47240f 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll index 78e1a39bb5..85e6c94bed 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll index 96bec07f91..48d8544bc9 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll index d613c5f820..d2cd9dbe77 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll index 4095f9f91f..7717a20137 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll index f69630077d..1c18689075 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll index 99b477f92b..63ff2ac9cd 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll index b88ab5dfe7..3564bf5b72 100644 --- a/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll index 230955bafa..709775021b 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll index 6e9e9b3ee1..74a9d054fc 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll index 19baa5f06f..6032325a89 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll index 0133bde57f..53f9cf0c74 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll index e4dcd66474..1a9253fbd9 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll index 6c19cdad41..71e5693710 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll index ee58d4837d..bf763df4ea 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll index 8c2b89be5b..e315c1b1bf 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll index 97bac36f86..acebeeb9eb 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll index 41b64f612c..1f9dd18dd6 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll index 903a071085..8e60e1ac6e 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll index dd2c2a4782..a3ebb22c46 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll index f14f075df1..1b37bf22af 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll index d93185a9d6..dae77b0c15 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll b/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll index 5fe775d3cd..548c836e36 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll index a544ecb655..abe763a131 100644 --- a/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll index 60a076fe83..6dbd997ae4 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll index 545930c809..e2e73af8c4 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll index 827771e3d5..d62c39a2ad 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll index 1721568cf6..c094d59a32 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll index 15d598cec6..cb16ac2657 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll index a698bdc75d..1174e5ae03 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll index ecedbf8056..6e03531652 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll index 6a83680da1..e553ad04c3 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll index beea9becc1..a82ab54e2e 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll index 9b42f9b0a4..5dce371b0f 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll index ccec22823a..50aa664b76 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll index c8b9ddc19c..980a689e26 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll index b717e71d43..5352d21f03 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll index 25ac6dd11b..bbf852a424 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll b/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll index 97b923d8a6..9bfe1af3a2 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll b/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll index b810bab3a5..3845472efe 100644 --- a/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll +++ b/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tcgen05_store_wait.ll b/intrinsics/probes/tcgen05_store_wait.ll index 7726764af4..20d1356a46 100644 --- a/intrinsics/probes/tcgen05_store_wait.ll +++ b/intrinsics/probes/tcgen05_store_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_box_dim.ll b/intrinsics/probes/tensormap_replace_box_dim.ll index 383215df50..9e45e9fe4e 100644 --- a/intrinsics/probes/tensormap_replace_box_dim.ll +++ b/intrinsics/probes/tensormap_replace_box_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_element_stride.ll b/intrinsics/probes/tensormap_replace_element_stride.ll index aa4f704382..6cfcc01610 100644 --- a/intrinsics/probes/tensormap_replace_element_stride.ll +++ b/intrinsics/probes/tensormap_replace_element_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_element_type.ll b/intrinsics/probes/tensormap_replace_element_type.ll index 10bd454b1a..c48f957533 100644 --- a/intrinsics/probes/tensormap_replace_element_type.ll +++ b/intrinsics/probes/tensormap_replace_element_type.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_fill_mode.ll b/intrinsics/probes/tensormap_replace_fill_mode.ll index 4a4137fb54..5180c86304 100644 --- a/intrinsics/probes/tensormap_replace_fill_mode.ll +++ b/intrinsics/probes/tensormap_replace_fill_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_global_address.ll b/intrinsics/probes/tensormap_replace_global_address.ll index 4ae2808853..6f8a08ec76 100644 --- a/intrinsics/probes/tensormap_replace_global_address.ll +++ b/intrinsics/probes/tensormap_replace_global_address.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_global_dim.ll b/intrinsics/probes/tensormap_replace_global_dim.ll index 61aa6c25e2..7c868fa5ac 100644 --- a/intrinsics/probes/tensormap_replace_global_dim.ll +++ b/intrinsics/probes/tensormap_replace_global_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_global_stride.ll b/intrinsics/probes/tensormap_replace_global_stride.ll index c0758977bb..f69e3ff73d 100644 --- a/intrinsics/probes/tensormap_replace_global_stride.ll +++ b/intrinsics/probes/tensormap_replace_global_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_interleave_layout.ll b/intrinsics/probes/tensormap_replace_interleave_layout.ll index cd114237d2..aad1077526 100644 --- a/intrinsics/probes/tensormap_replace_interleave_layout.ll +++ b/intrinsics/probes/tensormap_replace_interleave_layout.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_rank.ll b/intrinsics/probes/tensormap_replace_rank.ll index fa0adf79fc..130b8749e2 100644 --- a/intrinsics/probes/tensormap_replace_rank.ll +++ b/intrinsics/probes/tensormap_replace_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll b/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll index dc59ddd5bd..9a9052aafd 100644 --- a/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll +++ b/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/tensormap_replace_swizzle_mode.ll b/intrinsics/probes/tensormap_replace_swizzle_mode.ll index 79573631db..b0835552bd 100644 --- a/intrinsics/probes/tensormap_replace_swizzle_mode.ll +++ b/intrinsics/probes/tensormap_replace_swizzle_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/thread_idx_x.ll b/intrinsics/probes/thread_idx_x.ll index 70f54c724c..f88d1ecf2d 100644 --- a/intrinsics/probes/thread_idx_x.ll +++ b/intrinsics/probes/thread_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/thread_idx_y.ll b/intrinsics/probes/thread_idx_y.ll index d8623f0ce3..961942780c 100644 --- a/intrinsics/probes/thread_idx_y.ll +++ b/intrinsics/probes/thread_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/thread_idx_z.ll b/intrinsics/probes/thread_idx_z.ll index 94614f0fda..cdeeec7c5c 100644 --- a/intrinsics/probes/thread_idx_z.ll +++ b/intrinsics/probes/thread_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/threadfence.ll b/intrinsics/probes/threadfence.ll index 4dcb217402..fa822eadcc 100644 --- a/intrinsics/probes/threadfence.ll +++ b/intrinsics/probes/threadfence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/threadfence_block.ll b/intrinsics/probes/threadfence_block.ll index cbb9417b54..643e98b5b5 100644 --- a/intrinsics/probes/threadfence_block.ll +++ b/intrinsics/probes/threadfence_block.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/threadfence_system.ll b/intrinsics/probes/threadfence_system.ll index c6aa00abc1..53d02edd54 100644 --- a/intrinsics/probes/threadfence_system.ll +++ b/intrinsics/probes/threadfence_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/total_smem_size.lib_nvvm.ll b/intrinsics/probes/total_smem_size.lib_nvvm.ll index 56ddcc2e84..4d2ec3edbf 100644 --- a/intrinsics/probes/total_smem_size.lib_nvvm.ll +++ b/intrinsics/probes/total_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/total_smem_size.llvm_nvptx.ll b/intrinsics/probes/total_smem_size.llvm_nvptx.ll index 8ac62d912e..8655d75696 100644 --- a/intrinsics/probes/total_smem_size.llvm_nvptx.ll +++ b/intrinsics/probes/total_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/trap.ll b/intrinsics/probes/trap.ll index 50e301518a..8221dba208 100644 --- a/intrinsics/probes/trap.ll +++ b/intrinsics/probes/trap.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/uni_sync.ll b/intrinsics/probes/uni_sync.ll index c90bea50bf..23a5cef25b 100644 --- a/intrinsics/probes/uni_sync.ll +++ b/intrinsics/probes/uni_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/warpid.lib_nvvm.ll b/intrinsics/probes/warpid.lib_nvvm.ll index 647cb3118a..2e974878f0 100644 --- a/intrinsics/probes/warpid.lib_nvvm.ll +++ b/intrinsics/probes/warpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/warpid.llvm_nvptx.ll b/intrinsics/probes/warpid.llvm_nvptx.ll index a9453df671..56efa46b5c 100644 --- a/intrinsics/probes/warpid.llvm_nvptx.ll +++ b/intrinsics/probes/warpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/wgmma_commit_group.ll b/intrinsics/probes/wgmma_commit_group.ll index e6c502f5d5..d47098f2b0 100644 --- a/intrinsics/probes/wgmma_commit_group.ll +++ b/intrinsics/probes/wgmma_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/wgmma_fence.ll b/intrinsics/probes/wgmma_fence.ll index fc90766de2..e412d10a35 100644 --- a/intrinsics/probes/wgmma_fence.ll +++ b/intrinsics/probes/wgmma_fence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" diff --git a/intrinsics/probes/wgmma_wait_group.ll b/intrinsics/probes/wgmma_wait_group.ll index 8691bd31ae..55a584a0cd 100644 --- a/intrinsics/probes/wgmma_wait_group.ll +++ b/intrinsics/probes/wgmma_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 016c9d9b55b84d99d56fe2690dce4b3ba7ded5318890996a7459a142ddb5edf7. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. ; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. target triple = "nvptx64-nvidia-cuda" From 8ca523a14a7a6f2f971b231b472b9c287fb63c2a Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 18:02:41 +0530 Subject: [PATCH 2/7] fix(cuda-oxide-codegen): require LLVM 23 for PTX 9.1 The PTX vocabulary now contains +ptx91, so a 9.1 floor can reach an LLVM 22 llc. LLVM 22 ignores the feature with a warning and emits PTX 8.8; the build passed and the module failed to load. The gate refuses PTX 9.1 or newer on anything older than LLVM 23. The fixture for a floor newer than cuda-oxide can request moves to 9.2, and the floors test covers the 9.1 rule. Signed-off-by: nihalpasham --- .../cuda-oxide-codegen/src/target/select.rs | 30 ++++++++++++++----- .../src/target/tests/fixtures.rs | 10 +++---- .../src/target/tests/generated_floors.rs | 24 +++++++++++++-- 3 files changed, 48 insertions(+), 16 deletions(-) diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/target/select.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/target/select.rs index cdc258e6b8..efee412907 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/target/select.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/target/select.rs @@ -252,18 +252,32 @@ pub fn validate_target_for_llvm_major( Ok(()) } +/// Refuses a PTX ISA requirement that the LLVM backend cannot honor. +/// +/// Each PTX feature the compiler can request has an LLVM floor: `+ptx90` needs +/// LLVM 22 and `+ptx91` needs LLVM 23. An older llc does not fail on a feature +/// it lacks; LLVM 22 ignores `+ptx91` with a warning and emits PTX 8.8, so the +/// build would pass and the module would fail to load. The gate refuses such a +/// backend at compile time instead. An unknown backend version cannot prove +/// support and is refused whenever a floor applies. pub(crate) fn validate_ptx_isa_for_llvm_major( requirement: PtxIsaRequirement, llc_major: Option, ) -> Result<(), String> { - if requirement >= PtxIsaRequirement::new(90) && llc_major.is_none_or(|major| major < 22) { - let backend = llc_major.map_or_else( - || "an LLVM backend with an unknown version".to_string(), - |major| format!("LLVM {major}"), - ); - return Err(format!( - "PTX 9.0 or newer requires LLVM 22 or newer; {backend} does not support the required PTX feature" - )); + // Newest floor first, so the message names the rule that applies. + const LLVM_FLOORS: [(u16, u32, &str); 2] = [(91, 23, "9.1"), (90, 22, "9.0")]; + for (spelling, llvm_major, label) in LLVM_FLOORS { + if requirement >= PtxIsaRequirement::new(spelling) + && llc_major.is_none_or(|major| major < llvm_major) + { + let backend = llc_major.map_or_else( + || "an LLVM backend with an unknown version".to_string(), + |major| format!("LLVM {major}"), + ); + return Err(format!( + "PTX {label} or newer requires LLVM {llvm_major} or newer; {backend} does not support the required PTX feature" + )); + } } Ok(()) } diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/fixtures.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/fixtures.rs index e06a2abce7..17020cc268 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/fixtures.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/fixtures.rs @@ -54,14 +54,14 @@ pub(super) static PTX90: GeneratedIntrinsicTarget = GeneratedIntrinsicTarget { selections: &[], llvm: None, }; -pub(super) static PTX91_FUTURE: GeneratedIntrinsicTarget = GeneratedIntrinsicTarget { - marker: "test:ptx91", - id: "ptx91_future", +pub(super) static PTX92_FUTURE: GeneratedIntrinsicTarget = GeneratedIntrinsicTarget { + marker: "test:ptx92", + id: "ptx92_future", abi_id: "test", - dialect_op: "test.ptx91", + dialect_op: "test.ptx92", variant: GeneratedIntrinsicVariant::Scalar, requirement: GeneratedTargetRequirement { - minimum_ptx: GeneratedPtxVersion::from_encoded(91), + minimum_ptx: GeneratedPtxVersion::from_encoded(92), hardware: GeneratedHardwareTarget::All, }, backend_requirements: &[], diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/generated_floors.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/generated_floors.rs index 52ed6075ce..ef1079fb5d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/generated_floors.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/target/tests/generated_floors.rs @@ -3,7 +3,7 @@ * SPDX-License-Identifier: Apache-2.0 */ -use super::fixtures::{PTX87_EXACT_SM120A, PTX88, PTX90, PTX91_FUTURE, TCGEN_F16, TCGEN_I8}; +use super::fixtures::{PTX87_EXACT_SM120A, PTX88, PTX90, PTX92_FUTURE, TCGEN_F16, TCGEN_I8}; use crate::generated::GeneratedModuleRequirements; use crate::generated_intrinsic_targets::{ GeneratedHardwareAlternative, GeneratedHardwareTarget, GeneratedIntrinsicBackend, @@ -482,10 +482,28 @@ fn generated_ptx88_and_ptx90_floors_are_preserved() { assert!(validate_ptx_isa_for_llvm_major(PtxIsaRequirement::new(90), Some(21)).is_err()); validate_ptx_isa_for_llvm_major(PtxIsaRequirement::new(90), Some(22)).unwrap(); - let generated = GeneratedModuleRequirements::from_targets(vec![&PTX91_FUTURE]); + // PTX 9.1 is requestable and needs LLVM 23. LLVM 22 ignores `+ptx91` and + // emits PTX 8.8, so the gate has to refuse it at compile time. + assert_eq!( + ptx_isa_requirement_for_floor(91, "test", "test").unwrap(), + PtxIsaRequirement::new(91) + ); + assert_eq!( + required_ptx_feature(&"sm_120f".parse().unwrap(), PtxIsaRequirement::new(91)).unwrap(), + Some("+ptx91") + ); + assert!(validate_ptx_isa_for_llvm_major(PtxIsaRequirement::new(91), None).is_err()); + let error = validate_ptx_isa_for_llvm_major(PtxIsaRequirement::new(91), Some(22)).unwrap_err(); + assert!( + error.contains("PTX 9.1 or newer requires LLVM 23 or newer"), + "{error}" + ); + validate_ptx_isa_for_llvm_major(PtxIsaRequirement::new(91), Some(23)).unwrap(); + + let generated = GeneratedModuleRequirements::from_targets(vec![&PTX92_FUTURE]); let error = generated_ptx_isa_requirement(&generated).unwrap_err(); - assert!(error.contains("requires PTX 9.1"), "{error}"); + assert!(error.contains("requires PTX 9.2"), "{error}"); assert!( error.contains("newer than cuda-oxide can request"), "{error}" From d196775293a021415cd2053726b3a93b86ec7a0d Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 18:03:08 +0530 Subject: [PATCH 3/7] fix(intrinsics): record bf16x2 FP8 pack evidence under the current profiles The four bf16x2 to packed-FP8 records now meet the packed-conversion family contract instead of being exempted from it. The LLVM evidence is re-recorded with the pinned nightly-2026-10-11 llc and ptxas 13.3.73 at sm_120f with +ptx91: typed declaration canonicalization, typed and inline codegen, and ptxas assembly of both routes, which produce the same cubin. The libNVVM evidence is a real libNVVM 13.3.73 compile at compute_120f, a ptxas assembly and an nvJitLink 13.3 link of the LTOIR; the previous file reused another record's header around an llc run. The target list is the family-complete Blackwell set the catalog uses for its other family records, because cuda-oxide-codegen matches a family target only on the f suffix and the short list refused sm_120a. The admission test gains the fail-closed cases, the f16x2 field gets its serde default back, the generated outputs are regenerated, and the book catalog stamp moves with them. Signed-off-by: nihalpasham --- .../cuda-device/src/generated/async_copy.rs | 2 +- .../cuda-device/src/generated/atomic.rs | 2 +- .../crates/cuda-device/src/generated/bf16.rs | 2 +- .../cuda-device/src/generated/bf16x2.rs | 2 +- .../crates/cuda-device/src/generated/clc.rs | 2 +- .../src/generated/cluster_barrier.rs | 2 +- .../src/generated/cluster_memory.rs | 2 +- .../cuda-device/src/generated/cluster_sreg.rs | 2 +- .../cuda-device/src/generated/convert.rs | 38 +- .../src/generated/counted_barrier.rs | 2 +- .../src/generated/debug_control.rs | 2 +- .../cuda-device/src/generated/debug_sreg.rs | 2 +- .../cuda-device/src/generated/dotprod.rs | 2 +- .../crates/cuda-device/src/generated/f16.rs | 2 +- .../crates/cuda-device/src/generated/f16x2.rs | 2 +- .../crates/cuda-device/src/generated/f32x2.rs | 2 +- .../crates/cuda-device/src/generated/fence.rs | 2 +- .../crates/cuda-device/src/generated/float.rs | 2 +- .../src/generated/grid_dependency.rs | 2 +- .../cuda-device/src/generated/grid_sreg.rs | 2 +- .../crates/cuda-device/src/generated/i16x2.rs | 2 +- .../crates/cuda-device/src/generated/int.rs | 2 +- .../cuda-device/src/generated/ldmatrix.rs | 2 +- .../src/generated/mbarrier_basic.rs | 2 +- .../src/generated/mbarrier_extended.rs | 2 +- .../cuda-device/src/generated/movmatrix.rs | 2 +- .../crates/cuda-device/src/generated/prmt.rs | 2 +- .../src/generated/register_control.rs | 2 +- .../cuda-device/src/generated/register_mma.rs | 2 +- .../cuda-device/src/generated/shared_sreg.rs | 2 +- .../cuda-device/src/generated/sparse_mma.rs | 2 +- .../crates/cuda-device/src/generated/sreg.rs | 2 +- .../cuda-device/src/generated/stmatrix.rs | 2 +- .../cuda-device/src/generated/tcgen05.rs | 2 +- .../crates/cuda-device/src/generated/tma.rs | 2 +- .../cuda-device/src/generated/warp_sreg.rs | 2 +- .../src/generated/wgmma_control.rs | 2 +- .../cuda-intrinsics-gen/src/model/overlay.rs | 6 +- .../src/resolve/evidence.rs | 12 - .../src/resolve/families/packed_conversion.rs | 36 +- .../src/resolve/tests/scalar_packed.rs | 25 +- .../cuda-intrinsics/src/generated/abi_v1.rs | 50 +- .../cuda-intrinsics/src/generated/mod.rs | 30 +- .../active_mask.rs | 2 +- .../src/generated_intrinsic_targets/clc.rs | 2 +- .../cluster_barrier.rs | 2 +- .../cluster_memory.rs | 2 +- .../generated_intrinsic_targets/cp_async.rs | 2 +- .../debug_control.rs | 2 +- .../generated_intrinsic_targets/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/elect.rs | 2 +- .../execution_control.rs | 2 +- .../extended_minmax.rs | 2 +- .../integer_minmax.rs | 2 +- .../generated_intrinsic_targets/ldmatrix.rs | 2 +- .../mbarrier_basic.rs | 2 +- .../mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/mod.rs | 2 +- .../generated_intrinsic_targets/movmatrix.rs | 2 +- .../generated_intrinsic_targets/packed_alu.rs | 2 +- .../packed_atomic.rs | 2 +- .../packed_conversion.rs | 266 ++- .../src/generated_intrinsic_targets/prmt.rs | 2 +- .../src/generated_intrinsic_targets/redux.rs | 2 +- .../register_mma.rs | 2 +- .../scalar_arithmetic.rs | 2 +- .../scalar_conversion.rs | 2 +- .../scalar_math.rs | 2 +- .../generated_intrinsic_targets/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/sreg.rs | 2 +- .../generated_intrinsic_targets/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/sync.rs | 2 +- .../generated_intrinsic_targets/tcgen05_cp.rs | 2 +- .../generated_intrinsic_targets/tcgen05_ld.rs | 2 +- .../tcgen05_mma.rs | 2 +- .../tcgen05_mma_ws.rs | 2 +- .../tcgen05_other.rs | 2 +- .../generated_intrinsic_targets/tcgen05_st.rs | 2 +- .../tests/active_mask.rs | 2 +- .../generated_intrinsic_targets/tests/clc.rs | 2 +- .../tests/cluster_barrier.rs | 2 +- .../tests/cluster_memory.rs | 2 +- .../tests/cp_async.rs | 2 +- .../tests/debug_control.rs | 2 +- .../tests/dotprod.rs | 2 +- .../tests/elect.rs | 2 +- .../tests/execution_control.rs | 2 +- .../tests/extended_minmax.rs | 2 +- .../tests/integer_minmax.rs | 2 +- .../tests/ldmatrix.rs | 2 +- .../tests/mbarrier_basic.rs | 2 +- .../tests/mbarrier_extended.rs | 2 +- .../generated_intrinsic_targets/tests/mod.rs | 2 +- .../tests/movmatrix.rs | 2 +- .../tests/packed_alu.rs | 2 +- .../tests/packed_atomic.rs | 2 +- .../tests/packed_conversion.rs | 308 ++- .../generated_intrinsic_targets/tests/prmt.rs | 2 +- .../tests/redux.rs | 2 +- .../tests/register_mma_float.rs | 2 +- .../tests/register_mma_s32.rs | 2 +- .../tests/scalar_arithmetic.rs | 2 +- .../tests/scalar_conversion.rs | 2 +- .../tests/scalar_math.rs | 2 +- .../tests/sparse_mma.rs | 2 +- .../generated_intrinsic_targets/tests/sreg.rs | 2 +- .../tests/stmatrix.rs | 2 +- .../generated_intrinsic_targets/tests/sync.rs | 2 +- .../tests/tcgen05_cp.rs | 2 +- .../tests/tcgen05_ld.rs | 2 +- .../tests/tcgen05_mma.rs | 2 +- .../tests/tcgen05_mma_ws.rs | 2 +- .../tests/tcgen05_other.rs | 2 +- .../tests/tcgen05_st.rs | 2 +- .../generated_intrinsic_targets/tests/tma.rs | 2 +- .../generated_intrinsic_targets/tests/vote.rs | 2 +- .../tests/warp_barrier.rs | 2 +- .../tests/warp_match.rs | 2 +- .../tests/warp_shuffle.rs | 2 +- .../tests/wgmma_control.rs | 2 +- .../src/generated_intrinsic_targets/tma.rs | 2 +- .../src/generated_intrinsic_targets/vote.rs | 2 +- .../warp_barrier.rs | 2 +- .../generated_intrinsic_targets/warp_match.rs | 2 +- .../warp_shuffle.rs | 2 +- .../wgmma_control.rs | 2 +- .../src/ops/generated/active_mask.rs | 2 +- .../dialect-nvvm/src/ops/generated/clc.rs | 2 +- .../src/ops/generated/cluster_barrier.rs | 2 +- .../src/ops/generated/cluster_memory.rs | 2 +- .../src/ops/generated/cp_async.rs | 2 +- .../src/ops/generated/debug_control.rs | 2 +- .../dialect-nvvm/src/ops/generated/dotprod.rs | 2 +- .../dialect-nvvm/src/ops/generated/elect.rs | 2 +- .../src/ops/generated/execution_control.rs | 2 +- .../src/ops/generated/extended_minmax.rs | 2 +- .../src/ops/generated/integer_minmax.rs | 2 +- .../src/ops/generated/ldmatrix.rs | 2 +- .../src/ops/generated/mbarrier_basic.rs | 2 +- .../src/ops/generated/mbarrier_extended.rs | 2 +- .../dialect-nvvm/src/ops/generated/mod.rs | 2 +- .../src/ops/generated/movmatrix.rs | 2 +- .../src/ops/generated/packed_alu.rs | 2 +- .../src/ops/generated/packed_atomic.rs | 2 +- .../src/ops/generated/packed_conversion.rs | 202 +- .../dialect-nvvm/src/ops/generated/prmt.rs | 2 +- .../dialect-nvvm/src/ops/generated/redux.rs | 2 +- .../src/ops/generated/register_mma.rs | 2 +- .../src/ops/generated/scalar_arithmetic.rs | 2 +- .../src/ops/generated/scalar_conversion.rs | 2 +- .../src/ops/generated/scalar_math.rs | 2 +- .../src/ops/generated/sparse_mma.rs | 2 +- .../dialect-nvvm/src/ops/generated/sreg.rs | 2 +- .../src/ops/generated/stmatrix.rs | 2 +- .../dialect-nvvm/src/ops/generated/sync.rs | 2 +- .../dialect-nvvm/src/ops/generated/tcgen05.rs | 2 +- .../dialect-nvvm/src/ops/generated/tma.rs | 2 +- .../dialect-nvvm/src/ops/generated/vote.rs | 2 +- .../src/ops/generated/warp_barrier.rs | 2 +- .../src/ops/generated/warp_match.rs | 2 +- .../src/ops/generated/warp_shuffle.rs | 2 +- .../src/ops/generated/wgmma_control.rs | 2 +- .../intrinsics/generated/active_mask.rs | 2 +- .../terminator/intrinsics/generated/clc.rs | 2 +- .../intrinsics/generated/cluster_barrier.rs | 2 +- .../intrinsics/generated/cluster_memory.rs | 2 +- .../intrinsics/generated/cp_async.rs | 2 +- .../intrinsics/generated/debug_control.rs | 2 +- .../intrinsics/generated/dotprod.rs | 2 +- .../terminator/intrinsics/generated/elect.rs | 2 +- .../intrinsics/generated/execution_control.rs | 2 +- .../intrinsics/generated/extended_minmax.rs | 2 +- .../intrinsics/generated/helpers.rs | 2 +- .../intrinsics/generated/integer_minmax.rs | 2 +- .../intrinsics/generated/ldmatrix.rs | 2 +- .../intrinsics/generated/mbarrier_basic.rs | 2 +- .../intrinsics/generated/mbarrier_extended.rs | 2 +- .../terminator/intrinsics/generated/mod.rs | 22 +- .../intrinsics/generated/movmatrix.rs | 2 +- .../intrinsics/generated/packed_alu.rs | 2 +- .../intrinsics/generated/packed_atomic.rs | 2 +- .../intrinsics/generated/packed_conversion.rs | 164 +- .../terminator/intrinsics/generated/prmt.rs | 2 +- .../terminator/intrinsics/generated/redux.rs | 2 +- .../intrinsics/generated/register_mma.rs | 2 +- .../intrinsics/generated/scalar_arithmetic.rs | 2 +- .../intrinsics/generated/scalar_conversion.rs | 2 +- .../intrinsics/generated/scalar_math.rs | 2 +- .../intrinsics/generated/sparse_mma.rs | 2 +- .../terminator/intrinsics/generated/sreg.rs | 2 +- .../intrinsics/generated/stmatrix.rs | 2 +- .../terminator/intrinsics/generated/sync.rs | 2 +- .../intrinsics/generated/tcgen05_cp.rs | 2 +- .../intrinsics/generated/tcgen05_ld.rs | 2 +- .../intrinsics/generated/tcgen05_mma.rs | 2 +- .../intrinsics/generated/tcgen05_other.rs | 2 +- .../intrinsics/generated/tcgen05_st.rs | 2 +- .../intrinsics/generated/tests/active_mask.rs | 2 +- .../intrinsics/generated/tests/clc.rs | 2 +- .../generated/tests/cluster_barrier.rs | 2 +- .../generated/tests/cluster_memory.rs | 2 +- .../intrinsics/generated/tests/cp_async.rs | 2 +- .../generated/tests/debug_control.rs | 2 +- .../intrinsics/generated/tests/dotprod.rs | 2 +- .../intrinsics/generated/tests/elect.rs | 2 +- .../generated/tests/execution_control.rs | 2 +- .../generated/tests/extended_minmax.rs | 2 +- .../generated/tests/integer_minmax.rs | 2 +- .../intrinsics/generated/tests/ldmatrix.rs | 2 +- .../generated/tests/mbarrier_basic.rs | 2 +- .../generated/tests/mbarrier_extended.rs | 2 +- .../intrinsics/generated/tests/mod.rs | 2 +- .../intrinsics/generated/tests/movmatrix.rs | 2 +- .../intrinsics/generated/tests/packed_alu.rs | 2 +- .../generated/tests/packed_atomic.rs | 2 +- .../generated/tests/packed_conversion.rs | 146 +- .../intrinsics/generated/tests/prmt.rs | 2 +- .../intrinsics/generated/tests/redux.rs | 2 +- .../generated/tests/register_mma.rs | 2 +- .../generated/tests/scalar_arithmetic.rs | 2 +- .../generated/tests/scalar_conversion.rs | 2 +- .../intrinsics/generated/tests/scalar_math.rs | 2 +- .../intrinsics/generated/tests/sparse_mma.rs | 2 +- .../intrinsics/generated/tests/sreg.rs | 2 +- .../intrinsics/generated/tests/stmatrix.rs | 2 +- .../intrinsics/generated/tests/sync.rs | 2 +- .../intrinsics/generated/tests/tcgen05.rs | 2 +- .../intrinsics/generated/tests/tma.rs | 2 +- .../intrinsics/generated/tests/vote.rs | 2 +- .../generated/tests/warp_barrier.rs | 2 +- .../intrinsics/generated/tests/warp_match.rs | 2 +- .../generated/tests/warp_shuffle.rs | 2 +- .../generated/tests/wgmma_control.rs | 2 +- .../terminator/intrinsics/generated/tma.rs | 2 +- .../terminator/intrinsics/generated/vote.rs | 2 +- .../intrinsics/generated/warp_barrier.rs | 2 +- .../intrinsics/generated/warp_match.rs | 2 +- .../intrinsics/generated/warp_shuffle.rs | 2 +- .../intrinsics/generated/wgmma_control.rs | 2 +- .../generated_intrinsics/active_mask.rs | 2 +- .../src/convert/generated_intrinsics/clc.rs | 2 +- .../generated_intrinsics/cluster_barrier.rs | 2 +- .../generated_intrinsics/cluster_memory.rs | 2 +- .../convert/generated_intrinsics/cp_async.rs | 2 +- .../generated_intrinsics/debug_control.rs | 2 +- .../convert/generated_intrinsics/dotprod.rs | 2 +- .../src/convert/generated_intrinsics/elect.rs | 2 +- .../generated_intrinsics/execution_control.rs | 2 +- .../generated_intrinsics/extended_minmax.rs | 2 +- .../generated_intrinsics/integer_minmax.rs | 2 +- .../convert/generated_intrinsics/ldmatrix.rs | 2 +- .../generated_intrinsics/mbarrier_basic.rs | 2 +- .../generated_intrinsics/mbarrier_extended.rs | 2 +- .../src/convert/generated_intrinsics/mod.rs | 2 +- .../convert/generated_intrinsics/movmatrix.rs | 2 +- .../generated_intrinsics/packed_alu.rs | 2 +- .../generated_intrinsics/packed_atomic.rs | 2 +- .../generated_intrinsics/packed_conversion.rs | 84 +- .../src/convert/generated_intrinsics/prmt.rs | 2 +- .../src/convert/generated_intrinsics/redux.rs | 2 +- .../generated_intrinsics/register_mma.rs | 2 +- .../generated_intrinsics/scalar_arithmetic.rs | 2 +- .../generated_intrinsics/scalar_conversion.rs | 2 +- .../generated_intrinsics/scalar_math.rs | 2 +- .../generated_intrinsics/sparse_mma.rs | 2 +- .../src/convert/generated_intrinsics/sreg.rs | 2 +- .../convert/generated_intrinsics/stmatrix.rs | 2 +- .../src/convert/generated_intrinsics/sync.rs | 2 +- .../convert/generated_intrinsics/tcgen05.rs | 2 +- .../src/convert/generated_intrinsics/tma.rs | 2 +- .../src/convert/generated_intrinsics/vote.rs | 2 +- .../generated_intrinsics/warp_barrier.rs | 2 +- .../generated_intrinsics/warp_match.rs | 2 +- .../generated_intrinsics/warp_shuffle.rs | 2 +- .../generated_intrinsics/wgmma_control.rs | 2 +- .../src/generated_intrinsics.rs | 18 +- .../cuda-oxide-book/compiler/mlir-dialects.md | 10 +- cuda-oxide/intrinsics/catalog.json | 1994 +++++++++++++++-- ...-13.3.33-packed-conversion-fp8-bf16x2.json | 200 -- ...-13.3.73-packed-conversion-fp8-bf16x2.json | 320 +++ ...16696adc-packed-conversion-fp8-bf16x2.json | 196 -- ...aab11ee0-packed-conversion-fp8-bf16x2.json | 432 ++++ cuda-oxide/intrinsics/generated-reference.md | 74 +- .../overlay/packed_conversion_fp8_bf16x2.toml | 10 +- cuda-oxide/intrinsics/probes/abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/active_mask.ll | 2 +- cuda-oxide/intrinsics/probes/add_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f64.ll | 2 +- .../intrinsics/probes/add_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f64.ll | 2 +- .../intrinsics/probes/add_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f64.ll | 2 +- .../intrinsics/probes/add_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f64.ll | 2 +- .../intrinsics/probes/add_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/add_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/add_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/all_sync.ll | 2 +- cuda-oxide/intrinsics/probes/any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/ballot_sync.ll | 2 +- .../probes/barrier_cluster_arrive.ll | 2 +- .../probes/barrier_cluster_arrive_aligned.ll | 2 +- .../probes/barrier_cluster_arrive_relaxed.ll | 2 +- .../barrier_cluster_arrive_relaxed_aligned.ll | 2 +- .../intrinsics/probes/barrier_cluster_wait.ll | 2 +- .../probes/barrier_cluster_wait_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_arrive.ll | 2 +- .../probes/barrier_cta_arrive_aligned.ll | 2 +- .../intrinsics/probes/barrier_cta_sync.ll | 2 +- .../probes/barrier_cta_sync_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/breakpoint.ll | 2 +- .../probes/clc_query_get_first_ctaid_x.ll | 2 +- .../probes/clc_query_get_first_ctaid_y.ll | 2 +- .../probes/clc_query_get_first_ctaid_z.ll | 2 +- .../probes/clc_query_is_canceled.ll | 2 +- .../intrinsics/probes/clc_try_cancel.ll | 2 +- .../probes/clc_try_cancel_multicast.ll | 2 +- .../intrinsics/probes/clock.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock.llvm_nvptx.ll | 2 +- .../intrinsics/probes/clock64.lib_nvvm.ll | 2 +- .../intrinsics/probes/clock64.llvm_nvptx.ll | 2 +- .../intrinsics/probes/cluster_block_count.ll | 2 +- .../intrinsics/probes/cluster_block_idx_x.ll | 2 +- .../intrinsics/probes/cluster_block_idx_y.ll | 2 +- .../intrinsics/probes/cluster_block_idx_z.ll | 2 +- .../intrinsics/probes/cluster_block_rank.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_z.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_x.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_y.ll | 2 +- .../intrinsics/probes/cluster_grid_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_z.ll | 2 +- .../intrinsics/probes/cos_approx_f32.ll | 2 +- .../intrinsics/probes/cos_approx_ftz_f32.ll | 2 +- .../probes/cp_async_bulk_commit_group.ll | 2 +- .../cp_async_bulk_prefetch_tensor_1d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_1d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_2d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_2d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_3d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_3d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_4d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_4d_l2_cache_hint.ll | 2 +- .../cp_async_bulk_prefetch_tensor_5d_l2.ll | 2 +- ...c_bulk_prefetch_tensor_5d_l2_cache_hint.ll | 2 +- ...sync_bulk_prefetch_tensor_gather4_2d_l2.ll | 2 +- ...refetch_tensor_gather4_2d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_1d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s.ll | 2 +- .../cp_async_bulk_tensor_2d_g2s_multicast.ll | 2 +- ..._async_bulk_tensor_2d_g2s_multicast_cg2.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_3d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_4d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_5d_s2g.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_add_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_add_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_and_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_and_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_dec_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_dec_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_inc_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_inc_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_max_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_max_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_min_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_min_tile_5d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_3d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_4d.ll | 2 +- ...p_async_bulk_tensor_reduce_or_im2col_5d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_1d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_2d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_3d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_4d.ll | 2 +- .../cp_async_bulk_tensor_reduce_or_tile_5d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_3d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_4d.ll | 2 +- ..._async_bulk_tensor_reduce_xor_im2col_5d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_1d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_2d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_3d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_4d.ll | 2 +- ...cp_async_bulk_tensor_reduce_xor_tile_5d.ll | 2 +- .../probes/cp_async_bulk_wait_group.ll | 2 +- .../probes/cp_async_bulk_wait_group_read.ll | 2 +- .../intrinsics/probes/cp_async_ca_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_4.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_8.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_16.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_4.ll | 2 +- .../intrinsics/probes/cp_async_ca_zfill_8.ll | 2 +- .../intrinsics/probes/cp_async_cg_16.ll | 2 +- .../intrinsics/probes/cp_async_cg_zfill_16.ll | 2 +- .../probes/cp_async_commit_group.ll | 2 +- .../probes/cp_async_mbarrier_arrive.ll | 2 +- .../probes/cp_async_mbarrier_arrive_noinc.ll | 2 +- .../cp_async_mbarrier_arrive_noinc_shared.ll | 2 +- .../probes/cp_async_mbarrier_arrive_shared.ll | 2 +- .../intrinsics/probes/cp_async_wait_all.ll | 2 +- .../intrinsics/probes/cp_async_wait_group.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_f32x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll | 2 +- .../intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_bf16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e4m3x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_e5m2x2.ll | 2 +- .../probes/cvt_rn_relu_f16x2_f32.ll | 2 +- .../probes/cvt_rn_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_relu_tf32_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll | 4 +- .../probes/cvt_rn_satfinite_e4m3x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e4m3x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll | 4 +- .../probes/cvt_rn_satfinite_e5m2x2_f16x2.ll | 2 +- .../probes/cvt_rn_satfinite_e5m2x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll | 4 +- .../cvt_rn_satfinite_relu_e4m3x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e4m3x2_f32.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll | 4 +- .../cvt_rn_satfinite_relu_e5m2x2_f16x2.ll | 2 +- .../cvt_rn_satfinite_relu_e5m2x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_tf32_f32.ll | 2 +- .../probes/cvt_rna_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rna_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_bf16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_f16x2_f32.ll | 2 +- .../probes/cvt_rz_relu_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_relu_tf32_f32.ll | 2 +- .../probes/cvt_rz_satfinite_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f64.ll | 2 +- .../intrinsics/probes/div_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f64.ll | 2 +- .../intrinsics/probes/div_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f64.ll | 2 +- .../intrinsics/probes/div_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f64.ll | 2 +- .../intrinsics/probes/div_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_u32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_u32.ll | 2 +- .../intrinsics/probes/dsmem_read_u32.ll | 2 +- .../probes/dynamic_smem_size.lib_nvvm.ll | 2 +- .../probes/dynamic_smem_size.llvm_nvptx.ll | 2 +- .../intrinsics/probes/elect_sync.lib_nvvm.ll | 2 +- .../probes/elect_sync.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg1.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg1.llvm_nvptx.ll | 2 +- .../intrinsics/probes/envreg2.lib_nvvm.ll | 2 +- .../intrinsics/probes/envreg2.llvm_nvptx.ll | 2 +- .../intrinsics/probes/ex2_approx_f16.ll | 2 +- .../intrinsics/probes/ex2_approx_f32.ll | 2 +- .../intrinsics/probes/ex2_approx_ftz_f32.ll | 2 +- .../fence_mbarrier_init_release_cluster.ll | 2 +- ..._generic_acquire_shared_cluster_cluster.ll | 2 +- ...sync_generic_release_shared_cta_cluster.ll | 2 +- .../probes/fence_proxy_async_shared_cta.ll | 2 +- ...proxy_tensormap_generic_acquire_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_cta.ll | 2 +- ...nce_proxy_tensormap_generic_acquire_gpu.ll | 2 +- ..._proxy_tensormap_generic_acquire_system.ll | 2 +- ...proxy_tensormap_generic_release_cluster.ll | 2 +- ...nce_proxy_tensormap_generic_release_cta.ll | 2 +- ...nce_proxy_tensormap_generic_release_gpu.ll | 2 +- ..._proxy_tensormap_generic_release_system.ll | 2 +- cuda-oxide/intrinsics/probes/fma_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll | 2 +- .../intrinsics/probes/fma_ftz_relu_f16x2.ll | 2 +- .../intrinsics/probes/fma_ftz_sat_f16x2.ll | 2 +- .../intrinsics/probes/fma_relu_bf16x2.ll | 2 +- .../intrinsics/probes/fma_relu_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f64.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rm_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f64.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rn_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f64.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rp_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f64.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/fma_rz_ftz_sat_f32.ll | 2 +- .../intrinsics/probes/fma_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll | 2 +- .../intrinsics/probes/globaltimer.lib_nvvm.ll | 2 +- .../probes/globaltimer.llvm_nvptx.ll | 2 +- .../grid_dependency_launch_dependents.ll | 2 +- .../intrinsics/probes/grid_dependency_wait.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_z.ll | 2 +- .../intrinsics/probes/gridid.lib_nvvm.ll | 2 +- .../intrinsics/probes/gridid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/lane_id.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_eq.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_ge.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_gt.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_le.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_lt.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll | 2 +- ...dmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll | 2 +- .../ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x1_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x1_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x2_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n8_x4_b16.ll | 2 +- .../probes/ldmatrix_m8n8_x4_trans_b16.ll | 2 +- .../intrinsics/probes/lg2_approx_f32.ll | 2 +- .../intrinsics/probes/lg2_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/map_shared_rank.ll | 2 +- .../intrinsics/probes/match_all_i64_sync.ll | 2 +- .../intrinsics/probes/match_all_sync.ll | 2 +- .../intrinsics/probes/match_any_i64_sync.ll | 2 +- .../intrinsics/probes/match_any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/max_ftz_nan_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16.ll | 2 +- .../probes/max_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/max_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_bf16.ll | 2 +- .../intrinsics/probes/max_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16.ll | 2 +- .../probes/max_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f16.ll | 2 +- .../probes/max_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/max_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/max_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/max_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_u16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_bf16.ll | 2 +- .../probes/max_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f16.ll | 2 +- .../probes/max_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/max_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/mbarrier_arrive.ll | 2 +- .../probes/mbarrier_arrive_cluster.ll | 2 +- .../probes/mbarrier_arrive_expect_tx.ll | 2 +- .../mbarrier_arrive_expect_tx_cluster.ll | 2 +- .../probes/mbarrier_arrive_no_complete.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_init.ll | 2 +- .../intrinsics/probes/mbarrier_inval.ll | 2 +- .../intrinsics/probes/mbarrier_test_wait.ll | 2 +- .../intrinsics/probes/mbarrier_try_wait.ll | 2 +- .../probes/mbarrier_try_wait_parity.ll | 2 +- .../mbarrier_try_wait_parity_cluster.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16.ll | 2 +- .../intrinsics/probes/min_ftz_nan_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_nan_xorsign_abs_f32.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16.ll | 2 +- .../probes/min_ftz_xorsign_abs_f16x2.ll | 2 +- .../probes/min_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_bf16.ll | 2 +- .../intrinsics/probes/min_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16.ll | 2 +- .../probes/min_nan_xorsign_abs_bf16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f16.ll | 2 +- .../probes/min_nan_xorsign_abs_f16x2.ll | 2 +- .../probes/min_nan_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/min_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/min_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_u16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_bf16.ll | 2 +- .../probes/min_xorsign_abs_bf16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f16.ll | 2 +- .../probes/min_xorsign_abs_f16x2.ll | 2 +- .../intrinsics/probes/min_xorsign_abs_f32.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f16_f16.ll | 2 +- .../probes/mma_m16n8k16_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_f32_f16.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_s8.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_s8_u8.ll | 2 +- .../mma_m16n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_u8.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_s8.ll | 2 +- .../mma_m16n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k16_s32_u8_satfinite.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_and_popc.ll | 2 +- .../probes/mma_m16n8k256_s32_b1_xor_popc.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_f32_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll | 2 +- .../mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s4.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s4_u4.ll | 2 +- .../mma_m16n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s8.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u4.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_s4.ll | 2 +- .../mma_m16n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m16n8k32_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k4_f32_tf32.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_s4.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m16n8k64_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f16_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_bf16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k8_f32_tf32.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_and_popc.ll | 2 +- .../probes/mma_m8n8k128_s32_b1_xor_popc.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_s8_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_u8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_s8_satfinite.ll | 2 +- .../probes/mma_m8n8k16_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_s4_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_u4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_s4_satfinite.ll | 2 +- .../probes/mma_m8n8k32_s32_u4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k4_f64.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k128_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k32_s32_u8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll | 2 +- .../mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_sp_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8_u8.ll | 2 +- .../mma_sp_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8_s8.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_s4.ll | 2 +- ...red_metadata_m16n8k128_s32_s4_satfinite.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_s4_u4.ll | 2 +- ..._metadata_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k128_s32_u4.ll | 2 +- ...sp_ordered_metadata_m16n8k128_s32_u4_s4.ll | 2 +- ..._metadata_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- ...red_metadata_m16n8k128_s32_u4_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k16_f32_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k16_f32_tf32.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f16_f16.ll | 2 +- ...a_sp_ordered_metadata_m16n8k32_f32_bf16.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k32_f32_f16.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_s8.ll | 2 +- ...ered_metadata_m16n8k32_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k32_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k32_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k32_s32_u8_satfinite.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll | 2 +- ..._m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s4.ll | 2 +- ...ered_metadata_m16n8k64_s32_s4_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s4_u4.ll | 2 +- ...d_metadata_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_s8.ll | 2 +- ...ered_metadata_m16n8k64_s32_s8_satfinite.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_s8_u8.ll | 2 +- ...d_metadata_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u4.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u4_s4.ll | 2 +- ...d_metadata_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u4_satfinite.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k64_s32_u8.ll | 2 +- ..._sp_ordered_metadata_m16n8k64_s32_u8_s8.ll | 2 +- ...d_metadata_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- ...ered_metadata_m16n8k64_s32_u8_satfinite.ll | 2 +- ...ma_sp_ordered_metadata_m16n8k8_f32_tf32.ll | 2 +- .../intrinsics/probes/movmatrix_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/mul_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f64.ll | 2 +- .../intrinsics/probes/mul_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f64.ll | 2 +- .../intrinsics/probes/mul_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f64.ll | 2 +- .../intrinsics/probes/mul_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f64.ll | 2 +- .../intrinsics/probes/mul_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/nanosleep.ll | 2 +- cuda-oxide/intrinsics/probes/neg_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/neg_f16x2.ll | 2 +- .../intrinsics/probes/nsmid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nsmid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/nwarpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/nwarpid.llvm_nvptx.ll | 2 +- .../probes/packed_atomic_add_bf16x2.ll | 2 +- .../probes/packed_atomic_add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/pmevent.ll | 2 +- .../probes/prefetch_tma_descriptor.ll | 2 +- cuda-oxide/intrinsics/probes/prmt.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_b4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecl.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecr.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_f4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc16.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc8.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/rcp_approx_ftz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f64.ll | 2 +- .../intrinsics/probes/rcp_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f64.ll | 2 +- .../intrinsics/probes/rcp_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f64.ll | 2 +- .../intrinsics/probes/rcp_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f64.ll | 2 +- .../intrinsics/probes/rcp_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/redux_sync_add.ll | 2 +- .../intrinsics/probes/redux_sync_and.ll | 2 +- .../probes/redux_sync_max_abs_f32.ll | 2 +- .../probes/redux_sync_max_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_i32.ll | 2 +- .../probes/redux_sync_max_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_max_u32.ll | 2 +- .../probes/redux_sync_min_abs_f32.ll | 2 +- .../probes/redux_sync_min_abs_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_i32.ll | 2 +- .../probes/redux_sync_min_nan_f32.ll | 2 +- .../intrinsics/probes/redux_sync_min_u32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_or.ll | 2 +- .../intrinsics/probes/redux_sync_xor.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f32.ll | 2 +- .../intrinsics/probes/rsqrt_approx_f64.ll | 2 +- .../intrinsics/probes/rsqrt_approx_ftz_f32.ll | 2 +- .../intrinsics/probes/setmaxnreg_dec.ll | 2 +- .../intrinsics/probes/setmaxnreg_inc.ll | 2 +- .../probes/shuffle_down_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_down_sync.ll | 2 +- .../probes/shuffle_down_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_sync.ll | 2 +- .../intrinsics/probes/shuffle_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_sync.ll | 2 +- .../intrinsics/probes/shuffle_up_u64_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_f32_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_sync.ll | 2 +- .../intrinsics/probes/shuffle_xor_u64_sync.ll | 2 +- .../intrinsics/probes/sin_approx_f32.ll | 2 +- .../intrinsics/probes/sin_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll | 2 +- .../intrinsics/probes/smid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/sqrt_approx_f32.ll | 2 +- .../intrinsics/probes/sqrt_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll | 2 +- .../intrinsics/probes/sqrt_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll | 2 +- .../intrinsics/probes/sqrt_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll | 2 +- .../intrinsics/probes/sqrt_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll | 2 +- .../intrinsics/probes/sqrt_rz_ftz_f32.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x2_b16.ll | 2 +- .../probes/stmatrix_m8n8_x2_trans_b16.ll | 2 +- .../intrinsics/probes/stmatrix_m8n8_x4_b16.ll | 2 +- .../probes/stmatrix_m8n8_x4_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/sub_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sync_mask.ll | 2 +- cuda-oxide/intrinsics/probes/sync_threads.ll | 2 +- .../intrinsics/probes/tanh_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_alloc.ll | 2 +- .../intrinsics/probes/tcgen05_alloc_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_commit.ll | 2 +- .../intrinsics/probes/tcgen05_commit_cg2.ll | 2 +- .../probes/tcgen05_commit_multicast.ll | 2 +- .../probes/tcgen05_commit_multicast_cg2.ll | 2 +- .../probes/tcgen05_commit_shared_cluster.ll | 2 +- .../tcgen05_commit_shared_cluster_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b.ll | 2 +- .../probes/tcgen05_cp_128x128b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x128b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x128b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_128x128b_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b4x16_p64.ll | 2 +- .../tcgen05_cp_128x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_128x256b_b6x16_p32.ll | 2 +- .../tcgen05_cp_128x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b4x16_p64.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll | 2 +- .../tcgen05_cp_32x128b_warpx4_b6x16_p32.ll | 2 +- ...tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_4x256b.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_4x256b_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_01_23_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll | 2 +- ...gen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll | 2 +- ...5_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll | 2 +- .../tcgen05_cp_64x128b_warpx2_02_13_cg2.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem.ll | 2 +- .../probes/tcgen05_cp_smem_to_tmem_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc.ll | 2 +- .../intrinsics/probes/tcgen05_dealloc_cg2.ll | 2 +- .../probes/tcgen05_fence_after_thread_sync.ll | 2 +- .../tcgen05_fence_before_thread_sync.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_pure.ll | 2 +- .../probes/tcgen05_ld_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_pack16.ll | 2 +- .../probes/tcgen05_ld_32x32b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_load_wait.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_f16_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_shared.ll | 2 +- .../probes/tcgen05_mma_sp_shared.ll | 2 +- .../probes/tcgen05_mma_sp_tensor.ll | 2 +- .../probes/tcgen05_mma_sp_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_tensor.ll | 2 +- .../probes/tcgen05_mma_tensor_ashift.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_bf16.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m1.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e2m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e3m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e4m3.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_e5m2.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_f16.ll | 2 +- .../probes/tcgen05_mma_ws_shared.ll | 2 +- .../tcgen05_mma_ws_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_shared.ll | 2 +- .../tcgen05_mma_ws_sp_shared_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_sp_tensor.ll | 2 +- .../tcgen05_mma_ws_sp_tensor_zero_col_mask.ll | 2 +- .../probes/tcgen05_mma_ws_tensor.ll | 2 +- .../tcgen05_mma_ws_tensor_zero_col_mask.ll | 2 +- .../intrinsics/probes/tcgen05_mma_ws_tf32.ll | 2 +- .../probes/tcgen05_relinquish_alloc_permit.ll | 2 +- .../tcgen05_relinquish_alloc_permit_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_shift_down.ll | 2 +- .../probes/tcgen05_shift_down_cg2.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x128b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x256b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x128_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x16_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x32_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x64_raw.ll | 2 +- .../tcgen05_st_16x32bx2_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x32bx2_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_raw.ll | 2 +- .../probes/tcgen05_st_16x64b_x8_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x128_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x16_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x1_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x2_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x32_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x4_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x64_unpack16.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_raw.ll | 2 +- .../probes/tcgen05_st_32x32b_x8_unpack16.ll | 2 +- .../intrinsics/probes/tcgen05_store_wait.ll | 2 +- .../probes/tensormap_replace_box_dim.ll | 2 +- .../tensormap_replace_element_stride.ll | 2 +- .../probes/tensormap_replace_element_type.ll | 2 +- .../probes/tensormap_replace_fill_mode.ll | 2 +- .../tensormap_replace_global_address.ll | 2 +- .../probes/tensormap_replace_global_dim.ll | 2 +- .../probes/tensormap_replace_global_stride.ll | 2 +- .../tensormap_replace_interleave_layout.ll | 2 +- .../probes/tensormap_replace_rank.ll | 2 +- .../tensormap_replace_swizzle_atomicity.ll | 2 +- .../probes/tensormap_replace_swizzle_mode.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/threadfence.ll | 2 +- .../intrinsics/probes/threadfence_block.ll | 2 +- .../intrinsics/probes/threadfence_system.ll | 2 +- .../probes/total_smem_size.lib_nvvm.ll | 2 +- .../probes/total_smem_size.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/trap.ll | 2 +- cuda-oxide/intrinsics/probes/uni_sync.ll | 2 +- .../intrinsics/probes/warpid.lib_nvvm.ll | 2 +- .../intrinsics/probes/warpid.llvm_nvptx.ll | 2 +- .../intrinsics/probes/wgmma_commit_group.ll | 2 +- cuda-oxide/intrinsics/probes/wgmma_fence.ll | 2 +- .../intrinsics/probes/wgmma_wait_group.ll | 2 +- 1330 files changed, 5294 insertions(+), 1971 deletions(-) delete mode 100644 cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json create mode 100644 cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-packed-conversion-fp8-bf16x2.json delete mode 100644 cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json create mode 100644 cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json diff --git a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs index 10e6c3e0b2..0128559fea 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::async_copy` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs index 425aedc4ac..80a721f3ce 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::atomic` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs index 57652a5228..0708a9edc8 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs index aa95140971..cefb1c6d5d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/clc.rs b/cuda-oxide/crates/cuda-device/src/generated/clc.rs index 0c9b7ff9aa..2ed6634024 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/clc.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::clc` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs index a4b905b524..8af44556d6 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs index 04d544f30c..989446243b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs index 6be477a270..de20b34d47 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::cluster`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/convert.rs b/cuda-oxide/crates/cuda-device/src/generated/convert.rs index 36a19cb533..d6ba966953 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/convert.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/convert.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::convert`. @@ -81,6 +81,15 @@ pub fn cvt_rn_relu_f16x2_f32(lo: f32, hi: f32) -> u32 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_e4m3x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_e4m3x2_f16x2(packed: u32) -> u16 { @@ -99,6 +108,15 @@ pub fn cvt_rn_satfinite_e4m3x2_f32(lo: f32, hi: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_e5m2x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_e5m2x2_f16x2(packed: u32) -> u16 { @@ -117,6 +135,15 @@ pub fn cvt_rn_satfinite_e5m2x2_f32(lo: f32, hi: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_relu_e4m3x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_relu_e4m3x2_f16x2(packed: u32) -> u16 { @@ -135,6 +162,15 @@ pub fn cvt_rn_satfinite_relu_e4m3x2_f32(lo: f32, hi: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. +#[inline(never)] +pub fn cvt_rn_satfinite_relu_e5m2x2_bf16x2(packed: u32) -> u16 { + let _ = packed; + unreachable!( + "generated CUDA intrinsic `cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. #[inline(never)] pub fn cvt_rn_satfinite_relu_e5m2x2_f16x2(packed: u32) -> u16 { diff --git a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs index 872d91cee2..87e3367bb4 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs index d6828a8aba..abeecdf5c0 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs index ee39d4257d..a62e6e9b68 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs index 37d8248126..63efdfa946 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::dotprod` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16.rs b/cuda-oxide/crates/cuda-device/src/generated/f16.rs index 2ec7953393..649458a29b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs index d985923438..c043673ede 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs index 349e4a4d35..ec41a6b7ff 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f32x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/fence.rs b/cuda-oxide/crates/cuda-device/src/generated/fence.rs index 88ee2432f3..87ada2f7fb 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/fence.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/fence.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::fence` so existing paths stay stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/float.rs b/cuda-oxide/crates/cuda-device/src/generated/float.rs index 4dcf1edb9b..7ca4e794c7 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/float.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::float`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs index 3fb0c6490c..4ec97e0e96 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs index 9b98009d81..dfd627ccb6 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs index e99a610b45..a08a48c30f 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::i16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/int.rs b/cuda-oxide/crates/cuda-device/src/generated/int.rs index 3788f17e4b..b8765678db 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/int.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/int.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::int` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs index c90b321c91..5c3be5db31 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs index f7e8c4e54a..e5c1c4763a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs index 0f1714ad05..495ed52b2b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs index 3b0f283d9b..4ba4a9decd 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs index d7a9131ae3..12676ce0b0 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::prmt` to keep the public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs index 86e1231142..b40f7ec321 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::thread` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs index e643c970b3..d587bc0f8a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs index 037c21f721..6a097b9195 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::shared` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs index 08980edf29..d3d12eeb1d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs index 412e03bf24..982a9f8997 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included lexically inside `cuda_device::thread` so existing diff --git a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs index ece2a3992a..1d1d158e26 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs index 0ad1f01a09..a671cac636 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tma.rs b/cuda-oxide/crates/cuda-device/src/generated/tma.rs index 92226b0c88..bd010dab68 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs index b8f8444bac..68b8084af3 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::warp` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs index 6a734aa7fb..4ae022e8cc 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wgmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs index 4f3e94cddb..bea0d9e44b 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs @@ -82,6 +82,7 @@ pub struct OverlayShardFile { pub packed_conversion_fp8: Option, #[serde(default)] pub packed_conversion_fp8_bf16x2: Option, + #[serde(default)] pub packed_conversion_fp8_f16x2: Option, #[serde(default)] pub scalar_conversion: Option, @@ -508,8 +509,9 @@ pub enum PackedConversionFp8Format { /// Compact admission for packing `bf16x2` into `e4m3x2` or `e5m2x2`. /// -/// Pack only. The imported reverse conversions are block-scaled -/// (`scale_n2_ue8m0`) and are not part of this family. +/// Pack only. The pinned LLVM import has no unscaled unpack record for +/// `bf16x2`; its reverse records are the block-scaled `scale_n2_ue8m0` forms. +/// The PTX ISA 9.2 unpack forms are a later admission. #[derive(Debug, Clone, Deserialize)] #[serde(deny_unknown_fields)] pub struct PackedConversionFp8Bf16x2Admission { diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/evidence.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/evidence.rs index 694c2579fc..30387a6c9a 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/evidence.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/evidence.rs @@ -831,18 +831,6 @@ pub(super) fn validate_packed_conversion_backend_evidence( } match lowering.backend { IntrinsicBackend::LlvmNvptx => { - // `bf16x2` pack records have empty selections, so there is no typed - // NVVM route to assemble. Inline PTX is the admitted lowering. - if policy.packed_conversion.as_ref().is_some_and(|conversion| { - conversion.source_format == crate::model::PackedConversionSourceFormat::Bf16x2 - }) { - ensure!( - lowering.mechanism == BackendLoweringMechanism::InlinePtx, - "{} bf16x2 packed conversion must lower through inline PTX", - policy.id - ); - return Ok(()); - } validate_typed_llvm_evidence(policy, record)?; for stage in [ EvidenceStageKind::DeclarationCanonicalization, diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs index d3e2220b8f..ee75422523 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/families/packed_conversion.rs @@ -17,6 +17,17 @@ use std::collections::BTreeSet; use crate::resolve::guards::*; +// PTX ISA 9.1 lists `cvt.rn.satfinite{.relu}.{e4m3x2,e5m2x2}.bf16x2` for +// sm_100f, sm_110f and sm_120f "or higher in the same family". The ISA's +// `.target` section lets family-specific features run on f-targets and on +// a-targets of the same family, nvcc defines `__CUDA_ARCH_FAMILY_SPECIFIC__` +// for `compute_100a` as well, and ptxas 13.4 and the pinned llc accept every +// member below while rejecting the unsuffixed targets. cuda-oxide-codegen +// matches a family target only on the literal `f` suffix, so the a-targets +// are enumerated here, as the other Blackwell family records do. +pub(in crate::resolve) const PACKED_CONVERSION_BF16X2_TARGETS: &str = + "sm_100a|sm_100f|sm_103a|sm_103f|sm_110a|sm_110f|sm_120a|sm_120f|sm_121a|sm_121f"; + pub(in crate::resolve) struct PackedConversionRecipe { pub(in crate::resolve) id: &'static str, pub(in crate::resolve) abi_id: &'static str, @@ -45,9 +56,13 @@ pub(in crate::resolve) fn packed_conversion_recipe( /// Recipes for packing `bf16x2` into `e4m3x2` or `e5m2x2`. /// -/// These imported records have empty selections, so the lowering is inline PTX. -/// PTX ISA 9.1 introduces them on the `sm_100f`, `sm_110f`, and `sm_120f` -/// families, not on the Ada `sm_89` floor used by `f16x2`. +/// Like the `f16x2` packs, these lower through inline PTX: the MIR side carries +/// both halves in one integer register while the typed intrinsics are declared +/// over `<2 x bfloat>` (see [`packed_conversion_uses_typed_nvvm`]). The +/// recorded evidence assembles the typed and inline routes to the same cubin. +/// PTX ISA 9.1 introduces the instruction on the Blackwell families named by +/// [`PACKED_CONVERSION_BF16X2_TARGETS`], not on the Ada `sm_89` floor used by +/// `f16x2`. pub(in crate::resolve) fn packed_conversion_recipe_fp8_bf16x2( conversion: &crate::model::PackedConversion, ) -> Option { @@ -555,7 +570,7 @@ pub(in crate::resolve) fn packed_conversion_floor( // `bf16x2` to packed FP8 is a PTX ISA 9.1 family-specific conversion. // It is not available on the Ada `sm_89` floor used by `f16x2`. if conversion.source_format == PackedConversionSourceFormat::Bf16x2 { - return ("9.1", None, "sm_100f|sm_110f|sm_120f"); + return ("9.1", None, PACKED_CONVERSION_BF16X2_TARGETS); } // FP8 on either side carries the Ada floor, including when FP8 is the // source and the destination is the older `f16x2`. @@ -584,12 +599,13 @@ pub(in crate::resolve) fn packed_conversion_backend_mechanism( /// Whether the conversion lowers through a typed NVVM intrinsic call. /// -/// Only the scalar-`f32` pair does. The `f16x2` sources and destinations carry -/// their halves in one integer register on the MIR side, while the typed -/// intrinsics are declared over `<2 x half>`; routing them through inline PTX -/// avoids a packed-half MIR type for no change in emitted code. The recorded -/// evidence assembles both routes to a byte-identical cubin, so the typed route -/// can be switched on later without changing the PTX contract. +/// Only the scalar-`f32` pair does. The `f16x2` and `bf16x2` sources and the +/// `f16x2` destinations carry their halves in one integer register on the MIR +/// side, while the typed intrinsics are declared over `<2 x half>` or +/// `<2 x bfloat>`; routing them through inline PTX avoids a packed-half MIR +/// type for no change in emitted code. The recorded evidence assembles both +/// routes to a byte-identical cubin, so the typed route can be switched on +/// later without changing the PTX contract. pub(in crate::resolve) fn packed_conversion_uses_typed_nvvm( conversion: &crate::model::PackedConversion, ) -> bool { diff --git a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs index ac4cfbae1b..a1ad55b83b 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/scalar_packed.rs @@ -172,7 +172,7 @@ fn compact_fp8_f16x2_conversion_axes_require_the_exact_closed_product() { } #[test] -fn compact_fp8_bf16x2_pack_admits_rn_satfinite_forms() { +fn compact_fp8_bf16x2_pack_axes_require_the_closed_rn_satfinite_product() { let records = expand_packed_conversion_fp8_bf16x2_admission(&test_fp8_bf16x2_conversion_admission()) .unwrap(); @@ -207,12 +207,33 @@ fn compact_fp8_bf16x2_pack_admits_rn_satfinite_forms() { && record.expected_ptx.operands.len() == 2 && record.minimum_ptx == "9.1" && record.minimum_sm.is_none() - && record.targets == "sm_100f|sm_110f|sm_120f" + && record.targets + == "sm_100a|sm_100f|sm_103a|sm_103f|sm_110a|sm_110f|sm_120a|sm_120f|sm_121a|sm_121f" && record .source_record .as_deref() .is_some_and(|source| source.starts_with("int_nvvm_bf16x2_to_")) })); + + let mut missing_format = test_fp8_bf16x2_conversion_admission(); + missing_format.fp8_formats.pop(); + assert!(expand_packed_conversion_fp8_bf16x2_admission(&missing_format).is_err()); + + let mut reversed_formats = test_fp8_bf16x2_conversion_admission(); + reversed_formats.fp8_formats.reverse(); + assert!(expand_packed_conversion_fp8_bf16x2_admission(&reversed_formats).is_err()); + + let mut without_relu = test_fp8_bf16x2_conversion_admission(); + without_relu.relu_variants = false; + assert!(expand_packed_conversion_fp8_bf16x2_admission(&without_relu).is_err()); + + let mut wrong_count = test_fp8_bf16x2_conversion_admission(); + wrong_count.product_count = 2; + assert!(expand_packed_conversion_fp8_bf16x2_admission(&wrong_count).is_err()); + + let mut executed = test_fp8_bf16x2_conversion_admission(); + executed.runtime_validation = RuntimeValidation::Executed; + assert!(expand_packed_conversion_fp8_bf16x2_admission(&executed).is_err()); } #[test] diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs index 3eccfe34de..23e0148903 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Raw ABI functions recognized by cuda-oxide. @@ -3002,6 +3002,18 @@ pub fn i0371(_arg0: f32) -> u32 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_e4m3x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_satfinite`; expects PTX `cvt.rn.satfinite.e4m3x2.bf16x2 , ;`. +/// Available on `sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1030(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_e4m3x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn` from `int_nvvm_f16x2_to_e4m3x2_rn`; expects PTX `cvt.rn.satfinite.e4m3x2.f16x2 , ;`. @@ -3026,6 +3038,18 @@ pub fn i0259(_arg0: f32, _arg1: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_e5m2x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_satfinite`; expects PTX `cvt.rn.satfinite.e5m2x2.bf16x2 , ;`. +/// Available on `sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1032(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_e5m2x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn` from `int_nvvm_f16x2_to_e5m2x2_rn`; expects PTX `cvt.rn.satfinite.e5m2x2.f16x2 , ;`. @@ -3050,6 +3074,18 @@ pub fn i0261(_arg0: f32, _arg1: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_relu_e4m3x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite`; expects PTX `cvt.rn.satfinite.relu.e4m3x2.bf16x2 , ;`. +/// Available on `sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1031(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_relu_e4m3x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn.relu` from `int_nvvm_f16x2_to_e4m3x2_rn_relu`; expects PTX `cvt.rn.satfinite.relu.e4m3x2.f16x2 , ;`. @@ -3074,6 +3110,18 @@ pub fn i0260(_arg0: f32, _arg1: f32) -> u16 { ) } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// Catalog ID: `cvt_rn_satfinite_relu_e5m2x2_bf16x2`. Source: LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite`; expects PTX `cvt.rn.satfinite.relu.e5m2x2.bf16x2 , ;`. +/// Available on `sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f` targets from PTX 9.1. +/// Safe because This conversion has no caller obligations. +#[inline(never)] +pub fn i1033(_arg0: u32) -> u16 { + unreachable!( + "generated CUDA intrinsic `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` executed outside device compilation" + ) +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// Catalog ID: `cvt_rn_satfinite_relu_e5m2x2_f16x2`. Source: LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn.relu` from `int_nvvm_f16x2_to_e5m2x2_rn_relu`; expects PTX `cvt.rn.satfinite.relu.e5m2x2.f16x2 , ;`. diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs index fb3435f9b8..230c301192 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. #[doc(hidden)] @@ -143,6 +143,10 @@ pub mod convert { pub use crate::__cuda_oxide_intrinsic_abi_v1::i0827 as cvt_rn_relu_f16x2_e4m3x2; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0828 as cvt_rn_f16x2_e5m2x2; pub use crate::__cuda_oxide_intrinsic_abi_v1::i0829 as cvt_rn_relu_f16x2_e5m2x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1030 as cvt_rn_satfinite_e4m3x2_bf16x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1031 as cvt_rn_satfinite_relu_e4m3x2_bf16x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1032 as cvt_rn_satfinite_e5m2x2_bf16x2; + pub use crate::__cuda_oxide_intrinsic_abi_v1::i1033 as cvt_rn_satfinite_relu_e5m2x2_bf16x2; } /// Generated `debug` intrinsic source API. @@ -2465,6 +2469,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_e4m3x2_bf16x2_reexports_abi_i1030() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e4m3x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1030; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_e4m3x2_f16x2_reexports_abi_i0822() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e4m3x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0822; @@ -2477,6 +2487,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_e5m2x2_bf16x2_reexports_abi_i1032() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e5m2x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1032; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_e5m2x2_f16x2_reexports_abi_i0824() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_e5m2x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0824; @@ -2489,6 +2505,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_relu_e4m3x2_bf16x2_reexports_abi_i1031() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1031; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_relu_e4m3x2_f16x2_reexports_abi_i0823() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0823; @@ -2501,6 +2523,12 @@ mod tests { assert_eq!(public as usize, canonical as usize); } #[test] + fn public_cvt_rn_satfinite_relu_e5m2x2_bf16x2_reexports_abi_i1033() { + let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2; + let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i1033; + assert_eq!(public as usize, canonical as usize); + } + #[test] fn public_cvt_rn_satfinite_relu_e5m2x2_f16x2_reexports_abi_i0825() { let public: fn(u32) -> u16 = super::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2; let canonical: fn(u32) -> u16 = super::__cuda_oxide_intrinsic_abi_v1::i0825; diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs index 6dfd69169a..0239d0f93d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs index aadf572896..14017a9360 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs index fe36a8aff8..651e146c2b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs index 1e1509cef6..5082d81d49 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs index ead108a430..78c0cefb35 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs index 8cf09104c6..1879b0b81e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs index 7020044fa2..5120baef23 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs index 72c7864a4f..e7ccbc0653 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs index 1953d9b216..640730af0c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs index fcb91feaf8..6f575c2c3b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs index 89d677a028..209a3b3c8b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs index c6c7586c8c..17fb644cd0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs index 3e1feecc57..c65cbfffbc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs index 5b15bf968e..f81417f854 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs index fb9ad47daa..3c3e0f45fc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated target requirements and separately imported LLVM/selection facts. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs index 9530f36fbc..e661174324 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs index 9dca2e8bfa..f247bfd93c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs index 76f6352d37..b800661cd3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs index 41849141b1..d3b266cd9a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_conversion` intrinsics. @@ -718,4 +718,268 @@ pub(super) const TARGETS: &[GeneratedIntrinsicTarget] = &[ result_range: None, }), }, + GeneratedIntrinsicTarget { + marker: "v1:i1030", + id: "cvt_rn_satfinite_e4m3x2_bf16x2", + abi_id: "i1030", + dialect_op: "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1031", + id: "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + abi_id: "i1031", + dialect_op: "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1032", + id: "cvt_rn_satfinite_e5m2x2_bf16x2", + abi_id: "i1032", + dialect_op: "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, + GeneratedIntrinsicTarget { + marker: "v1:i1033", + id: "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + abi_id: "i1033", + dialect_op: "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2", + variant: GeneratedIntrinsicVariant::Scalar, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + backend_requirements: &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ], + selections: &[], + llvm: Some(GeneratedLlvmFacts { + properties: &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"], + result_no_undef: false, + result_range: None, + }), + }, ]; diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs index 26829231f7..9b4536096b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs index 643b0122d8..e872cdfe36 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs index d21b5037bf..22d53bb67d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs index f714ace2f2..505c35ee94 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs index 26c2b9101e..88bd43dd92 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs index ca977feaf9..306ff4506f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs index 60066e6a95..bf4b466b9b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs index ced2ff35e1..9ed4bee138 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs index 6987c58ffe..df6be66828 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs index e6589fd47f..287b63910d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs index e6dfa4e62d..2b35f7db35 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs index a71b4e90ce..ea7fd7a808 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs index 2c5b7977ea..8c0a753757 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs index 7519ac479c..ca52c76201 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs index 8101ef71f2..255e84d5d6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs index 942b0d8d20..414e9fdb5c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs index b4388a4208..b27eca3b68 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs index b05ce7cbf9..e7d193b812 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs index d13cbf86fb..f7bcd434cf 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs index a35c18829c..0ed058d9ab 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs index 4ea27e9777..a5e77d62f3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs index be7f8fbead..ff9d6085e7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs index e56d77b272..c4e30eebe6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs index c681aa4b93..57d66496e0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs index 458a3d4b8f..e209e0efcc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs index f1503abbd0..e1d2893771 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs index 9f808f3507..f1c70ffed8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs index e6d33d9513..44bee24026 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs index 678bbd7350..c16d2378da 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs index 971f157365..ab7cc9ad4c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs index eacc5c5122..04b43a40c0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests, grouped per family shard. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs index f0e0a4f733..379e9355a0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs index 0945d8c1f6..241ccff625 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs index 7ed4a44ef6..f7e234ac84 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs index 6f86271313..3101685bec 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_conversion` intrinsics. @@ -802,3 +802,309 @@ fn cvt_rn_relu_f16x2_e5m2x2_target_matches_the_catalog() { &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] ); } + +#[test] +fn cvt_rn_satfinite_e4m3x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1030").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_e4m3x2_bf16x2"); + assert_eq!(target.abi_id, "i1030"); + assert_eq!(target.dialect_op, "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2"); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} + +#[test] +fn cvt_rn_satfinite_relu_e4m3x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1031").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_relu_e4m3x2_bf16x2"); + assert_eq!(target.abi_id, "i1031"); + assert_eq!( + target.dialect_op, + "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} + +#[test] +fn cvt_rn_satfinite_e5m2x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1032").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_e5m2x2_bf16x2"); + assert_eq!(target.abi_id, "i1032"); + assert_eq!(target.dialect_op, "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2"); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} + +#[test] +fn cvt_rn_satfinite_relu_e5m2x2_bf16x2_target_matches_the_catalog() { + let target = generated_intrinsic_target_by_marker("v1:i1033").unwrap(); + assert_eq!(target.id, "cvt_rn_satfinite_relu_e5m2x2_bf16x2"); + assert_eq!(target.abi_id, "i1033"); + assert_eq!( + target.dialect_op, + "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2" + ); + assert_eq!(target.variant, GeneratedIntrinsicVariant::Scalar); + assert_eq!(target.requirement.minimum_ptx.encoded(), 91); + assert_eq!( + target.requirement.hardware, + const { + GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]) + } + ); + assert_eq!( + target.backend_requirements, + const { + &[ + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LlvmNvptx, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + GeneratedBackendRequirement { + backend: GeneratedIntrinsicBackend::LibNvvm, + requirement: GeneratedTargetRequirement { + minimum_ptx: GeneratedPtxVersion::from_encoded(91), + hardware: GeneratedHardwareTarget::AnyOf(&[ + GeneratedHardwareAlternative::ExactArchitecture(100), + GeneratedHardwareAlternative::FamilyTarget(100), + GeneratedHardwareAlternative::ExactArchitecture(103), + GeneratedHardwareAlternative::FamilyTarget(103), + GeneratedHardwareAlternative::ExactArchitecture(110), + GeneratedHardwareAlternative::FamilyTarget(110), + GeneratedHardwareAlternative::ExactArchitecture(120), + GeneratedHardwareAlternative::FamilyTarget(120), + GeneratedHardwareAlternative::ExactArchitecture(121), + GeneratedHardwareAlternative::FamilyTarget(121), + ]), + }, + }, + ] + } + ); + assert_eq!(target.selections, &[]); + assert_eq!( + target.llvm.unwrap().properties, + &["IntrNoCreateUndefOrPoison", "IntrNoMem", "IntrSpeculatable"] as &[&str] + ); +} diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs index 491b841ab5..aa2de4858f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs index 8647954b1c..6837fc80e4 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs index 9865a8fa47..6c019aff4b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_float` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs index 478197ec78..c475ea05a9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_s32` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs index 60e4b4a4f0..3327bd58b9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs index d775d2d9c0..a116824677 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs index da31b53b2c..c981eae547 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs index b2477c9d74..62f21ab2c2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs index 8a1758029b..b268a94e97 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs index c136d5c526..7a9f6bd5a4 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs index aa7f15a2bb..346e40c5d1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs index 0659c6ce7e..f19ba6de56 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs index 67318df319..f808964666 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs index 322035b874..2340ea3d4f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs index 1aecf409b0..d25e0ec6cd 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs index 5a2af67e32..0b48060bd9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs index 0d5606e517..b1cfd22959 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs index 764bc401f0..38e7d1eed6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs index 04c7f5e3a7..156718acbc 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs index ce2d1695bb..f148c48874 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs index f20f4c4491..605d30ab64 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs index 36db24882a..b923ae9c9b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs index 447579f72e..df33b68cd6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs index 129f12b861..f4651cbde3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs index 033064166b..69b2f96e1d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs index d45478d0ce..59b6c2ae6d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs index 7e1c4672ee..e041b342b2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs index f44e7d1502..6c3bddfa4d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs index 90d30ad393..bbd66f64b3 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs index 2c3f93f347..90f7916a8f 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for the generated active warp mask. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs index 616d65ca24..5475615b82 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Cluster Launch Control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs index 1e84152b60..ed6cc1381a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated operations for the closed cluster-barrier family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs index f475f9c8a1..20aeaebdfe 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for cluster address mapping and remote shared reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs index 373648c5be..5eee58ea50 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for classic `cp.async` instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs index 812c00a924..78e38fcd7c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated PTX debug controls. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs index b4eed5cbbd..fe732aae38 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed integer dot products. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs index 0975f2887d..40f02ec97a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated warp leader-election operation. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs index 7ef65fad25..00d1514966 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated counted-barrier, grid-dependency, and register-control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs index e8d6750774..b7411e2d32 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for extended floating-point min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs index 6bc73ad2ec..4174cddf08 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated extended integer min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs index 5643285302..f6dcda6e8a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation and compatibility carriers for generated `ldmatrix`. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs index 8d2447c25d..bae82bb165 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the basic shared-memory mbarrier lifecycle. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs index 37a6d2de5f..c44e9fd3fe 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for extended mbarrier and async-proxy instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs index 6346816724..c36b45f001 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. mod active_mask; diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs index 82ebc991f6..36800bcc18 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated in-register matrix transpose. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs index db1dbf890f..05428f0b16 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed floating-point arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs index 96379c594d..479315ec2a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated packed-atomic family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs index fe611f1857..9633dcfd80 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated packed conversion. @@ -431,6 +431,55 @@ impl Verify for CvtRnReluF16x2F32Op { } } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteE4m3x2Bf16x2Op; + +impl CvtRnSatfiniteE4m3x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteE4m3x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -530,6 +579,55 @@ impl Verify for CvtRnSatfiniteE4m3x2F32Op { } } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteE5m2x2Bf16x2Op; + +impl CvtRnSatfiniteE5m2x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteE5m2x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -629,6 +727,55 @@ impl Verify for CvtRnSatfiniteE5m2x2F32Op { } } +/// Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteReluE4m3x2Bf16x2Op; + +impl CvtRnSatfiniteReluE4m3x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteReluE4m3x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -728,6 +875,55 @@ impl Verify for CvtRnSatfiniteReluE4m3x2F32Op { } } +/// Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. +/// +/// The single input carries both packed lanes, and lane order is preserved. +#[pliron_op( + name = "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2", + format, + interfaces = [NOpdsInterface<1>, NResultsInterface<1>], +)] +pub struct CvtRnSatfiniteReluE5m2x2Bf16x2Op; + +impl CvtRnSatfiniteReluE5m2x2Bf16x2Op { + pub fn new(op: Ptr) -> Self { + Self { op } + } + + pub fn build(ctx: &mut Context, packed: Value) -> Ptr { + let result_ty = IntegerType::get(ctx, 16, Signedness::Unsigned); + Operation::new( + ctx, + Self::get_concrete_op_info(), + vec![result_ty.into()], + vec![packed], + vec![], + 0, + ) + } +} + +impl Verify for CvtRnSatfiniteReluE5m2x2Bf16x2Op { + fn verify(&self, ctx: &Context) -> Result<(), Error> { + let op = self.get_operation().deref(ctx); + if op.get_num_operands() != 1 || op.get_num_results() != 1 { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2 requires one operand and one result" + ); + } + if !is_integer_width(ctx, op.get_operand(0).get_type(ctx), 32) + || !is_integer_width(ctx, op.get_result(0).get_type(ctx), 16) + { + return verify_err!( + op.loc(), + "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2 requires one 32-bit integer operand and one 16-bit integer result" + ); + } + Ok(()) + } +} + /// Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order. /// /// The single input carries both packed lanes, and lane order is preserved. @@ -936,12 +1132,16 @@ pub(super) fn register(ctx: &mut Context) { CvtRnReluF16x2E4m3x2Op::register(ctx); CvtRnReluF16x2E5m2x2Op::register(ctx); CvtRnReluF16x2F32Op::register(ctx); + CvtRnSatfiniteE4m3x2Bf16x2Op::register(ctx); CvtRnSatfiniteE4m3x2F16x2Op::register(ctx); CvtRnSatfiniteE4m3x2F32Op::register(ctx); + CvtRnSatfiniteE5m2x2Bf16x2Op::register(ctx); CvtRnSatfiniteE5m2x2F16x2Op::register(ctx); CvtRnSatfiniteE5m2x2F32Op::register(ctx); + CvtRnSatfiniteReluE4m3x2Bf16x2Op::register(ctx); CvtRnSatfiniteReluE4m3x2F16x2Op::register(ctx); CvtRnSatfiniteReluE4m3x2F32Op::register(ctx); + CvtRnSatfiniteReluE5m2x2Bf16x2Op::register(ctx); CvtRnSatfiniteReluE5m2x2F16x2Op::register(ctx); CvtRnSatfiniteReluE5m2x2F32Op::register(ctx); CvtRzBf16x2F32Op::register(ctx); diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs index 84759078dc..e193f30c27 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated `prmt` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs index 9692546253..889fcfee97 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for the closed generated `redux.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs index 8fef6c41f9..490bdc8225 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated register-only `mma.sync` variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs index e30a341c2f..de2948ccae 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs index f13122c5cd..d8564f2a78 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated `f32` to TF32 conversions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs index ff68d7ff74..781cf1fb11 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar math. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs index 0ab9e5d669..7b39943b66 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated sparse MMA variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs index 19a5061577..05bd65db3a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for generated special-register reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs index ebb671fe9a..784b724660 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the four generated `stmatrix` stores. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs index 90a3220586..997075c845 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated CTA synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs index f59d9a2360..b380cab8a0 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Core Generation 5 operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs index ea10d67f42..94626d7d14 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Memory Accelerator operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs index 082a1c4a04..1539afbe91 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `vote.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs index 28109e6e97..afb62c4aa6 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated warp synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs index b0e5ae390f..8dae7791b7 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `match.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs index 8066f000b3..51cb6b7994 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `shfl.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs index 079d2f4753..d937a15ccb 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated WGMMA controls. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs index dbf3e84396..3f6aeee7bb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs index ade0190400..49c78c077b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs index 54b62778ae..bbf3a8c0f5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs index e3996216fd..351178c7d1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs index b3af25763b..176d7f4b64 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs index a04d390784..0df994d857 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs index 9fc04194b7..4406504628 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs index 3b6afc8dde..3a8d723925 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs index 265c0112f4..00547b63e7 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs index 4ab825e358..ceb084a673 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs index dca4ce9c0b..d9cf5bf71c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Shared glue for the generated intrinsic dispatch shards. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs index 40595ac88b..01e235616d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs index 3178c8f319..ac8828b5b9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs index d6a9cdc83f..f8cbbade8d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs index b74612cfa4..979e76f5db 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs index 531e5c92af..6a09c4c3fb 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility path dispatch for CUDA intrinsics. @@ -517,18 +517,26 @@ pub fn is_generated_intrinsic_path(name: &str) -> bool { | "cuda_device::convert::cvt_rn_relu_satfinite_tf32_f32" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" | "cuda_device::convert::cvt_rn_relu_tf32_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" @@ -2403,12 +2411,16 @@ pub fn is_raw_generated_intrinsic_path(name: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0083" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0373" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0372" @@ -3635,18 +3647,26 @@ pub fn generated_intrinsic_marker(name: &str) -> Option<&'static str> { | "cuda_device::convert::cvt_rn_relu_satfinite_tf32_f32" => Some("v1:i0373"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" | "cuda_device::convert::cvt_rn_relu_tf32_f32" => Some("v1:i0371"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" => Some("v1:i1030"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" => Some("v1:i0822"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" => Some("v1:i0259"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" => Some("v1:i1032"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" => Some("v1:i0824"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" => Some("v1:i0261"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" => Some("v1:i1031"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" => Some("v1:i0823"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" => Some("v1:i0260"), + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" => Some("v1:i1033"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" => Some("v1:i0825"), "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs index 1510f4da7a..f9e8b85921 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs index 2f359a9d70..0be9ab755d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs index f66702fb4f..af0a248b90 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs index 346a6badbc..54eba69310 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_conversion` intrinsics. @@ -15,8 +15,10 @@ use crate::translator::{rvalue, terminator::helpers, values::ValueMap}; use dialect_nvvm::ops::{ CvtF16x2F32Op, CvtF32x2Bf16x2Op, CvtRnF16x2E4m3x2Op, CvtRnF16x2E5m2x2Op, CvtRnReluBf16x2F32Op, CvtRnReluF16x2E4m3x2Op, CvtRnReluF16x2E5m2x2Op, CvtRnReluF16x2F32Op, - CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, CvtRnSatfiniteE5m2x2F16x2Op, - CvtRnSatfiniteE5m2x2F32Op, CvtRnSatfiniteReluE4m3x2F16x2Op, CvtRnSatfiniteReluE4m3x2F32Op, + CvtRnSatfiniteE4m3x2Bf16x2Op, CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, + CvtRnSatfiniteE5m2x2Bf16x2Op, CvtRnSatfiniteE5m2x2F16x2Op, CvtRnSatfiniteE5m2x2F32Op, + CvtRnSatfiniteReluE4m3x2Bf16x2Op, CvtRnSatfiniteReluE4m3x2F16x2Op, + CvtRnSatfiniteReluE4m3x2F32Op, CvtRnSatfiniteReluE5m2x2Bf16x2Op, CvtRnSatfiniteReluE5m2x2F16x2Op, CvtRnSatfiniteReluE5m2x2F32Op, CvtRzBf16x2F32Op, CvtRzF16x2F32Op, }; @@ -389,6 +391,45 @@ pub(super) fn try_dispatch( "cvt_rn_relu_f16x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteE4m3x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1030"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_e4m3x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; @@ -476,6 +517,45 @@ pub(super) fn try_dispatch( "cvt_rn_satfinite_e4m3x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteE5m2x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1032"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_e5m2x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; @@ -563,6 +643,45 @@ pub(super) fn try_dispatch( "cvt_rn_satfinite_e5m2x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteReluE4m3x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1031"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_relu_e4m3x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; @@ -650,6 +769,45 @@ pub(super) fn try_dispatch( "cvt_rn_satfinite_relu_e4m3x2_f32 call without target block", )?)) } + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" => { + require_arity(name, args.len(), 1, &loc)?; + let (arg0, last_op) = rvalue::translate_operand( + ctx, + body, + &args[0], + value_map, + block_ptr, + prev_op, + loc.clone(), + )?; + let intrinsic = CvtRnSatfiniteReluE5m2x2Bf16x2Op::build(ctx, arg0); + intrinsic.deref_mut(ctx).set_loc(loc.clone()); + helpers::set_generated_intrinsic_marker(ctx, intrinsic, "v1:i1033"); + let (prepared_destination, prepared_last_op) = helpers::prepare_destination_write( + ctx, + body, + destination, + value_map, + block_ptr, + last_op, + loc.clone(), + )?; + helpers::insert_op(ctx, intrinsic, block_ptr, prepared_last_op); + let result = intrinsic.deref(ctx).get_result(0); + Ok(Some(helpers::emit_prepared_result_and_goto( + ctx, + prepared_destination, + result, + target, + block_ptr, + intrinsic, + value_map, + block_map, + loc, + "cvt_rn_satfinite_relu_e5m2x2_bf16x2 call without target block", + )?)) + } "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" => { require_arity(name, args.len(), 1, &loc)?; diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs index 69eaf65b5d..bee646d75f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs index 269ecd2f67..49df85ea6a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs index a741546fde..2b1b3a2b3d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs index 8db6935f06..910c5ab14c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs index 0ae8377b8d..4bce17f6dd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs index 62df911084..d26be5029e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs index 86c8b04476..475dc8fcd5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs index 042d22fe9a..0c1737f3fc 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs index 4ec1b4dc49..691aeb3413 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs index a9533ccfd6..141dc6b556 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs index 959c46be61..e1dbe30e4e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs index 2514ff7c0b..662cd78b23 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs index 024eacb481..46fd1a1bb2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs index b48fe1d28e..c6400fff3a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs index ffb9d6def0..5f273cf54a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs index c4ab0cc802..f39ec2e19d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs index 84bf97d723..0b531d4792 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs index 9f0bcb6654..bb0a33e75f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs index b0a5c03ee5..1b606f30c8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs index 75edecc7e8..cfff78569c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs index c0b0a46333..079705b6bf 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs index ebf3d5ae86..e6801830e9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs index 4f2a4b0dba..5aa654aef2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs index bfe29b8e9a..91c42098b8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs index fc08e24bd9..cb190aa51e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs index 4e168a9be0..c7393d698b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs index bcd1ae8633..47a1f09a81 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs index 4c4e7a84fc..c6804c4e81 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs index 2cfdad1ac9..58d0ca9486 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs index 773f7f2f0c..4c11633be2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests, grouped per intrinsic family. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs index dada114a26..67c1cbd993 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs index 41d58a16a0..5104d5f877 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs index c0c0d792ba..ebe0a721aa 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs index 7c515139fa..784c00add8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_conversion` intrinsics. @@ -300,6 +300,42 @@ fn cvt_rn_relu_f16x2_f32_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cvt_rn_satfinite_e4m3x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030"), + Some("v1:i1030") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2"), + Some("v1:i1030") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1030" + )); +} +#[test] fn cvt_rn_satfinite_e4m3x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" @@ -372,6 +408,42 @@ fn cvt_rn_satfinite_e4m3x2_f32_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cvt_rn_satfinite_e5m2x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032"), + Some("v1:i1032") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2"), + Some("v1:i1032") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1032" + )); +} +#[test] fn cvt_rn_satfinite_e5m2x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" @@ -444,6 +516,42 @@ fn cvt_rn_satfinite_e5m2x2_f32_uses_only_canonical_or_compatibility_defpaths() { )); } #[test] +fn cvt_rn_satfinite_relu_e4m3x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031"), + Some("v1:i1031") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2"), + Some("v1:i1031") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1031" + )); +} +#[test] fn cvt_rn_satfinite_relu_e4m3x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" @@ -516,6 +624,42 @@ fn cvt_rn_satfinite_relu_e4m3x2_f32_uses_only_canonical_or_compatibility_defpath )); } #[test] +fn cvt_rn_satfinite_relu_e5m2x2_bf16x2_uses_only_canonical_or_compatibility_defpaths() { + assert!(is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert!(is_raw_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" + )); + assert_eq!( + generated_intrinsic_marker("cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033"), + Some("v1:i1033") + ); + assert!(matches!( + classify_raw_intrinsic_path( + "cuda_intrinsics", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033".into() + ), + RawIntrinsicIdentity::Known(_) + )); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + )); + assert!(is_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + )); + assert!(!is_raw_generated_intrinsic_path( + "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + )); + assert_eq!( + generated_intrinsic_marker("cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2"), + Some("v1:i1033") + ); + assert!(!is_generated_intrinsic_path( + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v2::i1033" + )); +} +#[test] fn cvt_rn_satfinite_relu_e5m2x2_f16x2_uses_only_canonical_or_compatibility_defpaths() { assert!(is_generated_intrinsic_path( "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs index e614d48ac4..b9c4979cdc 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs index 3539b87869..c590ba15a9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs index 766e88b2f5..f34812b2da 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs index 43676af093..7363ce86a8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs index 45ccb2d478..bf90a576c1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs index ea1a765cb8..6c8bb2a9b1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs index 3e48ff5889..2398d0be33 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs index 4e2fa3ace6..6f1359c171 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs index 77cf005293..56ecaaeba4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs index 7c67982816..1eaba2ee1e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs index 4dc9842f60..99e36d57f8 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs index 2acc775bc3..750590f4f5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs index 2a3002383a..3e8f73da4e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs index 98ad159499..9b497724fa 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs index f33fee4ce7..51f04ca4ea 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs index fe43a1c911..da0961b523 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs index 5da7982e82..5e4f8b7bed 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs index 6fb539a5f9..46e487db10 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs index 171dae8c5a..2b386d159a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs index 5b440dda69..6e235b57ca 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs index 7f967c28c7..d54b3f3591 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs index 9e852b7a6c..ff9bb1b22e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs index 99f9c2491c..3b14eb12cd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs index 8406c947cb..47afeda00f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs index 7234063462..0ae755b455 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs index b3a91fd5e6..efea0f0dc6 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs index 3e1aa64684..76b36a28fa 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs index c57daa967f..393fa04892 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs index 576e50e4e9..7a43837d90 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs index 10bb6fe166..ca35f32b72 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs index e4d3bd35ca..d9c1efce64 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs index 89dbae811c..8d776c191e 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs index 969c0b61a3..da0ed6cc69 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs index 435dd821d3..2c73ad62c8 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs index 4f53978669..aa28f4755b 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs index 3555905110..02da2ac579 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs index 75541afe5d..185d3e6157 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs index 3215b3beee..571e2ca51c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces for admitted CUDA intrinsic families. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs index 22f41b615a..4cce932d5e 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs index 1c15e63abb..a9c7610f67 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs index ab105b7637..f8630d9c01 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs index cd88761a89..ec8d3baa5d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_conversion` intrinsics. @@ -16,8 +16,10 @@ use crate::convert::intrinsics::{ use dialect_nvvm::ops::{ CvtF16x2F32Op, CvtF32x2Bf16x2Op, CvtRnF16x2E4m3x2Op, CvtRnF16x2E5m2x2Op, CvtRnReluBf16x2F32Op, CvtRnReluF16x2E4m3x2Op, CvtRnReluF16x2E5m2x2Op, CvtRnReluF16x2F32Op, - CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, CvtRnSatfiniteE5m2x2F16x2Op, - CvtRnSatfiniteE5m2x2F32Op, CvtRnSatfiniteReluE4m3x2F16x2Op, CvtRnSatfiniteReluE4m3x2F32Op, + CvtRnSatfiniteE4m3x2Bf16x2Op, CvtRnSatfiniteE4m3x2F16x2Op, CvtRnSatfiniteE4m3x2F32Op, + CvtRnSatfiniteE5m2x2Bf16x2Op, CvtRnSatfiniteE5m2x2F16x2Op, CvtRnSatfiniteE5m2x2F32Op, + CvtRnSatfiniteReluE4m3x2Bf16x2Op, CvtRnSatfiniteReluE4m3x2F16x2Op, + CvtRnSatfiniteReluE4m3x2F32Op, CvtRnSatfiniteReluE5m2x2Bf16x2Op, CvtRnSatfiniteReluE5m2x2F16x2Op, CvtRnSatfiniteReluE5m2x2F32Op, CvtRzBf16x2F32Op, CvtRzF16x2F32Op, }; @@ -179,6 +181,25 @@ impl MirToLlvmConversion for CvtRnReluF16x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteE4m3x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.e4m3x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteE4m3x2F16x2Op { fn convert( @@ -217,6 +238,25 @@ impl MirToLlvmConversion for CvtRnSatfiniteE4m3x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteE5m2x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.e5m2x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteE5m2x2F16x2Op { fn convert( @@ -255,6 +295,25 @@ impl MirToLlvmConversion for CvtRnSatfiniteE5m2x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteReluE4m3x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.relu.e4m3x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteReluE4m3x2F16x2Op { fn convert( @@ -293,6 +352,25 @@ impl MirToLlvmConversion for CvtRnSatfiniteReluE4m3x2F32Op { } } +#[op_interface_impl] +impl MirToLlvmConversion for CvtRnSatfiniteReluE5m2x2Bf16x2Op { + fn convert( + &self, + ctx: &mut Context, + rewriter: &mut DialectConversionRewriter, + _operands_info: &OperandsInfo, + ) -> Result<()> { + convert_generated_packed_unary( + ctx, + rewriter, + self.get_operation(), + "cvt.rn.satfinite.relu.e5m2x2.bf16x2", + 16, + 32, + ) + } +} + #[op_interface_impl] impl MirToLlvmConversion for CvtRnSatfiniteReluE5m2x2F16x2Op { fn convert( diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs index 065a2989e9..638d253285 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs index ba3730d318..d336c85a70 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs index edb62cf830..415ed85300 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs index a2a9b9bc70..d9fdb8d3c6 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs index a579f8bb6a..0b3196f3c5 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs index 3e23e8f32c..87a83cf928 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs index a1469bfddf..09f7a16c2f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs index 441e699695..c478472b0b 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs index 8e465421a3..9e99d1772d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs index 05fd89a47e..24d6739251 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs index 45cfb5e9c1..153f30e4ff 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs index 3ba049ba8f..be7622b45d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs index 9de55faf9b..5c84bc3bc5 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs index 7e15564c79..c1fdc553d8 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs index b29139f764..febf994684 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs index b84d82d837..547b472e34 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs index 5b6f6313f4..0d1129d530 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs index d0f9932736..5215dc378c 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs +++ b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated collector predicates for intrinsic placeholder functions. @@ -201,12 +201,16 @@ pub(crate) const GENERATED_INTRINSIC_CANONICAL_PATHS: &[&str] = &[ "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0083", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0373", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260", + "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262", "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0372", @@ -1233,12 +1237,16 @@ pub(crate) const GENERATED_INTRINSIC_PUBLIC_PATHS: &[&str] = &[ "cuda_intrinsics::convert::cvt_rn_relu_f16x2_f32", "cuda_intrinsics::convert::cvt_rn_relu_satfinite_tf32_f32", "cuda_intrinsics::convert::cvt_rn_relu_tf32_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32", + "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2", "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f32", "cuda_intrinsics::convert::cvt_rn_satfinite_tf32_f32", @@ -2271,12 +2279,16 @@ pub(crate) fn is_generated_intrinsic_canonical_path(path: &str) -> bool { | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0083" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0373" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260" + | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0262" | "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0372" @@ -3310,12 +3322,16 @@ pub(crate) fn is_generated_intrinsic_compatibility_path(path: &str) -> bool { | "cuda_device::convert::cvt_rn_relu_f16x2_f32" | "cuda_device::convert::cvt_rn_relu_satfinite_tf32_f32" | "cuda_device::convert::cvt_rn_relu_tf32_f32" + | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" + | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" + | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32" + | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2" | "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_f32" | "cuda_device::convert::cvt_rn_satfinite_tf32_f32" diff --git a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md index 76fb786e38..ecf6910e32 100644 --- a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md +++ b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md @@ -303,11 +303,11 @@ they become `call` instructions to `@llvm.nvvm.*` intrinsics. ### Architecture Coverage -At catalog SHA-256 `c243e568` (the stamp in every `ops/generated/` file -header), the dialect holds 576 operations across 42 modules, and they come +At catalog SHA-256 `3f49098f` (the stamp in every `ops/generated/` file +header), the dialect holds 580 operations across 42 modules, and they come from two different places. The split is the first thing to know about it, because it decides where -- and whether -- you would add one. If the header -stamp no longer starts with `c243e568`, the counts on this page predate the +stamp no longer starts with `3f49098f`, the counts on this page predate the catalog you are reading. **Hand-written**, directly under `cuda-oxide/crates/dialect-nvvm/src/ops/`. These are the @@ -327,7 +327,7 @@ describe. There are seven modules and 26 operations: **Generated**, under `ops/generated/`, from `intrinsics/catalog.json` by `cuda-intrinsics-gen`. Every file there opens with `// @generated ... DO NOT EDIT.`, and editing one by hand is undone by the next generator run. This is -the large majority -- 35 modules and 550 operations, resolved from 1029 catalog +the large majority -- 35 modules and 554 operations, resolved from 1033 catalog entries, since several intrinsics can share one structural op: | Area | Modules | Ops | @@ -335,7 +335,7 @@ entries, since several intrinsics can share one structural op: | Tensor Core Gen 5 + TMEM | `tcgen05` | 210 | | Tensor Memory Accelerator | `tma` | 111 | | Special registers | `sreg` | 44 | -| Packed (SIMD-in-register) | `packed_alu`, `packed_conversion`, `packed_atomic` | 51 | +| Packed (SIMD-in-register) | `packed_alu`, `packed_conversion`, `packed_atomic` | 55 | | Async copy and barriers | `cp_async`, `mbarrier_extended`, `mbarrier_basic`, `sync` | 35 | | Warp-level | `warp_shuffle`, `redux`, `vote`, `warp_match`, `warp_barrier`, `active_mask`, `elect` | 39 | | Matrix fragment movement | `ldmatrix`, `register_mma`, `stmatrix`, `wgmma_control`, `movmatrix`, `sparse_mma` | 22 | diff --git a/cuda-oxide/intrinsics/catalog.json b/cuda-oxide/intrinsics/catalog.json index 9c57efa92b..a984bc87c4 100644 --- a/cuda-oxide/intrinsics/catalog.json +++ b/cuda-oxide/intrinsics/catalog.json @@ -11,8 +11,8 @@ }, "inputs": { "imported_sha256": "54d47b8e281dd953dc95fe16f834ff38164d9809171c9a1af097231b0bc45cab", - "overlay_sha256": "08f7a25510022bf48902cb4889b9f711a1cff31d1c5e8ce21c669d61afbb0310", - "abi_ledger_sha256": "fc378e74ef6da50df31de62c3fafcb3264dfb414034e5852147359fcad4184eb", + "overlay_sha256": "9a9ba81bc0e7a27a16ccf6a5254123c36ad98b47a8829bbc222cfb77173ff31e", + "abi_ledger_sha256": "0e8cec7fdd52ad28ff4365bea512101aef06f29ad7781dcde6b867ad93a0a5bc", "evidence_sha256": [ "cuda-12.0-libnvvm-12.0.140-ex2-f16.json:7102ea7d567ff2d2eb75a0384148d2e5eb6e605c26a5b98c71371fa3c46adef6", "cuda-12.9-libnvvm-12.9.86-f32x2.json:bc5bb505677989affa9ac1044da772b4eb349700dc65e16f09fffffa1f5c493e", @@ -68,6 +68,7 @@ "cuda-13.3-libnvvm-13.3.33-wgmma-control.json:6afe57cb717496e040ed81f12d28f1b8a3fe3603e98180626de4bfb265d02c61", "cuda-13.3-libnvvm-13.3.33.json:683386bf1139bff78c82da76c4b985c17f8c7d0e1e02f80689a54ccb066f044f", "cuda-13.3-libnvvm-13.3.73-execution-control.json:f66216bc7aaccca9ea66fc557e19fdeeaac8c277781f669ee969e727e203aac5", + "cuda-13.3-libnvvm-13.3.73-packed-conversion-fp8-bf16x2.json:7cc62aa25f9d696eaefcaedf104f53da9d08f24385f18d1539308f34d56d38f8", "cuda-13.3-libnvvm-13.3.73-register-mma-mxf8f6f4-f32.json:e437bf84570f4df0a28a743a132e0b93f359f9431d1af03b443f324f62b715ba", "cuda-13.3-libnvvm-13.3.73-sparse-mma-fp8.json:8a57d207ffd510b8c1bb40643bdb828d3d22e5d1c5ab24e1bc1b697538e9bc21", "cuda-13.3-libnvvm-13.3.73-tma-descriptor-control.json:ac44f485122404f9a900cfb94bff658160a6c23d9e2f8ef74bb751f044b848fc", @@ -88,6 +89,7 @@ "rust-llvm-23.1.3-aab11ee0-integer-minmax.json:94bfd0d84bf57737acda19278a10cb566c25677c5688e6a918c83da822100480", "rust-llvm-23.1.3-aab11ee0-mbarrier-basic.json:193966219733100689a43dfde7e9f8aba1f3e3ba2cf80b14bd4893504f900489", "rust-llvm-23.1.3-aab11ee0-mbarrier-extended.json:c92addb73ebd8ef4f3bb2d5902a9e7f280ba898d10b54682d44f26cd2968fcd8", + "rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json:4ee208e5c7bf8891f58e193593f763dd32f1005dbbcb153e99fd65e13de2177c", "rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-f16x2.json:20124138104fe9f71a583a73cb8656c7d91aa10366e9e041499e024578c1a2e3", "rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8.json:42561b5ee707545ccabe2b3b78de12a31614893611e0deccbb498a3ba2967fd1", "rust-llvm-23.1.3-aab11ee0-prmt.json:c77aaaa7fed2d55c2319e286f092bc440e34ef6e6aaf08abe712a3179745d990", @@ -45202,18 +45204,18 @@ "summary": "Converts one f32 value with rn and returns raw TF32 bits." }, { - "id": "cvt_rn_satfinite_e4m3x2_f16x2", - "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e4m3x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_f16x2_to_e4m3x2_rn" + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite" }, "selections": [], "rust": { - "abi_id": "i0822", + "abi_id": "i1030", "module": "convert", - "name": "cvt_rn_satfinite_e4m3x2_f16x2", + "name": "cvt_rn_satfinite_e4m3x2_bf16x2", "arguments": [ "u32" ], @@ -45221,15 +45223,15 @@ "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE4m3x2F16x2Op", - "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f16x2", + "op_type": "CvtRnSatfiniteE4m3x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_bf16x2", "operands": [ "i32" ], @@ -45238,9 +45240,9 @@ ] }, "llvm": { - "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn", + "symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", "arguments": [ - "v2f16" + "v2bf16" ], "results": [ "i16" @@ -45262,18 +45264,54 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 } ] }, "ptx_result": "u16", - "targets": "all", + "targets": "sm_100a|sm_100f|sm_103a|sm_103f|sm_110a|sm_110f|sm_120a|sm_120f|sm_121a|sm_121f", "ptx_isa_version": "9.3", "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" @@ -45284,8 +45322,8 @@ "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_89", - "ptx_feature": "+ptx81" + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91" }, "backend_lowerings": [ { @@ -45293,13 +45331,49 @@ "mechanism": "inline_ptx", "evidence_profile": "rust-llvm-23.1.3-aab11ee0", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 } ] } @@ -45312,14 +45386,14 @@ "stages": [ { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate typed LLVM declaration probes", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 777bd7a5362a935cb9b45c99eb5c7912d8296fd322f232be5feb4dcb71afe958.", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 189-byte declaration-only fixture (SHA-256 eed732da3bb65f7612f2146d1c60ae15fbace894a167c2fde5a4960cb450c262) to a 447-byte module (SHA-256 037d22467fb57cf069c0ab49b3d503e190446292c119acda734bd80445e2cedf) whose declaration of @llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45327,14 +45401,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate probes for the recorded lowering mechanism", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 8843a1e5b9fee11892601ac9688d35a5195bdd8c88122bb6ab5b4fa21cf76e29.", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 367-byte typed fixture (SHA-256 6c31c3240a710f61269c924f0f9efe46537f417c3ce954fdc99402ad8336595f), one call of @llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite on a dynamic <2 x bfloat> argument, to 745-byte PTX (SHA-256 e012440161e64091eec89c929a50d9d7567fccf63f87f5d17822a8b3447112d6; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e4m3x2.bf16x2 with register operands.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45342,14 +45416,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate PTX outputs assembled for the recorded target", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 ed9c5390039d4cfeed4d5708005fc89938d3c0b2b9ff580d4d4d86a004deb5bb.", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.E4M3.BF16 instruction.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -45357,14 +45431,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate probes for the recorded lowering mechanism", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 8d344820320e9db759a6db9eebd370929159ce71d483a76cb70f846d27abf426.", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 210-byte inline-PTX fixture (SHA-256 9f11d88381826f98151d13a0f04189bab6f6870223472c1a55c2980303df5dbf; the body of intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll without its comment header) to 749-byte PTX (SHA-256 67affa4ebac67d992e18927e0c99d3f12a131aba71bb7eec020ba0b0cc1d3e43; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e4m3x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45372,14 +45446,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate PTX outputs assembled for the recorded target", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 ed9c5390039d4cfeed4d5708005fc89938d3c0b2b9ff580d4d4d86a004deb5bb.", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -45390,35 +45464,71 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 } ] } }, - "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73", + "sha256": "e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", - "build_id_prefix": "2330c0", + "build_id_prefix": "88b802", "status": "validated", "stages": [ { "targets": [ - "compute_89", + "compute_120f", "ptx93" ], "representation": "pure inline PTX with a single packed source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4).", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45426,39 +45536,39 @@ }, { "targets": [ - "sm_89", + "sm_120f", "ptx93" ], - "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.E4M3.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" }, { "targets": [ - "sm_89", + "sm_120f", "ptx93" ], - "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "CUDA 13.3 nvJitLink 13.3", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f16x2", + "source_format": "bf16x2", "destination_format": "e4m3x2", "rounding": "nearest_even", "saturation": "satfinite", @@ -45471,7 +45581,7 @@ "rn", "satfinite", "e4m3x2", - "f16x2" + "bf16x2" ], "operands": [ { @@ -45482,51 +45592,48 @@ } ] }, - "summary": "Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order." + "summary": "Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order." }, { - "id": "cvt_rn_satfinite_e4m3x2_f32", - "operation_key": "packed.convert.f32x2.e4m3x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e4m3x2_f16x2", + "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_ff_to_e4m3x2_rn" + "source_record": "int_nvvm_f16x2_to_e4m3x2_rn" }, "selections": [], "rust": { - "abi_id": "i0259", + "abi_id": "i0822", "module": "convert", - "name": "cvt_rn_satfinite_e4m3x2_f32", + "name": "cvt_rn_satfinite_e4m3x2_f16x2", "arguments": [ - "f32", - "f32" + "u32" ], "result": "u16", "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE4m3x2F32Op", - "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f32", + "op_type": "CvtRnSatfiniteE4m3x2F16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f16x2", "operands": [ - "f32", - "f32" + "i32" ], "results": [ "i16" ] }, "llvm": { - "symbol": "llvm.nvvm.ff.to.e4m3x2.rn", + "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn", "arguments": [ - "f32", - "f32" + "v2f16" ], "results": [ "i16" @@ -45576,7 +45683,7 @@ "backend_lowerings": [ { "backend": "llvm_nvptx", - "mechanism": "typed_nvvm", + "mechanism": "inline_ptx", "evidence_profile": "rust-llvm-23.1.3-aab11ee0", "target": { "minimum_ptx": "8.1", @@ -45605,7 +45712,7 @@ "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 bd31e276f07818fdd0a11d4334bf456235366a9c46422edc617ef0140d6997ab.", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 777bd7a5362a935cb9b45c99eb5c7912d8296fd322f232be5feb4dcb71afe958.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45620,7 +45727,7 @@ "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 fa496a4aa988e4fd9ab8baf8b1ad392ebe27bbacb8c7baafc198d96c6756993a.", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 8843a1e5b9fee11892601ac9688d35a5195bdd8c88122bb6ab5b4fa21cf76e29.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45635,7 +45742,7 @@ "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 ed9c5390039d4cfeed4d5708005fc89938d3c0b2b9ff580d4d4d86a004deb5bb.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -45650,7 +45757,7 @@ "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 9050d525b077779e4e5e0ddfa8ab75608b22adab5731337e1925cb992440558c.", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 8d344820320e9db759a6db9eebd370929159ce71d483a76cb70f846d27abf426.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45665,7 +45772,7 @@ "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 ed9c5390039d4cfeed4d5708005fc89938d3c0b2b9ff580d4d4d86a004deb5bb.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -45690,7 +45797,7 @@ } }, "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", + "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", "build_id_prefix": "2330c0", "status": "validated", @@ -45700,11 +45807,11 @@ "compute_89", "ptx93" ], - "representation": "pure inline PTX with u16 output and public low, high inputs", + "representation": "pure inline PTX with a single packed source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", + "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45715,11 +45822,11 @@ "sm_89", "ptx93" ], - "representation": "libNVVM-emitted packed-FP8 PTX", + "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", + "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -45730,89 +45837,89 @@ "sm_89", "ptx93" ], - "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", + "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", + "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "V13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3", "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f32x2", + "source_format": "f16x2", "destination_format": "e4m3x2", "rounding": "nearest_even", "saturation": "satfinite", - "adapter": "reverse_high_low_operands" + "adapter": "identity" }, - "lowering": "generated_packed_conversion_backend", + "lowering": "generated_packed_conversion_inline_ptx", "expected_ptx": { "mnemonic": "cvt", "modifiers": [ "rn", "satfinite", "e4m3x2", - "f32" + "f16x2" ], "operands": [ { "kind": "register" }, - { - "kind": "register" - }, { "kind": "register" } ] }, - "summary": "Converts two f32 values to packed e4m3x2 with nearest-even finite saturation and the first argument in the low byte." + "summary": "Converts packed f16x2 to packed e4m3x2 with nearest-even finite saturation, preserving half order." }, { - "id": "cvt_rn_satfinite_e5m2x2_f16x2", - "operation_key": "packed.convert.f16x2.e5m2x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e4m3x2_f32", + "operation_key": "packed.convert.f32x2.e4m3x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_f16x2_to_e5m2x2_rn" + "source_record": "int_nvvm_ff_to_e4m3x2_rn" }, "selections": [], "rust": { - "abi_id": "i0824", + "abi_id": "i0259", "module": "convert", - "name": "cvt_rn_satfinite_e5m2x2_f16x2", + "name": "cvt_rn_satfinite_e4m3x2_f32", "arguments": [ - "u32" + "f32", + "f32" ], "result": "u16", "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" + "cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE5m2x2F16x2Op", - "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f16x2", + "op_type": "CvtRnSatfiniteE4m3x2F32Op", + "op_name": "nvvm.cvt_rn_satfinite_e4m3x2_f32", "operands": [ - "i32" + "f32", + "f32" ], "results": [ "i16" ] }, "llvm": { - "symbol": "llvm.nvvm.f16x2.to.e5m2x2.rn", + "symbol": "llvm.nvvm.ff.to.e4m3x2.rn", "arguments": [ - "v2f16" + "f32", + "f32" ], "results": [ "i16" @@ -45862,7 +45969,7 @@ "backend_lowerings": [ { "backend": "llvm_nvptx", - "mechanism": "inline_ptx", + "mechanism": "typed_nvvm", "evidence_profile": "rust-llvm-23.1.3-aab11ee0", "target": { "minimum_ptx": "8.1", @@ -45891,7 +45998,7 @@ "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 5b6fcb8ea52dc533221fc3a953835ab9fb2cdb8a3c82a1dc9527dc2b980cb096.", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 bd31e276f07818fdd0a11d4334bf456235366a9c46422edc617ef0140d6997ab.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45906,7 +46013,7 @@ "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6707ccb77a07e8b004032798bbfd07afba3af30d15568ac2c3d08d84d8b10b3.", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 fa496a4aa988e4fd9ab8baf8b1ad392ebe27bbacb8c7baafc198d96c6756993a.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45921,7 +46028,7 @@ "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 f3263d6f3b707aebc45c8e4608b974d23fd5769c7d9886716e35456201ee254a.", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -45936,7 +46043,7 @@ "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 6064f89491e6f84f9ac56c6373eec145a9b8d55507b30a4ee2ffcbe2deed7dd1.", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 9050d525b077779e4e5e0ddfa8ab75608b22adab5731337e1925cb992440558c.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -45951,7 +46058,7 @@ "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 f3263d6f3b707aebc45c8e4608b974d23fd5769c7d9886716e35456201ee254a.", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -45976,7 +46083,7 @@ } }, "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", "build_id_prefix": "2330c0", "status": "validated", @@ -45986,11 +46093,11 @@ "compute_89", "ptx93" ], - "representation": "pure inline PTX with a single packed source operand", + "representation": "pure inline PTX with u16 output and public low, high inputs", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", + "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -46001,11 +46108,11 @@ "sm_89", "ptx93" ], - "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", + "representation": "libNVVM-emitted packed-FP8 PTX", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", + "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", @@ -46016,89 +46123,89 @@ "sm_89", "ptx93" ], - "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", + "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", + "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "CUDA 13.3 nvJitLink 13.3", + "tool_version": "V13.3.33", "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f16x2", - "destination_format": "e5m2x2", + "source_format": "f32x2", + "destination_format": "e4m3x2", "rounding": "nearest_even", "saturation": "satfinite", - "adapter": "identity" + "adapter": "reverse_high_low_operands" }, - "lowering": "generated_packed_conversion_inline_ptx", + "lowering": "generated_packed_conversion_backend", "expected_ptx": { "mnemonic": "cvt", "modifiers": [ "rn", "satfinite", - "e5m2x2", - "f16x2" + "e4m3x2", + "f32" ], "operands": [ { "kind": "register" }, + { + "kind": "register" + }, { "kind": "register" } ] }, - "summary": "Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order." + "summary": "Converts two f32 values to packed e4m3x2 with nearest-even finite saturation and the first argument in the low byte." }, { - "id": "cvt_rn_satfinite_e5m2x2_f32", - "operation_key": "packed.convert.f32x2.e5m2x2.nearest_even.satfinite", + "id": "cvt_rn_satfinite_e5m2x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_ff_to_e5m2x2_rn" + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite" }, "selections": [], "rust": { - "abi_id": "i0261", + "abi_id": "i1032", "module": "convert", - "name": "cvt_rn_satfinite_e5m2x2_f32", + "name": "cvt_rn_satfinite_e5m2x2_bf16x2", "arguments": [ - "f32", - "f32" + "u32" ], "result": "u16", "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteE5m2x2F32Op", - "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f32", + "op_type": "CvtRnSatfiniteE5m2x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_bf16x2", "operands": [ - "f32", - "f32" + "i32" ], "results": [ "i16" ] }, "llvm": { - "symbol": "llvm.nvvm.ff.to.e5m2x2.rn", + "symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", "arguments": [ - "f32", - "f32" + "v2bf16" ], "results": [ "i16" @@ -46120,18 +46227,54 @@ "execution_scope": "thread" }, "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 } ] }, "ptx_result": "u16", - "targets": "all", + "targets": "sm_100a|sm_100f|sm_103a|sm_103f|sm_110a|sm_110f|sm_120a|sm_120f|sm_121a|sm_121f", "ptx_isa_version": "9.3", "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" @@ -46142,22 +46285,58 @@ "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", "status": "validated", "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_89", - "ptx_feature": "+ptx81" + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91" }, "backend_lowerings": [ { "backend": "llvm_nvptx", - "mechanism": "typed_nvvm", + "mechanism": "inline_ptx", "evidence_profile": "rust-llvm-23.1.3-aab11ee0", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 } ] } @@ -46170,14 +46349,14 @@ "stages": [ { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate typed LLVM declaration probes", "stage": "declaration_canonicalization", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 bd31e276f07818fdd0a11d4334bf456235366a9c46422edc617ef0140d6997ab.", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 189-byte declaration-only fixture (SHA-256 dcf348779f9dde49249381b61a3141b1d36f5665ecbb2bc53ebc0be7e430c075) to a 447-byte module (SHA-256 63eaec75fcb9115321790aa90787b46064e7d007952182040a761407edb5f60c) whose declaration of @llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -46185,14 +46364,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate probes for the recorded lowering mechanism", "stage": "backend_codegen", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 fa496a4aa988e4fd9ab8baf8b1ad392ebe27bbacb8c7baafc198d96c6756993a.", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 367-byte typed fixture (SHA-256 7b4acbf67e3c0f83b1099150d70876603349db6db1b1502dce79da9b702b04f6), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite on a dynamic <2 x bfloat> argument, to 745-byte PTX (SHA-256 3334d10e01ea5e8f531e764a85e85f87cc2d0bc378f9bb38f9ef8a934ac54614; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with register operands.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -46200,14 +46379,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate PTX outputs assembled for the recorded target", "stage": "ptx_assembly", "mechanism": "typed_nvvm", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.E5M2.BF16 instruction.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -46215,14 +46394,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate probes for the recorded lowering mechanism", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 9050d525b077779e4e5e0ddfa8ab75608b22adab5731337e1925cb992440558c.", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 210-byte inline-PTX fixture (SHA-256 fff4ab63076b1daff9772ebd8744e62bebb0f71d8c9b573111157b2125c4e96b; the body of intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll without its comment header) to 749-byte PTX (SHA-256 bdfc6e8b024796f9b5081893885b7d3a34fb36bc97e07fad8481bab56b9eab79; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result.", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -46230,14 +46409,14 @@ }, { "targets": [ - "sm_89", - "ptx81" + "sm_120f", + "ptx91" ], "representation": "separate PTX outputs assembled for the recorded target", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -46248,35 +46427,71 @@ { "backend": "lib_nvvm", "mechanism": "inline_ptx", - "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73", "target": { - "minimum_ptx": "8.1", + "minimum_ptx": "9.1", "hardware": { "kind": "any_of", "alternatives": [ { - "kind": "minimum_sm", - "sm": 89 + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 } ] } }, - "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73", + "sha256": "e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77", "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", - "build_id_prefix": "2330c0", + "build_id_prefix": "88b802", "status": "validated", "stages": [ { "targets": [ - "compute_89", + "compute_120f", "ptx93" ], - "representation": "pure inline PTX with u16 output and public low, high inputs", + "representation": "pure inline PTX with a single packed source operand", "stage": "backend_codegen", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4).", "artifact_kind": null, "tool_path": null, "tool_version": null, @@ -46284,80 +46499,77 @@ }, { "targets": [ - "sm_89", + "sm_120f", "ptx93" ], - "representation": "libNVVM-emitted packed-FP8 PTX", + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", - "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", - "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" }, { "targets": [ - "sm_89", + "sm_120f", "ptx93" ], - "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", "stage": "device_link", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", - "tool_version": "V13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" } ] } ], "packed_conversion": { - "source_format": "f32x2", + "source_format": "bf16x2", "destination_format": "e5m2x2", "rounding": "nearest_even", "saturation": "satfinite", - "adapter": "reverse_high_low_operands" + "adapter": "identity" }, - "lowering": "generated_packed_conversion_backend", + "lowering": "generated_packed_conversion_inline_ptx", "expected_ptx": { "mnemonic": "cvt", "modifiers": [ "rn", "satfinite", "e5m2x2", - "f32" + "bf16x2" ], "operands": [ { "kind": "register" }, - { - "kind": "register" - }, { "kind": "register" } ] }, - "summary": "Converts two f32 values to packed e5m2x2 with nearest-even finite saturation and the first argument in the low byte." + "summary": "Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order." }, { - "id": "cvt_rn_satfinite_relu_e4m3x2_f16x2", - "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite.relu", + "id": "cvt_rn_satfinite_e5m2x2_f16x2", + "operation_key": "packed.convert.f16x2.e5m2x2.nearest_even.satfinite", "family": "packed_conversion", "source": { "kind": "llvm_imported", - "source_record": "int_nvvm_f16x2_to_e4m3x2_rn_relu" + "source_record": "int_nvvm_f16x2_to_e5m2x2_rn" }, "selections": [], "rust": { - "abi_id": "i0823", + "abi_id": "i0824", "module": "convert", - "name": "cvt_rn_satfinite_relu_e4m3x2_f16x2", + "name": "cvt_rn_satfinite_e5m2x2_f16x2", "arguments": [ "u32" ], @@ -46365,15 +46577,15 @@ "safe": true, "must_use": false, "safe_allowlist_reason": "This conversion has no caller obligations.", - "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823", - "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2", "compatibility_paths": [ - "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f16x2" ] }, "dialect": { - "op_type": "CvtRnSatfiniteReluE4m3x2F16x2Op", - "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_f16x2", + "op_type": "CvtRnSatfiniteE5m2x2F16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f16x2", "operands": [ "i32" ], @@ -46382,7 +46594,971 @@ ] }, "llvm": { - "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn.relu", + "symbol": "llvm.nvvm.f16x2.to.e5m2x2.rn", + "arguments": [ + "v2f16" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + }, + "ptx_result": "u16", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_89", + "ptx_feature": "+ptx81" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "inline_ptx", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 5b6fcb8ea52dc533221fc3a953835ab9fb2cdb8a3c82a1dc9527dc2b980cb096.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6707ccb77a07e8b004032798bbfd07afba3af30d15568ac2c3d08d84d8b10b3.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 f3263d6f3b707aebc45c8e4608b974d23fd5769c7d9886716e35456201ee254a.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 6064f89491e6f84f9ac56c6373eec145a9b8d55507b30a4ee2ffcbe2deed7dd1.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 f3263d6f3b707aebc45c8e4608b974d23fd5769c7d9886716e35456201ee254a.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", + "sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "2330c0", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_89", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,774-byte NVVM IR fixture (SHA-256 177071f5289fef1bade95e29f12afa689c91dc8f4c6f4d89cca810527c1317da) to 1,628-byte PTX (SHA-256 54595aa338e74747cf5ec117a0570f112c673a3d0266b7d2a5d850d3ba6c8ba4) and 2,268-byte LTOIR (SHA-256 dae644d035c13e6afec18e39d26ba16c8326caecab2d6425158b563d3d91a4ba). All eight forms retained their single packed source operand.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM-emitted f16x2 and packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas accepted all eight forms at sm_89 and produced a 3,624-byte cubin (SHA-256 c761f544931d83935d10c7965037bad1536fd8d1f30cea4ae52f3826b4e8821e). The same PTX emitted for compute_87 was rejected with \"Feature 'cvt.e4m3x2.f16x2' requires .target sm_89 or higher\".", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all eight f16x2 and packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced a 1,464-byte cubin (SHA-256 aadc245b2a8c0d63d2a3d5337bff8cb7c778022514b8f32d1e1efd61bf91ac3f) from the sm_89 LTOIR. Linking that same LTOIR for sm_87 failed with 'LTOIR arch newer than target arch'.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ] + } + ], + "packed_conversion": { + "source_format": "f16x2", + "destination_format": "e5m2x2", + "rounding": "nearest_even", + "saturation": "satfinite", + "adapter": "identity" + }, + "lowering": "generated_packed_conversion_inline_ptx", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e5m2x2", + "f16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts packed f16x2 to packed e5m2x2 with nearest-even finite saturation, preserving half order." + }, + { + "id": "cvt_rn_satfinite_e5m2x2_f32", + "operation_key": "packed.convert.f32x2.e5m2x2.nearest_even.satfinite", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_ff_to_e5m2x2_rn" + }, + "selections": [], + "rust": { + "abi_id": "i0261", + "module": "convert", + "name": "cvt_rn_satfinite_e5m2x2_f32", + "arguments": [ + "f32", + "f32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteE5m2x2F32Op", + "op_name": "nvvm.cvt_rn_satfinite_e5m2x2_f32", + "operands": [ + "f32", + "f32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.ff.to.e5m2x2.rn", + "arguments": [ + "f32", + "f32" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + }, + "ptx_result": "u16", + "targets": "all", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_89", + "ptx_feature": "+ptx81" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "typed_nvvm", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 bd31e276f07818fdd0a11d4334bf456235366a9c46422edc617ef0140d6997ab.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 fa496a4aa988e4fd9ab8baf8b1ad392ebe27bbacb8c7baafc198d96c6756993a.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): the exact reviewed PTX instruction shapes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 9050d525b077779e4e5e0ddfa8ab75608b22adab5731337e1925cb992440558c.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx81" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Revalidated 4 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): terminal PTX assembly passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a6dded5e03b1ce7eb2e2bfe0e88f4f3a77fb6b5c87f3e0fa2f5312cb1358b0b2.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.33", + "target": { + "minimum_ptx": "8.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "minimum_sm", + "sm": 89 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", + "sha256": "b6a5db233133fb0ed693f8cb8aeb627e6f8e2597ba015a8fc53f58f4956dee58", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "2330c0", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_89", + "ptx93" + ], + "representation": "pure inline PTX with u16 output and public low, high inputs", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM-emitted packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7).", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0", + "tool_sha256": "7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb" + }, + { + "targets": [ + "sm_89", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "V13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ] + } + ], + "packed_conversion": { + "source_format": "f32x2", + "destination_format": "e5m2x2", + "rounding": "nearest_even", + "saturation": "satfinite", + "adapter": "reverse_high_low_operands" + }, + "lowering": "generated_packed_conversion_backend", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e5m2x2", + "f32" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts two f32 values to packed e5m2x2 with nearest-even finite saturation and the first argument in the low byte." + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e4m3x2.nearest_even.satfinite.relu", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite" + }, + "selections": [], + "rust": { + "abi_id": "i1031", + "module": "convert", + "name": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "arguments": [ + "u32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteReluE4m3x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "operands": [ + "i32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", + "arguments": [ + "v2bf16" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 + } + ] + }, + "ptx_result": "u16", + "targets": "sm_100a|sm_100f|sm_103a|sm_103f|sm_110a|sm_110f|sm_120a|sm_120f|sm_121a|sm_121f", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "inline_ptx", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 194-byte declaration-only fixture (SHA-256 318784297cbf5626850918899261c40db020b1cf755939c774dbdfcc0f4e67f2) to a 462-byte module (SHA-256 5237fee14f66ec4f52406118999e5656f0624e15918febf001630cd8c142f522) whose declaration of @llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 382-byte typed fixture (SHA-256 f374f8c786899f7e794a325136a9b411add44e6ca053a0d5a659ce2b62fcc013), one call of @llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite on a dynamic <2 x bfloat> argument, to 780-byte PTX (SHA-256 ed6663e57024ca9435b851169d11a9a4fee64eb7f94899ef02f20d3d886839dc; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e4m3x2.bf16x2 with register operands.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.RELU.E4M3.BF16 instruction.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 220-byte inline-PTX fixture (SHA-256 dd4c8e4f33b8463bc6adf33bf59fe47cb6054daea21b7cf3a84dd634fb0f201f; the body of intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll without its comment header) to 784-byte PTX (SHA-256 3cfe93699632a0d050f21ba46f5e94dba53fa7b9f2ffbb369c1bc06190b7ce80; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e4m3x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73", + "sha256": "e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_120f", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4).", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.RELU.E4M3.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ] + } + ], + "packed_conversion": { + "source_format": "bf16x2", + "destination_format": "e4m3x2", + "rounding": "nearest_even", + "saturation": "satfinite_relu", + "adapter": "identity" + }, + "lowering": "generated_packed_conversion_inline_ptx", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts packed bf16x2 to packed e4m3x2 with nearest-even finite saturation and ReLU, preserving half order." + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_f16x2", + "operation_key": "packed.convert.f16x2.e4m3x2.nearest_even.satfinite.relu", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_f16x2_to_e4m3x2_rn_relu" + }, + "selections": [], + "rust": { + "abi_id": "i0823", + "module": "convert", + "name": "cvt_rn_satfinite_relu_e4m3x2_f16x2", + "arguments": [ + "u32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteReluE4m3x2F16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_relu_e4m3x2_f16x2", + "operands": [ + "i32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.f16x2.to.e4m3x2.rn.relu", "arguments": [ "v2f16" ], @@ -46919,6 +48095,398 @@ }, "summary": "Converts two f32 values to packed e4m3x2 with nearest-even finite saturation, ReLU, and the first argument in the low byte." }, + { + "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "operation_key": "packed.convert.bf16x2.e5m2x2.nearest_even.satfinite.relu", + "family": "packed_conversion", + "source": { + "kind": "llvm_imported", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite" + }, + "selections": [], + "rust": { + "abi_id": "i1033", + "module": "convert", + "name": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "arguments": [ + "u32" + ], + "result": "u16", + "safe": true, + "must_use": false, + "safe_allowlist_reason": "This conversion has no caller obligations.", + "canonical_path": "cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033", + "public_path": "cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "compatibility_paths": [ + "cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2" + ] + }, + "dialect": { + "op_type": "CvtRnSatfiniteReluE5m2x2Bf16x2Op", + "op_name": "nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "operands": [ + "i32" + ], + "results": [ + "i16" + ] + }, + "llvm": { + "symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", + "arguments": [ + "v2bf16" + ], + "results": [ + "i16" + ], + "properties": [ + "IntrNoCreateUndefOrPoison", + "IntrNoMem", + "IntrSpeculatable" + ], + "result_facts": { + "no_undef": false, + "range": null + } + }, + "semantics": { + "pure": true, + "memory": "none", + "convergent": false, + "execution_scope": "thread" + }, + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 + } + ] + }, + "ptx_result": "u16", + "targets": "sm_100a|sm_100f|sm_103a|sm_103f|sm_110a|sm_110f|sm_120a|sm_120f|sm_121a|sm_121f", + "ptx_isa_version": "9.3", + "ptx_isa_section": "9.7.9.22 Data Movement and Conversion Instructions: cvt", + "ptx_isa_url": "https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt" + }, + "backend": { + "profile": "rust-llvm-23.1.3-aab11ee0", + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "status": "validated", + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91" + }, + "backend_lowerings": [ + { + "backend": "llvm_nvptx", + "mechanism": "inline_ptx", + "evidence_profile": "rust-llvm-23.1.3-aab11ee0", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 + } + ] + } + }, + "version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "artifact_path": null, + "build_id_prefix": null, + "status": "validated", + "stages": [ + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 194-byte declaration-only fixture (SHA-256 12c7df1f1031c28863d8993452d940b6ed6affc024534d32de66272e70490048) to a 462-byte module (SHA-256 5cfbd96861014107c50c79030418eb8400ee1e6ec25dfc0d6679739e8997de5f) whose declaration of @llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 382-byte typed fixture (SHA-256 531ef4ab3bc1e551c3a2b5fc5d646a67f522a454b64cb47bcd0ac7617c73a1d3), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite on a dynamic <2 x bfloat> argument, to 780-byte PTX (SHA-256 460efd987919d91caf04ca68e9223b4d22c4d37872164593799122bd0ea8d764; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with register operands.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.RELU.E5M2.BF16 instruction.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 220-byte inline-PTX fixture (SHA-256 052e0049faabfaae45bfd2f5e74607a3a39c2f40cd108c6656a00104c933e207; the body of intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll without its comment header) to 784-byte PTX (SHA-256 8946110f2d12fa0239e4132675f8f6facd71e80dfbba7e6cf40d010cdba91e96; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result.", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ] + }, + { + "backend": "lib_nvvm", + "mechanism": "inline_ptx", + "evidence_profile": "cuda-13.3-libnvvm-13.3.73", + "target": { + "minimum_ptx": "9.1", + "hardware": { + "kind": "any_of", + "alternatives": [ + { + "kind": "exact_architecture", + "sm": 100 + }, + { + "kind": "family_target", + "sm": 100 + }, + { + "kind": "exact_architecture", + "sm": 103 + }, + { + "kind": "family_target", + "sm": 103 + }, + { + "kind": "exact_architecture", + "sm": 110 + }, + { + "kind": "family_target", + "sm": 110 + }, + { + "kind": "exact_architecture", + "sm": 120 + }, + { + "kind": "family_target", + "sm": 120 + }, + { + "kind": "exact_architecture", + "sm": 121 + }, + { + "kind": "family_target", + "sm": 121 + } + ] + } + }, + "version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73", + "sha256": "e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "88b802", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_120f", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4).", + "artifact_kind": null, + "tool_path": null, + "tool_version": null, + "tool_sha256": null + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.RELU.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", + "artifact_kind": "cubin", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ] + } + ], + "packed_conversion": { + "source_format": "bf16x2", + "destination_format": "e5m2x2", + "rounding": "nearest_even", + "saturation": "satfinite_relu", + "adapter": "identity" + }, + "lowering": "generated_packed_conversion_inline_ptx", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + }, + "summary": "Converts packed bf16x2 to packed e5m2x2 with nearest-even finite saturation and ReLU, preserving half order." + }, { "id": "cvt_rn_satfinite_relu_e5m2x2_f16x2", "operation_key": "packed.convert.f16x2.e5m2x2.nearest_even.satfinite.relu", diff --git a/cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json b/cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json deleted file mode 100644 index 033c904a10..0000000000 --- a/cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.33-packed-conversion-fp8-bf16x2.json +++ /dev/null @@ -1,200 +0,0 @@ -{ - "schema": 6, - "backend_profile": "cuda-13.3-libnvvm-13.3.33", - "backend_kind": "lib_nvvm", - "llvm_revision": "not-applicable-libnvvm-ir-2.0", - "backend_version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.33", - "backend_sha256": "b6a5db233133fb0ed693f8cb8aeb1d405254ce0ab10364478985292ec98cfcf0", - "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", - "build_id_prefix": "2330c0", - "nvvm_ir_version": "2.0", - "debug_ir_version": "3.2", - "records": [ - { - "id": "cvt_rn_satfinite_e4m3x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "e4m3x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - }, - { - "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "relu", - "e4m3x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - }, - { - "id": "cvt_rn_satfinite_e5m2x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "e5m2x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - }, - { - "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "relu", - "e5m2x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - } - ] -} diff --git a/cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-packed-conversion-fp8-bf16x2.json b/cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-packed-conversion-fp8-bf16x2.json new file mode 100644 index 0000000000..627ceba64c --- /dev/null +++ b/cuda-oxide/intrinsics/evidence/cuda-13.3-libnvvm-13.3.73-packed-conversion-fp8-bf16x2.json @@ -0,0 +1,320 @@ +{ + "schema": 6, + "backend_profile": "cuda-13.3-libnvvm-13.3.73", + "backend_kind": "lib_nvvm", + "llvm_revision": "not-applicable-libnvvm-ir-2.0", + "backend_version": "CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73", + "backend_sha256": "e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77", + "artifact_path": "/usr/local/cuda-13.3/nvvm/lib64/libnvvm.so.4.0.0", + "build_id_prefix": "88b802", + "nvvm_ir_version": "2.0", + "debug_ir_version": "3.2", + "records": [ + { + "id": "cvt_rn_satfinite_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "compute_120f", + "ptx_feature": "PTX ISA 9.1 minimum; CUDA 13.3 emits PTX 9.3", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_120f", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4)." + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.E4M3.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "compute_120f", + "ptx_feature": "PTX ISA 9.1 minimum; CUDA 13.3 emits PTX 9.3", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_120f", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4)." + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.RELU.E4M3.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "compute_120f", + "ptx_feature": "PTX ISA 9.1 minimum; CUDA 13.3 emits PTX 9.3", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_120f", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4)." + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "compute_120f", + "ptx_feature": "PTX ISA 9.1 minimum; CUDA 13.3 emits PTX 9.3", + "status": "validated", + "stages": [ + { + "targets": [ + "compute_120f", + "ptx93" + ], + "representation": "pure inline PTX with a single packed source operand", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4)." + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM-emitted bf16x2 to packed-FP8 PTX", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.RELU.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx93" + ], + "representation": "libNVVM LTOIR retaining all four bf16x2 to packed-FP8 conversions", + "stage": "device_link", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error.", + "tool_path": "/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33", + "tool_version": "CUDA 13.3 nvJitLink 13.3.33", + "tool_sha256": "3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8" + } + ], + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + } + ] +} diff --git a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json deleted file mode 100644 index aac4bc8454..0000000000 --- a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.0-16696adc-packed-conversion-fp8-bf16x2.json +++ /dev/null @@ -1,196 +0,0 @@ -{ - "schema": 6, - "backend_profile": "rust-llvm-23.1.0-16696adc", - "backend_kind": "llvm_nvptx", - "llvm_revision": "16696adcd119e6ba9cc175207d984d7021211acb", - "backend_version": "LLVM version 23.1.0-rust-1.100.0-nightly", - "backend_sha256": "2e5bc688625ca7d98bca9b5f0f0d14525d2eb59e8a9842964e480df3ffd1f921", - "records": [ - { - "id": "cvt_rn_satfinite_e4m3x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "e4m3x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - }, - { - "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "relu", - "e4m3x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - }, - { - "id": "cvt_rn_satfinite_e5m2x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "e5m2x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - }, - { - "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", - "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite", - "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", - "llvm_arguments": [ - "v2bf16" - ], - "llvm_results": [ - "i16" - ], - "target_triple": "nvptx64-nvidia-cuda", - "gpu_target": "sm_100f", - "ptx_feature": "+ptx91", - "status": "lowered", - "stages": [ - { - "targets": [ - "sm_100f", - "ptx91" - ], - "representation": "pure inline PTX with a single packed bf16x2 source operand", - "stage": "backend_codegen", - "mechanism": "inline_ptx", - "outcome": "succeeded", - "detail": "Host llc from rustc 1.100.0-nightly (e457a7b0d 2026-08-27) lowered a 743-byte inline-PTX fixture (SHA-256 7cad46e32924ad631c6cdd05fc8859123ddfa4280b0ccee48f897292315cdc45) to 2,789-byte PTX (SHA-256 d0ad803094d87ebaee0a1e4163fe311a11555381de02e3f08de6a0c6441d6f05) at sm_100 +ptx91. Each form is one packed i32 source and an i16 result with constraint =h,r. ptxas and libNVVM are not installed here, so this record is lowered rather than validated and runtime_validation stays unexecuted." - } - ], - "runtime_validation": "unexecuted", - "expected_ptx": { - "mnemonic": "cvt", - "modifiers": [ - "rn", - "satfinite", - "relu", - "e5m2x2", - "bf16x2" - ], - "operands": [ - { - "kind": "register" - }, - { - "kind": "register" - } - ] - } - } - ] -} diff --git a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json new file mode 100644 index 0000000000..67f1816041 --- /dev/null +++ b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json @@ -0,0 +1,432 @@ +{ + "schema": 6, + "backend_profile": "rust-llvm-23.1.3-aab11ee0", + "backend_kind": "llvm_nvptx", + "llvm_revision": "aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019", + "backend_version": "LLVM version 23.1.3-rust-1.101.0-nightly", + "backend_sha256": "7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449", + "records": [ + { + "id": "cvt_rn_satfinite_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "concrete_llvm_arguments": [ + "v2bf16" + ], + "concrete_llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91", + "status": "validated", + "stages": [ + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 189-byte declaration-only fixture (SHA-256 eed732da3bb65f7612f2146d1c60ae15fbace894a167c2fde5a4960cb450c262) to a 447-byte module (SHA-256 037d22467fb57cf069c0ab49b3d503e190446292c119acda734bd80445e2cedf) whose declaration of @llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 367-byte typed fixture (SHA-256 6c31c3240a710f61269c924f0f9efe46537f417c3ce954fdc99402ad8336595f), one call of @llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite on a dynamic <2 x bfloat> argument, to 745-byte PTX (SHA-256 e012440161e64091eec89c929a50d9d7567fccf63f87f5d17822a8b3447112d6; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e4m3x2.bf16x2 with register operands." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.E4M3.BF16 instruction.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 210-byte inline-PTX fixture (SHA-256 9f11d88381826f98151d13a0f04189bab6f6870223472c1a55c2980303df5dbf; the body of intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll without its comment header) to 749-byte PTX (SHA-256 67affa4ebac67d992e18927e0c99d3f12a131aba71bb7eec020ba0b0cc1d3e43; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e4m3x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ], + "declaration_attributes_canonicalized": true, + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e4m3x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "concrete_llvm_arguments": [ + "v2bf16" + ], + "concrete_llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91", + "status": "validated", + "stages": [ + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 194-byte declaration-only fixture (SHA-256 318784297cbf5626850918899261c40db020b1cf755939c774dbdfcc0f4e67f2) to a 462-byte module (SHA-256 5237fee14f66ec4f52406118999e5656f0624e15918febf001630cd8c142f522) whose declaration of @llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 382-byte typed fixture (SHA-256 f374f8c786899f7e794a325136a9b411add44e6ca053a0d5a659ce2b62fcc013), one call of @llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite on a dynamic <2 x bfloat> argument, to 780-byte PTX (SHA-256 ed6663e57024ca9435b851169d11a9a4fee64eb7f94899ef02f20d3d886839dc; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e4m3x2.bf16x2 with register operands." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.RELU.E4M3.BF16 instruction.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 220-byte inline-PTX fixture (SHA-256 dd4c8e4f33b8463bc6adf33bf59fe47cb6054daea21b7cf3a84dd634fb0f201f; the body of intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll without its comment header) to 784-byte PTX (SHA-256 3cfe93699632a0d050f21ba46f5e94dba53fa7b9f2ffbb369c1bc06190b7ce80; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e4m3x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ], + "declaration_attributes_canonicalized": true, + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e4m3x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "concrete_llvm_arguments": [ + "v2bf16" + ], + "concrete_llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91", + "status": "validated", + "stages": [ + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 189-byte declaration-only fixture (SHA-256 dcf348779f9dde49249381b61a3141b1d36f5665ecbb2bc53ebc0be7e430c075) to a 447-byte module (SHA-256 63eaec75fcb9115321790aa90787b46064e7d007952182040a761407edb5f60c) whose declaration of @llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 367-byte typed fixture (SHA-256 7b4acbf67e3c0f83b1099150d70876603349db6db1b1502dce79da9b702b04f6), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite on a dynamic <2 x bfloat> argument, to 745-byte PTX (SHA-256 3334d10e01ea5e8f531e764a85e85f87cc2d0bc378f9bb38f9ef8a934ac54614; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with register operands." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.E5M2.BF16 instruction.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 210-byte inline-PTX fixture (SHA-256 fff4ab63076b1daff9772ebd8744e62bebb0f71d8c9b573111157b2125c4e96b; the body of intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll without its comment header) to 749-byte PTX (SHA-256 bdfc6e8b024796f9b5081893885b7d3a34fb36bc97e07fad8481bab56b9eab79; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ], + "declaration_attributes_canonicalized": true, + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + }, + { + "id": "cvt_rn_satfinite_relu_e5m2x2_bf16x2", + "source_record": "int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite", + "llvm_symbol": "llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite", + "llvm_arguments": [ + "v2bf16" + ], + "llvm_results": [ + "i16" + ], + "concrete_llvm_arguments": [ + "v2bf16" + ], + "concrete_llvm_results": [ + "i16" + ], + "target_triple": "nvptx64-nvidia-cuda", + "gpu_target": "sm_120f", + "ptx_feature": "+ptx91", + "status": "validated", + "stages": [ + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate typed LLVM declaration probes", + "stage": "declaration_canonicalization", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 194-byte declaration-only fixture (SHA-256 12c7df1f1031c28863d8993452d940b6ed6affc024534d32de66272e70490048) to a 462-byte module (SHA-256 5cfbd96861014107c50c79030418eb8400ee1e6ec25dfc0d6679739e8997de5f) whose declaration of @llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 382-byte typed fixture (SHA-256 531ef4ab3bc1e551c3a2b5fc5d646a67f522a454b64cb47bcd0ac7617c73a1d3), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite on a dynamic <2 x bfloat> argument, to 780-byte PTX (SHA-256 460efd987919d91caf04ca68e9223b4d22c4d37872164593799122bd0ea8d764; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with register operands." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "typed_nvvm", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.RELU.E5M2.BF16 instruction.", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate probes for the recorded lowering mechanism", + "stage": "backend_codegen", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "detail": "Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 220-byte inline-PTX fixture (SHA-256 052e0049faabfaae45bfd2f5e74607a3a39c2f40cd108c6656a00104c933e207; the body of intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll without its comment header) to 784-byte PTX (SHA-256 8946110f2d12fa0239e4132675f8f6facd71e80dfbba7e6cf40d010cdba91e96; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result." + }, + { + "targets": [ + "sm_120f", + "ptx91" + ], + "representation": "separate PTX outputs assembled for the recorded target", + "stage": "ptx_assembly", + "mechanism": "inline_ptx", + "outcome": "succeeded", + "artifact_kind": "cubin", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "tool_path": "/usr/local/cuda-13.3/bin/ptxas", + "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", + "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" + } + ], + "declaration_attributes_canonicalized": true, + "runtime_validation": "unexecuted", + "expected_ptx": { + "mnemonic": "cvt", + "modifiers": [ + "rn", + "satfinite", + "relu", + "e5m2x2", + "bf16x2" + ], + "operands": [ + { + "kind": "register" + }, + { + "kind": "register" + } + ] + } + } + ] +} diff --git a/cuda-oxide/intrinsics/generated-reference.md b/cuda-oxide/intrinsics/generated-reference.md index d9fa06d84b..6cc83d5aa3 100644 --- a/cuda-oxide/intrinsics/generated-reference.md +++ b/cuda-oxide/intrinsics/generated-reference.md @@ -2,7 +2,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0 @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. --> @@ -199,12 +199,16 @@ This table is generated from the resolved catalog. Evidence stages below disting | `cuda_intrinsics::convert::cvt_rn_relu_f16x2_f32` | `nvvm.cvt_rn_relu_f16x2_f32` | LLVM `llvm.nvvm.ff2f16x2.rn.relu` from `int_nvvm_ff2f16x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 7.0 on sm_80+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.relu.f16x2.f32 , , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_relu_satfinite_tf32_f32` | `nvvm.scalar_conversion` | LLVM `llvm.nvvm.f2tf32.rn.relu.satfinite` from `int_nvvm_f2tf32_rn_relu_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 8.6 on sm_100+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.relu.satfinite.tf32.f32 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_relu_tf32_f32` | `nvvm.scalar_conversion` | LLVM `llvm.nvvm.f2tf32.rn.relu` from `int_nvvm_f2tf32_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 7.8 on sm_90+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.relu.tf32.f32 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2` | `nvvm.cvt_rn_satfinite_e4m3x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.e4m3x2.bf16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2` | `nvvm.cvt_rn_satfinite_e4m3x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn` from `int_nvvm_f16x2_to_e4m3x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.e4m3x2.f16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32` | `nvvm.cvt_rn_satfinite_e4m3x2_f32` | LLVM `llvm.nvvm.ff.to.e4m3x2.rn` from `int_nvvm_ff_to_e4m3x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.e4m3x2.f32 , , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2` | `nvvm.cvt_rn_satfinite_e5m2x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.e5m2x2.bf16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2` | `nvvm.cvt_rn_satfinite_e5m2x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn` from `int_nvvm_f16x2_to_e5m2x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.e5m2x2.f16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32` | `nvvm.cvt_rn_satfinite_e5m2x2_f32` | LLVM `llvm.nvvm.ff.to.e5m2x2.rn` from `int_nvvm_ff_to_e5m2x2_rn` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.e5m2x2.f32 , , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` | `nvvm.cvt_rn_satfinite_relu_e4m3x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e4m3x2_rn_relu_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.relu.e4m3x2.bf16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2` | `nvvm.cvt_rn_satfinite_relu_e4m3x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e4m3x2.rn.relu` from `int_nvvm_f16x2_to_e4m3x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.relu.e4m3x2.f16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32` | `nvvm.cvt_rn_satfinite_relu_e4m3x2_f32` | LLVM `llvm.nvvm.ff.to.e4m3x2.rn.relu` from `int_nvvm_ff_to_e4m3x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.relu.e4m3x2.f32 , , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | +| `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` | `nvvm.cvt_rn_satfinite_relu_e5m2x2_bf16x2` | LLVM `llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite` from `int_nvvm_bf16x2_to_e5m2x2_rn_relu_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 9.1 on sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.relu.e5m2x2.bf16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2` | `nvvm.cvt_rn_satfinite_relu_e5m2x2_f16x2` | LLVM `llvm.nvvm.f16x2.to.e5m2x2.rn.relu` from `int_nvvm_f16x2_to_e5m2x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.relu.e5m2x2.f16x2 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f32` | `nvvm.cvt_rn_satfinite_relu_e5m2x2_f32` | LLVM `llvm.nvvm.ff.to.e5m2x2.rn.relu` from `int_nvvm_ff_to_e5m2x2_rn_relu` | safe; scope thread; pure; memory none; convergent false | PTX 8.1 on sm_89+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.relu.e5m2x2.f32 , , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | | `cuda_intrinsics::convert::cvt_rn_satfinite_tf32_f32` | `nvvm.scalar_conversion` | LLVM `llvm.nvvm.f2tf32.rn.satfinite` from `int_nvvm_f2tf32_rn_satfinite` | safe; scope thread; pure; memory none; convergent false | PTX 8.6 on sm_100+ ([PTX ISA 9.3, 9.7.9.22 Data Movement and Conversion Instructions: cvt](https://docs.nvidia.com/cuda/parallel-thread-execution/#data-movement-and-conversion-instructions-cvt)) | validated on `rust-llvm-23.1.3-aab11ee0`; expects `cvt.rn.satfinite.tf32.f32 , ;` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) | @@ -1796,6 +1800,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0371` - public source path: `cuda_intrinsics::convert::cvt_rn_relu_tf32_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_relu_tf32_f32` +- `cvt_rn_satfinite_e4m3x2_bf16x2` (`i1030`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1030` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e4m3x2_bf16x2` - `cvt_rn_satfinite_e4m3x2_f16x2` (`i0822`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0822` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f16x2` @@ -1804,6 +1812,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0259` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e4m3x2_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e4m3x2_f32` +- `cvt_rn_satfinite_e5m2x2_bf16x2` (`i1032`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1032` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e5m2x2_bf16x2` - `cvt_rn_satfinite_e5m2x2_f16x2` (`i0824`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0824` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f16x2` @@ -1812,6 +1824,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0261` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_e5m2x2_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_e5m2x2_f32` +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2` (`i1031`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1031` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_bf16x2` - `cvt_rn_satfinite_relu_e4m3x2_f16x2` (`i0823`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0823` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f16x2` @@ -1820,6 +1836,10 @@ This table is generated from the resolved catalog. Evidence stages below disting - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0260` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e4m3x2_f32` - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_relu_e4m3x2_f32` +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2` (`i1033`) + - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i1033` + - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` + - compatibility compiler path: `cuda_device::convert::cvt_rn_satfinite_relu_e5m2x2_bf16x2` - `cvt_rn_satfinite_relu_e5m2x2_f16x2` (`i0825`) - canonical compiler path: `cuda_intrinsics::__cuda_oxide_intrinsic_abi_v1::i0825` - public source path: `cuda_intrinsics::convert::cvt_rn_satfinite_relu_e5m2x2_f16x2` @@ -5527,12 +5547,16 @@ This table is generated from the resolved catalog. Evidence stages below disting - `cvt_rn_relu_f16x2_e4m3x2` converts two `f32` inputs to packed `f16x2` using nearest-even rounding with ReLU. It lowers to pure `cvt.rn.relu.f16x2.e4m3x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_relu_f16x2_e5m2x2` converts two `f32` inputs to packed `f16x2` using nearest-even rounding with ReLU. It lowers to pure `cvt.rn.relu.f16x2.e5m2x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_relu_f16x2_f32` converts two `f32` inputs to packed `f16x2` using nearest-even rounding with ReLU. It lowers to pure `cvt.rn.relu.f16x2.f32` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. +- `cvt_rn_satfinite_e4m3x2_bf16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e4m3x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e4m3x2_f16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e4m3x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e4m3x2_f32` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e4m3x2.rn` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.e4m3x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. +- `cvt_rn_satfinite_e5m2x2_bf16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e5m2x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e5m2x2_f16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation. It lowers to pure `cvt.rn.satfinite.e5m2x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_e5m2x2_f32` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e5m2x2.rn` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.e5m2x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e4m3x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e4m3x2_f16x2` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e4m3x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e4m3x2_f32` converts two `f32` inputs to packed `e4m3x2` using nearest-even rounding with finite saturation and ReLU. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e4m3x2.rn.relu` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.relu.e4m3x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e5m2x2.bf16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e5m2x2_f16x2` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation and ReLU. It lowers to pure `cvt.rn.satfinite.relu.e5m2x2.f16x2` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. - `cvt_rn_satfinite_relu_e5m2x2_f32` converts two `f32` inputs to packed `e5m2x2` using nearest-even rounding with finite saturation and ReLU. LLVM-NVPTX uses typed `llvm.nvvm.ff.to.e5m2x2.rn.relu` with `[high, low]` inputs; libNVVM uses pure `cvt.rn.satfinite.relu.e5m2x2.f32` inline PTX. The public first input becomes the low lane and the second becomes the high lane. - `cvt_rz_bf16x2_f32` converts two `f32` inputs to packed `bf16x2` using toward-zero rounding without saturation. It lowers to pure `cvt.rz.bf16x2.f32` inline PTX. The first input becomes the low lane and the second becomes the high lane, so PTX prints the inputs in reverse order. @@ -5905,12 +5929,16 @@ This table is generated from the resolved catalog. Evidence stages below disting - `cvt_rn_relu_f16x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_relu_satfinite_tf32_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records `CVT_to_tf32_rn_relu_satf`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 100` - `cvt_rn_relu_tf32_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records `CVT_to_tf32_rn_relu`; selection predicates `Subtarget->getPTXVersion() >= 78`, `Subtarget->getSmVersion() >= 90` +- `cvt_rn_satfinite_e4m3x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e4m3x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e4m3x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none +- `cvt_rn_satfinite_e5m2x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e5m2x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_e5m2x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e4m3x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e4m3x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e5m2x2_f16x2`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_relu_e5m2x2_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records ; selection predicates none - `cvt_rn_satfinite_tf32_f32`: properties `IntrNoCreateUndefOrPoison`, `IntrNoMem`, `IntrSpeculatable`; result `NoUndef` false; half-open result range `none`; selection records `CVT_to_tf32_rn_satf`; selection predicates `Subtarget->getPTXVersion() >= 86`, `Subtarget->getSmVersion() >= 100` @@ -8126,6 +8154,17 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_90, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the fresh 1,203-byte NVVM IR fixture (SHA-256 4b5b61cbde895ad4cca3777fb86c33c06e96d35b70d9fc589d3e0126201dd222) to 956-byte PTX 9.3 (SHA-256 aca3fe16c949e7614f988c423f06912fd05a431babafe2def5af07b03abaaea9) and 2,040-byte LTOIR (SHA-256 d0affd25b909632cee088d988a7778bde56d065edbd5c1324c3e1a8cf957bb23), retaining all four exact instructions and =r,f constraints. - PTX assembly on `sm_90, ptx78`: `succeeded` — CUDA 13.3 ptxas accepted the 469-byte exact sm_90/PTX 7.8 fixture (SHA-256 7455770182c5f46203de2170777b7bce100e445a41562bbc34945256325fa93a) and produced a 3,704-byte cubin (SHA-256 33f8941a912c4580054123f3ce55e599143c13cdfa161f0b3de1c55158a1780f). It also assembled the libNVVM-emitted PTX to a 3,704-byte cubin (SHA-256 1302e6a89ef1629ad728d8459ca785a8dc0a8901231869bbb56c0435e78922bb). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.33; build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_90, ptx93`: `succeeded` — CUDA 13.3 nvJitLink 13.3 consumed the fresh LTOIR and produced a 3,704-byte sm_90 cubin (SHA-256 7758fa36cb80e9d0f3bd35c3084f710b4dd47870ffb3e9c4f4f1f378a2beefab). Tool `/usr/local/cuda-13.3/targets/x86_64-linux/lib/libnvJitLink.so.13.3.33` reports `nvJitLinkVersion 13.3; CUDA 13.3 V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_e4m3x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 189-byte declaration-only fixture (SHA-256 eed732da3bb65f7612f2146d1c60ae15fbace894a167c2fde5a4960cb450c262) to a 447-byte module (SHA-256 037d22467fb57cf069c0ab49b3d503e190446292c119acda734bd80445e2cedf) whose declaration of @llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 367-byte typed fixture (SHA-256 6c31c3240a710f61269c924f0f9efe46537f417c3ce954fdc99402ad8336595f), one call of @llvm.nvvm.bf16x2.to.e4m3x2.rn.satfinite on a dynamic <2 x bfloat> argument, to 745-byte PTX (SHA-256 e012440161e64091eec89c929a50d9d7567fccf63f87f5d17822a8b3447112d6; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e4m3x2.bf16x2 with register operands. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.E4M3.BF16 instruction. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 210-byte inline-PTX fixture (SHA-256 9f11d88381826f98151d13a0f04189bab6f6870223472c1a55c2980303df5dbf; the body of intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll without its comment header) to 749-byte PTX (SHA-256 67affa4ebac67d992e18927e0c99d3f12a131aba71bb7eec020ba0b0cc1d3e43; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e4m3x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73`, SHA-256 `e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77`) + - backend codegen on `compute_120f, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4). + - PTX assembly on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.E4M3.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - device link on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. - `cvt_rn_satfinite_e4m3x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 777bd7a5362a935cb9b45c99eb5c7912d8296fd322f232be5feb4dcb71afe958. @@ -8148,6 +8187,17 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_89, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order. - PTX assembly on `sm_89, ptx93`: `succeeded` — CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_89, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_e5m2x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 189-byte declaration-only fixture (SHA-256 dcf348779f9dde49249381b61a3141b1d36f5665ecbb2bc53ebc0be7e430c075) to a 447-byte module (SHA-256 63eaec75fcb9115321790aa90787b46064e7d007952182040a761407edb5f60c) whose declaration of @llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 367-byte typed fixture (SHA-256 7b4acbf67e3c0f83b1099150d70876603349db6db1b1502dce79da9b702b04f6), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite on a dynamic <2 x bfloat> argument, to 745-byte PTX (SHA-256 3334d10e01ea5e8f531e764a85e85f87cc2d0bc378f9bb38f9ef8a934ac54614; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with register operands. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.E5M2.BF16 instruction. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 210-byte inline-PTX fixture (SHA-256 fff4ab63076b1daff9772ebd8744e62bebb0f71d8c9b573111157b2125c4e96b; the body of intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll without its comment header) to 749-byte PTX (SHA-256 bdfc6e8b024796f9b5081893885b7d3a34fb36bc97e07fad8481bab56b9eab79; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73`, SHA-256 `e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77`) + - backend codegen on `compute_120f, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4). + - PTX assembly on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - device link on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. - `cvt_rn_satfinite_e5m2x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 5b6fcb8ea52dc533221fc3a953835ab9fb2cdb8a3c82a1dc9527dc2b980cb096. @@ -8170,6 +8220,17 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_89, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order. - PTX assembly on `sm_89, ptx93`: `succeeded` — CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_89, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_relu_e4m3x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 194-byte declaration-only fixture (SHA-256 318784297cbf5626850918899261c40db020b1cf755939c774dbdfcc0f4e67f2) to a 462-byte module (SHA-256 5237fee14f66ec4f52406118999e5656f0624e15918febf001630cd8c142f522) whose declaration of @llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 382-byte typed fixture (SHA-256 f374f8c786899f7e794a325136a9b411add44e6ca053a0d5a659ce2b62fcc013), one call of @llvm.nvvm.bf16x2.to.e4m3x2.rn.relu.satfinite on a dynamic <2 x bfloat> argument, to 780-byte PTX (SHA-256 ed6663e57024ca9435b851169d11a9a4fee64eb7f94899ef02f20d3d886839dc; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e4m3x2.bf16x2 with register operands. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.RELU.E4M3.BF16 instruction. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 220-byte inline-PTX fixture (SHA-256 dd4c8e4f33b8463bc6adf33bf59fe47cb6054daea21b7cf3a84dd634fb0f201f; the body of intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll without its comment header) to 784-byte PTX (SHA-256 3cfe93699632a0d050f21ba46f5e94dba53fa7b9f2ffbb369c1bc06190b7ce80; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e4m3x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73`, SHA-256 `e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77`) + - backend codegen on `compute_120f, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4). + - PTX assembly on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.RELU.E4M3.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - device link on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. - `cvt_rn_satfinite_relu_e4m3x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 a17f8309a9b1436ceae14d34278cbc25993c1607850ec87962d7fef9fd6c3215. @@ -8192,6 +8253,17 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `compute_89, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,184-byte NVVM IR fixture (SHA-256 1f48f717966e2249dc5f2d82492c9fd945cc7bf1a87becef38df48881995fd88) to 1,068-byte PTX (SHA-256 371b97f7ec8b94fbe36b4406dc227fe62e4df8dbba8541b512b5e7c998e13323) and 2,080-byte LTOIR (SHA-256 f29c9bdfd76cb78bdd90783f31c97d9a984379f171df011ed97926faaebc5596). All four forms retained high, low PTX source order. - PTX assembly on `sm_89, ptx93`: `succeeded` — CUDA 13.3 ptxas accepted all four forms at sm_89 and produced a 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). Tool `/usr/local/cuda-13.3/bin/ptxas` reports `CUDA 13.3 V13.3.33 build cuda_13.3.r13.3/compiler.37862127_0` (SHA-256 `7fdd01a4cf50e30746da98989c9272a907f491e6fd7fecfda14642e4375f88fb`). Artifact `cubin`. - device link on `sm_89, ptx93`: `succeeded` — CUDA 13.3 nvJitLink produced the 3,368-byte cubin (SHA-256 e52e9372da7213161b9c6987cb718ca9c508e325c91a6f76538eff35fda911e7). The same route at sm_87 failed terminal compilation with the sm_89 requirement for every form. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `V13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. +- `cvt_rn_satfinite_relu_e5m2x2_bf16x2`: runtime `not recorded` + - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) + - declaration canonicalization on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llvm-as and llvm-dis canonicalized the 194-byte declaration-only fixture (SHA-256 12c7df1f1031c28863d8993452d940b6ed6affc024534d32de66272e70490048) to a 462-byte module (SHA-256 5cfbd96861014107c50c79030418eb8400ee1e6ec25dfc0d6679739e8997de5f) whose declaration of @llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite over <2 x bfloat> with an i16 result carries speculatable and memory(none), the attributes the imported IntrSpeculatable and IntrNoMem properties require. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 382-byte typed fixture (SHA-256 531ef4ab3bc1e551c3a2b5fc5d646a67f522a454b64cb47bcd0ac7617c73a1d3), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite on a dynamic <2 x bfloat> argument, to 780-byte PTX (SHA-256 460efd987919d91caf04ca68e9223b4d22c4d37872164593799122bd0ea8d764; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with register operands. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.RELU.E5M2.BF16 instruction. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 220-byte inline-PTX fixture (SHA-256 052e0049faabfaae45bfd2f5e74607a3a39c2f40cd108c6656a00104c933e207; the body of intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll without its comment header) to 784-byte PTX (SHA-256 8946110f2d12fa0239e4132675f8f6facd71e80dfbba7e6cf40d010cdba91e96; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73`, SHA-256 `e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77`) + - backend codegen on `compute_120f, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4). + - PTX assembly on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.RELU.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - device link on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 nvJitLink linked the sm_120f LTOIR with -lto into a 13,896-byte cubin (SHA-256 d11dba66e56c2124a3ce2b8f9ef6c5c54a8930a9ac20ec32735b64e3355d0b71) that keeps all four kernels and their F2FP instructions. Linking the same LTOIR for sm_120 fails with the ptxas feature error. Tool `/usr/local/cuda-13.3/lib64/libnvJitLink.so.13.3.33` reports `CUDA 13.3 nvJitLink 13.3.33` (SHA-256 `3ba1e744347cd68617b862eccfd98b125482e882b7a6319f42abc9a768513db8`). Artifact `cubin`. - `cvt_rn_satfinite_relu_e5m2x2_f16x2`: runtime `not recorded` - `llvm_nvptx` uses `inline_ptx` from profile `rust-llvm-23.1.3-aab11ee0` at PTX 8.1 / sm_89+: status `validated` (`LLVM version 23.1.3-rust-1.101.0-nightly`, SHA-256 `7534b0236b9b26c7e6910548038b2285ee1bb5908e4700764dc78a7a1a54b449`) - declaration canonicalization on `sm_89, ptx81`: `succeeded` — Revalidated 1 separate probes with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): canonical typed declarations and their exact imported attributes passed. The sorted ID/artifact-kind/SHA-256 ledger in refresh-2026-10-11/stages.jsonl has SHA-256 2a8ce890aff0fa2df126aba76cf0dcc3fbda114de759ad215043e6ad0d400d23. diff --git a/cuda-oxide/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml b/cuda-oxide/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml index bd52006d40..21c092f2a9 100644 --- a/cuda-oxide/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml +++ b/cuda-oxide/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml @@ -3,15 +3,17 @@ # # Reviewed bf16x2-to-packed-FP8 conversions. # -# Pack only. The source arrives already packed in one register. The imported -# reverse conversions carry a block scale and are not admitted here. +# Pack only. The source arrives already packed in one register. The pinned +# LLVM import has no unscaled unpack record for bf16x2 (its reverse records are +# the block-scaled `scale_n2_ue8m0` forms); the PTX ISA 9.2 unpack forms are a +# later admission. schema = 65 family = "packed_conversion" [packed_conversion_fp8_bf16x2] -llvm_evidence_profile = "rust-llvm-23.1.0-16696adc" -libnvvm_evidence_profile = "cuda-13.3-libnvvm-13.3.33" +llvm_evidence_profile = "rust-llvm-23.1.3-aab11ee0" +libnvvm_evidence_profile = "cuda-13.3-libnvvm-13.3.73" runtime_validation = "unexecuted" fp8_formats = ["e4m3x2", "e5m2x2"] relu_variants = true diff --git a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll index e4f90dbf4f..3e0f9d76c9 100644 --- a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/abs_f16x2.ll b/cuda-oxide/intrinsics/probes/abs_f16x2.ll index d282bddf00..b6bc6d7d37 100644 --- a/cuda-oxide/intrinsics/probes/abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/active_mask.ll b/cuda-oxide/intrinsics/probes/active_mask.ll index cfb28da0a0..6330e0855e 100644 --- a/cuda-oxide/intrinsics/probes/active_mask.ll +++ b/cuda-oxide/intrinsics/probes/active_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_bf16x2.ll b/cuda-oxide/intrinsics/probes/add_bf16x2.ll index 9e8b4b623c..03da325f33 100644 --- a/cuda-oxide/intrinsics/probes/add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f16x2.ll b/cuda-oxide/intrinsics/probes/add_f16x2.ll index 9c9be96eed..3a49e6fa50 100644 --- a/cuda-oxide/intrinsics/probes/add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f32x2.ll b/cuda-oxide/intrinsics/probes/add_f32x2.ll index 8bd18c88a6..ee6a2e9c23 100644 --- a/cuda-oxide/intrinsics/probes/add_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll index 15b70e709e..a4c7bb236d 100644 --- a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_f32.ll index 6db9ffcbc0..4ba9271e68 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f64.ll b/cuda-oxide/intrinsics/probes/add_rm_f64.ll index 969c43674e..35e9de9fc8 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll index aa286f0f62..65bfa0e99d 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll index b6b6565745..1b82320433 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll index a03d96f4a1..d9339dff5e 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_f32.ll index 1c618b8bff..1eb51f7336 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f64.ll b/cuda-oxide/intrinsics/probes/add_rn_f64.ll index cdc0104e4b..bf3668f6d4 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll index 5af9974339..a1012daaaa 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll index 12355cdf98..4341bc7d5f 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll index 31e4fc10a6..ab4add9da4 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_f32.ll index 2e8a575e27..a9c41dc1f2 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f64.ll b/cuda-oxide/intrinsics/probes/add_rp_f64.ll index 5e098e67bc..0c8ddd6454 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll index 6daf152ec6..3c28880f31 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll index 731fb3e96a..42e3cd380c 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll index 4922207ad2..66ef200a5d 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_f32.ll index 1d6bc06f27..493bdc68ad 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f64.ll b/cuda-oxide/intrinsics/probes/add_rz_f64.ll index 66bfa0676a..831859fc46 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll index de5dedd64b..6678cc4b68 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll index 9182470492..83aee309e7 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll index cceab23750..8a04fa233f 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/all_sync.ll b/cuda-oxide/intrinsics/probes/all_sync.ll index cb66ad10f4..98304e37e4 100644 --- a/cuda-oxide/intrinsics/probes/all_sync.ll +++ b/cuda-oxide/intrinsics/probes/all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/any_sync.ll b/cuda-oxide/intrinsics/probes/any_sync.ll index 97b18cd57a..6bc242fd8b 100644 --- a/cuda-oxide/intrinsics/probes/any_sync.ll +++ b/cuda-oxide/intrinsics/probes/any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ballot_sync.ll b/cuda-oxide/intrinsics/probes/ballot_sync.ll index 96081557da..341c46747a 100644 --- a/cuda-oxide/intrinsics/probes/ballot_sync.ll +++ b/cuda-oxide/intrinsics/probes/ballot_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll index 304516396b..218967e072 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll index cfc169549d..6eaa0cb693 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll index c03df80219..764c755711 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll index 08e8e7e515..b553621ab5 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll index 4c91be89ba..28a4402acd 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll index ad7e518834..db275c5ef1 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll index ccaed8aec9..1e2aa12703 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll index 97fb497976..975fd5eb7a 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll index 24e0cf60ac..d58f366099 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll index b1d298dd80..2788a846e7 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_x.ll b/cuda-oxide/intrinsics/probes/block_dim_x.ll index ad845532eb..0f78ffd65b 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_y.ll b/cuda-oxide/intrinsics/probes/block_dim_y.ll index 3d78afddb1..ec11f16ef9 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_z.ll b/cuda-oxide/intrinsics/probes/block_dim_z.ll index 7048c46326..6a6d48d592 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_x.ll b/cuda-oxide/intrinsics/probes/block_idx_x.ll index 2066c67b2a..982679de95 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_y.ll b/cuda-oxide/intrinsics/probes/block_idx_y.ll index 56b24a2ca6..208459e386 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_z.ll b/cuda-oxide/intrinsics/probes/block_idx_z.ll index 69ec9340ab..e26a0c5125 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/breakpoint.ll b/cuda-oxide/intrinsics/probes/breakpoint.ll index fee6f3fa0c..8c1cde8219 100644 --- a/cuda-oxide/intrinsics/probes/breakpoint.ll +++ b/cuda-oxide/intrinsics/probes/breakpoint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll index 4920f565fa..cc415c4334 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll index d3dc26d79b..10cb0c67d4 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll index 418a1c73fa..7c1cc0c076 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll index aec4b27149..900b76251a 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll index a44a1cfa20..d4f441c139 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll index 38ffb0ce73..d4cad92630 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll index 56dfe3d6b4..363c8d3e46 100644 --- a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll index aed88470f6..ec08138839 100644 --- a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll index f51209ab7f..6f0dea61de 100644 --- a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll index 3148c69953..85f3b12ece 100644 --- a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_count.ll b/cuda-oxide/intrinsics/probes/cluster_block_count.ll index 39fcd3f775..1280cedf7c 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_count.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_count.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll index 9384df8463..e52fa68835 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll index 72edce8ff2..7e9b045703 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll index d3ee68d839..57b3773500 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll index b5e1ebc766..6faa8e0b0b 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll index dcb424b2de..63432523b6 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll index 08cb319243..c78ca2663c 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll index 0f89613939..66eae31678 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll index d3007ef56d..ff06904ee0 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll index be7a7a08aa..ed98bfce87 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll index 51b8879cae..db2904ec18 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll index b0f443649b..7cb52c49f2 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll index 69fc747df9..2ae587971f 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll index c72d22ab0d..f8d220a379 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll index 0e3115de83..bcb1d5751a 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll index 6e16ebc12f..0b7a304b04 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll index 405c6b104e..03a79696ce 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll index a1bfe4e31c..1ea8185454 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll index 72da0da001..61a0410a5c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll index f9247ae03f..0c79f8cd6b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll index a49fab07b3..3084413e82 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll index 734a7ae722..0cd4810dee 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll index a927796f63..42c683f82f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll index 51a44d497d..bb10981d07 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll index 5962e7a5a3..0fc1433183 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll index b4e0d8cfd5..d743f2e5cd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll index 42ba75f968..a5bda09637 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll index 0c97d6057b..08d3af97be 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll index dcf767ab90..7a30b43ffa 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll index 03600b606f..2cdf67cefa 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll index 899b146871..bf8cc52010 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll index 1f7da55135..4d1ea76468 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll index c990b39f80..0becd87d2d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll index b2266ee055..907e2987fb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll index 8b9b870c67..457b677f32 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll index 39f9357b2c..8ab9a6cfd9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll index 83dd948659..8e89b8d8aa 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll index f726d34eb9..52e47b4c86 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll index bf08d0169b..ccd207661a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll index 7515b63d4c..25d9b4418c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll index 8b3c1839f9..0662f26544 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll index 5c3dedff9e..e45c58974b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll index 96feaf9b06..a2059a313e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll index 045c4388c4..3b20c4a703 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll index 35062c7b02..9a2d168277 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll index a529f2b873..dd11181238 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll index d2977dfc28..07202036e0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll index e18e0ecee3..6cdfdde1da 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll index 9471b296a5..ea4feba388 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll index b7ea799c65..2f35825a32 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll index cd741e34cd..cf2a3a4d44 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll index 91d6f48365..980cd55253 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll index 07e0943450..19cb808a91 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll index ca6edd125f..995ad309ed 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll index 0aa4d4b10c..60d9f2f995 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll index 0b1a2fe822..fe7db8df8d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll index 3806624727..2de372324f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll index a922f25e39..63b7a7b372 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll index 0252d6f7e0..bb99f7e743 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll index 1b81e77aeb..7f27588763 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll index d5c2922c22..5eec5367d4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll index 1486b3d709..5609ec2e09 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll index 41dc2529ea..aad1f12fe6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll index ae03b1fceb..b6a94fb21e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll index 9619620a5c..2938f65e91 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll index 2dd47ef1bc..a43b169994 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll index b2793b839a..2691a872e7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll index f59d0ab3dd..d10df69058 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll index bb49fcd1a4..135cb9a328 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll index a6c45925a5..90ee7c4178 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll index c885e75016..de77107cbe 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll index 6a1a53cdaf..a08d407ae7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll index 9a84e9a095..744cbcf557 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll index d1ed727d67..8bd5f7dace 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll index db43ebc6d7..2226a2b728 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll index c65c97f089..d034017638 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll index 180c30a72c..ae38e12568 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll index 9e49134245..f42f506f71 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll index 6888b3f9fc..d90bebfdec 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll index 2727eb9958..4862fe3c4a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll index b26850ff5f..ec60a9befb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll index 4a303f2f91..b4f000f1a6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll index f8d523f2c7..44970f40b6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll index f1104484c8..83d3e67ed8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll index d25cd2aa46..3b6901ec98 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll index c2fdb514b6..f6381878cf 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll index a77f118b64..daf7bd8a6e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll index a171fba27f..d645b7a39c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll index 1538221ea6..e25c10a3f8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll index d3832a0185..0b9ec53151 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll index 9d7a79214c..ac63182dbe 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll index 2b8f449c46..cd067a2749 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll index ea00c76c5b..f75fec1b56 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll index b259dd25a2..b953fd88e0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll index d485f438b6..dcec639cf6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll index 2f33dd8e47..2b10caac7c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll index 6238356da1..618dac9cef 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll index a9cf6db63d..2add1f6c51 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll index d14f710cef..40eae07289 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll index 299203b666..cf4ee7c744 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll index 8d20bef888..83e1755b32 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll index e3ea390a0b..55fb9cbcea 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll index 7c785b4521..e526a68e88 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll index 5f0fadf127..6e7b8b213e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll index cfa609bfc4..4d4926f08f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll index d107ebada2..8a37eaa584 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll index 9a8167a5fa..5090032eba 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll index fd68791d7b..7be8135688 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll index 4d7856699e..78da69865f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll index e002dcbaf2..3ce4021f1f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll index c7dd69139f..39ddc777c4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll index 73ea5503d5..4065240a3a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll index 378d524a1d..c06f455a2a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll index 9ed6a52162..b5031711fe 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll index e285d27077..b993a12337 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll index 0f8eb28414..7cd619c5ba 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll index 0addb10c1a..bdd2499c59 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll index 61f4b8c68b..d605207ed3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll index 99de175278..856c94a505 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll index d49656c4e0..a8b5c36fa7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll index 26ce456bc6..4d65bdf6c1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll index 347c99ea5c..1b564ae5a1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll index e68eafcffa..b03d0ed753 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll index aa9ccb605e..42b38ac1f5 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll index 5feca63319..71f27aba98 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll index daab8d78d9..0bf5c6fd97 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll index 9e8688dea8..a6deb3a827 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll index 920123763b..0f55629b8c 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll index 6e4b6697e5..77f50cb1ab 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll index 2c100257bb..4398157574 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll index 72f2068054..dca550b55b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll index 798279c8b3..7d8f7a0af4 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll index 04556ff2e8..1cd2fd3ae3 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index d90d9e288a..5c67d7990a 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll index 2df0cc6f42..e1dc2fe2c7 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll index 9dc6f41d3a..2ad7cc2dcf 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index fc62e942db..b319dd2700 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll index 694cc51b18..448fb62cfc 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll index e6bf6bb677..99e467eb9d 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll index 1df9b892fc..d83ad5ca7c 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll index 45cdaeb7a7..456fc35fe0 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll index a882c1163f..dc98c0ee78 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll @@ -1,8 +1,8 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: bb8f2f358faeed56f5323e8766e7f4428f99eae248357e43ae4b648be743f3c2. -; LLVM source: 16696adcd119e6ba9cc175207d984d7021211acb. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll index b006511c9a..7c1bc741ff 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll index cac2a150ff..8c0dc3c13d 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll index fa711fd95f..687216d915 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll index a1ddc82b97..66302e842b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll index 764d2a6e71..f95175648a 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll index a5c282a4c9..295360c483 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll index 6a946aed71..36a1f8d8e2 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll index 000b7054c0..ed8b4e3ff6 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll index f7a12ec46c..6abe89727c 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll index 733dda6360..43a4f2ea37 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll index 5b6c24128c..9f41aff4cd 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll index f8f39abb08..b5f03b3ac4 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_f32.ll index e8f91ad578..474edc3d24 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f64.ll b/cuda-oxide/intrinsics/probes/div_rm_f64.ll index 09f00888e4..886d44e0d9 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll index 0ea9c58769..dbc9679295 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_f32.ll index 228cc0aa87..c4dc8cdd0f 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f64.ll b/cuda-oxide/intrinsics/probes/div_rn_f64.ll index bd90372579..cd82155f61 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll index 985d40669e..2bfe028017 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_f32.ll index 264c51acac..ca0ea87891 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f64.ll b/cuda-oxide/intrinsics/probes/div_rp_f64.ll index 2a63516b60..06dffdb5ac 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll index 2afa9efd1d..a349395bb4 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_f32.ll index e0e3aa902a..aecba7e73b 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f64.ll b/cuda-oxide/intrinsics/probes/div_rz_f64.ll index a0efcae75c..734b475a56 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll index 656c119c87..673fc0cbac 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_s32.ll b/cuda-oxide/intrinsics/probes/dp2a_s32.ll index c876a0c3b8..e704dc8408 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_u32.ll b/cuda-oxide/intrinsics/probes/dp2a_u32.ll index f7c3e2f850..568c6d6b21 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_s32.ll b/cuda-oxide/intrinsics/probes/dp4a_s32.ll index ac0329ea6f..84347a5193 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_u32.ll b/cuda-oxide/intrinsics/probes/dp4a_u32.ll index 9a6bbb84c9..d05020f0d8 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll index 693c323247..037936b01e 100644 --- a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll +++ b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll index 89000848ba..293040fac1 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll index b0ae0bf16e..4336a92fdf 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll index 3ceb7bacc1..3fb815716b 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll index 56c243c066..0b45c6e502 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll index 2084941712..3f15df0ab5 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll index c26aa21244..070ea46978 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll index 57c4d39162..6a1d97147d 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll index af884de851..f7305f533e 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll index dc1d7a4a5b..dc024a3991 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll index b0d7997a38..b6cb344b3e 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll index 7306e10b93..17aa0c8c6f 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll index ad98279bac..b89437f96b 100644 --- a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll index 54f2db264f..f3552319b9 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll index fd539cbdcd..7e21a11133 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll index 07cdb52a2a..b50b50d5f7 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll index 1af7bc56ea..35f3223fed 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll index a1e3f909a1..037df6fb76 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll index b2b7da6b0a..597a4417ee 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll index d02c4efd94..2a74b5fa26 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll index d63f549758..1534d50108 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll index dbdb54654f..2b9342a8dc 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll index c63b505f08..b598bbcdf6 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll index a48d2cb019..760b62b2b2 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll index 0227a87cf5..6025afa0f1 100644 --- a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_f16x2.ll index 3006ca5867..2149d5cd1a 100644 --- a/cuda-oxide/intrinsics/probes/fma_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_f32x2.ll index 4178195aff..00bdd486e5 100644 --- a/cuda-oxide/intrinsics/probes/fma_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll index 1dab638241..7d6f50de10 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll index 88f236546c..6f8be8c53d 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll index e6ce5730d2..7f9dfaaed1 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll index 69c4856a3f..a11ccb6828 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll index 1f06311c75..d1116c437c 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll index 85fc23c5e9..da30780743 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll index bd78e7fec2..e0568e391a 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll index 172f5a198f..4500a275eb 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll index 15c7ff5a84..4a17faa04c 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll index d5d90f504e..d76442ec15 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll index f6a0afb978..5c1e8a1ade 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll index fe95c6605a..e67634f9fe 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll index aa8de7458c..4bb416ef7b 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll index 7f8f6bcdb5..90ba2621bb 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll index 7675704ac7..f907ce5d43 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll index abd1aa265f..ab20d0dd1f 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll index c2c322df2a..f82c12243c 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll index 64bc22b5ab..d27695eaf1 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll index e938739361..944456bf65 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll index 39143c43a8..a864a7b10d 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll index 795d4c65bd..864ca38d29 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll index 4d0f422f23..4df946f63c 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll index 22b01235a1..2f0df8f8b0 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll index ed8b184bc3..8c983aba40 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll index 03b8c41659..9ea06ed6c5 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll index ef0b109944..4675ce8267 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll index bb61385a0f..6112290dbf 100644 --- a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll index 341ecde82b..581935764f 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll index c562b17969..18aeb16f78 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll index b90c665711..f3fd9aeeb2 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll index 2cbb8fa061..d2d0aa8f62 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_x.ll b/cuda-oxide/intrinsics/probes/grid_dim_x.ll index d46c40c165..865ff9cd4b 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_y.ll b/cuda-oxide/intrinsics/probes/grid_dim_y.ll index 00341c087b..488186cafa 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_z.ll b/cuda-oxide/intrinsics/probes/grid_dim_z.ll index 7cfe56e7bc..5ea78d17fe 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll index a2b3c2796c..c5281b45b8 100644 --- a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll index 4bf2bb73f3..e8b3d28d87 100644 --- a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lane_id.ll b/cuda-oxide/intrinsics/probes/lane_id.ll index 3810a36af7..7ca38e501c 100644 --- a/cuda-oxide/intrinsics/probes/lane_id.ll +++ b/cuda-oxide/intrinsics/probes/lane_id.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_eq.ll b/cuda-oxide/intrinsics/probes/lanemask_eq.ll index 20799aed61..c079280e62 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_eq.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_eq.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_ge.ll b/cuda-oxide/intrinsics/probes/lanemask_ge.ll index f29b87e61f..8c86ea7505 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_ge.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_ge.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_gt.ll b/cuda-oxide/intrinsics/probes/lanemask_gt.ll index 53a7542517..4348e071ef 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_gt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_gt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_le.ll b/cuda-oxide/intrinsics/probes/lanemask_le.ll index 924858d48a..7c8dead4f4 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_le.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_le.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_lt.ll b/cuda-oxide/intrinsics/probes/lanemask_lt.ll index 189d4d8182..6d0410a13c 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_lt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_lt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll index 2ba6ad2898..a966592ddd 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll index 6056d52646..6452915221 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll index 7eeffa64bd..efee94aacc 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll index c64347b04d..cf013da91f 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll index 0333251475..9d3d8816ed 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll index 5f2ccf3356..fff4692803 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll index 1819f1a48a..3bed716b2e 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll index fb06db92da..444f53c970 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll index d610e0ceba..4962be0d13 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll index 4ddddf4817..bfd146333a 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll index e6ea8ec6e2..d5461078d0 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll index 216a433048..782ecc56e6 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll index 3f3b32a4b7..3d800e89c8 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll index 4c4f562852..f34f69f1f8 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll index 5437865344..6390afffbd 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll index 38c1e31d97..40d511cb90 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll index f823d1f56d..4d88abca4f 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll index 2c7a254706..b8de72f3cc 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll index d20f9b8100..ba4a7e687a 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll index 3150cdcf51..332723cab1 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/map_shared_rank.ll b/cuda-oxide/intrinsics/probes/map_shared_rank.ll index 17ee6027fd..5ff91fc381 100644 --- a/cuda-oxide/intrinsics/probes/map_shared_rank.ll +++ b/cuda-oxide/intrinsics/probes/map_shared_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll index c0c2733487..e29e0e32b9 100644 --- a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_sync.ll b/cuda-oxide/intrinsics/probes/match_all_sync.ll index b076a21e01..33d79900ef 100644 --- a/cuda-oxide/intrinsics/probes/match_all_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll index cfc9acc3e1..94a6d8fdc5 100644 --- a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_sync.ll b/cuda-oxide/intrinsics/probes/match_any_sync.ll index e4e65e322a..3f0628acae 100644 --- a/cuda-oxide/intrinsics/probes/match_any_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16.ll b/cuda-oxide/intrinsics/probes/max_bf16.ll index 77be96d793..3902fc5218 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_bf16x2.ll index 7150683ad3..0b299d1173 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16.ll b/cuda-oxide/intrinsics/probes/max_f16.ll index d12d32818b..d834592c34 100644 --- a/cuda-oxide/intrinsics/probes/max_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16x2.ll b/cuda-oxide/intrinsics/probes/max_f16x2.ll index eb099b1f93..c2274faa4b 100644 --- a/cuda-oxide/intrinsics/probes/max_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll index 9ce110e2e6..a92b855da7 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll index 99431d33f4..71f41b9ca3 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll index 5b041b2ded..5d1f72142a 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll index 3fce4e7b67..29533efb19 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll index 5ccb70e93a..e0ddc9a286 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll index d1b271ce91..031961e344 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll index 0663c3c250..bba0a433f5 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll index 6c218589e8..8eb5f6b0e3 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll index 289a6f955e..9af8f76aa8 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll index 3fa48796ac..64f6a14d32 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll index 9239b9fe52..aad24fb5fe 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll index 0da393f4ff..89b1fa92d8 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_f16.ll index f91129746f..8fa660f5b9 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll index 4656697104..9d02d437b3 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll index 974357c43b..c70f89c623 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll index 437daaa6bb..c8d09d3fec 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll index 5d73b53155..d1913555c7 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll index ab4af618e8..5081b9a271 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll index 4f558e905c..379e69a4c2 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll index bd8265fe52..2f340fcf5f 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s32.ll b/cuda-oxide/intrinsics/probes/max_relu_s32.ll index 5e8e6ed0d7..ab7f9fcc30 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_s16x2.ll b/cuda-oxide/intrinsics/probes/max_s16x2.ll index db5370ba75..c51a9756c3 100644 --- a/cuda-oxide/intrinsics/probes/max_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_u16x2.ll b/cuda-oxide/intrinsics/probes/max_u16x2.ll index 6a2e6c31c9..792375a35b 100644 --- a/cuda-oxide/intrinsics/probes/max_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll index 03e42f432b..f08e2f62bf 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll index a73ef96512..f25d1ad4a1 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll index f053059555..5198f93023 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll index 04aefd53a3..eb75cd56f6 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll index ca09146c29..4c67a39488 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll index 1b0d581793..e86225e1ab 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll index aff83d50be..a4fb417a93 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll index 4e35deeb04..0ed5da251a 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll index 65fe02b980..cd69e0d7f0 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll index 3cf7ee744f..7faa0a8371 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_init.ll b/cuda-oxide/intrinsics/probes/mbarrier_init.ll index 910ac349d8..ed4203accc 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_init.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_init.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll index 2c7fae458f..6fe7acc786 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll index 0e21dc951d..1f989d89c7 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll index df497af274..eaa1ef5e9e 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll index 8374649481..bd7930530f 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll index 128985b3ea..a149046cda 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16.ll b/cuda-oxide/intrinsics/probes/min_bf16.ll index 3bbd2641ca..64de583ad3 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_bf16x2.ll index f57634c72f..84dcfb53b1 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16.ll b/cuda-oxide/intrinsics/probes/min_f16.ll index a944cfb4dc..263d99e167 100644 --- a/cuda-oxide/intrinsics/probes/min_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16x2.ll b/cuda-oxide/intrinsics/probes/min_f16x2.ll index f295381bb6..38578710d3 100644 --- a/cuda-oxide/intrinsics/probes/min_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll index 3d9104e413..3966303fd9 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll index b06fa76805..b3e0a6a239 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll index 00d289910f..de697ae966 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll index b583273e3c..f655342bb0 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll index aa8217ccf9..d33e190b68 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll index 53ceb1f1a5..dfbde7e06a 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll index 037e2349a9..23e434141d 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll index 153ba0e6d6..61a7d2b802 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll index d2e8fa9463..bb8963794c 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll index 27fd6d5ccc..e02521f2b0 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll index c533c9544b..eb71a3acf6 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll index 20513a48e4..66216bebe1 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_f16.ll index 1a5d56cda6..919a04c666 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll index 16d01c1151..819fa57f95 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll index 1607ff0343..3b702ff059 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll index e65f385040..6d907f674c 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll index 35b19c4fd1..a6bc124bcd 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll index f87889e445..057dca1b5a 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll index 653e8bf18d..21ce1da164 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll index c08d69fe08..6110af7a5a 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s32.ll b/cuda-oxide/intrinsics/probes/min_relu_s32.ll index ab32ac384a..59428ad2e3 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_s16x2.ll b/cuda-oxide/intrinsics/probes/min_s16x2.ll index 2f9c14a0d2..d2cd401912 100644 --- a/cuda-oxide/intrinsics/probes/min_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_u16x2.ll b/cuda-oxide/intrinsics/probes/min_u16x2.ll index 211614bc5d..b19f7b2ccf 100644 --- a/cuda-oxide/intrinsics/probes/min_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll index c4e6e964e5..c01c2de600 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll index 3b46362412..7ae84944ae 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll index 9e9415875f..5273270966 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll index 0c5a76f1d3..95a2b995b7 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll index 9486508752..725b8a6723 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll index 1ea6db4179..cd44e16db7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll index f343c80ca5..321fe4e71f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll index f147489e24..83c3d4a671 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll index 7776405699..2b8acf90b1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll index 8e3187734e..db2e707f95 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll index 2f60560537..bdee8cdff1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll index 0107ddc2be..37a320ecfd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll index 348ca5a36b..882209aaef 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll index 044208ada0..29ad808028 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll index da44fee12a..55bd6f8fde 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll index cb86d24427..8da77c0ac3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll index 10859f722d..1879df218b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll index 0ba1467481..0d12c38d3f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll index 92879daa3a..88e61aa6e8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll index 07c76884b2..88d9d5d053 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll index 3fa61efac6..6f2d9b88a1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll index a5032ec5cb..3d4b929689 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll index 70397939ac..3484d91f64 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll index 39b5948174..01d5e4c74d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll index f6f7f63e30..060c64aef6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll index 3038cd2516..c461a1ff52 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll index e7ed2d289f..8892fa8f40 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll index 7db1a4bb35..70248e4b97 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll index 4f85a99466..53c67c0ce2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll index d2104e7fb4..857a6741b5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll index 8ca6cb55a3..e5cb84fc94 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll index db9953f57d..38a9855e75 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll index 87a760c4ff..6d912dd0a6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll index dfd197e064..cf39b34121 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll index 735b58026a..d5fb3c38d4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll index eaa10620d8..f8a60e545f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll index e44cce86e8..b6c5149673 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll index 8fbc3419a0..0fd90e8afc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll index e3f1714a86..f7c8eb2eb2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll index 76c3c32edf..9ea3c9b99e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll index 3c8be80062..6698d76523 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll index eee0fafb8f..e81951f3fa 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll index fd52e2f349..88a4bc5098 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll index 3baa32de0e..edd82746f7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll index add7cc4475..a7bade8a20 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll index 9e7c95d969..ef0992657f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll index 4b7122373b..257f5d75de 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll index c4ec2080fb..11da7bab91 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll index 630f56bf62..ce5085c5da 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll index fecfeab6b8..7e21193bd8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll index 1a0bc3f45b..a25414bcf5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll index bc46a827c3..de1bfe8689 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll index 951a05b0f6..a28af820d1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll index 96c95b15a5..943ee87347 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll index b0b2bcc4a1..cb60b19aea 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll index 52db2b582e..6b3c0307f4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll index 84c6daf029..6d03a3faad 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll index 022877ae35..6b9ecd5328 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll index 96c70b0d9e..39fe857548 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll index 97fe74ef02..9a1730295a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll index 69f1fabc6e..f4c7742cac 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll index de067d5495..1ad93da507 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll index ed9138286c..836b0a6a96 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll index 1c775c4e24..8bf14f3dce 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll index 4cc5ed515c..9ca367101b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll index 6f432aa964..ce2deba6df 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll index 5141ec0786..a777a2e24e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll index ffccc1412f..5d8e72661d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll index b4d7276108..062693d1ee 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll index 1221b46de6..949b27c5a3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll index 779b69c37b..c312b6065e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll index b51a697a76..57eea13823 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll index 767763cb26..621c792582 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll index 92c8be40b4..d44cd85fba 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll index 980aa082e0..3066e64ec6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll index 2101d232ac..017546e43d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll index 46d5d57412..ce5d271387 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll index 4bec7816f6..d777544bab 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll index c6f7ee644a..b83080ba98 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll index 38c8822a13..9f90bff2f2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll index 823495e8ab..1b70f47887 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll index 00aa64998e..62961dfa47 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll index fd74bfbf88..2d5719d757 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll index 751d6de387..64dc030fc1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll index e62f38b979..69e81a6ed4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll index 9bb91ff4be..07d37001af 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll index 6ac77efe23..f2d70ed570 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll index a29c45e41e..73e2e53718 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll index 8b6865a36d..5329adf168 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll index f1a47b1431..74400564e3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll index c141a5ed75..3de7736738 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll index bcd74f3058..ab4b98f5df 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll index 15c81cafb8..003f8af757 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll index 1090808214..c710a5f90f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll index d8a60527d0..c7f1b14a06 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll index 9a10044a3c..4aa4bb2525 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll index c9839d36d0..e37c3f8b11 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll index 0ad35af407..48b22b5142 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll index b3433362fc..b596b116c4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll index d009f0a87d..123f3ef049 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll index 7f33262008..d33948ea9a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll index c3de14683a..b42e0288a3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll index db4ba8e839..e9acc9840d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll index e220edb8f2..39b9da7b68 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll index 7900e8e9e6..a4e9bff2dd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll index f879eaf258..15332bb717 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll index 7afaf17ef0..93e7483d20 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll index 6b40349352..be88ca87cd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll index 699d72600e..bec8310322 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll index df4a91ccd7..f9c1f8a864 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll index 7af2a65bb3..e32d67ed8b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll index 7b171cbe43..b5015a1f6c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll index 7eb4d7daae..bde945dc21 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll index c6017a7c15..5b2d2cd47b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll index 051cc6c569..bca8cce2d8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll index b7f1b1db9e..d23124c3f3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll index b206c7932c..e805e22ee0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll index ac519e39fc..f2d3d19276 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll index d4f89e99bb..57d83fb7ff 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll index 209a3d9989..43e1dad364 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll index a681419efb..56f57c8d76 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll index 5c96b8193e..d0c87ca0a7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll index cecdf1f30f..ce4b7d2b1f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll index 7c350cf31b..a213fa4792 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll index 0bebe9da87..ee50dd7eeb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll index 980f0166b2..7a696b7236 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll index 0efe91ff83..60079188cd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll index 6e70e7d1ec..99653bd66e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll index e9890ce30d..e5d20b70c4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll index 9ee23211c7..0ab16bfb5b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll index 14d8201f7e..874655323d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll index 19343fb6de..67885e09e8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll index 8826984e26..252e92d878 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll index 3344055106..b2754b553a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll index a33954d5ee..fa78ada27d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll index 934c12dff5..ecb10661a9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll index b021c14fac..cab04420f6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll index e64504fff5..f3de39a8dd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll index 7caff5c91c..378ce07f88 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll index d030101330..186acd72d4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll index 473c2e79de..258a89042c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll index 3a432cfb15..b19a973c44 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll index adddb22d13..b3bb4ab6f8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll index 4a8d18d3da..0b9ebb91c3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll index 054ea5e279..0327cd2fc6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll index 9bf0b3be33..1b39221322 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll index 50bc3245d2..e73cc8b843 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll index 6e4ffb7eb5..cb112e2795 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll index e2160e822c..f8b7971070 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll index dd72002d02..7cd09d56a7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll index b1d3623e4d..60cb52e873 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll index 2ad3a82c38..1a677bf974 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll index 530fd956cb..13733f50a6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll index 53b117b1c6..9f88dff37c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll index 7b524c3780..642dd662cc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll index 598c261e2d..8e0e9a3db5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll index b69f73456e..2cbb6cc467 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll index d3892576bf..41e7508b94 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll index 2e369de46c..7607113e87 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll index 96d692c1ac..a1c26a8fca 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll index 87cb08ca39..42376ee6bc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll index cfe144b4f5..51f7d154f1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll index 07e3b09849..5ca5923d23 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll index 54beec52dd..09c1db0801 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll index 3ae0c3ab34..066ae9dbff 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll index 33b8a3e468..e5613903ff 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll index 3d7fdf7538..efc85ed2d5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll index c476d84a31..ab4f4791c7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll index 7114cc098e..89dc2554e9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll index ee3ac32bed..a90372038d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll index 5cc5b3a4f4..ffaee874b5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll index 97dda1e812..fcccf992d5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll index f05b0a4760..c4d2243bb0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll index 00a0005305..1c81033278 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll index f60dcffd1b..ea26b6de06 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll index 7352440cd0..b926c032e0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll index 64747e411f..ed0d6aa920 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll index 99f94d3258..a2b89ad7c3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll index 5cebce211f..05060391c6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll index 1a2729aa84..7bdbc1e2ac 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll index 4ffcd5c1b4..8c2759d30f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll index 74262a9bd4..9f1fc9cdfb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll index c798abd261..2f913716c0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll index f685ca86ab..ea9a05d040 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll index 6d046a2fa9..d7b4387097 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll index b0be097754..7cd996192d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll index 1f9c965c2d..e47c43c803 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll index 4240c1d1fa..bab474279a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll index 3011a63159..e70f6f05aa 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll index 82307523e8..a4139c0d24 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll index 35d18671bf..8c43a439cd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll index d70ff12bf7..5e38f3cd20 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll index 83026e728b..9cd5a9b070 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll index 1598d12615..73fea51268 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll index 4332a7521b..90d22e821e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll index 684f3a2e70..4f57f735b0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll index 21bb6d0c6d..1a8f543452 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll index c54884e450..a5eff6c3d3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll index cb2ddfa7aa..62afa3840e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll index 01817551d1..8132406012 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll index 525df8de25..9d5a7e9c9a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll index b8fd3dde7e..4848f36e89 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll index 31643145aa..dcf768487b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll index 4c4533956b..cac229d40e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll index b284e897b7..b1eb3d7a8e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll index f0f1dbd4a4..de3bf01b9e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll index be5dde36bc..e6060062de 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll index 85056d3498..7c708cd6ea 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll index cbba60c4ba..caf27dcee3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll index 15c849298e..cb341c1467 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll index 8c42464754..5fffffc516 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll index b1050defb6..aa2795de6f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll index d88fe3e3eb..3e7bedbf61 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll index 0fd3ba60e1..cd5ca5ebc2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll index 4d1c1830ae..fb9ae969d0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll index b3dab246b3..eaebd92019 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll index c9e1a911e9..d9b9546b8a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll index 021358fd1d..9bc18a7cbf 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll index 70b7321527..6ad1dc3c3f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll index 4914fd926b..58cb90c339 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll index 6fd7141112..0c45109156 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll index 6e85ddd82a..4d2b8fee55 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll index 02a23a1169..8f6675c115 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll index 618904c371..e0ee31e501 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll index 2d9b3994b7..2523957c4b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll index 456c9a0483..a4a93b11b0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll index c5f468014c..fcd5fe7bf8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll index 9fc3d9b536..00b4d61a0e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll index 30466576ba..4e64370269 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll index e56812c604..987573f46d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll index 83c991c331..b6bdfd8728 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll index 5ba1ecb190..19ded5a79f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll index c3fbfa9f94..6125280516 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll index d32068231a..1372e1adbf 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll index da1a18a093..38cff699b9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll index 4e2a6b84ec..7301c82898 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll index 60d12b5bba..3ea077a7ce 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll index a0b26fc7e0..1062056ed7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll index 7f1f38bee5..f4efe5b3b9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll index e51fcbf0e9..904b059c60 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll index 78bffdbec0..43e9aa80db 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll index eb254e0845..f554b6747d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll index 961ec8b463..bf723fe739 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll index cccadb49d9..7374091317 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll index 13c28b8e69..f06123d3b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll index 87433ab492..76a4ebf952 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll index 377d250f54..12fc4370dd 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll index 4274cd36df..03505223e6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll index 2f26e87cc5..2d45ea4401 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll index 78a15d17ca..9ab5fdb0fa 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll index 3b5459eae9..b876e7d07b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll index 080082879d..7b51521b64 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll index 2973d38341..833b3febf7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll index 5fd43812e4..bbd176dc6e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll index eb3ff37c98..6ece6a57aa 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll index 25662838f7..2cc4765392 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll index e4554753f0..f20dfab5d1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll index 86de360fe1..0fd65bdbfc 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll index 017859b219..ec70736a73 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll index e1e2970290..2f6380b2b9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll index ede1412f6d..0f72f09644 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll index 6e33033b3a..2eebffc7c0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll index ed5694e626..0999acaffb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll index b034f065e4..fc30cfa11f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll index 3f4d201382..3cc2a08c15 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll index bca458a7c3..6562efc51a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll index 3a2cc89340..eb889176b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll index 2173d4db76..0ed96d019e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll index 4a9fdedf9a..bee68344b0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll index 20e4d53b8a..30edeb35af 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll index ddbe82e11f..06457dc22f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll index 6aff5c6f88..de428774f3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll index 0421f0ffc6..2142289885 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll index 1d59d7a529..ede089b1f7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll index 1a9cce88d0..062d0190e3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll index 17d4263041..007b24267a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll index 9c8a310db5..ea682776ed 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll index 491a194d31..86d9af6f87 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll index 1d1a5e1c84..c622a59f34 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll index d4e56c6b32..b1314db4ba 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll index 254cf0452f..ef29a7fa31 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll index a03a94023e..20f69bb2c0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll index f235abf757..75e7b3b24a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll index 0e68928c1e..e4807e796b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll index 10b9927d32..de86f78650 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll index 204e98b791..a985af3c9d 100644 --- a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll index 726b7ea310..ef377e5b3b 100644 --- a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f16x2.ll b/cuda-oxide/intrinsics/probes/mul_f16x2.ll index 092afdea4a..eb906f9921 100644 --- a/cuda-oxide/intrinsics/probes/mul_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_f32x2.ll index 7a4daaeeb6..37976ac13d 100644 --- a/cuda-oxide/intrinsics/probes/mul_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll index e0e2b089d4..4cd1a6bdd4 100644 --- a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll index 1e194d6780..49dfebde02 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll index 513893e676..99d1bfa070 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll index 86b0f26e72..54b5d782d4 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll index c936466501..0b62e43f28 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll index d6b9062a8a..105b9904a4 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll index e43e931133..43a7fedc6e 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll index 98b013e2c0..7da685d65a 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll index a0b25e06c2..86a45504f4 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll index b5932f240f..9c951a97e3 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll index 1edbbaaef0..d9e4596b3f 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll index a8f00b6bd3..1b567938e3 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll index 9fa2605615..45475415b0 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nanosleep.ll b/cuda-oxide/intrinsics/probes/nanosleep.ll index ef3c013ae5..276ece2ef1 100644 --- a/cuda-oxide/intrinsics/probes/nanosleep.ll +++ b/cuda-oxide/intrinsics/probes/nanosleep.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll index dc2600aad3..76258fba4b 100644 --- a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_f16x2.ll b/cuda-oxide/intrinsics/probes/neg_f16x2.ll index ee5840d2d9..7bdee19d56 100644 --- a/cuda-oxide/intrinsics/probes/neg_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll index 7d0749f632..f0e67d1b38 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll index 69cffffa8a..e0e366c6d5 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll index b9f83e5cb8..ded450886c 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll index ee1c99f101..4fbc24b344 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll index 1111cab0e8..546b11b015 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll index 0ec238e0da..f7b4e80421 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/pmevent.ll b/cuda-oxide/intrinsics/probes/pmevent.ll index ff9af628e0..2ac6c42605 100644 --- a/cuda-oxide/intrinsics/probes/pmevent.ll +++ b/cuda-oxide/intrinsics/probes/pmevent.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll index dc36ea55be..2bb9e1342b 100644 --- a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll +++ b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt.ll b/cuda-oxide/intrinsics/probes/prmt.ll index e7d238a19f..048b55bbf9 100644 --- a/cuda-oxide/intrinsics/probes/prmt.ll +++ b/cuda-oxide/intrinsics/probes/prmt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_b4e.ll b/cuda-oxide/intrinsics/probes/prmt_b4e.ll index a1218e54b3..c34ab4b4fb 100644 --- a/cuda-oxide/intrinsics/probes/prmt_b4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_b4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecl.ll b/cuda-oxide/intrinsics/probes/prmt_ecl.ll index bbb2e73e77..d38597f87e 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecl.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecl.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecr.ll b/cuda-oxide/intrinsics/probes/prmt_ecr.ll index 48b9949bd5..761d439b73 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecr.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecr.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_f4e.ll b/cuda-oxide/intrinsics/probes/prmt_f4e.ll index bbcd5d8d54..01ff20eb5d 100644 --- a/cuda-oxide/intrinsics/probes/prmt_f4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_f4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc16.ll b/cuda-oxide/intrinsics/probes/prmt_rc16.ll index aae7467217..edc156b77f 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc16.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc8.ll b/cuda-oxide/intrinsics/probes/prmt_rc8.ll index 8830088f2e..856d6ec0dd 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc8.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll index cb9e316c01..a0c99ecc71 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll index ad7ad7aaef..7131e8bbdb 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll index e16fed6685..53b4e257b6 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll index e52fc61d12..82aa47691c 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll index 818212973b..e5f566f96b 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll index e0357d9a35..98af3b5b1b 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll index f1cf96b1e4..3d842a7fc8 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll index d534303ff7..cfc02003fd 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll index d116a4b0fd..5d3b9c6221 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll index 4e6231b9b1..3de48205ab 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll index 91617991d7..fb52fb88c7 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll index 42337793ae..c7e348a788 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll index 211d1bb57f..9948810f8c 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll index ce89a31458..22699f6896 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_add.ll b/cuda-oxide/intrinsics/probes/redux_sync_add.ll index 0c2e753bb3..0668528f29 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_add.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_add.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_and.ll b/cuda-oxide/intrinsics/probes/redux_sync_and.ll index b8c214ab64..3cf523f6cc 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_and.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_and.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll index 71ed83b2fc..29dfca9b7b 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll index 26e5f081cb..4977262ec6 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll index fbc23a8b97..72d511d4a4 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll index 119729a5fb..15c4542186 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll index 602124e3a7..56c0c90208 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll index 26e1140870..5e55d82089 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll index e9770680f2..0802c93ad4 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll index 0ca86b7845..58dfa74d36 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll index f96e6ddf2f..006f1d87b1 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll index ffe5fd9168..06da2103fc 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll index ee0fc0d5a4..66cadfde8a 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll index 457e80576a..fdfafbdc28 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_or.ll b/cuda-oxide/intrinsics/probes/redux_sync_or.ll index ed161fd7e1..f387da6b4a 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_or.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_or.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll index c03c50cf02..88815c8349 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll index e4d7228376..492c43498f 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll index 4f48c804f1..f6c66601bf 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll index 828ad30b00..b5caa515c0 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll index d2cba3359d..8d10403591 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll index e1953f3cfa..6357954c9b 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll index db2087224f..48706c2880 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll index 02a95a332e..b9959f6a96 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll index 6b0e005a30..46b687e91d 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll index c627065725..69dd6205ba 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_sync.ll index e6f72c28ae..28b2213411 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll index 2a7a9a7a70..4b5b169f31 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll index c14e1bc459..a8b9647cb1 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll index bb43fd0b6d..33bcaec9c2 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll index 13be404b92..1634efe068 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll index 37030a5272..342b7df9fc 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll index b25cc907c1..2841b507b7 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll index 46b9532127..e649754501 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll index 56238c93d7..dea6ef5902 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll index e46f7410da..0d421a5457 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll index 4d5c293ef1..655c224f6f 100644 --- a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll index 597b0dcc44..d4e14ab760 100644 --- a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll index 66cae7bf2d..f4320e94a9 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll index f30d3a602f..b07789af50 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll index e4f66f1f98..812fea62dc 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll index 2ffca499c0..ffa1ab715d 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll index ae698906d1..9d04714adb 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll index a0f3e67b69..762b7a5685 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll index 84f69a4c2b..cc089f4674 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll index 2d18411bf8..c7a43bfef8 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll index e582bd4225..90a1feac0b 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll index e36ad71f3d..316fbeb150 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll index d77884c98c..caafffb00c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll index 08308e002c..7ec27a332a 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll index abe907ac33..623e182995 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll index e6498b0c18..4fdb40fbd0 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll index 8b1fbcef6e..1370ae34af 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll index 32b43dd8c2..66a51c5098 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll index 5db2851f97..2103cc29d7 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll index 4f7d355392..7ecf23b4af 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll index 8970e79012..c29beb88b8 100644 --- a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f16x2.ll b/cuda-oxide/intrinsics/probes/sub_f16x2.ll index d77114a98d..95080ca4b3 100644 --- a/cuda-oxide/intrinsics/probes/sub_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_f32x2.ll index 686258e982..ac564be93f 100644 --- a/cuda-oxide/intrinsics/probes/sub_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll index a402a2c3f7..4a99843635 100644 --- a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_mask.ll b/cuda-oxide/intrinsics/probes/sync_mask.ll index 683b1be50a..53a93c41a7 100644 --- a/cuda-oxide/intrinsics/probes/sync_mask.ll +++ b/cuda-oxide/intrinsics/probes/sync_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_threads.ll b/cuda-oxide/intrinsics/probes/sync_threads.ll index bfa844406b..12cfdc03bb 100644 --- a/cuda-oxide/intrinsics/probes/sync_threads.ll +++ b/cuda-oxide/intrinsics/probes/sync_threads.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll index 5f0d7e4493..edbac2acd4 100644 --- a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll index b823ac413b..fc31f58597 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll index dfb302cecb..395576b8b2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll index 779a70d70e..3e7912ec34 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll index cb0f8096d4..a90b14cc68 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll index 2e30ccc923..2e4f51197d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll index f92286e1f9..7bc1d24e6c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll index 59af03b61b..6489a4d025 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll index df4b9c0626..43f09bf278 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll index a56fb45429..f67c9f785b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll index 31a987f4e0..3622af6d2d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll index 1b5002737d..d1518d14da 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll index b2ff7cd266..6a4d5cb912 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll index 5a92491c7c..be037f8638 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll index 40c0f29f78..ea68010c16 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll index 12f23ad13d..5429eb2bbc 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll index 09e05d6ef3..1b54296d5c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll index 431b81da7a..9a681d7415 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll index 09774f7173..2631496007 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll index e6128ec64b..9745d16185 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll index a729bc01d9..c2e7fb314f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll index e63edfab81..54a4a3f400 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll index dca98ba632..30aff21895 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll index 3f05d3a6e2..4bfd7be094 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll index c09e87a586..f126754627 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll index 8d4d634cbb..097d0030c6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll index 822e77604e..be596cc4a2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll index 1d376316dc..fcf02ebf93 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll index 272e8ef0e1..33a470d014 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll index bf72fbf488..69992024f9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll index 5b4bd94c8d..135a6149ed 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll index 7837962693..ff4c64b02e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll index 0c8ef2cfd3..9af1d29f04 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll index a78f8aa6d6..ca0de79732 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll index a0b7e81098..bc10081434 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll index 02a4d9562c..4863c57053 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll index 14d660a9cc..4df9e6c767 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll index 8656971789..dfd0d4ec85 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll index 7bd6d95497..1b60b85733 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll index d646da8df8..51b592a056 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll index 82a1b27449..0d08def162 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll index 69ca6f3776..56672bf829 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll index 82d525ada3..9be04b6015 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll index 21adeb6aec..fc3042245b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll index f335cec4d1..ef4c0454aa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll index 34170acc77..5602949fd5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll index 6c4b3c8e84..d9447055da 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll index 463b96a822..27c89cb897 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll index 7eb0ca8654..0700594a29 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll index 5ae9bd0f15..ccd96e765b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll index 586087f6a9..8b3c3052fb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll index fed4990a1b..26361b45ad 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll index 647f31d933..a08fbae133 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll index 650cea233f..b234d265a9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll index 3c1b0e92fd..4dd0d027e1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll index 910354cb3f..d021f5b30c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll index eca0f558cb..044d0d9d82 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll index e9bbbe0942..81cc0844ff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll index 0231c8cee7..aa0996b49c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll index 18d52449a3..f88dfeb055 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll index 52c3931b6d..a59be23b32 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll index 5c14240d1c..0eb53f8172 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll index d5780d2b58..527987765c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll index e473742565..1f49030f0a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll index 4446a354ea..3e60613e4f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll index b45bcd8f64..1362605978 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll index bfd93db9ea..627d4b93ff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll index 61e93d7b86..6e4d18fee8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll index b970b32f99..6a274c6fc4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll index 5e6ec4657f..3dae478f5f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll index f869e87f30..7348802335 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll index 197489f82c..2315f1f192 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll index 070622fe59..7331b4fb6f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll index 03fb61e7fb..5892dfc8d8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll index 4a51274d3b..704d87b5f8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll index 90cb0d5fe4..fac30abc83 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll index 2a94f9813d..1c0b584f10 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll index 4e2ab19a27..1d03e49cfe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll index b034aee05b..4d2769046c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll index 5fe0a8efed..b968b4852a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll index 83cac7b1e1..36b862fe3d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll index 71aa4fc632..6e04cf5ef5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll index 43fac69148..2a28f65245 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll index 22f0043a56..d8fbd85570 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll index 261adbdbc9..3f3de404d6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll index e5f9e24f90..8162131bab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll index 02be0e8cad..1d6994185a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll index e74d6957ac..73bb9e3577 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll index 371d6f78b0..e379507c79 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll index 71c6cc8313..0e8ded2591 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll index 722bbb9938..51cf37df62 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll index 68d013a0dc..8e56c17978 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll index d9db4b69fd..ccb4c0dc96 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll index 578be1d4dd..477411bac9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll index 730e54dbfb..88d1bc8ea1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll index 71c61e2deb..76db69b1b2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll index 183b1767e6..9dcd518fab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll index bb25fecc3c..41f4dbfeba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll index e527fb2524..9205d1a381 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll index 15dcf1f873..bf0ea0fcfa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll index f59cd79143..25f0487143 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll index 8947077001..4c32a0ce14 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll index 23f0103b48..45776a4007 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll index e2003259a6..a90bf8b47e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll index 83f8e8f89c..21b6215750 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll index 01c6816cfb..6800ef72e1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll index 1f54053c1c..f025369505 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll index 8c24485dba..618c8a05ac 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll index 4165c98728..aa33e3af6c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll index dcf3eb1268..d7842555ea 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll index bd0569970f..5db306e2de 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll index 830163dda2..7cdbca2058 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll index 41bf253cc1..5b96a05380 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll index fa2725a12d..5946c8122f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll index 89ca9a137d..048856188c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll index f290ab329c..2d16d3c3bb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll index f1bbfeef13..0ebe513204 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll index 1f4b9e89b4..6664e2749c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll index e46f0ca74e..ade7ee312b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll index a7d62aaf59..a8900d6cf9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll index c94cd41efc..df8bb3b916 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll index dbf59bafdd..3c8df30c84 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll index 6dead57bca..edbbb7696c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll index 9ea197a58a..9785d254e6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll index 74df38e7c9..73977f774e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll index d3fd44cf7b..9862624480 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll index 2c1773e38b..73753dcd9c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll index e975159ebf..8da08a6003 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll index 54fbc8fca1..6b903bd942 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll index a160caea4e..7f5a70026d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll index fa6921722a..83568d75eb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll index 8de63e138d..d1ad1d6417 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll index 2f264a9b12..7edfa35d48 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll index 0f32015fb1..5d340d6635 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll index 35f7bc7167..696c5de431 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll index 9117cdfd9e..bdc91ed5e2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll index 5facb4658a..afe2f6e2eb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll index 748f593bc6..d129e09365 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll index e3a2ddf422..40e444cbf4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll index 6bb3038653..993d7e2526 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll index aedbde911b..2773ed4292 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll index c9211ebf93..35278436c0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll index ff4d689ed2..768bd6e8aa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll index 6119c1b78a..bb5ff188c2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll index 3587b55966..ab842bdeab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll index 3d1513573e..6e93f1da48 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll index db6f697ace..343efc6617 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll index 94ffd5f6aa..3a5b6a3a01 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll index e9445ec605..253c461403 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll index e6d9cdf552..d9f7674921 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll index 13650f29e6..370809bbca 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll index 9881f97c0a..73daf95af6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll index 981f53b0f8..827254daf9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll index e80ecc14d9..663bd0f7fe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll index a341605a9e..f8fc0ce249 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll index 61a4985f62..2994ab1e6d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll index 405b42e46a..6be38afd39 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll index b6cd43c4ed..2ec5dfae75 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll index edf0f0f845..d1c3467bee 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll index ba656768e3..43bfc6f846 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll index 325dcf58ec..e6614e2c13 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll index 06939a001b..510a3c08f6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll index fe2ce75b77..197d7d53f7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll index 3f7835672d..2340587798 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll index 9ce2e04316..76bd27ba1e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll index e0399b1b80..dd329528e4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll index ea382a3e4c..caef6cccba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll index eb14712d1a..8a0cb086e2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll index cef55806f6..f526de725b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll index cce1d76678..6aa3bcc2f8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll index ae71ecb02d..f8fabede28 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll index 1309cf9a45..6387cc43fe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll index 561ad2b226..69e8a07c8f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll index 4a7a422acf..4df2b6b459 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll index 102e728e71..17e085f9de 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll index 77e2d52b11..1638f95864 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll index 2d6e832e41..9be641f894 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll index 8f76bfac3b..3f74be72ed 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll index 8ed6bcc5aa..fe66ba4b24 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll index c3d4ae1412..bebf3ef42f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll index 88b8146e7f..fd8038ec00 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll index e7c2de1788..ad9799b076 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll index 734f56972d..196e49b726 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll index 9368684e3e..b47fa7c7e2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll index cd978dbea9..f91fb35a65 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll index b452b20c0d..721ed471d5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll index 291d882495..30936267b9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll index 334d590622..ae6418a3d5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll index a1cc78d36b..4ea479203e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll index 3fc331675c..d858525d25 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll index 351a96ea42..62ce095c48 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll index 33179a9019..b54190360f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll index c9e00cdf2e..7e7d419a94 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll index 8969ab99f5..ccefd8fa71 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll index 13ae0c1322..c4e0c4ba92 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll index 5615fdc594..0a353fd4ae 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll index ccc1986b6d..b180e2a1ce 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll index 76ba355871..e6a264bca3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll index 93ce5bfb45..5bccc23776 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll index 1d98697b7d..2a285acd10 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll index 6ec802a078..a3125783ec 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll index 54408310a9..618adf2bc2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll index fce5aab34f..2385efe8fd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll index 579190ce50..e08168d62d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll index 945c41b3a1..5e97ab129c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll index 6c3d0cbc2b..ddcfedadf4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll index ebe4861d94..fe5122d36f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll index cc93a7369e..2d03c1a55f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll index 8d9474b696..e5642c1c51 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll index f5a16d9133..0ac53dfc0e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll index 693fc0b853..4b160f2227 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll index 2c73baa378..aa66d99016 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll index 5d0cf33455..3a3791e947 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll index 4f3dbb60ef..21a51f9159 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll index 57ce4c0886..3b1579f29e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll index 5e85f3a09d..23b0bd9092 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll index 9b709f422e..22eba73b99 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll index 5a9c9373e0..e124078150 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll index bee9bffa8c..234053ecea 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll index 8c77ac107a..c2b8849f6a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll index c8cc9d74ed..d2328953c6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll index 333f2b5205..611a540a91 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll index ac0554d50e..62ea9464d6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll index d9eb449d59..8e0f1c1a47 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll index 610f0af819..1eeb06b6a2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll index 33a2ee4102..bcaa3e3349 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll index 548bd59e8d..9824934037 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll index baad23fe6a..ba67fdd5de 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll index 98d7bbbb7e..247a2227cf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll index 0ec73445c8..740462be72 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll index 5a3eb0984c..6ce4972b5d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll index bca6fc8d35..deb7d47cfa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll index b811beb1f0..865b8ccb67 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll index 43185ffd6a..1b7b956a0e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll index d0f0f31dd9..492568717b 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll index cb489ffa7c..67b7affd7e 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll index b65e7f27aa..95d08cc831 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll index 02e4af589f..db514e32a4 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll index 8c289fca96..b88aeb493b 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll index 1e1f8ae263..f3aa874dc7 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll index aee07673e4..e847ad0cc9 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll index ce6ff316b2..c9d3b1b71a 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll index e335b2b467..39964e69ca 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll index afb08c6673..f73b3b8e61 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll index d63b34a14c..f04918f255 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_x.ll b/cuda-oxide/intrinsics/probes/thread_idx_x.ll index 406589b9ea..16eac4a1c8 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_y.ll b/cuda-oxide/intrinsics/probes/thread_idx_y.ll index df4ebf993a..ba0ce02788 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_z.ll b/cuda-oxide/intrinsics/probes/thread_idx_z.ll index 04e6cb9ae7..854c169db5 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence.ll b/cuda-oxide/intrinsics/probes/threadfence.ll index 3e40aeed62..881cd4d8d8 100644 --- a/cuda-oxide/intrinsics/probes/threadfence.ll +++ b/cuda-oxide/intrinsics/probes/threadfence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_block.ll b/cuda-oxide/intrinsics/probes/threadfence_block.ll index c75307a800..18637438f6 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_block.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_block.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_system.ll b/cuda-oxide/intrinsics/probes/threadfence_system.ll index 826e9530a3..7dc31e7a53 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_system.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll index 4d92137b25..1ddd1e6aeb 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll index 5707a1d084..83a3246633 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/trap.ll b/cuda-oxide/intrinsics/probes/trap.ll index 6da7e92ab2..5aa2b7ee63 100644 --- a/cuda-oxide/intrinsics/probes/trap.ll +++ b/cuda-oxide/intrinsics/probes/trap.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/uni_sync.ll b/cuda-oxide/intrinsics/probes/uni_sync.ll index a33eef123c..1cde13d790 100644 --- a/cuda-oxide/intrinsics/probes/uni_sync.ll +++ b/cuda-oxide/intrinsics/probes/uni_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll index 3106842b64..d26c218655 100644 --- a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll index 5a050dcd77..3f532a5816 100644 --- a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll index e9bf779030..50a9a6296b 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_fence.ll b/cuda-oxide/intrinsics/probes/wgmma_fence.ll index 0f230c57dd..d4195c11ee 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_fence.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_fence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll index 25ac4f95cc..294106451e 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" From d7a20f02d020f96ed4f442ef8d0bc3bf200ab704 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 18:03:08 +0530 Subject: [PATCH 4/7] test(examples): check the bf16x2 FP8 packs in generated_intrinsics_blackwell Adds a kernel that calls the four bf16x2 to packed-FP8 conversions and four host cases checked bit for bit, including +inf saturation and NaN. Verified on an RTX 5090 at sm_120a; compile-only in CI as before. Signed-off-by: nihalpasham --- .../src/main.rs | 110 ++++++++++++++++-- 1 file changed, 101 insertions(+), 9 deletions(-) diff --git a/cuda-oxide/crates/rustc-codegen-cuda/examples/generated_intrinsics_blackwell/src/main.rs b/cuda-oxide/crates/rustc-codegen-cuda/examples/generated_intrinsics_blackwell/src/main.rs index 5454804be5..0d29f842d9 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/examples/generated_intrinsics_blackwell/src/main.rs +++ b/cuda-oxide/crates/rustc-codegen-cuda/examples/generated_intrinsics_blackwell/src/main.rs @@ -5,11 +5,12 @@ //! Coverage for generated high-target intrinsics. //! -//! The packed-FP8 and TF32 conversions are launched and their results checked -//! against the two formats, so the example reports what the device computed -//! rather than that it compiled. Both are bit-exact, which leaves no tolerance -//! to choose. The sparse MMA and TMA kernels stay compile-only: they need -//! operands and descriptors an example cannot supply. +//! The packed-FP8 conversions, from f32 pairs and from packed `bf16x2`, and +//! the TF32 conversions are launched and their results checked against the +//! formats, so the example reports what the device computed rather than that +//! it compiled. All are bit-exact, which leaves no tolerance to choose. The +//! sparse MMA and TMA kernels stay compile-only: they need operands and +//! descriptors an example cannot supply. //! //! Compile-only in CI, since it pins `sm_120a` and no runner has that device. //! Run it on one and the conversions are verified: @@ -22,10 +23,11 @@ use cuda_device::{ DisjointSlice, barrier::Barrier, convert::{ - cvt_rn_relu_satfinite_tf32_f32, cvt_rn_relu_tf32_f32, cvt_rn_satfinite_tf32_f32, - cvt_rn_tf32_f32, cvt_rna_satfinite_tf32_f32, cvt_rna_tf32_f32, - cvt_rz_relu_satfinite_tf32_f32, cvt_rz_relu_tf32_f32, cvt_rz_satfinite_tf32_f32, - cvt_rz_tf32_f32, + cvt_rn_relu_satfinite_tf32_f32, cvt_rn_relu_tf32_f32, cvt_rn_satfinite_e4m3x2_bf16x2, + cvt_rn_satfinite_e5m2x2_bf16x2, cvt_rn_satfinite_relu_e4m3x2_bf16x2, + cvt_rn_satfinite_relu_e5m2x2_bf16x2, cvt_rn_satfinite_tf32_f32, cvt_rn_tf32_f32, + cvt_rna_satfinite_tf32_f32, cvt_rna_tf32_f32, cvt_rz_relu_satfinite_tf32_f32, + cvt_rz_relu_tf32_f32, cvt_rz_satfinite_tf32_f32, cvt_rz_tf32_f32, }, cuda_module, kernel, thread, tma::{self, TmaDescriptor}, @@ -60,6 +62,28 @@ mod kernels { } } + /// Keeps every generated `bf16x2` to packed-FP8 conversion in device code. + /// + /// Launched, and its four results checked against the two formats. The + /// source is one register holding two bf16 values, low half first, and + /// the low half lands in the low byte of the result. + #[kernel] + pub fn compile_bf16x2_fp8_conversions(mut output: DisjointSlice, packed: u32) { + let values = [ + cvt_rn_satfinite_e4m3x2_bf16x2(packed), + cvt_rn_satfinite_relu_e4m3x2_bf16x2(packed), + cvt_rn_satfinite_e5m2x2_bf16x2(packed), + cvt_rn_satfinite_relu_e5m2x2_bf16x2(packed), + ]; + let start = thread::index_1d().get() * values.len(); + if start + values.len() <= output.len() { + for (offset, value) in values.into_iter().enumerate() { + // SAFETY: the bounds check covers this thread's unique slots. + unsafe { *output.get_unchecked_mut(start + offset) = value }; + } + } + } + /// Keeps every generated TF32 conversion in device code. /// /// Launched, and its ten results checked against the format. @@ -517,6 +541,53 @@ const FP8_CASES: [Fp8Case; 3] = [ }, ]; +/// One `bf16x2` case: the packed bf16 pair, low half in the low 16 bits, and +/// what each of the four conversions must return. +struct Bf16x2Case { + packed: u32, + e4m3: u16, + relu_e4m3: u16, + e5m2: u16, + relu_e5m2: u16, +} + +/// bf16 keeps the f32 exponent with a 7-bit mantissa, so the values of the +/// FP8 table are bf16-exact and the expected bytes carry over: 1.0 = 0x3f80, +/// 2.0 = 0x4000, -1.0 = 0xbf80, 448.0 = 0x43e0, 0.0 = 0x0000. The last two +/// cases add the bf16-only inputs: +inf (0x7f80) saturates to the largest +/// finite value, and a quiet NaN (0x7fc0) becomes 0x7f in both formats. +const BF16X2_CASES: [Bf16x2Case; 4] = [ + Bf16x2Case { + packed: 0x4000_3f80, + e4m3: 0x4038, + relu_e4m3: 0x4038, + e5m2: 0x403c, + relu_e5m2: 0x403c, + }, + // relu replaces the negative low lane with zero and leaves the high lane. + Bf16x2Case { + packed: 0x43e0_bf80, + e4m3: 0x7eb8, + relu_e4m3: 0x7e00, + e5m2: 0x5fbc, + relu_e5m2: 0x5f00, + }, + Bf16x2Case { + packed: 0x0000_7f80, + e4m3: 0x007e, + relu_e4m3: 0x007e, + e5m2: 0x007b, + relu_e5m2: 0x007b, + }, + Bf16x2Case { + packed: 0x3f80_7fc0, + e4m3: 0x387f, + relu_e4m3: 0x387f, + e5m2: 0x3c7f, + relu_e5m2: 0x3c7f, + }, +]; + /// TF32 keeps the f32 exponent and truncates the mantissa to 10 bits, so each /// value here is already representable and every rounding mode agrees. The /// four relu forms sit at indices 3, 5, 7 and 9 of the kernel's output. @@ -564,6 +635,27 @@ fn main() { } } + for case in &BF16X2_CASES { + let mut out = DeviceBuffer::::zeroed(&stream, 4).expect("alloc bf16x2 output"); + // SAFETY: the two arguments match the kernel's slice and scalar, and + // `out` holds the four results one thread writes. + unsafe { module.compile_bf16x2_fp8_conversions(&stream, cfg, &mut out, case.packed) } + .expect("bf16x2 launch failed"); + let got = out.to_host_vec(&stream).expect("bf16x2 readback failed"); + let want = [case.e4m3, case.relu_e4m3, case.e5m2, case.relu_e5m2]; + let names = ["e4m3x2", "relu_e4m3x2", "e5m2x2", "relu_e5m2x2"]; + for ((got, want), name) in got.iter().zip(want).zip(names) { + let mark = if *got == want { "ok" } else { "MISMATCH" }; + if *got != want { + failures += 1; + } + println!( + " bf16x2 {name:<12} (0x{:08x}) = 0x{got:04x} want 0x{want:04x} {mark}", + case.packed + ); + } + } + for (value, want_plain, want_relu) in TF32_CASES { let mut out = DeviceBuffer::::zeroed(&stream, 10).expect("alloc tf32 output"); // SAFETY: the two arguments match the kernel's slice and scalar, and From 71a6c4eb561e6d58824f56951fb4b9588c11619a Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 18:44:54 +0530 Subject: [PATCH 5/7] test(smoketest): pin the Blackwell lane to PTX 9.1 with bf16x2 counts generated_intrinsics_blackwell now calls the four bf16x2 to packed-FP8 conversions, which are PTX 9.1 forms, so its PTX header moved from .version 8.7 to .version 9.1 and the blackwell-compile lane's header pin went red while every other assertion still held. Pin the new floor and prove the new instructions the way the lane proves the f32 packs: four distinct cvt.rn.satfinite[.relu].{e4m3x2,e5m2x2}.bf16x2 instructions in the llc-route PTX and four "=h,r" inline-asm calls in the libNVVM-route NVVM IR. Signed-off-by: nihalpasham --- cuda-oxide/scripts/smoketest.sh | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-) diff --git a/cuda-oxide/scripts/smoketest.sh b/cuda-oxide/scripts/smoketest.sh index 0d23fca93e..46201ed90a 100755 --- a/cuda-oxide/scripts/smoketest.sh +++ b/cuda-oxide/scripts/smoketest.sh @@ -1455,6 +1455,7 @@ run_cargo() { local unresolved_sparse_f16_re='llvm\.nvvm\.mma\.sp\.ordered\.metadata\.m16n8k64\.row\.col\.kind\.f8f6f4\.f16\.' local unresolved_standard_fp8_re='llvm\.nvvm\.mma\.m16n8k(16|32)\.row\.col\.f(16|32)\.(e4m3|e5m2)\.' local fp8_re='cvt\.rn\.satfinite(\.relu)?\.(e4m3x2|e5m2x2)\.f32' + local bf16x2_fp8_re='cvt\.rn\.satfinite(\.relu)?\.(e4m3x2|e5m2x2)\.bf16x2' local tf32_re='cvt\.(rna|rn|rz)(\.relu)?(\.satfinite)?\.tf32\.f32' local ldmatrix_re='ldmatrix\.sync\.aligned\.(m16n16\.x(1|2)\.trans\.shared\.(b8|b8x16\.(b4x16_p64|b6x16_p32))|m8n16\.x(1|2|4)\.shared\.b8x16\.(b4x16_p64|b6x16_p32))' local instruction_count unique_instruction_count @@ -1464,6 +1465,7 @@ run_cargo() { local standard_fp8_f32_instruction_count unique_standard_fp8_f32_instruction_count local standard_fp8_f16_instruction_count unique_standard_fp8_f16_instruction_count local fp8_count unique_fp8_count + local bf16x2_fp8_count unique_bf16x2_fp8_count local tf32_count unique_tf32_count local ldmatrix_count unique_ldmatrix_count instruction_count="$(grep -oE "${instruction_re}" "${llvm_ptx}" 2>/dev/null | wc -l)" @@ -1480,12 +1482,17 @@ run_cargo() { unique_standard_fp8_f16_instruction_count="$(grep -oE "${standard_fp8_f16_instruction_re}" "${llvm_ptx}" 2>/dev/null | sort -u | wc -l)" fp8_count="$(grep -oE "${fp8_re}" "${llvm_ptx}" 2>/dev/null | wc -l)" unique_fp8_count="$(grep -oE "${fp8_re}" "${llvm_ptx}" 2>/dev/null | sort -u | wc -l)" + bf16x2_fp8_count="$(grep -oE "${bf16x2_fp8_re}" "${llvm_ptx}" 2>/dev/null | wc -l)" + unique_bf16x2_fp8_count="$(grep -oE "${bf16x2_fp8_re}" "${llvm_ptx}" 2>/dev/null | sort -u | wc -l)" tf32_count="$(grep -oE "${tf32_re}" "${llvm_ptx}" 2>/dev/null | wc -l)" unique_tf32_count="$(grep -oE "${tf32_re}" "${llvm_ptx}" 2>/dev/null | sort -u | wc -l)" ldmatrix_count="$(grep -oE "${ldmatrix_re}" "${llvm_ptx}" 2>/dev/null | wc -l)" unique_ldmatrix_count="$(grep -oE "${ldmatrix_re}" "${llvm_ptx}" 2>/dev/null | sort -u | wc -l)" + # The header's .version is the floor of the newest form in the + # module: PTX 9.1 since the bf16x2 to packed-FP8 conversions joined + # the example (the ldmatrix and kind::f8f6f4 MMA forms need 8.7). if [[ ! -s "${llvm_ptx}" ]] \ - || ! grep -qx '\.version 8\.7' "${llvm_ptx}" \ + || ! grep -qx '\.version 9\.1' "${llvm_ptx}" \ || ! grep -qx '\.target sm_120a' "${llvm_ptx}" \ || [[ ${instruction_count} -ne 25 || ${unique_instruction_count} -ne 25 ]] \ || [[ ${sparse_f16_instruction_count} -ne 25 || ${unique_sparse_f16_instruction_count} -ne 25 ]] \ @@ -1494,6 +1501,7 @@ run_cargo() { || [[ ${standard_fp8_f32_instruction_count} -ne 8 || ${unique_standard_fp8_f32_instruction_count} -ne 8 ]] \ || [[ ${standard_fp8_f16_instruction_count} -ne 8 || ${unique_standard_fp8_f16_instruction_count} -ne 8 ]] \ || [[ ${fp8_count} -ne 4 || ${unique_fp8_count} -ne 4 ]] \ + || [[ ${bf16x2_fp8_count} -ne 4 || ${unique_bf16x2_fp8_count} -ne 4 ]] \ || [[ ${tf32_count} -ne 10 || ${unique_tf32_count} -ne 10 ]] \ || [[ ${ldmatrix_count} -ne 12 || ${unique_ldmatrix_count} -ne 12 ]] \ || grep -qE "${unresolved_sparse_f16_re}|${unresolved_dense_f32_re}|${unresolved_dense_f16_re}|${unresolved_standard_fp8_re}" "${llvm_ptx}"; then @@ -1516,8 +1524,10 @@ run_cargo() { fi local nvvm_ll="crates/rustc-codegen-cuda/examples/${ex}/${ex}.ll" local inline_fp8_re='call i16 asm "cvt\.rn\.satfinite(\.relu)?\.(e4m3x2|e5m2x2)\.f32 \$0, \$2, \$1;", "=h,f,f"' + local inline_bf16x2_fp8_re='call i16 asm "cvt\.rn\.satfinite(\.relu)?\.(e4m3x2|e5m2x2)\.bf16x2 \$0, \$1;", "=h,r"' local inline_tf32_re='call i32 asm "cvt\.(rna|rn|rz)(\.relu)?(\.satfinite)?\.tf32\.f32 \$0, \$1;", "=r,f"' local inline_fp8_count unique_inline_fp8_count + local inline_bf16x2_fp8_count unique_inline_bf16x2_fp8_count local inline_tf32_count unique_inline_tf32_count local inline_ldmatrix_count unique_inline_ldmatrix_count local inline_sparse_f16_count unique_inline_sparse_f16_count @@ -1527,6 +1537,8 @@ run_cargo() { local inline_standard_fp8_f16_count unique_inline_standard_fp8_f16_count inline_fp8_count="$(grep -oE "${inline_fp8_re}" "${nvvm_ll}" 2>/dev/null | wc -l)" unique_inline_fp8_count="$(grep -oE "${inline_fp8_re}" "${nvvm_ll}" 2>/dev/null | sort -u | wc -l)" + inline_bf16x2_fp8_count="$(grep -oE "${inline_bf16x2_fp8_re}" "${nvvm_ll}" 2>/dev/null | wc -l)" + unique_inline_bf16x2_fp8_count="$(grep -oE "${inline_bf16x2_fp8_re}" "${nvvm_ll}" 2>/dev/null | sort -u | wc -l)" inline_tf32_count="$(grep -oE "${inline_tf32_re}" "${nvvm_ll}" 2>/dev/null | wc -l)" unique_inline_tf32_count="$(grep -oE "${inline_tf32_re}" "${nvvm_ll}" 2>/dev/null | sort -u | wc -l)" inline_ldmatrix_count="$(grep -oE "${ldmatrix_re}" "${nvvm_ll}" 2>/dev/null | wc -l)" @@ -1543,6 +1555,7 @@ run_cargo() { unique_inline_standard_fp8_f16_count="$(grep -oE "${standard_fp8_f16_instruction_re}" "${nvvm_ll}" 2>/dev/null | sort -u | wc -l)" if [[ ! -s "${nvvm_ll}" ]] \ || [[ ${inline_fp8_count} -ne 4 || ${unique_inline_fp8_count} -ne 4 ]] \ + || [[ ${inline_bf16x2_fp8_count} -ne 4 || ${unique_inline_bf16x2_fp8_count} -ne 4 ]] \ || [[ ${inline_tf32_count} -ne 10 || ${unique_inline_tf32_count} -ne 10 ]] \ || [[ ${inline_ldmatrix_count} -ne 12 || ${unique_inline_ldmatrix_count} -ne 12 ]] \ || [[ ${inline_sparse_f16_count} -ne 25 || ${unique_inline_sparse_f16_count} -ne 25 ]] \ From 0ed724f558292f9e660d458711455a7c589973aa Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 18:44:54 +0530 Subject: [PATCH 6/7] ci(examples): move the examples floor gate to LLVM 23 generated_intrinsics_blackwell now needs +ptx91 for the bf16x2 to packed-FP8 conversions. llc-22 knows +ptx90 at most, and the codegen gate refuses a PTX 9.1 request below LLVM 23 at compile time, so the compile-only lane cannot build that example under its LLVM 22 pin. Install LLVM 23 from apt.llvm.org instead (the llvm-toolchain-*-23 suites serve 23.1.4 for noble and resolute) and point both CUDA_OXIDE_LLC sites at llc-23, following the previous floor bump from 21 to 22. The floor comment names the new reason and keeps the old one. Signed-off-by: nihalpasham --- .github/workflows/examples-compile.yml | 35 +++++++++++++------------- 1 file changed, 18 insertions(+), 17 deletions(-) diff --git a/.github/workflows/examples-compile.yml b/.github/workflows/examples-compile.yml index 0f8544cb63..ee70d3a956 100644 --- a/.github/workflows/examples-compile.yml +++ b/.github/workflows/examples-compile.yml @@ -76,30 +76,31 @@ jobs: # The device pipeline shells out to LLVM's llc/opt for PTX # generation. This pin is the FLOOR GATE: if someone lands IR that # needs a newer llc, this job is where the floor bump becomes - # visible instead of on contributors' machines. LLVM 22 is the true - # floor: the ldmatrix/stmatrix family (llvm.nvvm.stmatrix.*, used by - # gemm_sol and gemm_sol_final) postdates llc-21, which "lowers" an - # unknown intrinsic by leaking its dotted LLVM name into the PTX as - # an `.extern .func` (exit 0, no diagnostic): unassemblable output - # that only surfaced once the ptxas verdict gate started assembling - # what this job produces. The pipeline now rejects such leaks with a - # named error, so an under-floor llc fails loudly here. - # If a future ubuntu-latest codename has no llvm-toolchain-*-22 + # visible instead of on contributors' machines. LLVM 23 is the true + # floor: the bf16x2 to packed-FP8 conversions (PTX 9.1 forms, used + # by generated_intrinsics_blackwell) need `+ptx91`, which llc-22 + # does not know. It ignored the unknown feature and emitted + # `.version 8.8` PTX that failed only at JIT; the pipeline now + # refuses a PTX 9.1 request below LLVM 23 with a named error, so an + # under-floor llc fails loudly here. The previous floor, LLVM 22 for + # the ldmatrix/stmatrix family (llc-21 leaked unknown intrinsics + # into the PTX as `.extern .func`), is covered by the same gate. + # If a future ubuntu-latest codename has no llvm-toolchain-*-23 # suite, apt-get update fails loudly; recover by pinning runs-on to # the previous image or bumping the LLVM major here. - - name: Install LLVM 22 (llc/opt for the NVPTX backend) + - name: Install LLVM 23 (llc/opt for the NVPTX backend) run: | set -euo pipefail source /etc/os-release sudo install -d -m 0755 /etc/apt/keyrings wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key \ | sudo tee /etc/apt/keyrings/apt-llvm-org.asc >/dev/null - echo "deb [signed-by=/etc/apt/keyrings/apt-llvm-org.asc] https://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-22 main" \ - | sudo tee /etc/apt/sources.list.d/llvm-22.list + echo "deb [signed-by=/etc/apt/keyrings/apt-llvm-org.asc] https://apt.llvm.org/${VERSION_CODENAME}/ llvm-toolchain-${VERSION_CODENAME}-23 main" \ + | sudo tee /etc/apt/sources.list.d/llvm-23.list sudo apt-get update - sudo apt-get install -y --no-install-recommends llvm-22 - llc-22 --version | head -2 - opt-22 --version | head -2 + sudo apt-get install -y --no-install-recommends llvm-23 + llc-23 --version | head -2 + opt-23 --version | head -2 - name: Compile every example (smoketest --compile-only) # Step-level timeout, below the job's 75: a job-level timeout @@ -108,7 +109,7 @@ jobs: timeout-minutes: 50 env: CUDA_TOOLKIT_PATH: ${{ steps.cuda-toolkit.outputs.CUDA_PATH }} - CUDA_OXIDE_LLC: /usr/bin/llc-22 + CUDA_OXIDE_LLC: /usr/bin/llc-23 # Each example is its own cargo workspace; without a shared # target dir the dep tree rebuilds ~85 times and blows the # runner's ~14 GB disk (same trick as clippy.yml). smoketest.sh @@ -136,7 +137,7 @@ jobs: working-directory: cuda-oxide env: CUDA_TOOLKIT_PATH: ${{ steps.cuda-toolkit.outputs.CUDA_PATH }} - CUDA_OXIDE_LLC: /usr/bin/llc-22 + CUDA_OXIDE_LLC: /usr/bin/llc-23 CARGO_TARGET_DIR: ${{ runner.temp }}/examples-shared-target CARGO_INCREMENTAL: "0" run: | From 89e24e263aeb1c233252ba0f893e96725f158fd6 Mon Sep 17 00:00:00 2001 From: nihalpasham Date: Mon, 19 Oct 2026 18:44:55 +0530 Subject: [PATCH 7/7] fix(intrinsics): quote the e5m2x2 feature name in the bf16x2 evidence The inline-route ptx_assembly detail of the two e5m2x2 records repeated the e4m3x2 records' negative control. Retargeting each form's inline PTX to sm_120 with ptxas 13.3.73 rejects the e5m2x2 forms with "Feature 'cvt.e5m2x2.bf16x2' not supported on .target 'sm_120'", and the two details now quote that message. The catalog carries the detail text, so the generated outputs and the book's catalog stamp move with it; no route, target or signature changes. Signed-off-by: nihalpasham --- cuda-oxide/crates/cuda-device/src/generated/async_copy.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/atomic.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/bf16.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/clc.rs | 2 +- .../crates/cuda-device/src/generated/cluster_barrier.rs | 2 +- .../crates/cuda-device/src/generated/cluster_memory.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/convert.rs | 2 +- .../crates/cuda-device/src/generated/counted_barrier.rs | 2 +- .../crates/cuda-device/src/generated/debug_control.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/dotprod.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/f16.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/f16x2.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/f32x2.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/fence.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/float.rs | 2 +- .../crates/cuda-device/src/generated/grid_dependency.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/i16x2.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/int.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs | 2 +- .../crates/cuda-device/src/generated/mbarrier_basic.rs | 2 +- .../crates/cuda-device/src/generated/mbarrier_extended.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/prmt.rs | 2 +- .../crates/cuda-device/src/generated/register_control.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/register_mma.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/sreg.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/tma.rs | 2 +- cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs | 2 +- .../crates/cuda-device/src/generated/wgmma_control.rs | 2 +- cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs | 2 +- cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs | 2 +- .../src/generated_intrinsic_targets/active_mask.rs | 2 +- .../src/generated_intrinsic_targets/clc.rs | 2 +- .../src/generated_intrinsic_targets/cluster_barrier.rs | 2 +- .../src/generated_intrinsic_targets/cluster_memory.rs | 2 +- .../src/generated_intrinsic_targets/cp_async.rs | 2 +- .../src/generated_intrinsic_targets/debug_control.rs | 2 +- .../src/generated_intrinsic_targets/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/elect.rs | 2 +- .../src/generated_intrinsic_targets/execution_control.rs | 2 +- .../src/generated_intrinsic_targets/extended_minmax.rs | 2 +- .../src/generated_intrinsic_targets/integer_minmax.rs | 2 +- .../src/generated_intrinsic_targets/ldmatrix.rs | 2 +- .../src/generated_intrinsic_targets/mbarrier_basic.rs | 2 +- .../src/generated_intrinsic_targets/mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/mod.rs | 2 +- .../src/generated_intrinsic_targets/movmatrix.rs | 2 +- .../src/generated_intrinsic_targets/packed_alu.rs | 2 +- .../src/generated_intrinsic_targets/packed_atomic.rs | 2 +- .../src/generated_intrinsic_targets/packed_conversion.rs | 2 +- .../src/generated_intrinsic_targets/prmt.rs | 2 +- .../src/generated_intrinsic_targets/redux.rs | 2 +- .../src/generated_intrinsic_targets/register_mma.rs | 2 +- .../src/generated_intrinsic_targets/scalar_arithmetic.rs | 2 +- .../src/generated_intrinsic_targets/scalar_conversion.rs | 2 +- .../src/generated_intrinsic_targets/scalar_math.rs | 2 +- .../src/generated_intrinsic_targets/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/sreg.rs | 2 +- .../src/generated_intrinsic_targets/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/sync.rs | 2 +- .../src/generated_intrinsic_targets/tcgen05_cp.rs | 2 +- .../src/generated_intrinsic_targets/tcgen05_ld.rs | 2 +- .../src/generated_intrinsic_targets/tcgen05_mma.rs | 2 +- .../src/generated_intrinsic_targets/tcgen05_mma_ws.rs | 2 +- .../src/generated_intrinsic_targets/tcgen05_other.rs | 2 +- .../src/generated_intrinsic_targets/tcgen05_st.rs | 2 +- .../src/generated_intrinsic_targets/tests/active_mask.rs | 2 +- .../src/generated_intrinsic_targets/tests/clc.rs | 2 +- .../generated_intrinsic_targets/tests/cluster_barrier.rs | 2 +- .../src/generated_intrinsic_targets/tests/cluster_memory.rs | 2 +- .../src/generated_intrinsic_targets/tests/cp_async.rs | 2 +- .../src/generated_intrinsic_targets/tests/debug_control.rs | 2 +- .../src/generated_intrinsic_targets/tests/dotprod.rs | 2 +- .../src/generated_intrinsic_targets/tests/elect.rs | 2 +- .../generated_intrinsic_targets/tests/execution_control.rs | 2 +- .../generated_intrinsic_targets/tests/extended_minmax.rs | 2 +- .../src/generated_intrinsic_targets/tests/integer_minmax.rs | 2 +- .../src/generated_intrinsic_targets/tests/ldmatrix.rs | 2 +- .../src/generated_intrinsic_targets/tests/mbarrier_basic.rs | 2 +- .../generated_intrinsic_targets/tests/mbarrier_extended.rs | 2 +- .../src/generated_intrinsic_targets/tests/mod.rs | 2 +- .../src/generated_intrinsic_targets/tests/movmatrix.rs | 2 +- .../src/generated_intrinsic_targets/tests/packed_alu.rs | 2 +- .../src/generated_intrinsic_targets/tests/packed_atomic.rs | 2 +- .../generated_intrinsic_targets/tests/packed_conversion.rs | 2 +- .../src/generated_intrinsic_targets/tests/prmt.rs | 2 +- .../src/generated_intrinsic_targets/tests/redux.rs | 2 +- .../generated_intrinsic_targets/tests/register_mma_float.rs | 2 +- .../generated_intrinsic_targets/tests/register_mma_s32.rs | 2 +- .../generated_intrinsic_targets/tests/scalar_arithmetic.rs | 2 +- .../generated_intrinsic_targets/tests/scalar_conversion.rs | 2 +- .../src/generated_intrinsic_targets/tests/scalar_math.rs | 2 +- .../src/generated_intrinsic_targets/tests/sparse_mma.rs | 2 +- .../src/generated_intrinsic_targets/tests/sreg.rs | 2 +- .../src/generated_intrinsic_targets/tests/stmatrix.rs | 2 +- .../src/generated_intrinsic_targets/tests/sync.rs | 2 +- .../src/generated_intrinsic_targets/tests/tcgen05_cp.rs | 2 +- .../src/generated_intrinsic_targets/tests/tcgen05_ld.rs | 2 +- .../src/generated_intrinsic_targets/tests/tcgen05_mma.rs | 2 +- .../src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs | 2 +- .../src/generated_intrinsic_targets/tests/tcgen05_other.rs | 2 +- .../src/generated_intrinsic_targets/tests/tcgen05_st.rs | 2 +- .../src/generated_intrinsic_targets/tests/tma.rs | 2 +- .../src/generated_intrinsic_targets/tests/vote.rs | 2 +- .../src/generated_intrinsic_targets/tests/warp_barrier.rs | 2 +- .../src/generated_intrinsic_targets/tests/warp_match.rs | 2 +- .../src/generated_intrinsic_targets/tests/warp_shuffle.rs | 2 +- .../src/generated_intrinsic_targets/tests/wgmma_control.rs | 2 +- .../src/generated_intrinsic_targets/tma.rs | 2 +- .../src/generated_intrinsic_targets/vote.rs | 2 +- .../src/generated_intrinsic_targets/warp_barrier.rs | 2 +- .../src/generated_intrinsic_targets/warp_match.rs | 2 +- .../src/generated_intrinsic_targets/warp_shuffle.rs | 2 +- .../src/generated_intrinsic_targets/wgmma_control.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/active_mask.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs | 2 +- .../dialect-nvvm/src/ops/generated/cluster_barrier.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/cluster_memory.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/cp_async.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/debug_control.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs | 2 +- .../dialect-nvvm/src/ops/generated/execution_control.rs | 2 +- .../dialect-nvvm/src/ops/generated/extended_minmax.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/integer_minmax.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/ldmatrix.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs | 2 +- .../dialect-nvvm/src/ops/generated/mbarrier_extended.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/movmatrix.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/packed_alu.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/packed_atomic.rs | 2 +- .../dialect-nvvm/src/ops/generated/packed_conversion.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/register_mma.rs | 2 +- .../dialect-nvvm/src/ops/generated/scalar_arithmetic.rs | 2 +- .../dialect-nvvm/src/ops/generated/scalar_conversion.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/scalar_math.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/sparse_mma.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/stmatrix.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs | 2 +- cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/warp_barrier.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/warp_match.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs | 2 +- .../crates/dialect-nvvm/src/ops/generated/wgmma_control.rs | 2 +- .../terminator/intrinsics/generated/active_mask.rs | 2 +- .../src/translator/terminator/intrinsics/generated/clc.rs | 2 +- .../terminator/intrinsics/generated/cluster_barrier.rs | 2 +- .../terminator/intrinsics/generated/cluster_memory.rs | 2 +- .../translator/terminator/intrinsics/generated/cp_async.rs | 2 +- .../terminator/intrinsics/generated/debug_control.rs | 2 +- .../translator/terminator/intrinsics/generated/dotprod.rs | 2 +- .../src/translator/terminator/intrinsics/generated/elect.rs | 2 +- .../terminator/intrinsics/generated/execution_control.rs | 2 +- .../terminator/intrinsics/generated/extended_minmax.rs | 2 +- .../translator/terminator/intrinsics/generated/helpers.rs | 2 +- .../terminator/intrinsics/generated/integer_minmax.rs | 2 +- .../translator/terminator/intrinsics/generated/ldmatrix.rs | 2 +- .../terminator/intrinsics/generated/mbarrier_basic.rs | 2 +- .../terminator/intrinsics/generated/mbarrier_extended.rs | 2 +- .../src/translator/terminator/intrinsics/generated/mod.rs | 2 +- .../translator/terminator/intrinsics/generated/movmatrix.rs | 2 +- .../terminator/intrinsics/generated/packed_alu.rs | 2 +- .../terminator/intrinsics/generated/packed_atomic.rs | 2 +- .../terminator/intrinsics/generated/packed_conversion.rs | 2 +- .../src/translator/terminator/intrinsics/generated/prmt.rs | 2 +- .../src/translator/terminator/intrinsics/generated/redux.rs | 2 +- .../terminator/intrinsics/generated/register_mma.rs | 2 +- .../terminator/intrinsics/generated/scalar_arithmetic.rs | 2 +- .../terminator/intrinsics/generated/scalar_conversion.rs | 2 +- .../terminator/intrinsics/generated/scalar_math.rs | 2 +- .../terminator/intrinsics/generated/sparse_mma.rs | 2 +- .../src/translator/terminator/intrinsics/generated/sreg.rs | 2 +- .../translator/terminator/intrinsics/generated/stmatrix.rs | 2 +- .../src/translator/terminator/intrinsics/generated/sync.rs | 2 +- .../terminator/intrinsics/generated/tcgen05_cp.rs | 2 +- .../terminator/intrinsics/generated/tcgen05_ld.rs | 2 +- .../terminator/intrinsics/generated/tcgen05_mma.rs | 2 +- .../terminator/intrinsics/generated/tcgen05_other.rs | 2 +- .../terminator/intrinsics/generated/tcgen05_st.rs | 2 +- .../terminator/intrinsics/generated/tests/active_mask.rs | 2 +- .../translator/terminator/intrinsics/generated/tests/clc.rs | 2 +- .../intrinsics/generated/tests/cluster_barrier.rs | 2 +- .../terminator/intrinsics/generated/tests/cluster_memory.rs | 2 +- .../terminator/intrinsics/generated/tests/cp_async.rs | 2 +- .../terminator/intrinsics/generated/tests/debug_control.rs | 2 +- .../terminator/intrinsics/generated/tests/dotprod.rs | 2 +- .../terminator/intrinsics/generated/tests/elect.rs | 2 +- .../intrinsics/generated/tests/execution_control.rs | 2 +- .../intrinsics/generated/tests/extended_minmax.rs | 2 +- .../terminator/intrinsics/generated/tests/integer_minmax.rs | 2 +- .../terminator/intrinsics/generated/tests/ldmatrix.rs | 2 +- .../terminator/intrinsics/generated/tests/mbarrier_basic.rs | 2 +- .../intrinsics/generated/tests/mbarrier_extended.rs | 2 +- .../translator/terminator/intrinsics/generated/tests/mod.rs | 2 +- .../terminator/intrinsics/generated/tests/movmatrix.rs | 2 +- .../terminator/intrinsics/generated/tests/packed_alu.rs | 2 +- .../terminator/intrinsics/generated/tests/packed_atomic.rs | 2 +- .../intrinsics/generated/tests/packed_conversion.rs | 2 +- .../terminator/intrinsics/generated/tests/prmt.rs | 2 +- .../terminator/intrinsics/generated/tests/redux.rs | 2 +- .../terminator/intrinsics/generated/tests/register_mma.rs | 2 +- .../intrinsics/generated/tests/scalar_arithmetic.rs | 2 +- .../intrinsics/generated/tests/scalar_conversion.rs | 2 +- .../terminator/intrinsics/generated/tests/scalar_math.rs | 2 +- .../terminator/intrinsics/generated/tests/sparse_mma.rs | 2 +- .../terminator/intrinsics/generated/tests/sreg.rs | 2 +- .../terminator/intrinsics/generated/tests/stmatrix.rs | 2 +- .../terminator/intrinsics/generated/tests/sync.rs | 2 +- .../terminator/intrinsics/generated/tests/tcgen05.rs | 2 +- .../translator/terminator/intrinsics/generated/tests/tma.rs | 2 +- .../terminator/intrinsics/generated/tests/vote.rs | 2 +- .../terminator/intrinsics/generated/tests/warp_barrier.rs | 2 +- .../terminator/intrinsics/generated/tests/warp_match.rs | 2 +- .../terminator/intrinsics/generated/tests/warp_shuffle.rs | 2 +- .../terminator/intrinsics/generated/tests/wgmma_control.rs | 2 +- .../src/translator/terminator/intrinsics/generated/tma.rs | 2 +- .../src/translator/terminator/intrinsics/generated/vote.rs | 2 +- .../terminator/intrinsics/generated/warp_barrier.rs | 2 +- .../terminator/intrinsics/generated/warp_match.rs | 2 +- .../terminator/intrinsics/generated/warp_shuffle.rs | 2 +- .../terminator/intrinsics/generated/wgmma_control.rs | 2 +- .../src/convert/generated_intrinsics/active_mask.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/clc.rs | 2 +- .../src/convert/generated_intrinsics/cluster_barrier.rs | 2 +- .../src/convert/generated_intrinsics/cluster_memory.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/cp_async.rs | 2 +- .../src/convert/generated_intrinsics/debug_control.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/dotprod.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/elect.rs | 2 +- .../src/convert/generated_intrinsics/execution_control.rs | 2 +- .../src/convert/generated_intrinsics/extended_minmax.rs | 2 +- .../src/convert/generated_intrinsics/integer_minmax.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/ldmatrix.rs | 2 +- .../src/convert/generated_intrinsics/mbarrier_basic.rs | 2 +- .../src/convert/generated_intrinsics/mbarrier_extended.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/mod.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/movmatrix.rs | 2 +- .../src/convert/generated_intrinsics/packed_alu.rs | 2 +- .../src/convert/generated_intrinsics/packed_atomic.rs | 2 +- .../src/convert/generated_intrinsics/packed_conversion.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/prmt.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/redux.rs | 2 +- .../src/convert/generated_intrinsics/register_mma.rs | 2 +- .../src/convert/generated_intrinsics/scalar_arithmetic.rs | 2 +- .../src/convert/generated_intrinsics/scalar_conversion.rs | 2 +- .../src/convert/generated_intrinsics/scalar_math.rs | 2 +- .../src/convert/generated_intrinsics/sparse_mma.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/sreg.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/stmatrix.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/sync.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/tcgen05.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/tma.rs | 2 +- .../mir-lower/src/convert/generated_intrinsics/vote.rs | 2 +- .../src/convert/generated_intrinsics/warp_barrier.rs | 2 +- .../src/convert/generated_intrinsics/warp_match.rs | 2 +- .../src/convert/generated_intrinsics/warp_shuffle.rs | 2 +- .../src/convert/generated_intrinsics/wgmma_control.rs | 2 +- .../crates/rustc-codegen-cuda/src/generated_intrinsics.rs | 2 +- cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md | 4 ++-- cuda-oxide/intrinsics/catalog.json | 6 +++--- ...t-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json | 4 ++-- cuda-oxide/intrinsics/generated-reference.md | 6 +++--- cuda-oxide/intrinsics/probes/abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/active_mask.ll | 2 +- cuda-oxide/intrinsics/probes/add_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_f64.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_f64.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_f64.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/all_sync.ll | 2 +- cuda-oxide/intrinsics/probes/any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/ballot_sync.ll | 2 +- cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll | 2 +- .../intrinsics/probes/barrier_cluster_arrive_aligned.ll | 2 +- .../intrinsics/probes/barrier_cluster_arrive_relaxed.ll | 2 +- .../probes/barrier_cluster_arrive_relaxed_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll | 2 +- .../intrinsics/probes/barrier_cluster_wait_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll | 2 +- cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/barrier_cta_sync.ll | 2 +- cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/block_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/breakpoint.ll | 2 +- cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll | 2 +- cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll | 2 +- cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll | 2 +- cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll | 2 +- cuda-oxide/intrinsics/probes/clc_try_cancel.ll | 2 +- cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll | 2 +- cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_block_count.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_block_rank.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/cluster_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/cos_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll | 2 +- .../probes/cp_async_bulk_prefetch_tensor_1d_l2.ll | 2 +- .../cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_prefetch_tensor_2d_l2.ll | 2 +- .../cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_prefetch_tensor_3d_l2.ll | 2 +- .../cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_prefetch_tensor_4d_l2.ll | 2 +- .../cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_prefetch_tensor_5d_l2.ll | 2 +- .../cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll | 2 +- .../probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll | 2 +- ...p_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s_multicast.ll | 2 +- .../probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll | 2 +- .../probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll | 2 +- .../intrinsics/probes/cp_async_bulk_wait_group_read.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_4.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_8.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_cg_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_commit_group.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll | 2 +- .../intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll | 2 +- .../probes/cp_async_mbarrier_arrive_noinc_shared.ll | 2 +- .../intrinsics/probes/cp_async_mbarrier_arrive_shared.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_wait_all.ll | 2 +- cuda-oxide/intrinsics/probes/cp_async_wait_group.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll | 2 +- .../probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll | 2 +- .../intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll | 2 +- .../intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_f64.ll | 2 +- cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_f64.ll | 2 +- cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_f64.ll | 2 +- cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp2a_u32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_s32.ll | 2 +- cuda-oxide/intrinsics/probes/dp4a_u32.ll | 2 +- cuda-oxide/intrinsics/probes/dsmem_read_u32.ll | 2 +- cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll | 2 +- .../intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/ex2_approx_f16.ll | 2 +- cuda-oxide/intrinsics/probes/ex2_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll | 2 +- .../probes/fence_mbarrier_init_release_cluster.ll | 2 +- ...ce_proxy_async_generic_acquire_shared_cluster_cluster.ll | 2 +- .../fence_proxy_async_generic_release_shared_cta_cluster.ll | 2 +- .../intrinsics/probes/fence_proxy_async_shared_cta.ll | 2 +- .../probes/fence_proxy_tensormap_generic_acquire_cluster.ll | 2 +- .../probes/fence_proxy_tensormap_generic_acquire_cta.ll | 2 +- .../probes/fence_proxy_tensormap_generic_acquire_gpu.ll | 2 +- .../probes/fence_proxy_tensormap_generic_acquire_system.ll | 2 +- .../probes/fence_proxy_tensormap_generic_release_cluster.ll | 2 +- .../probes/fence_proxy_tensormap_generic_release_cta.ll | 2 +- .../probes/fence_proxy_tensormap_generic_release_gpu.ll | 2 +- .../probes/fence_proxy_tensormap_generic_release_system.ll | 2 +- cuda-oxide/intrinsics/probes/fma_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_f64.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_f64.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_f64.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll | 2 +- cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll | 2 +- .../intrinsics/probes/grid_dependency_launch_dependents.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dependency_wait.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_x.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_y.ll | 2 +- cuda-oxide/intrinsics/probes/grid_dim_z.ll | 2 +- cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/lane_id.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_eq.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_ge.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_gt.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_le.ll | 2 +- cuda-oxide/intrinsics/probes/lanemask_lt.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll | 2 +- .../probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll | 2 +- .../probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll | 2 +- .../intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll | 2 +- cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/lg2_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/map_shared_rank.ll | 2 +- cuda-oxide/intrinsics/probes/match_all_i64_sync.ll | 2 +- cuda-oxide/intrinsics/probes/match_all_sync.ll | 2 +- cuda-oxide/intrinsics/probes/match_any_i64_sync.ll | 2 +- cuda-oxide/intrinsics/probes/match_any_sync.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll | 2 +- .../intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/max_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/max_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_u16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll | 2 +- cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_arrive.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll | 2 +- .../intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_init.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_inval.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll | 2 +- cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll | 2 +- .../intrinsics/probes/mbarrier_try_wait_parity_cluster.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll | 2 +- .../intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/min_relu_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_relu_s32.ll | 2 +- cuda-oxide/intrinsics/probes/min_s16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_u16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll | 2 +- cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll | 2 +- .../intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll | 2 +- .../intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll | 2 +- .../intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll | 2 +- .../probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll | 2 +- .../intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll | 2 +- .../intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll | 2 +- .../intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll | 2 +- .../probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll | 2 +- .../probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll | 2 +- cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll | 2 +- .../probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll | 2 +- ...mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll | 2 +- ...dered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll | 2 +- .../mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll | 2 +- .../probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll | 2 +- cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/mul_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_f64.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_f64.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_f64.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/nanosleep.ll | 2 +- cuda-oxide/intrinsics/probes/neg_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/neg_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/pmevent.ll | 2 +- cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll | 2 +- cuda-oxide/intrinsics/probes/prmt.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_b4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecl.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_ecr.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_f4e.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc16.ll | 2 +- cuda-oxide/intrinsics/probes/prmt_rc8.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_add.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_and.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_or.ll | 2 +- cuda-oxide/intrinsics/probes/redux_sync_xor.ll | 2 +- cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll | 2 +- cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll | 2 +- cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_down_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_up_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll | 2 +- cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll | 2 +- cuda-oxide/intrinsics/probes/sin_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll | 2 +- cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll | 2 +- cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll | 2 +- cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll | 2 +- cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll | 2 +- cuda-oxide/intrinsics/probes/sub_bf16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f16x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll | 2 +- cuda-oxide/intrinsics/probes/sync_mask.ll | 2 +- cuda-oxide/intrinsics/probes/sync_threads.ll | 2 +- cuda-oxide/intrinsics/probes/tanh_approx_f32.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_alloc.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_commit.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll | 2 +- .../intrinsics/probes/tcgen05_commit_multicast_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_commit_shared_cluster.ll | 2 +- .../intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll | 2 +- .../intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll | 2 +- .../intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll | 2 +- .../intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll | 2 +- .../probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll | 2 +- .../intrinsics/probes/tcgen05_fence_after_thread_sync.ll | 2 +- .../intrinsics/probes/tcgen05_fence_before_thread_sync.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll | 2 +- .../intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll | 2 +- .../probes/tcgen05_mma_ws_shared_zero_col_mask.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll | 2 +- .../probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll | 2 +- .../probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll | 2 +- .../probes/tcgen05_mma_ws_tensor_zero_col_mask.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll | 2 +- .../intrinsics/probes/tcgen05_relinquish_alloc_permit.ll | 2 +- .../probes/tcgen05_relinquish_alloc_permit_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll | 2 +- .../intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll | 2 +- cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll | 2 +- cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll | 2 +- .../intrinsics/probes/tensormap_replace_element_stride.ll | 2 +- .../intrinsics/probes/tensormap_replace_element_type.ll | 2 +- cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll | 2 +- .../intrinsics/probes/tensormap_replace_global_address.ll | 2 +- .../intrinsics/probes/tensormap_replace_global_dim.ll | 2 +- .../intrinsics/probes/tensormap_replace_global_stride.ll | 2 +- .../probes/tensormap_replace_interleave_layout.ll | 2 +- cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll | 2 +- .../probes/tensormap_replace_swizzle_atomicity.ll | 2 +- .../intrinsics/probes/tensormap_replace_swizzle_mode.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_x.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_y.ll | 2 +- cuda-oxide/intrinsics/probes/thread_idx_z.ll | 2 +- cuda-oxide/intrinsics/probes/threadfence.ll | 2 +- cuda-oxide/intrinsics/probes/threadfence_block.ll | 2 +- cuda-oxide/intrinsics/probes/threadfence_system.ll | 2 +- cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/trap.ll | 2 +- cuda-oxide/intrinsics/probes/uni_sync.ll | 2 +- cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll | 2 +- cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll | 2 +- cuda-oxide/intrinsics/probes/wgmma_commit_group.ll | 2 +- cuda-oxide/intrinsics/probes/wgmma_fence.ll | 2 +- cuda-oxide/intrinsics/probes/wgmma_wait_group.ll | 2 +- 1322 files changed, 1328 insertions(+), 1328 deletions(-) diff --git a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs index 0128559fea..03e050e5c8 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/async_copy.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::async_copy` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs index 80a721f3ce..2dae6c0395 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/atomic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::atomic` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs index 0708a9edc8..f452534a7a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs index cefb1c6d5d..13faf64f3d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/bf16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::bf16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/clc.rs b/cuda-oxide/crates/cuda-device/src/generated/clc.rs index 2ed6634024..70cf419843 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/clc.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::clc` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs index 8af44556d6..e227ffaee8 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs index 989446243b..fbff9c9153 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::cluster` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs index de20b34d47..54ac3d92f6 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/cluster_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::cluster`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/convert.rs b/cuda-oxide/crates/cuda-device/src/generated/convert.rs index d6ba966953..883d367a2d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/convert.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/convert.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::convert`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs index 87e3367bb4..a51ac2a765 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/counted_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs index abeecdf5c0..fc4f94874a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs index a62e6e9b68..1c38bef616 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/debug_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::debug` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs index 63efdfa946..11bcebcedd 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::dotprod` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16.rs b/cuda-oxide/crates/cuda-device/src/generated/f16.rs index 649458a29b..37bb5d2051 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs index c043673ede..e0ba7d1157 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs index ec41a6b7ff..82010ea7d7 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/f32x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::f32x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/fence.rs b/cuda-oxide/crates/cuda-device/src/generated/fence.rs index 87ada2f7fb..7025a6e77a 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/fence.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/fence.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included inside `cuda_device::fence` so existing paths stay stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/float.rs b/cuda-oxide/crates/cuda-device/src/generated/float.rs index 7ca4e794c7..bb17949ced 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/float.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::float`. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs index 4ec97e0e96..7660a74197 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_dependency.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs index dfd627ccb6..e80f63b6cc 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/grid_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::grid` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs index a08a48c30f..8079ed880c 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/i16x2.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::i16x2` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/int.rs b/cuda-oxide/crates/cuda-device/src/generated/int.rs index b8765678db..91e59c67f2 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/int.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/int.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::int` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs index 5c3be5db31..15c90d0aea 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs index e5c1c4763a..42420ffe43 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs index 495ed52b2b..499be8e9ed 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::barrier` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs index 4ba4a9decd..958ac2da4d 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs index 12676ce0b0..ebbac73f06 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/prmt.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::prmt` to keep the public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs index b40f7ec321..f45cc3ae7f 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::thread` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs index d587bc0f8a..654e3954ed 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs index 6a097b9195..8e2dd4cde0 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/shared_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::shared` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs index d3d12eeb1d..80526e716b 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wmma` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs index 982a9f8997..6f178f6014 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // This file is included lexically inside `cuda_device::thread` so existing diff --git a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs index 1d1d158e26..25577a893e 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep public paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs index a671cac636..508f5eada4 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tcgen05` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/tma.rs b/cuda-oxide/crates/cuda-device/src/generated/tma.rs index bd010dab68..8800c3a641 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/tma.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::tma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs index 68b8084af3..8f1aa708c4 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/warp_sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::warp` to keep existing paths stable. diff --git a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs index 4ae022e8cc..750c0d2d44 100644 --- a/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-device/src/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. // Included inside `cuda_device::wgmma` to keep its public API stable. diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs index 23e0148903..698a64daf9 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Raw ABI functions recognized by cuda-oxide. diff --git a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs index 230c301192..814cc0e023 100644 --- a/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs +++ b/cuda-oxide/crates/cuda-intrinsics/src/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. #[doc(hidden)] diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs index 0239d0f93d..093d3bf025 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs index 14017a9360..688a055fe9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs index 651e146c2b..2c88800e44 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs index 5082d81d49..49d2527e37 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs index 78c0cefb35..7224a18db8 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs index 1879b0b81e..48b8883a5b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs index 5120baef23..f451ca2cd1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs index e7ccbc0653..537884cd43 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs index 640730af0c..e1b206d735 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs index 6f575c2c3b..58981a1e8d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs index 209a3b3c8b..89807769de 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs index 17fb644cd0..352537c282 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs index c65cbfffbc..c8c5f39b6d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs index f81417f854..efc2c5af98 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs index 3c3e0f45fc..a9f16d3a87 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated target requirements and separately imported LLVM/selection facts. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs index e661174324..75742c9f56 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs index f247bfd93c..1d0c819e57 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs index b800661cd3..47a457cd82 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs index d3b266cd9a..d32c5ca869 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs index 9b4536096b..6e2ea55160 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs index e872cdfe36..cb61ea288e 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs index 22d53bb67d..0944315314 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs index 505c35ee94..22caa376f5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs index 88bd43dd92..45b6c0a6c0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs index 306ff4506f..12ee214154 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs index bf4b466b9b..3fad449c57 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs index 9ed4bee138..702bb761a2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs index df6be66828..ee826a9198 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs index 287b63910d..90a5a6f5cf 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs index 2b35f7db35..2eaa0883ba 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs index ea7fd7a808..3ba5b6bc94 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs index 8c0a753757..5b33dcaab4 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs index ca52c76201..966f5d1253 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs index 255e84d5d6..4391014223 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs index 414e9fdb5c..b26d920818 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs index b27eca3b68..a91888a9e9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs index e7d193b812..28ac52b10c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs index f7bcd434cf..8bfd36c97c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs index 0ed058d9ab..25eaa798b7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs index a5e77d62f3..ebcf09356a 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs index ff9d6085e7..0fba32d40b 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs index c4e30eebe6..bde2e5e030 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs index 57d66496e0..e76d955ec1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs index e209e0efcc..75ddd78bd0 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs index e1d2893771..035edab564 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs index f1c70ffed8..8e53cab669 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs index 44bee24026..02ebdb7bd2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs index c16d2378da..1dec957e1f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs index ab7cc9ad4c..6af07e4e16 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs index 04b43a40c0..8a1c0f797f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests, grouped per family shard. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs index 379e9355a0..b4d35cc175 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs index 241ccff625..4f85f04af4 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs index f7e234ac84..49d50ea99c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs index 3101685bec..e8995a5f9c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs index aa2de4858f..ab4efef6dd 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs index 6837fc80e4..fdbce3ea50 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs index 6c019aff4b..ff14c1f0f1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_float.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_float` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs index c475ea05a9..a3142f7585 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/register_mma_s32.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `register_mma_s32` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs index 3327bd58b9..5a905c322f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs index a116824677..a752dbd24f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs index c981eae547..30bf7c6ad7 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs index 62f21ab2c2..699539ac93 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs index b268a94e97..1d3657b881 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs index 7a9f6bd5a4..948959dcac 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs index 346e40c5d1..f4730fecbe 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs index f19ba6de56..dc2676f8b2 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs index f808964666..3e29504925 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs index 2340ea3d4f..8ae2ee5ada 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs index d25e0ec6cd..7ed5e540eb 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_mma_ws.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_mma_ws` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs index 0b48060bd9..4c5fc7ce95 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs index b1cfd22959..155020b735 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs index 38e7d1eed6..6468154b5c 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `tma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs index 156718acbc..7aadebb14d 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs index f148c48874..d1faf65d46 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs index 605d30ab64..404e804d43 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs index b923ae9c9b..d69610fcb5 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs index df33b68cd6..86ec25cef1 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs index f4651cbde3..ac7d58f3ee 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `tma` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs index 69b2f96e1d..4bfa775523 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `vote` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs index 59b6c2ae6d..cccaf4dbd6 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs index e041b342b2..000d21b379 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs index 6c3bddfa4d..878925151f 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs index bbd66f64b3..30055018e9 100644 --- a/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs +++ b/cuda-oxide/crates/cuda-oxide-codegen/src/generated_intrinsic_targets/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated intrinsic target records: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs index 90f7916a8f..a515bdbbff 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for the generated active warp mask. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs index 5475615b82..4d251c3589 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Cluster Launch Control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs index ed6cc1381a..73b38afd48 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated operations for the closed cluster-barrier family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs index 20aeaebdfe..b6a6c99840 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for cluster address mapping and remote shared reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs index 5eee58ea50..553c161a53 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for classic `cp.async` instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs index 78e38fcd7c..101fc85f8a 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated PTX debug controls. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs index fe732aae38..4acdd93b37 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed integer dot products. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs index 40f02ec97a..a8612264f5 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated warp leader-election operation. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs index 00d1514966..f998bbb13f 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated counted-barrier, grid-dependency, and register-control operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs index b7411e2d32..e9f503bb0b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for extended floating-point min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs index 4174cddf08..7013c998e5 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated extended integer min/max. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs index f6dcda6e8a..74ec271c15 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation and compatibility carriers for generated `ldmatrix`. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs index bae82bb165..ea844c8cf2 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the basic shared-memory mbarrier lifecycle. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs index c44e9fd3fe..d5de90d50f 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for extended mbarrier and async-proxy instructions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs index c36b45f001..2cb7aff92e 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. mod active_mask; diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs index 36800bcc18..88980724ce 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated in-register matrix transpose. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs index 05428f0b16..51e660245f 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated packed floating-point arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs index 479315ec2a..8c19c30ae9 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated packed-atomic family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs index 9633dcfd80..fa83a5345f 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated packed conversion. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs index e193f30c27..f147341d0b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for the closed generated `prmt` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs index 889fcfee97..c6bcdbf3bc 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for the closed generated `redux.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs index 490bdc8225..4767e487c6 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated register-only `mma.sync` variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs index de2948ccae..fa498fbe9b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar arithmetic. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs index d8564f2a78..d6a767a80c 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated `f32` to TF32 conversions. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs index 781cf1fb11..0cdce1daac 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated scalar math. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs index 7b39943b66..9aaa2f04c4 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! One structural operation for generated sparse MMA variants. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs index 05bd65db3a..f63a3efacb 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural NVVM operations for generated special-register reads. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs index 784b724660..35fad6b583 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the four generated `stmatrix` stores. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs index 997075c845..212c36f829 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated CTA synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs index b380cab8a0..88b913dea1 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Core Generation 5 operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs index 94626d7d14..de5e56756b 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated Tensor Memory Accelerator operations. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs index 1539afbe91..02e65111a9 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `vote.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs index afb62c4aa6..0698e06b23 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operation for generated warp synchronization. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs index 8dae7791b7..4472cf1ca4 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `match.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs index 51cb6b7994..611cf8e798 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for the generated `shfl.sync` family. diff --git a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs index d937a15ccb..5733dd83cb 100644 --- a/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs +++ b/cuda-oxide/crates/dialect-nvvm/src/ops/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Structural operations for generated WGMMA controls. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs index 3f6aeee7bb..ce0a280e2c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs index 49c78c077b..a1e5adb5a9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs index bbf3a8c0f5..05c3918a4a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs index 351178c7d1..c11ddf6108 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs index 176d7f4b64..3c840f9b5a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs index 0df994d857..a9e4e19ff4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs index 4406504628..3e7c7620da 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs index 3a8d723925..59f961cb33 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs index 00547b63e7..ee0c6b7551 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs index ceb084a673..967a901883 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs index d9cf5bf71c..9a1899a62f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/helpers.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Shared glue for the generated intrinsic dispatch shards. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs index 01e235616d..6f3fdd86f0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs index ac8828b5b9..7dfe1ccfb0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs index f8cbbade8d..eb53af9b1e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs index 979e76f5db..ad552194d0 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs index 6a09c4c3fb..9c980fca9e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility path dispatch for CUDA intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs index f9e8b85921..2384dc6f85 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs index 0be9ab755d..480526765e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs index af0a248b90..32377c7d40 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs index 54eba69310..b7cbacf93b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs index bee646d75f..a15e3196ad 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs index 49df85ea6a..9a630b3ed3 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs index 2b1b3a2b3d..47ad59a700 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs index 910c5ab14c..cbdf2bbda6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs index 4bce17f6dd..9e08cbdc77 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs index d26be5029e..93fc0784ea 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs index 475dc8fcd5..993ac1a84f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs index 0c1737f3fc..97c70d8889 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs index 691aeb3413..a0f41e97f6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs index 141dc6b556..484c1a6a74 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs index e1dbe30e4e..9aa3a1c92e 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_cp.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_cp` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs index 662cd78b23..d47e6abb1d 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_ld.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_ld` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs index 46fd1a1bb2..eec859a000 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs index c6400fff3a..223a7e2117 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_other.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_other` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs index 5f273cf54a..599f8c5dfe 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tcgen05_st.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tcgen05_st` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs index f39ec2e19d..c24ccc905f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs index 0b531d4792..da599ce05c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs index bb0a33e75f..a96747bd8c 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs index 1b606f30c8..6218b18f60 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs index cfff78569c..5b14990157 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs index 079705b6bf..160cdd2547 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs index e6801830e9..481dcfbe08 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs index 5aa654aef2..c73be2a520 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs index 91c42098b8..57351d2b92 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs index cb190aa51e..b7ce30d142 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs index c7393d698b..87f98b4dfd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs index 47a1f09a81..da129938f1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs index c6804c4e81..a6235ba466 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs index 58d0ca9486..45f494fe2f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs index 4c11633be2..65795232e5 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests, grouped per intrinsic family. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs index 67c1cbd993..cb55540952 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs index 5104d5f877..1f01e82de2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs index ebe0a721aa..05b1990903 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs index 784c00add8..311c4f9703 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs index b9c4979cdc..0928fdfbcc 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs index c590ba15a9..cbe3751ad6 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs index f34812b2da..38a8944891 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs index 7363ce86a8..11eebceed1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs index bf90a576c1..35a1416a22 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs index 6c8bb2a9b1..a6efe97fe2 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs index 2398d0be33..375f4cddd1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs index 6f1359c171..62cb47f775 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs index 56ecaaeba4..7d2caa1ac9 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs index 1eaba2ee1e..0dad5bf19f 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs index 99e36d57f8..3eb2414b71 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs index 750590f4f5..7cac99888a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `tma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs index 3e8f73da4e..26513fe1f1 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs index 9b497724fa..1f5103a1e4 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs index 51f04ca4ea..8db832c682 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs index da0961b523..f8f1fd992a 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs index 5e4f8b7bed..97d4b2cacd 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tests/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated dispatch-path tests: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs index 46e487db10..8059ee6882 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `tma` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs index 2b386d159a..5cbe5f53af 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs index 6e235b57ca..c4e3307b00 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs index d54b3f3591..8b562adb4b 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs index ff9bb1b22e..bbb15ace98 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs index 3b14eb12cd..a9d7fadebc 100644 --- a/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs +++ b/cuda-oxide/crates/mir-importer/src/translator/terminator/intrinsics/generated/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated raw/compatibility dispatch arms: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs index 47afeda00f..30c8776679 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/active_mask.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `active_mask` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs index 0ae755b455..8fa8586470 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/clc.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `clc` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs index efea0f0dc6..4e2734e8d0 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs index 76b36a28fa..2829f83d49 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cluster_memory.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cluster_memory` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs index 393fa04892..2bc5b88829 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/cp_async.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `cp_async` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs index 7a43837d90..5813b36f72 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/debug_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `debug_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs index ca35f32b72..264a05534e 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/dotprod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `dotprod` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs index d9c1efce64..efc66dd8b1 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/elect.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `elect` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs index 8d776c191e..73face4f76 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/execution_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `execution_control` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs index da0ed6cc69..f128e318b5 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/extended_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `extended_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs index 2c73ad62c8..a35c8be28e 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/integer_minmax.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `integer_minmax` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs index aa28f4755b..580155604d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/ldmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `ldmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs index 02da2ac579..f54e78bfa2 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_basic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_basic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs index 185d3e6157..916f51a07e 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mbarrier_extended.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `mbarrier_extended` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs index 571e2ca51c..bfb759ffac 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/mod.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces for admitted CUDA intrinsic families. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs index 4cce932d5e..49e86c6d9c 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/movmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `movmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs index a9c7610f67..a86511b63d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_alu.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_alu` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs index f8630d9c01..87d5e3e381 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_atomic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_atomic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs index ec8d3baa5d..5f778e31c9 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/packed_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `packed_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs index 638d253285..8e843de47f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/prmt.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `prmt` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs index d336c85a70..e8106051b9 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/redux.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `redux` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs index 415ed85300..719f6e4c06 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/register_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `register_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs index d9fdb8d3c6..8fe82b112f 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_arithmetic.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_arithmetic` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs index 0b3196f3c5..278f2e00d4 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_conversion.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_conversion` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs index 87a83cf928..eeda1b07fb 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/scalar_math.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `scalar_math` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs index 09f7a16c2f..300f1ed21d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sparse_mma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sparse_mma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs index c478472b0b..1989dcc8b6 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sreg.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sreg` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs index 9e99d1772d..7b6f25b870 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/stmatrix.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `stmatrix` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs index 24d6739251..f6bea5c034 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/sync.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `sync` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs index 153f30e4ff..cdc12343b7 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tcgen05.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tcgen05` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs index be7622b45d..d7306ca5bb 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/tma.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `tma` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs index 5c84bc3bc5..8ec94c0b3d 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/vote.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `vote` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs index c1fdc553d8..cfebdc1122 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_barrier.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_barrier` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs index febf994684..3516d587b3 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_match.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_match` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs index 547b472e34..2a8c876e71 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/warp_shuffle.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `warp_shuffle` intrinsics. diff --git a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs index 0d1129d530..8fb47633da 100644 --- a/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs +++ b/cuda-oxide/crates/mir-lower/src/convert/generated_intrinsics/wgmma_control.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated conversion interfaces: `wgmma_control` intrinsics. diff --git a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs index 5215dc378c..9ee8fb794a 100644 --- a/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs +++ b/cuda-oxide/crates/rustc-codegen-cuda/src/generated_intrinsics.rs @@ -4,7 +4,7 @@ */ // @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. //! Generated collector predicates for intrinsic placeholder functions. diff --git a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md index ecf6910e32..96edf23871 100644 --- a/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md +++ b/cuda-oxide/cuda-oxide-book/compiler/mlir-dialects.md @@ -303,11 +303,11 @@ they become `call` instructions to `@llvm.nvvm.*` intrinsics. ### Architecture Coverage -At catalog SHA-256 `3f49098f` (the stamp in every `ops/generated/` file +At catalog SHA-256 `31b52388` (the stamp in every `ops/generated/` file header), the dialect holds 580 operations across 42 modules, and they come from two different places. The split is the first thing to know about it, because it decides where -- and whether -- you would add one. If the header -stamp no longer starts with `3f49098f`, the counts on this page predate the +stamp no longer starts with `31b52388`, the counts on this page predate the catalog you are reading. **Hand-written**, directly under `cuda-oxide/crates/dialect-nvvm/src/ops/`. These are the diff --git a/cuda-oxide/intrinsics/catalog.json b/cuda-oxide/intrinsics/catalog.json index a984bc87c4..52aea22d8d 100644 --- a/cuda-oxide/intrinsics/catalog.json +++ b/cuda-oxide/intrinsics/catalog.json @@ -89,7 +89,7 @@ "rust-llvm-23.1.3-aab11ee0-integer-minmax.json:94bfd0d84bf57737acda19278a10cb566c25677c5688e6a918c83da822100480", "rust-llvm-23.1.3-aab11ee0-mbarrier-basic.json:193966219733100689a43dfde7e9f8aba1f3e3ba2cf80b14bd4893504f900489", "rust-llvm-23.1.3-aab11ee0-mbarrier-extended.json:c92addb73ebd8ef4f3bb2d5902a9e7f280ba898d10b54682d44f26cd2968fcd8", - "rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json:4ee208e5c7bf8891f58e193593f763dd32f1005dbbcb153e99fd65e13de2177c", + "rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json:2e4a3ce7e895c4a39ea9fccd7b9580ce5e87d2e900cc5ea5853588d836bd12e3", "rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-f16x2.json:20124138104fe9f71a583a73cb8656c7d91aa10366e9e041499e024578c1a2e3", "rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8.json:42561b5ee707545ccabe2b3b78de12a31614893611e0deccbb498a3ba2967fd1", "rust-llvm-23.1.3-aab11ee0-prmt.json:c77aaaa7fed2d55c2319e286f092bc440e34ef6e6aaf08abe712a3179745d990", @@ -46416,7 +46416,7 @@ "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e5m2x2.bf16x2' not supported on .target 'sm_120'\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", @@ -48345,7 +48345,7 @@ "stage": "ptx_assembly", "mechanism": "inline_ptx", "outcome": "succeeded", - "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e5m2x2.bf16x2' not supported on .target 'sm_120'\".", "artifact_kind": "cubin", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", diff --git a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json index 67f1816041..8f7e7258b8 100644 --- a/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json +++ b/cuda-oxide/intrinsics/evidence/rust-llvm-23.1.3-aab11ee0-packed-conversion-fp8-bf16x2.json @@ -296,7 +296,7 @@ "mechanism": "inline_ptx", "outcome": "succeeded", "artifact_kind": "cubin", - "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e5m2x2.bf16x2' not supported on .target 'sm_120'\".", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" @@ -401,7 +401,7 @@ "mechanism": "inline_ptx", "outcome": "succeeded", "artifact_kind": "cubin", - "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'\".", + "detail": "CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with \"Feature 'cvt.e5m2x2.bf16x2' not supported on .target 'sm_120'\".", "tool_path": "/usr/local/cuda-13.3/bin/ptxas", "tool_version": "Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0", "tool_sha256": "afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc" diff --git a/cuda-oxide/intrinsics/generated-reference.md b/cuda-oxide/intrinsics/generated-reference.md index 6cc83d5aa3..fbcbe1298e 100644 --- a/cuda-oxide/intrinsics/generated-reference.md +++ b/cuda-oxide/intrinsics/generated-reference.md @@ -2,7 +2,7 @@ SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. SPDX-License-Identifier: Apache-2.0 @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. --> @@ -8193,7 +8193,7 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 367-byte typed fixture (SHA-256 7b4acbf67e3c0f83b1099150d70876603349db6db1b1502dce79da9b702b04f6), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.satfinite on a dynamic <2 x bfloat> argument, to 745-byte PTX (SHA-256 3334d10e01ea5e8f531e764a85e85f87cc2d0bc378f9bb38f9ef8a934ac54614; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with register operands. - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.E5M2.BF16 instruction. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 210-byte inline-PTX fixture (SHA-256 fff4ab63076b1daff9772ebd8744e62bebb0f71d8c9b573111157b2125c4e96b; the body of intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll without its comment header) to 749-byte PTX (SHA-256 bdfc6e8b024796f9b5081893885b7d3a34fb36bc97e07fad8481bab56b9eab79; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result. - - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e5m2x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73`, SHA-256 `e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77`) - backend codegen on `compute_120f, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4). - PTX assembly on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. @@ -8259,7 +8259,7 @@ This table is generated from the resolved catalog. Evidence stages below disting - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 382-byte typed fixture (SHA-256 531ef4ab3bc1e551c3a2b5fc5d646a67f522a454b64cb47bcd0ac7617c73a1d3), one call of @llvm.nvvm.bf16x2.to.e5m2x2.rn.relu.satfinite on a dynamic <2 x bfloat> argument, to 780-byte PTX (SHA-256 460efd987919d91caf04ca68e9223b4d22c4d37872164593799122bd0ea8d764; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with register operands. - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the typed-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d). ptxas keeps no code for an unreferenced non-entry function, so this cubin is byte-identical to the inline route's; the relocatable objects from ptxas -c keep the function, and both routes carry the same F2FP.SATFINITE.RELU.E5M2.BF16 instruction. Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. - backend codegen on `sm_120f, ptx91`: `succeeded` — Recorded with nightly-2026-10-11 (rustc 6f5da406b56c2db6607202b63daee73aa8c61297; LLVM aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019): llc -march=nvptx64 -mcpu=sm_120f -mattr=+ptx91 lowered the 220-byte inline-PTX fixture (SHA-256 052e0049faabfaae45bfd2f5e74607a3a39c2f40cd108c6656a00104c933e207; the body of intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll without its comment header) to 784-byte PTX (SHA-256 8946110f2d12fa0239e4132675f8f6facd71e80dfbba7e6cf40d010cdba91e96; .version 9.1, .target sm_120f) holding one cvt.rn.satfinite.relu.e5m2x2.bf16x2 with the =h,r constraint: one packed i32 source and an i16 result. - - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. + - PTX assembly on `sm_120f, ptx91`: `succeeded` — CUDA 13.3 ptxas assembled the inline-route PTX at sm_120f into a 2,240-byte cubin (SHA-256 ca980ee914906d30713cf21fb303b36ab9266b98c76fc0637ac71dee040e294d), byte-identical to the typed route's cubin. The same PTX retargeted to sm_120 is rejected with "Feature 'cvt.e5m2x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. - `lib_nvvm` uses `inline_ptx` from profile `cuda-13.3-libnvvm-13.3.73` at PTX 9.1 / sm_100a or sm_100f or sm_103a or sm_103f or sm_110a or sm_110f or sm_120a or sm_120f or sm_121a or sm_121f: status `validated` (`CUDA toolkit 13.3; nvcc/libNVVM release V13.3.73`, SHA-256 `e55af1135e987372355d19d66cb74e1a1e99c32ac98ffad14fb567eef16e4f77`) - backend codegen on `compute_120f, ptx93`: `succeeded` — CUDA 13.3 libNVVM compiled the 1,807-byte NVVM IR fixture (SHA-256 83e41481b170703c62b0ff63f647cb04b4e3e9b4c12b7c16a9fe6ad2d0a64bc9; four kernels pinned through llvm.used, each one inline-PTX call with a single packed i32 source operand) with -arch=compute_120f to 2,485-byte PTX (SHA-256 858cee379e200e21b36c1fbe0b76af425f37cc4057300ac9da7d0be9d873b62d; .version 9.3, .target sm_120f) retaining all four instructions verbatim, and with -gen-lto to 3,228-byte LTOIR (SHA-256 8c47b71a4f042ccb6713381bd52c58c4d05e74d910b16f19cade2d0c688e4ce4). - PTX assembly on `sm_120f, ptx93`: `succeeded` — CUDA 13.3 ptxas assembled the libNVVM PTX at sm_120f into a 13,768-byte cubin (SHA-256 f0a9b794cfd3da2368b5576e6bd99f747d82c0f2d7d60dfb5fa1c2241f925076) whose SASS holds one F2FP.SATFINITE[.RELU].{E4M3,E5M2}.BF16 per kernel, F2FP.SATFINITE.RELU.E5M2.BF16 for this form. The same fixture compiled for compute_120 is rejected at sm_120 with "Feature 'cvt.e4m3x2.bf16x2' not supported on .target 'sm_120'". Tool `/usr/local/cuda-13.3/bin/ptxas` reports `Cuda compilation tools, release 13.3, V13.3.73; build cuda_13.3.r13.3/compiler.38244171_0` (SHA-256 `afd8d1e1fa6e310f7faee44f6621e4c1315fb7fd6da7d4d87414358e12a651dc`). Artifact `cubin`. diff --git a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll index 3e0f9d76c9..067d4c64f9 100644 --- a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/abs_f16x2.ll b/cuda-oxide/intrinsics/probes/abs_f16x2.ll index b6bc6d7d37..6dae85758d 100644 --- a/cuda-oxide/intrinsics/probes/abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/active_mask.ll b/cuda-oxide/intrinsics/probes/active_mask.ll index 6330e0855e..7fa13f210e 100644 --- a/cuda-oxide/intrinsics/probes/active_mask.ll +++ b/cuda-oxide/intrinsics/probes/active_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_bf16x2.ll b/cuda-oxide/intrinsics/probes/add_bf16x2.ll index 03da325f33..9a18789357 100644 --- a/cuda-oxide/intrinsics/probes/add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f16x2.ll b/cuda-oxide/intrinsics/probes/add_f16x2.ll index 3a49e6fa50..fc61781575 100644 --- a/cuda-oxide/intrinsics/probes/add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_f32x2.ll b/cuda-oxide/intrinsics/probes/add_f32x2.ll index ee6a2e9c23..37d3007c21 100644 --- a/cuda-oxide/intrinsics/probes/add_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll index a4c7bb236d..57b115569d 100644 --- a/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/add_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_f32.ll index 4ba9271e68..7052963bf2 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_f64.ll b/cuda-oxide/intrinsics/probes/add_rm_f64.ll index 35e9de9fc8..93e78e7647 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll index 65bfa0e99d..0f4fb6b035 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll index 1b82320433..90ee605e4d 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll index d9339dff5e..d69fbd14af 100644 --- a/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_f32.ll index 1eb51f7336..73ed7cd3b8 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_f64.ll b/cuda-oxide/intrinsics/probes/add_rn_f64.ll index bf3668f6d4..3a6786a3c5 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll index a1012daaaa..9973ab1f59 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll index 4341bc7d5f..4711626207 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll index ab4add9da4..a8030adfb4 100644 --- a/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_f32.ll index a9c41dc1f2..13a461bd4a 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_f64.ll b/cuda-oxide/intrinsics/probes/add_rp_f64.ll index 0c8ddd6454..dd49cf5236 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll index 3c28880f31..8cf544e9fc 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll index 42e3cd380c..6c24875b2c 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll index 66ef200a5d..4fc32f3f88 100644 --- a/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_f32.ll index 493bdc68ad..6ac13a58b0 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_f64.ll b/cuda-oxide/intrinsics/probes/add_rz_f64.ll index 831859fc46..2af8a3023f 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll index 6678cc4b68..2bcdc125ce 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll index 83aee309e7..355e301756 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll index 8a04fa233f..094f9f6deb 100644 --- a/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/add_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/all_sync.ll b/cuda-oxide/intrinsics/probes/all_sync.ll index 98304e37e4..29d10663de 100644 --- a/cuda-oxide/intrinsics/probes/all_sync.ll +++ b/cuda-oxide/intrinsics/probes/all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/any_sync.ll b/cuda-oxide/intrinsics/probes/any_sync.ll index 6bc242fd8b..576cf8dd34 100644 --- a/cuda-oxide/intrinsics/probes/any_sync.ll +++ b/cuda-oxide/intrinsics/probes/any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ballot_sync.ll b/cuda-oxide/intrinsics/probes/ballot_sync.ll index 341c46747a..8befb7faae 100644 --- a/cuda-oxide/intrinsics/probes/ballot_sync.ll +++ b/cuda-oxide/intrinsics/probes/ballot_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll index 218967e072..cd1c453396 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll index 6eaa0cb693..5199ada72c 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll index 764c755711..2c4ba7b688 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll index b553621ab5..2665f273a9 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_arrive_relaxed_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll index 28a4402acd..e0304517aa 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll index db275c5ef1..fb0f319061 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cluster_wait_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll index 1e2aa12703..f527f256a7 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll index 975fd5eb7a..a79bf1924d 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_arrive_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll index d58f366099..b53080cc4d 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll index 2788a846e7..8cfb36aa36 100644 --- a/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll +++ b/cuda-oxide/intrinsics/probes/barrier_cta_sync_aligned.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_x.ll b/cuda-oxide/intrinsics/probes/block_dim_x.ll index 0f78ffd65b..e27224c43b 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_y.ll b/cuda-oxide/intrinsics/probes/block_dim_y.ll index ec11f16ef9..078e0407fb 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_dim_z.ll b/cuda-oxide/intrinsics/probes/block_dim_z.ll index 6a6d48d592..328f912a0f 100644 --- a/cuda-oxide/intrinsics/probes/block_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/block_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_x.ll b/cuda-oxide/intrinsics/probes/block_idx_x.ll index 982679de95..d3f10ee9b5 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_y.ll b/cuda-oxide/intrinsics/probes/block_idx_y.ll index 208459e386..393af52146 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/block_idx_z.ll b/cuda-oxide/intrinsics/probes/block_idx_z.ll index e26a0c5125..7e46684f4d 100644 --- a/cuda-oxide/intrinsics/probes/block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/breakpoint.ll b/cuda-oxide/intrinsics/probes/breakpoint.ll index 8c1cde8219..06a2c469b1 100644 --- a/cuda-oxide/intrinsics/probes/breakpoint.ll +++ b/cuda-oxide/intrinsics/probes/breakpoint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll index cc415c4334..f743449a72 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll index 10cb0c67d4..03b30c409f 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll index 7c1cc0c076..f7e0cdfe75 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_get_first_ctaid_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll index 900b76251a..3fb1cef4c0 100644 --- a/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll +++ b/cuda-oxide/intrinsics/probes/clc_query_is_canceled.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll index d4f441c139..c4c08feb4d 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll index d4cad92630..68c4df50e9 100644 --- a/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll +++ b/cuda-oxide/intrinsics/probes/clc_try_cancel_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll index 363c8d3e46..386bf7683c 100644 --- a/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll index ec08138839..4c04d476de 100644 --- a/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll index 6f0dea61de..f29a9214a6 100644 --- a/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/clock64.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll index 85f3b12ece..0c7b38daa7 100644 --- a/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/clock64.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_count.ll b/cuda-oxide/intrinsics/probes/cluster_block_count.ll index 1280cedf7c..d244f044ae 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_count.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_count.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll index e52fa68835..7488d448ac 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll index 7e9b045703..7376e4d983 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll index 57b3773500..1a2fc34785 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll index 6faa8e0b0b..440c88061e 100644 --- a/cuda-oxide/intrinsics/probes/cluster_block_rank.ll +++ b/cuda-oxide/intrinsics/probes/cluster_block_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll index 63432523b6..7230c4358b 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll index c78ca2663c..7c4e5810ce 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll index 66eae31678..e7c273060f 100644 --- a/cuda-oxide/intrinsics/probes/cluster_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll index ff06904ee0..1dd35af076 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll index ed98bfce87..9508406af5 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll index db2904ec18..910ee9bc3c 100644 --- a/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll index 7cb52c49f2..8e799974cb 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll index 2ae587971f..a445fe11c2 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll index f8d220a379..64bab3b053 100644 --- a/cuda-oxide/intrinsics/probes/cluster_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/cluster_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll index bcb1d5751a..17c45671ba 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll index 0b7a304b04..c00c249aff 100644 --- a/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/cos_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll index 03a79696ce..aaec000da3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll index 1ea8185454..46e442df90 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll index 61a0410a5c..9f65921817 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_1d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll index 0c79f8cd6b..454680f011 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll index 3084413e82..68cde3cd85 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll index 0cd4810dee..3a129e2a15 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll index 42c683f82f..2b171015e8 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_3d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll index bb10981d07..739d1e76ae 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll index 0fc1433183..a7443c9ee5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_4d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll index d743f2e5cd..6f1160d85b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll index a5bda09637..bf99c97424 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_5d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll index 08d3af97be..c1adbd1564 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll index 7a30b43ffa..f499d5ff07 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_prefetch_tensor_gather4_2d_l2_cache_hint.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll index 2cdf67cefa..5f6cd7d1a6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll index bf8cc52010..26c100136e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_1d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll index 4d1ea76468..e5287efe79 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll index 0becd87d2d..8e432bfd63 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll index 907e2987fb..d4c22e1e0b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_g2s_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll index 457b677f32..ff4a9d96c7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_2d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll index 8ab9a6cfd9..f4e59a83b7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll index 8e89b8d8aa..e360ec5554 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_3d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll index 52e47b4c86..734c32c344 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll index ccd207661a..78aa88746a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_4d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll index 25d9b4418c..59c4581047 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_g2s.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll index 0662f26544..08a7b47bf5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_5d_s2g.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll index e45c58974b..a328c13b54 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll index a2059a313e..47a90c1f91 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll index 3b20c4a703..b440cc9a66 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll index 9a2d168277..42c8c3406e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll index dd11181238..af359d940b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll index 07202036e0..4ad7a9bbd9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll index 6cdfdde1da..cdf1812920 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll index ea4feba388..69068f4348 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_add_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll index 2f35825a32..39b8079ff1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll index cf2a3a4d44..a8488e950b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll index 980cd55253..e9db974bc0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll index 19cb808a91..a1bfb7b666 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll index 995ad309ed..e53fff68a4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll index 60d9f2f995..8edc1a7962 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll index fe7db8df8d..3454107c84 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll index 2de372324f..a93faef3d6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_and_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll index 63b7a7b372..81cd03fd9f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll index bb99f7e743..36750c05df 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll index 7f27588763..460a9bca6d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll index 5eec5367d4..4e2cc0c1e6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll index 5609ec2e09..a4dcb9af51 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll index aad1f12fe6..9911f2ff92 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll index b6a94fb21e..5d64d411f9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll index 2938f65e91..636780c003 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_dec_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll index a43b169994..fcfd9f8f71 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll index 2691a872e7..851b12833b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll index d10df69058..57eba85ca6 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll index 135cb9a328..d442347227 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll index 90ee7c4178..59875f1bd9 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll index de77107cbe..9aea0c597c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll index a08d407ae7..d1717d0542 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll index 744cbcf557..686b3a3452 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_inc_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll index 8bd5f7dace..8ac11f8adb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll index 2226a2b728..3eba84995e 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll index d034017638..82af3065da 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll index ae38e12568..18f95c6b3d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll index f42f506f71..e3f194e2e7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll index d90bebfdec..5e9dc67265 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll index 4862fe3c4a..bab5881b9a 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll index ec60a9befb..7ed9a51e1d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_max_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll index b4f000f1a6..b2f6a782cb 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll index 44970f40b6..e3ff3c3dbe 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll index 83d3e67ed8..95f3385238 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll index 3b6901ec98..a0f4588257 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll index f6381878cf..d02c1e355d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll index daf7bd8a6e..83f57a71a7 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll index d645b7a39c..adab16d057 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll index e25c10a3f8..d5b4f9f0c4 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_min_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll index 0b9ec53151..feaf9c4b52 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll index ac63182dbe..eb6f62f03c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll index cd067a2749..2910ba2c1c 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll index f75fec1b56..e07e14ea52 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll index b953fd88e0..87f260e836 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll index dcec639cf6..4fcf98c004 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll index 2b10caac7c..e4ef06fc31 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll index 618dac9cef..d803d51e02 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_or_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll index 2add1f6c51..a885575b6f 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll index 40eae07289..a8fdaadc12 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll index cf4ee7c744..4f20aabc36 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_im2col_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll index 83e1755b32..72f78bf376 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_1d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll index 55fb9cbcea..c57638e74b 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_2d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll index e526a68e88..0b3816e9da 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_3d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll index 6e7b8b213e..0313e9fcbd 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_4d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll index 4d4926f08f..572c3abda0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_tensor_reduce_xor_tile_5d.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll index 8a37eaa584..fcb98f7e95 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll index 5090032eba..ddd4e30212 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_bulk_wait_group_read.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll index 7be8135688..92065d24e1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll index 78da69865f..152c11dce3 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll index 3ce4021f1f..0e6b58d2b0 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll index 39ddc777c4..91eed5c738 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll index 4065240a3a..0b45e2547d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll index c06f455a2a..aa2bfb45af 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_ca_zfill_8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll index b5031711fe..6b81d8bd58 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll index b993a12337..c6fb788be1 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_cg_zfill_16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll index 7cd619c5ba..9ae3549b39 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll index bdd2499c59..4255a91037 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll index d605207ed3..dcbc9a3db5 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll index 856c94a505..41db650465 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_noinc_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll index a8b5c36fa7..66ce94af51 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_mbarrier_arrive_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll index 4d65bdf6c1..4429f3f902 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_all.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll index 1b564ae5a1..ed8825d63d 100644 --- a/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/cp_async_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll index b03d0ed753..e87cebd799 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll index 42b38ac1f5..64c4b60d4e 100644 --- a/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_f32x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll index 71f27aba98..7dbe4d58c6 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll index 0bf5c6fd97..180687268b 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll index a6deb3a827..9aaebadeae 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll index 0f55629b8c..8a82e4f569 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e4m3x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll index 77f50cb1ab..6109e6e7b1 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_e5m2x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll index 4398157574..ca58ea76c8 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll index dca550b55b..e5e9348b81 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll index 7d8f7a0af4..e6a5e34364 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll index 1cd2fd3ae3..081544c233 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index 5c67d7990a..5e934210d7 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll index e1dc2fe2c7..5891836047 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll index 2ad7cc2dcf..0bb0c24789 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index b319dd2700..f440693c28 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll index 448fb62cfc..1e7961cb68 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll index 99e467eb9d..0b467a2c90 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll index d83ad5ca7c..a286c6d1bf 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll index 456fc35fe0..7ab9c6003e 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e4m3x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll index dc98c0ee78..299f50b531 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll index 7c1bc741ff..84aa86de91 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll index 8c0dc3c13d..cc500b7317 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_relu_e5m2x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll index 687216d915..1713149464 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll index 66302e842b..f476333136 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rn_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll index f95175648a..3229ba9b78 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll index 295360c483..db51c2bab7 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rna_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll index 36a1f8d8e2..b76bdde10a 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_bf16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll index ed8b4e3ff6..fe7d3d6937 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_f16x2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll index 6abe89727c..82ffd726e7 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll index 43a4f2ea37..74f5fc2ba1 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_relu_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll index 9f41aff4cd..63eef2db4c 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_satfinite_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll index b5f03b3ac4..419b47bfc3 100644 --- a/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll +++ b/cuda-oxide/intrinsics/probes/cvt_rz_tf32_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_f32.ll index 474edc3d24..a1d3da6717 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_f64.ll b/cuda-oxide/intrinsics/probes/div_rm_f64.ll index 886d44e0d9..00c593b7cf 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll index dbc9679295..0f51f288c5 100644 --- a/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_f32.ll index c4dc8cdd0f..da743c6ea0 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_f64.ll b/cuda-oxide/intrinsics/probes/div_rn_f64.ll index cd82155f61..47f2055c65 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll index 2bfe028017..016ec9e93a 100644 --- a/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_f32.ll index ca0ea87891..206e3ad108 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_f64.ll b/cuda-oxide/intrinsics/probes/div_rp_f64.ll index 06dffdb5ac..6d5d138bff 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll index a349395bb4..3433f00e97 100644 --- a/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_f32.ll index aecba7e73b..278d9fb167 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_f64.ll b/cuda-oxide/intrinsics/probes/div_rz_f64.ll index 734b475a56..3369341aec 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll index 673fc0cbac..447898907d 100644 --- a/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/div_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_s32.ll b/cuda-oxide/intrinsics/probes/dp2a_s32.ll index e704dc8408..33b9322ca8 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp2a_u32.ll b/cuda-oxide/intrinsics/probes/dp2a_u32.ll index 568c6d6b21..5c4f90aac8 100644 --- a/cuda-oxide/intrinsics/probes/dp2a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp2a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_s32.ll b/cuda-oxide/intrinsics/probes/dp4a_s32.ll index 84347a5193..8030171ce9 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_s32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dp4a_u32.ll b/cuda-oxide/intrinsics/probes/dp4a_u32.ll index d05020f0d8..7259fd2a04 100644 --- a/cuda-oxide/intrinsics/probes/dp4a_u32.ll +++ b/cuda-oxide/intrinsics/probes/dp4a_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll index 037936b01e..fcd64174aa 100644 --- a/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll +++ b/cuda-oxide/intrinsics/probes/dsmem_read_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll index 293040fac1..c0947ece4e 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll index 4336a92fdf..b5b176f529 100644 --- a/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/dynamic_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll index 3fb815716b..b1113d8083 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll index 0b45c6e502..27907c641f 100644 --- a/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/elect_sync.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll index 3f15df0ab5..5a813a2d8c 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll index 070ea46978..70ecb71ba2 100644 --- a/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg1.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll index 6a1d97147d..3f3050d23f 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll index f7305f533e..e61b17ab3e 100644 --- a/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/envreg2.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll index dc024a3991..a19c13f319 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll index b6cb344b3e..56f292188f 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll index 17aa0c8c6f..64c080d159 100644 --- a/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/ex2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll index b89437f96b..21aa09cb06 100644 --- a/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_mbarrier_init_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll index f3552319b9..0ac2068b54 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_acquire_shared_cluster_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll index 7e21a11133..5c2acf1398 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_generic_release_shared_cta_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll index b50b50d5f7..49531a6d30 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_async_shared_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll index 35f3223fed..e1422568b7 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll index 037df6fb76..336e78599e 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll index 597a4417ee..2403be316d 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll index 2a74b5fa26..49812021bc 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_acquire_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll index 1534d50108..85452598d9 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll index 2b9342a8dc..d18301c77e 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_cta.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll index b598bbcdf6..8c93d75714 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_gpu.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll index 760b62b2b2..5e17dedb6a 100644 --- a/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll +++ b/cuda-oxide/intrinsics/probes/fence_proxy_tensormap_generic_release_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll index 6025afa0f1..7ade7c9298 100644 --- a/cuda-oxide/intrinsics/probes/fma_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_f16x2.ll index 2149d5cd1a..f98be54f13 100644 --- a/cuda-oxide/intrinsics/probes/fma_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_f32x2.ll index 00bdd486e5..dd332b10eb 100644 --- a/cuda-oxide/intrinsics/probes/fma_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll index 7d6f50de10..7053c40911 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll index 6f8be8c53d..67522a4e4d 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll index 7f9dfaaed1..8e3c9eb719 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll index a11ccb6828..f0aa509ed2 100644 --- a/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_ftz_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll index d1116c437c..ed3b27aec5 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll index da30780743..1c97d0f674 100644 --- a/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_relu_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll index e0568e391a..92c934285e 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll index 4500a275eb..343fa5d580 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll index 4a17faa04c..a6ad65b13b 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll index d76442ec15..6d25197ac5 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll index 5c1e8a1ade..075296a965 100644 --- a/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rm_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll index e67634f9fe..315fa265e9 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll index 4bb416ef7b..9da6f952d3 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll index 90ba2621bb..aff9f07ff7 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll index f907ce5d43..5c471fec57 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll index ab20d0dd1f..c8c4f68e47 100644 --- a/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rn_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll index f82c12243c..399094b7fe 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll index d27695eaf1..4213b63aad 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll index 944456bf65..98cea80bd4 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll index a864a7b10d..ae1ee61678 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll index 864ca38d29..7ca0261283 100644 --- a/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rp_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll index 4df946f63c..365dea58a5 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll index 2f0df8f8b0..0e5ec433d2 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll index 8c983aba40..3188b45136 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll index 9ea06ed6c5..3a0c7b627a 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_ftz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll index 4675ce8267..9ca05876fc 100644 --- a/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll +++ b/cuda-oxide/intrinsics/probes/fma_rz_sat_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll index 6112290dbf..78b9c47da7 100644 --- a/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/fma_sat_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll index 581935764f..7d351266ff 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll index 18aeb16f78..c204923f75 100644 --- a/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/globaltimer.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll index f3fd9aeeb2..5b0764b9c7 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_launch_dependents.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll index d2d0aa8f62..6c1202bfa0 100644 --- a/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll +++ b/cuda-oxide/intrinsics/probes/grid_dependency_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_x.ll b/cuda-oxide/intrinsics/probes/grid_dim_x.ll index 865ff9cd4b..07aa2ea0b1 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_x.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_y.ll b/cuda-oxide/intrinsics/probes/grid_dim_y.ll index 488186cafa..ad96d8b8df 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_y.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/grid_dim_z.ll b/cuda-oxide/intrinsics/probes/grid_dim_z.ll index 5ea78d17fe..c28d2b2cc8 100644 --- a/cuda-oxide/intrinsics/probes/grid_dim_z.ll +++ b/cuda-oxide/intrinsics/probes/grid_dim_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll index c5281b45b8..4d46855c63 100644 --- a/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/gridid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll index e8b3d28d87..4fdc43eb18 100644 --- a/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/gridid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lane_id.ll b/cuda-oxide/intrinsics/probes/lane_id.ll index 7ca38e501c..50f6c834cc 100644 --- a/cuda-oxide/intrinsics/probes/lane_id.ll +++ b/cuda-oxide/intrinsics/probes/lane_id.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_eq.ll b/cuda-oxide/intrinsics/probes/lanemask_eq.ll index c079280e62..61843c33c7 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_eq.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_eq.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_ge.ll b/cuda-oxide/intrinsics/probes/lanemask_ge.ll index 8c86ea7505..1acedaa624 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_ge.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_ge.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_gt.ll b/cuda-oxide/intrinsics/probes/lanemask_gt.ll index 4348e071ef..1295da46e3 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_gt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_gt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_le.ll b/cuda-oxide/intrinsics/probes/lanemask_le.ll index 7c8dead4f4..e1c82790f0 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_le.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_le.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lanemask_lt.ll b/cuda-oxide/intrinsics/probes/lanemask_lt.ll index 6d0410a13c..bb04d23139 100644 --- a/cuda-oxide/intrinsics/probes/lanemask_lt.ll +++ b/cuda-oxide/intrinsics/probes/lanemask_lt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll index a966592ddd..d6c0fc68ae 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll index 6452915221..ddc3b57aff 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll index efee94aacc..3c9723fc8b 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x1_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll index cf013da91f..9c2ee7fca6 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll index 9d3d8816ed..e2fa0648b2 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll index fff4692803..c27457afd4 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m16n16_x2_trans_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll index 3bed716b2e..56bda53886 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll index 444f53c970..027688ac4b 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x1_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll index 4962be0d13..12ffeeed29 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll index bfd146333a..809f7982c5 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x2_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll index d5461078d0..6de8a3e416 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll index 782ecc56e6..894afe32fc 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n16_x4_b8x16_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll index 3d800e89c8..453adbec66 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll index f34f69f1f8..53cf01df1c 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x1_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll index 6390afffbd..f96143f6eb 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll index 40d511cb90..8d7a76b4a3 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll index 4d88abca4f..dc7eb370f2 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll index b8de72f3cc..d677b6c375 100644 --- a/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/ldmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll index ba4a7e687a..e47328e66d 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll index 332723cab1..fb08af5e0b 100644 --- a/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/lg2_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/map_shared_rank.ll b/cuda-oxide/intrinsics/probes/map_shared_rank.ll index 5ff91fc381..5358a9d506 100644 --- a/cuda-oxide/intrinsics/probes/map_shared_rank.ll +++ b/cuda-oxide/intrinsics/probes/map_shared_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll index e29e0e32b9..fdd5192dcc 100644 --- a/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_all_sync.ll b/cuda-oxide/intrinsics/probes/match_all_sync.ll index 33d79900ef..868dd4db47 100644 --- a/cuda-oxide/intrinsics/probes/match_all_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_all_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll index 94a6d8fdc5..bd6701cd5f 100644 --- a/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_i64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/match_any_sync.ll b/cuda-oxide/intrinsics/probes/match_any_sync.ll index 3f0628acae..d130a1e5c2 100644 --- a/cuda-oxide/intrinsics/probes/match_any_sync.ll +++ b/cuda-oxide/intrinsics/probes/match_any_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16.ll b/cuda-oxide/intrinsics/probes/max_bf16.ll index 3902fc5218..4c4301746f 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_bf16x2.ll index 0b299d1173..4d54c4f928 100644 --- a/cuda-oxide/intrinsics/probes/max_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16.ll b/cuda-oxide/intrinsics/probes/max_f16.ll index d834592c34..0bc3bc5009 100644 --- a/cuda-oxide/intrinsics/probes/max_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_f16x2.ll b/cuda-oxide/intrinsics/probes/max_f16x2.ll index c2274faa4b..1ee7701c9b 100644 --- a/cuda-oxide/intrinsics/probes/max_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll index a92b855da7..db80c0b0df 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll index 71f41b9ca3..9108717041 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll index 5d1f72142a..1a2b90b3bf 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll index 29533efb19..009bc5b38d 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll index e0ddc9a286..b0d5d6da4b 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll index 031961e344..b082a62ef8 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll index bba0a433f5..2725f06bbb 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll index 8eb5f6b0e3..97a59e57eb 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll index 9af8f76aa8..6f366cd931 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll index 64f6a14d32..8775a69610 100644 --- a/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll index aad24fb5fe..915471f870 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll index 89b1fa92d8..3aa7dab90f 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_f16.ll index 8fa660f5b9..8bf3acc56f 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll index 9d02d437b3..408b2a979d 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll index c70f89c623..f32ff88d10 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll index c8d09d3fec..f1c1aba39f 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll index d1913555c7..74a5ac7431 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll index 5081b9a271..776b6820e3 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll index 379e69a4c2..7d64f9b05d 100644 --- a/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll index 2f340fcf5f..26a3d223c3 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_relu_s32.ll b/cuda-oxide/intrinsics/probes/max_relu_s32.ll index ab7f9fcc30..d94f6d2429 100644 --- a/cuda-oxide/intrinsics/probes/max_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/max_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_s16x2.ll b/cuda-oxide/intrinsics/probes/max_s16x2.ll index c51a9756c3..5a7722e611 100644 --- a/cuda-oxide/intrinsics/probes/max_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_u16x2.ll b/cuda-oxide/intrinsics/probes/max_u16x2.ll index 792375a35b..c5d4473cdc 100644 --- a/cuda-oxide/intrinsics/probes/max_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll index f08e2f62bf..f1ce7e9a88 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll index f25d1ad4a1..e2a4004747 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll index 5198f93023..bf91eb2045 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll index eb75cd56f6..8f9aae2372 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll index 4c67a39488..6389bee8a4 100644 --- a/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/max_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll index e86225e1ab..a56003abd7 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll index a4fb417a93..51de19fe1a 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll index 0ed5da251a..736bb73f9a 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll index cd69e0d7f0..cf1247a804 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_expect_tx_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll index 7faa0a8371..6210d63811 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_arrive_no_complete.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_init.ll b/cuda-oxide/intrinsics/probes/mbarrier_init.ll index ed4203accc..047b17588c 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_init.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_init.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll index 6fe7acc786..baae5bf776 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_inval.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_inval.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll index 1f989d89c7..0bbf799963 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_test_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll index eaa1ef5e9e..31af9e74bf 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll index bd7930530f..b4896de30d 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll index a149046cda..6c9bdaeabe 100644 --- a/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll +++ b/cuda-oxide/intrinsics/probes/mbarrier_try_wait_parity_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16.ll b/cuda-oxide/intrinsics/probes/min_bf16.ll index 64de583ad3..14ed965fc9 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_bf16x2.ll index 84dcfb53b1..2c6070ac9e 100644 --- a/cuda-oxide/intrinsics/probes/min_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16.ll b/cuda-oxide/intrinsics/probes/min_f16.ll index 263d99e167..404f6960cb 100644 --- a/cuda-oxide/intrinsics/probes/min_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_f16x2.ll b/cuda-oxide/intrinsics/probes/min_f16x2.ll index 38578710d3..f876d127d2 100644 --- a/cuda-oxide/intrinsics/probes/min_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll index 3966303fd9..bfc0cd2f62 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll index b3e0a6a239..85d02ee233 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll index de697ae966..50132f3274 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll index f655342bb0..3481f3f285 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll index d33e190b68..06b403ec3f 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll index dfbde7e06a..7b1ebde1f8 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll index 23e434141d..dbd6dd62f5 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll index 61a7d2b802..958a82600c 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll index bb8963794c..7c2e8409fe 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll index e02521f2b0..2e1827fadd 100644 --- a/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_ftz_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll index eb71a3acf6..c110bc3aa7 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll index 66216bebe1..c3a58dec08 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_f16.ll index 919a04c666..f8192f0932 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll index 819fa57f95..f6fac96c8d 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll index 3b702ff059..75ff0178de 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll index 6d907f674c..b9b99edca6 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll index a6bc124bcd..57cf031d17 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll index 057dca1b5a..8d7e746214 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll index 21ce1da164..a798962b2c 100644 --- a/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_nan_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll index 6110af7a5a..c27c250591 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_relu_s32.ll b/cuda-oxide/intrinsics/probes/min_relu_s32.ll index 59428ad2e3..cc7b6317dd 100644 --- a/cuda-oxide/intrinsics/probes/min_relu_s32.ll +++ b/cuda-oxide/intrinsics/probes/min_relu_s32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_s16x2.ll b/cuda-oxide/intrinsics/probes/min_s16x2.ll index d2cd401912..cd633cc6d1 100644 --- a/cuda-oxide/intrinsics/probes/min_s16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_s16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_u16x2.ll b/cuda-oxide/intrinsics/probes/min_u16x2.ll index b19f7b2ccf..f41b0c3327 100644 --- a/cuda-oxide/intrinsics/probes/min_u16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_u16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll index c01c2de600..f7bde0891d 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll index 7ae84944ae..4166605972 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll index 5273270966..64bf809f51 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll index 95a2b995b7..b2c308038b 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll index 725b8a6723..f8510d0693 100644 --- a/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/min_xorsign_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll index cd44e16db7..5fddf2f973 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll index 321fe4e71f..ef355d4797 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll index 83c3d4a671..0926220c49 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll index 2b8acf90b1..bf74d189b9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll index db2e707f95..77fda80213 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll index bdee8cdff1..dfcd67fa9c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll index 37a320ecfd..7cfbe1497d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll index 882209aaef..194677aa5a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll index 29ad808028..fb49b36c7a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll index 55bd6f8fde..6535323bb5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll index 8da77c0ac3..f32e24c495 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll index 1879df218b..cebb70f36f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll index 0d12c38d3f..b5d157fdc7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll index 88e61aa6e8..60d0e78603 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll index 88d9d5d053..6546046d16 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll index 6f2d9b88a1..383bd2770f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll index 3d4b929689..4d1a732b1f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll index 3484d91f64..e5dc531dd1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll index 01d5e4c74d..ba122b4e6c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll index 060c64aef6..03bee05bdd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll index c461a1ff52..dc1b9bbf05 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll index 8892fa8f40..b5201afde4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll index 70248e4b97..476a4038dd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k256_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll index 53c67c0ce2..16f22fd9d9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll index 857a6741b5..694133d6fe 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll index e5cb84fc94..1a3a7b7054 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll index 38a9855e75..df6658f4bb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll index 6d912dd0a6..abcedf9ca2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll index cf39b34121..7e993480ed 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll index d5fb3c38d4..ba8a3d2399 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll index f8a60e545f..d2f6566bc5 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll index b6c5149673..cb58731491 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll index 0fd90e8afc..e46dc15509 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll index f7c8eb2eb2..d15a414e46 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll index 9ea3c9b99e..9e9620c221 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll index 6698d76523..c5d2cf8119 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll index e81951f3fa..65eaf27db0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll index 88a4bc5098..93e80f881a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll index edd82746f7..0c38aec935 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll index a7bade8a20..b1b3a4b4de 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll index ef0992657f..9d04dc8c93 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll index 257f5d75de..56313495d6 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll index 11da7bab91..204d1b2f75 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll index ce5085c5da..e48d800b41 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll index 7e21193bd8..cc3b2056cc 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll index a25414bcf5..e9a53d8649 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll index de1bfe8689..bbba2ff0be 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll index a28af820d1..ae1653eb58 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll index 943ee87347..dba4a512b8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll index cb60b19aea..a15fa7dbe7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll index 6b3c0307f4..0d75ea5930 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll index 6d03a3faad..092f54ee29 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll index 6b9ecd5328..a8af62e486 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll index 39fe857548..d00fa98966 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll index 9a1730295a..e3a1580960 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll index f4c7742cac..cbac0e3773 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll index 1ad93da507..1ca4d09147 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll index 836b0a6a96..e86249b588 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll index 8bf14f3dce..bc87ee24c7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll index 9ca367101b..091a7c67b9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll index ce2deba6df..95b5342200 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll index a777a2e24e..4785e99b81 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll index 5d8e72661d..42547cf2da 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll index 062693d1ee..3d165cbae8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll index 949b27c5a3..811dcce5d2 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll index c312b6065e..ff6d6cd68b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll index 57eea13823..27d37e69cf 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll index 621c792582..63b97ffcd0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll index d44cd85fba..508a2510cd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll index 3066e64ec6..8b913fd68c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll index 017546e43d..3e337b5e1c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll index ce5d271387..56e5157434 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll index d777544bab..352fd5d01b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll index b83080ba98..0e07abce1e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll index 9f90bff2f2..c444bd307b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll index 1b70f47887..985769e694 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll index 62961dfa47..21487f3252 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f16_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll index 2d5719d757..8ef69b5e08 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll index 64dc030fc1..99351c9bfa 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll index 69e81a6ed4..c71d6021ec 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll index 07d37001af..95abd6f977 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_fp8_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll index f2d70ed570..71a61bd426 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll index 73e2e53718..984b309228 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll index 5329adf168..fc6f40f9c3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll index 74400564e3..b89004ed20 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll index 3de7736738..d8cccb3c87 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m1_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll index ab4b98f5df..c5b129b074 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll index 003f8af757..9f26f91a45 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll index c710a5f90f..9c8ed180b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll index c7f1b14a06..e12f17e2cb 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll index 4aa4bb2525..7b90b57cdd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e2m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll index e37c3f8b11..42350b5441 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll index 48b22b5142..4fd25d377e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll index b596b116c4..2bd447cdc4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll index 123f3ef049..0a7db968b8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll index d33948ea9a..36a4777c65 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e3m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll index b42e0288a3..662d835a9a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll index e9acc9840d..bab546777b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll index 39b9da7b68..fd7ff404de 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll index a4e9bff2dd..2d7af4dc62 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll index 15332bb717..eff1df7d5f 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e4m3_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll index 93e7483d20..f8b0ca4421 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll index be88ca87cd..075addc2fd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll index bec8310322..74cb16cfba 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll index f9c1f8a864..d4288b006c 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll index e32d67ed8b..d7a93b5131 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_mxf8f6f4_f32_e5m2_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll index b5015a1f6c..2b05e6ac46 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll index bde945dc21..a6d9df46fd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll index 5b2d2cd47b..88978c7925 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll index bca8cce2d8..7347863db4 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll index d23124c3f3..1b82a45cb3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll index e805e22ee0..7f456cee1d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll index f2d3d19276..c48bc4588d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll index 57d83fb7ff..86d578415d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll index 43e1dad364..100d8d02fa 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll index 56f57c8d76..8fb5e99e1e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll index d0c87ca0a7..363c34aa54 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll index ce4b7d2b1f..0a4700d71e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll index a213fa4792..1cc31f091e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll index ee50dd7eeb..a8f3e8edcd 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll index 7a696b7236..b75a96d9ce 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll index 60079188cd..e8cddafd20 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll index 99653bd66e..df0a653fb8 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k4_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll index e5d20b70c4..9910653161 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll index 0ab16bfb5b..cc6ad92c93 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll index 874655323d..bd201010c3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll index 67885e09e8..fc5351eaa1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll index 252e92d878..2cd84c241a 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll index b2754b553a..e9d3682d82 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll index fa78ada27d..45e334cca0 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll index ecb10661a9..94995327ab 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll index cab04420f6..4ddf6656ed 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll index f3de39a8dd..ca1249742e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll index 378ce07f88..31875b3d5e 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll index 186acd72d4..ff8cf0fd4b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll index 258a89042c..c2c2b92355 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_and_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll index b19a973c44..1703de2ea9 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k128_s32_b1_xor_popc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll index b3bb4ab6f8..56df23b767 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll index 0b9ebb91c3..3f2d45dfa1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll index 0327cd2fc6..88b613308d 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll index 1b39221322..35397f1026 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll index e73cc8b843..f64e01c530 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll index cb112e2795..393c6a46e3 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll index f8b7971070..a62d22771b 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll index 7cd09d56a7..4e084fdc64 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k16_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll index 60cb52e873..5779d1fa19 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll index 1a677bf974..9ef79c8b34 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll index 13733f50a6..c5d0fa2683 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll index 9f88dff37c..8f121eec83 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll index 642dd662cc..e5a8b93a30 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll index 8e0e9a3db5..75221e54a1 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll index 2cbb6cc467..021fb95090 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll index 41e7508b94..f0b419a655 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k32_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll index 7607113e87..3e3e7712c7 100644 --- a/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll +++ b/cuda-oxide/intrinsics/probes/mma_m8n8k4_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll index a1c26a8fca..8077529cd6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll index 42376ee6bc..aaf3101ba5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll index 51f7d154f1..cbf2e9dcb5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll index 5ca5923d23..6201f118e5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll index 09c1db0801..62a20ced1e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll index 066ae9dbff..24f0d33c71 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll index e5613903ff..5f39597b9d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll index efc85ed2d5..df71dc6d0d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll index ab4f4791c7..481a6cdfe5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll index 89dc2554e9..2dc97ff959 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll index a90372038d..37e2562d47 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll index ffaee874b5..476882e1bb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll index fcccf992d5..8e7f547103 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll index c4d2243bb0..95c75878ba 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll index 1c81033278..eeb6598400 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll index ea26b6de06..4b32e8fc81 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll index b926c032e0..7207939f4f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll index ed0d6aa920..ad54063aaf 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll index a2b89ad7c3..c126f91510 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll index 05060391c6..c9c1918d9f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll index 7bdbc1e2ac..619e06bc16 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll index 8c2759d30f..37e8a6f059 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll index 9f1fc9cdfb..c85c9292ad 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll index 2f913716c0..943cfced44 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll index ea9a05d040..64891baec8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll index d7b4387097..ef1f8cb45c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll index 7cd996192d..9350335181 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll index e47c43c803..2af2dbbea4 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll index bab474279a..0be4489728 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll index e70f6f05aa..50df175b70 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll index a4139c0d24..1f5e2846e2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll index 8c43a439cd..c82d70105a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll index 5e38f3cd20..e643082510 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll index 9cd5a9b070..4938e384b8 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll index 73fea51268..1f7d24771d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll index 90d22e821e..11616cf36e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll index 4f57f735b0..7f2fab77e0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll index 1a8f543452..61c63143b7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll index a5eff6c3d3..8379d7c7c3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll index 62afa3840e..a0f4f44d79 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll index 8132406012..536ceec7d3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll index 9d5a7e9c9a..0fd960de52 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll index 4848f36e89..f4d0c1e716 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll index dcf768487b..ac08ff4cbb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k128_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll index cac229d40e..d1889712da 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll index b1eb3d7a8e..82b0cbfcab 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll index de3bf01b9e..6d183e3edf 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll index e6060062de..64459b199d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k16_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll index 7c708cd6ea..cd041ca53c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f16_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll index caf27dcee3..eaa8c5c985 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll index cb341c1467..e38d992dc3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_f32_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll index 5fffffc516..c5274b850f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll index aa2795de6f..4b76de0392 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll index 3e7bedbf61..c68144b194 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll index cd5ca5ebc2..6b07fa97b3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll index fb9ae969d0..53b8ebb68c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll index eaebd92019..4625dd5167 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll index d9b9546b8a..fbd0f98935 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll index 9bc18a7cbf..d0f9940106 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k32_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll index 6ad1dc3c3f..18c39df4d1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll index 58cb90c339..badd9009b2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll index 0c45109156..2954787465 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll index 4d2b8fee55..e4f5f34ecb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll index 8f6675c115..d7cdfc3325 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m1_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll index e0ee31e501..49ab833b46 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll index 2523957c4b..5f5e552bf4 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll index a4a93b11b0..61216b1c23 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll index fcd5fe7bf8..37e2d843d2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll index 00b4d61a0e..e355988fac 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e2m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll index 4e64370269..a48ae84609 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll index 987573f46d..365d99861b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll index b6bdfd8728..0c10aa3c53 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll index 19ded5a79f..7d2bf0ecda 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll index 6125280516..6d20bc3001 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e3m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll index 1372e1adbf..04bcd8194a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll index 38cff699b9..dbc72511a9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll index 7301c82898..9bf94d05be 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll index 3ea077a7ce..744ede663e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll index 1062056ed7..d169bcc771 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e4m3_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll index f4efe5b3b9..896c6513a4 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m1_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll index 904b059c60..9921ae1078 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e2m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll index 43e9aa80db..3ad8555152 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e3m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll index f554b6747d..584298de52 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e4m3_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll index bf723fe739..ed9e47ebd5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f16_e5m2_e5m2_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll index 7374091317..4b94ed7fa2 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll index f06123d3b7..cb687c945e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll index 76a4ebf952..fa7b4aae5f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll index 12fc4370dd..f925ccf290 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll index 03505223e6..338a035645 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m1_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll index 2d45ea4401..523615ae7e 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll index 9ab5fdb0fa..741ae6b002 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll index b876e7d07b..8ccf28a967 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll index 7b51521b64..5c195b2493 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll index 833b3febf7..7eb7c201a5 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e2m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll index bbd176dc6e..07e4ea7068 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll index 6ece6a57aa..f0b4414207 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll index 2cc4765392..203ff011b6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll index f20dfab5d1..6eeeae590f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll index 0fd65bdbfc..40679ce7d1 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e3m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll index ec70736a73..7793fb1b6f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll index 2f6380b2b9..186b31988f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll index 0f72f09644..4ecfffa2d3 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll index 2eebffc7c0..00301b238d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll index 0999acaffb..4584a6b8d6 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e4m3_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll index fc30cfa11f..36f7dab63f 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m1_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll index 3cc2a08c15..5782ced891 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e2m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll index 6562efc51a..47e26c23ff 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e3m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll index eb889176b6..3be3033ccb 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e4m3_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll index 0ed96d019e..7fb9a35414 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_kind_f8f6f4_f32_e5m2_e5m2_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll index bee68344b0..b2ba7d3855 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll index 30edeb35af..b06d99d190 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll index 06457dc22f..02c0aa5154 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll index de428774f3..9c406a2cac 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s4_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll index 2142289885..64cf990e67 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll index ede089b1f7..9c1a0875a9 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll index 062d0190e3..22b414bd57 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll index 007b24267a..063d770895 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_s8_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll index ea682776ed..d41ee4bd57 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll index 86d9af6f87..ff58561d3b 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll index c622a59f34..ea82f4255a 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_s4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll index b1314db4ba..9a5019678d 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u4_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll index ef29a7fa31..ad871e7a4c 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll index 20f69bb2c0..1db5d5afc0 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll index 75e7b3b24a..011ce74c43 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_s8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll index e4807e796b..50ed30e6ee 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k64_s32_u8_satfinite.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll index de86f78650..292f2dced7 100644 --- a/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll +++ b/cuda-oxide/intrinsics/probes/mma_sp_ordered_metadata_m16n8k8_f32_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll index a985af3c9d..ac18f46e13 100644 --- a/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/movmatrix_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll index ef377e5b3b..b02cc4168b 100644 --- a/cuda-oxide/intrinsics/probes/mul_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f16x2.ll b/cuda-oxide/intrinsics/probes/mul_f16x2.ll index eb906f9921..f29398d954 100644 --- a/cuda-oxide/intrinsics/probes/mul_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_f32x2.ll index 37976ac13d..0e826e31cc 100644 --- a/cuda-oxide/intrinsics/probes/mul_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll index 4cd1a6bdd4..1dd4cb4bdb 100644 --- a/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/mul_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll index 49dfebde02..946d9ec9d8 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll index 99d1bfa070..38869a4d3c 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll index 54b5d782d4..930ba15bea 100644 --- a/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll index 0b62e43f28..8845444e20 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll index 105b9904a4..9800e872ec 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll index 43a7fedc6e..1d01f663e5 100644 --- a/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll index 7da685d65a..6b5594af01 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll index 86a45504f4..68da7fde6f 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll index 9c951a97e3..50d581c61a 100644 --- a/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll index d9e4596b3f..75dd4ff99b 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll index 1b567938e3..2bfa2652fd 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll index 45475415b0..5c3c0abd74 100644 --- a/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/mul_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nanosleep.ll b/cuda-oxide/intrinsics/probes/nanosleep.ll index 276ece2ef1..3d173156cc 100644 --- a/cuda-oxide/intrinsics/probes/nanosleep.ll +++ b/cuda-oxide/intrinsics/probes/nanosleep.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll index 76258fba4b..e9488c24bf 100644 --- a/cuda-oxide/intrinsics/probes/neg_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/neg_f16x2.ll b/cuda-oxide/intrinsics/probes/neg_f16x2.ll index 7bdee19d56..385e6df41c 100644 --- a/cuda-oxide/intrinsics/probes/neg_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/neg_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll index f0e67d1b38..f9a0421868 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll index e0e366c6d5..fdda0fb0d5 100644 --- a/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nsmid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll index ded450886c..37efd6dc21 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll index 4fbc24b344..83519dc2c0 100644 --- a/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/nwarpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll index 546b11b015..4e227d2738 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll index f7b4e80421..b31c4bb7fe 100644 --- a/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/packed_atomic_add_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/pmevent.ll b/cuda-oxide/intrinsics/probes/pmevent.ll index 2ac6c42605..0fdf57f45f 100644 --- a/cuda-oxide/intrinsics/probes/pmevent.ll +++ b/cuda-oxide/intrinsics/probes/pmevent.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll index 2bb9e1342b..f6ea49f51a 100644 --- a/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll +++ b/cuda-oxide/intrinsics/probes/prefetch_tma_descriptor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt.ll b/cuda-oxide/intrinsics/probes/prmt.ll index 048b55bbf9..25e4ee79fd 100644 --- a/cuda-oxide/intrinsics/probes/prmt.ll +++ b/cuda-oxide/intrinsics/probes/prmt.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_b4e.ll b/cuda-oxide/intrinsics/probes/prmt_b4e.ll index c34ab4b4fb..484a1dfdc4 100644 --- a/cuda-oxide/intrinsics/probes/prmt_b4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_b4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecl.ll b/cuda-oxide/intrinsics/probes/prmt_ecl.ll index d38597f87e..34e0535dd1 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecl.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecl.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_ecr.ll b/cuda-oxide/intrinsics/probes/prmt_ecr.ll index 761d439b73..f83f4a1c08 100644 --- a/cuda-oxide/intrinsics/probes/prmt_ecr.ll +++ b/cuda-oxide/intrinsics/probes/prmt_ecr.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_f4e.ll b/cuda-oxide/intrinsics/probes/prmt_f4e.ll index 01ff20eb5d..c4a7614224 100644 --- a/cuda-oxide/intrinsics/probes/prmt_f4e.ll +++ b/cuda-oxide/intrinsics/probes/prmt_f4e.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc16.ll b/cuda-oxide/intrinsics/probes/prmt_rc16.ll index edc156b77f..14192b0ebf 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc16.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/prmt_rc8.ll b/cuda-oxide/intrinsics/probes/prmt_rc8.ll index 856d6ec0dd..009328e41e 100644 --- a/cuda-oxide/intrinsics/probes/prmt_rc8.ll +++ b/cuda-oxide/intrinsics/probes/prmt_rc8.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll index a0c99ecc71..c8ba34faf5 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll index 7131e8bbdb..e742043f1a 100644 --- a/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_approx_ftz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll index 53b4e257b6..22099fbb9d 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll index 82aa47691c..9350e2b746 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll index e5f566f96b..3662480176 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll index 98af3b5b1b..9b321528a5 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll index 3d842a7fc8..4474b209b6 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll index cfc02003fd..0706a5dd87 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll index 5d3b9c6221..ed85e8c9de 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll index 3de48205ab..67e94691f7 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll index fb52fb88c7..c55efd33b8 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll index c7e348a788..8be6130b77 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll index 9948810f8c..78a380cd8d 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll index 22699f6896..01d8028fd2 100644 --- a/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rcp_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_add.ll b/cuda-oxide/intrinsics/probes/redux_sync_add.ll index 0668528f29..3c6fef28a4 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_add.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_add.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_and.ll b/cuda-oxide/intrinsics/probes/redux_sync_and.ll index 3cf523f6cc..d62cd46fb1 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_and.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_and.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll index 29dfca9b7b..1977da8b38 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll index 4977262ec6..3da772f242 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll index 72d511d4a4..cd25209c11 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll index 15c4542186..f25556e96f 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll index 56c0c90208..cac9c75c91 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll index 5e55d82089..6b3c0ab201 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_max_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll index 0802c93ad4..8fea47dc6f 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll index 58dfa74d36..6990a7d483 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_abs_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll index 006f1d87b1..6de0160a51 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll index 06da2103fc..5ef98bce2b 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_i32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll index 66cadfde8a..e375542b35 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_nan_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll index fdfafbdc28..8b23ffc72f 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_min_u32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_or.ll b/cuda-oxide/intrinsics/probes/redux_sync_or.ll index f387da6b4a..5def8ee0a3 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_or.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_or.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll index 88815c8349..ba934006d2 100644 --- a/cuda-oxide/intrinsics/probes/redux_sync_xor.ll +++ b/cuda-oxide/intrinsics/probes/redux_sync_xor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll index 492c43498f..c3e17919e7 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll index f6c66601bf..379c790503 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll index b5caa515c0..345905cacb 100644 --- a/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/rsqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll index 8d10403591..3db9583452 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_dec.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll index 6357954c9b..5b81fdbd03 100644 --- a/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll +++ b/cuda-oxide/intrinsics/probes/setmaxnreg_inc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll index 48706c2880..07931bc541 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll index b9959f6a96..565a566ecf 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll index 46b687e91d..9101ef9b0d 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_down_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll index 69dd6205ba..d08405a0c0 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_sync.ll index 28b2213411..c095166597 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll index 4b5b169f31..1598ac5170 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll index a8b9647cb1..968d5ae090 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll index 33bcaec9c2..9eab4166a3 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll index 1634efe068..c3bdf2b0ea 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_up_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll index 342b7df9fc..8306292e9f 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_f32_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll index 2841b507b7..f62c3be3fb 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll index e649754501..34180ef988 100644 --- a/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll +++ b/cuda-oxide/intrinsics/probes/shuffle_xor_u64_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll index dea6ef5902..bac49742c0 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll index 0d421a5457..4577db39c5 100644 --- a/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sin_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll index 655c224f6f..1a59d8018a 100644 --- a/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/smid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll index d4e14ab760..286922323f 100644 --- a/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/smid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll index f4320e94a9..65b0fbe72c 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll index b07789af50..cd8eda50d4 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_approx_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll index 812fea62dc..2f866fb9ae 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll index ffa1ab715d..dd08cfa8d1 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll index 9d04714adb..9a52fb6c17 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rm_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll index 762b7a5685..6cb73537f2 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll index cc089f4674..7f30e096c5 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll index c7a43bfef8..111aff77cd 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rn_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll index 90a1feac0b..7bf5b85fe8 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll index 316fbeb150..fcd654a822 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll index caafffb00c..e0064049b8 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rp_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll index 7ec27a332a..161e597e63 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll index 623e182995..1ac36b7dcd 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_f64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll index 4fdb40fbd0..ed9a2b35cf 100644 --- a/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll +++ b/cuda-oxide/intrinsics/probes/sqrt_rz_ftz_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll index 1370ae34af..23cecab80c 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll index 66a51c5098..ff62f080e7 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x2_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll index 2103cc29d7..7e927c768d 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll index 7ecf23b4af..8dfad7bbda 100644 --- a/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll +++ b/cuda-oxide/intrinsics/probes/stmatrix_m8n8_x4_trans_b16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll index c29beb88b8..fb4c2e2fe2 100644 --- a/cuda-oxide/intrinsics/probes/sub_bf16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_bf16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f16x2.ll b/cuda-oxide/intrinsics/probes/sub_f16x2.ll index 95080ca4b3..f7ce7ff1c4 100644 --- a/cuda-oxide/intrinsics/probes/sub_f16x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f16x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_f32x2.ll index ac564be93f..182ae4c0bd 100644 --- a/cuda-oxide/intrinsics/probes/sub_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll index 4a99843635..0facbbe458 100644 --- a/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll +++ b/cuda-oxide/intrinsics/probes/sub_ftz_f32x2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_mask.ll b/cuda-oxide/intrinsics/probes/sync_mask.ll index 53a93c41a7..f6a9812a13 100644 --- a/cuda-oxide/intrinsics/probes/sync_mask.ll +++ b/cuda-oxide/intrinsics/probes/sync_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/sync_threads.ll b/cuda-oxide/intrinsics/probes/sync_threads.ll index 12cfdc03bb..01a3bb1d2b 100644 --- a/cuda-oxide/intrinsics/probes/sync_threads.ll +++ b/cuda-oxide/intrinsics/probes/sync_threads.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll index edbac2acd4..725f001d13 100644 --- a/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll +++ b/cuda-oxide/intrinsics/probes/tanh_approx_f32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll index fc31f58597..4681ee92fe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll index 395576b8b2..188cef95f3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_alloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll index 3e7912ec34..b18d1467d6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll index a90b14cc68..de11e3ecba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll index 2e4f51197d..d6b22fd8aa 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll index 7bc1d24e6c..eb3f9cc385 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_multicast_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll index 6489a4d025..1db2dbdbad 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll index 43f09bf278..f5aa13b3e6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_commit_shared_cluster_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll index f67c9f785b..1e5fe4bb0c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll index 3622af6d2d..19a6929919 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll index d1518d14da..c8d680b1cb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll index 6a4d5cb912..915f8808dd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll index be037f8638..272b8348b3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll index ea68010c16..934cd697ee 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x128b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll index 5429eb2bbc..3f3585ca34 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll index 1b54296d5c..bd9c2a8a6e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll index 9a681d7415..6915054db7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll index 2631496007..2c93c4b8b6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_128x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll index 9745d16185..243e2f4457 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll index c2e7fb314f..d7318b8ecf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll index 54a4a3f400..49f8360d73 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll index 30aff21895..1b0b9c8428 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll index 4bfd7be094..fb9294fc00 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll index f126754627..1f7537312c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_32x128b_warpx4_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll index 097d0030c6..705b185f67 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll index be596cc4a2..5ad6aabca9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll index fcf02ebf93..6123b0e8f9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll index 33a470d014..19e703d60c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll index 69992024f9..0a28e83671 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll index 135a6149ed..8de4c9caf8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_4x256b_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll index ff4c64b02e..8b83fcc6e3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll index 9af1d29f04..80745c11f6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll index ca0de79732..e224d3f2fe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll index bc10081434..b5fad6438a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll index 4863c57053..b4803dae61 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll index 4df9e6c767..9d52094864 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_01_23_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll index dfd0d4ec85..dab75d57b4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll index 1b60b85733..4ada1ddd76 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll index 51b592a056..4e70b6a4cf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b4x16_p64_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll index 0d08def162..1033e04fb7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll index 56672bf829..9165307648 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_b6x16_p32_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll index 9be04b6015..2645ba6aa8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_64x128b_warpx2_02_13_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll index fc3042245b..0325551ea7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll index ef4c0454aa..493cc35b92 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_cp_smem_to_tmem_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll index 5602949fd5..fc9d5dd471 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll index d9447055da..834b5cd913 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_dealloc_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll index 27c89cb897..badff34939 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_after_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll index 0700594a29..da4f8c90e7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_fence_before_thread_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll index ccd96e765b..510ca76c46 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll index 8b3c3052fb..39793770f4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll index 26361b45ad..6558b29937 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll index a08fbae133..5943a61b0e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll index b234d265a9..2c617ef281 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll index 4dd0d027e1..3f16f4c1b0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll index d021f5b30c..dabd22d996 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll index 044d0d9d82..9a21034963 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll index 81cc0844ff..5404838fb9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll index aa0996b49c..e4e364709f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll index f88dfeb055..34e7e8520c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll index a59be23b32..d10bdcb5c2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll index 0eb53f8172..49f9fad5c4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll index 527987765c..259a47baba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll index 1f49030f0a..ecc652aaaf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll index 3e60613e4f..fedec7933d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll index 1362605978..501dac0681 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll index 627d4b93ff..c81c06e7bb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll index 6e4d18fee8..c77bd824dd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll index 6a274c6fc4..ac2162f914 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll index 3dae478f5f..8ba23ed5ab 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll index 7348802335..6e1f955780 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll index 2315f1f192..bb00237f03 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll index 7331b4fb6f..68482cc33a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll index 5892dfc8d8..59defae2ad 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll index 704d87b5f8..bef54212cc 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll index fac30abc83..8f5b6ac36f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_pure.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll index 1c0b584f10..070cbbc91c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll index 1d03e49cfe..ad594f196b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll index 4d2769046c..60d37021f0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll index b968b4852a..c57c9da0a7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll index 36b862fe3d..4d818e944a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll index 6e04cf5ef5..ece5446590 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll index 2a28f65245..376b7b8102 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll index d8fbd85570..4fdc4403e8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll index 3f3de404d6..680541778b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll index 8162131bab..be22d6b1ff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll index 1d6994185a..6a1ea536a7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll index 73bb9e3577..b584ba9d9b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll index e379507c79..34458c1887 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll index 0e8ded2591..3ca1383000 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll index 51cf37df62..0dad601858 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll index 8e56c17978..eea053e0e9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll index ccb4c0dc96..e56db75f78 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll index 477411bac9..f54a4b971e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll index 88d1bc8ea1..662a79b2a6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll index 76db69b1b2..3015ca3058 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll index 9dcd518fab..5c9a851cfe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll index 41f4dbfeba..548524be6e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll index 9205d1a381..ee8e288721 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll index bf0ea0fcfa..6f7edcec98 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll index 25f0487143..5098ca6a08 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll index 4c32a0ce14..9236a7e66f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll index 45776a4007..9e8ccc97e2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll index a90bf8b47e..c3dc0adefe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll index 21b6215750..5630c662c3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll index 6800ef72e1..18bf3328bd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll index f025369505..da5a8ac721 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll index 618c8a05ac..ea903e9e9d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll index aa33e3af6c..3ce9fe0bba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll index d7842555ea..7f5c30cfa2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll index 5db306e2de..6a9ba29f79 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll index 7cdbca2058..c86e7781ff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll index 5b96a05380..b90be4e669 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll index 5946c8122f..15b87faad5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll index 048856188c..33344e3c04 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll index 2d16d3c3bb..b17d114dc8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll index 0ebe513204..af1d53993b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll index 6664e2749c..8bee1d37af 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll index ade7ee312b..2c05ecb2ce 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll index a8900d6cf9..0e80981b8a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll index df8bb3b916..fa8fdfa09e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll index 3c8df30c84..c3cbb87b0a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll index edbbb7696c..ea36ca43df 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll index 9785d254e6..6216ada9ac 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_pack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll index 73977f774e..8da86c6721 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_ld_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll index 9862624480..595b69fa29 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_load_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll index 73753dcd9c..573fed3b43 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll index 8da08a6003..27262fdf38 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll index 6b903bd942..6a61de230a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll index 7f5a70026d..98d37ad50d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll index 83568d75eb..6c40df6ff0 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll index d1ad1d6417..5308f43130 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll index 7edfa35d48..47f8ca3dd5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_f16_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll index 5d340d6635..4cd8bba4af 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll index 696c5de431..8bb03d2af6 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll index bdc91ed5e2..bb1f776258 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll index afe2f6e2eb..fbd424203c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_sp_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll index d129e09365..9dbba42424 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll index 40e444cbf4..9c2515b6ae 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_tensor_ashift.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll index 993d7e2526..8db4ffbae2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_bf16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll index 2773ed4292..c362ea1a36 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m1.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll index 35278436c0..b92b3aec0a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e2m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll index 768bd6e8aa..042ce7ad76 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e3m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll index bb5ff188c2..9a620b8308 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e4m3.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll index ab842bdeab..9f3ce75d00 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_e5m2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll index 6e93f1da48..e0e6c7c07b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_f16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll index 343efc6617..09ea38c767 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll index 3a5b6a3a01..ed121fd4c4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll index 253c461403..9e0601053d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll index d9f7674921..1caaecb0f7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_shared_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll index 370809bbca..28ac79fae7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll index 73daf95af6..e502338674 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_sp_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll index 827254daf9..3906ca1f48 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll index 663bd0f7fe..3289dc7937 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tensor_zero_col_mask.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll index f8fc0ce249..cae6ddcfb1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_mma_ws_tf32.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll index 2994ab1e6d..fbad879555 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll index 6be38afd39..ddfcc64947 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_relinquish_alloc_permit_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll index 2ec5dfae75..5086100059 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll index d1c3467bee..4003987731 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_shift_down_cg2.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll index 43bfc6f846..1b81acc23a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll index e6614e2c13..0d35a68a1a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll index 510a3c08f6..aa5dcdad3d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll index 197d7d53f7..24ba3b65f2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll index 2340587798..7b375acf90 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll index 76bd27ba1e..1347bdbd3e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll index dd329528e4..8ce48b98ba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll index caef6cccba..7c707115e2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll index 8a0cb086e2..29fd5f8215 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll index f526de725b..24a83eeca9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll index 6aa3bcc2f8..7e08f3cc6c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll index f8fabede28..7359c15811 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll index 6387cc43fe..7508d9512c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll index 69e8a07c8f..81c394680e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x128b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll index 4df2b6b459..dd617432bb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll index 17e085f9de..b1431a608d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll index 1638f95864..dfd391a873 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll index 9be641f894..113cd6f2c5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll index 3f74be72ed..65278745cb 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll index fe66ba4b24..32219dc011 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll index bebf3ef42f..c2616feaa9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll index fd8038ec00..bab355423a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll index ad9799b076..ef892661c9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll index 196e49b726..1d16dd79e5 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll index b47fa7c7e2..b3c367164a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll index f91fb35a65..1649c07be1 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x256b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll index 721ed471d5..ea6a58618d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll index 30936267b9..5fc9cc0c34 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll index ae6418a3d5..76c9b5e126 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll index 4ea479203e..281a56d4e4 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll index d858525d25..2fa294582b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll index 62ce095c48..d0d8759e93 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll index b54190360f..6710bd6d41 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll index 7e7d419a94..fa54eb1721 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll index ccefd8fa71..ddda0f9b83 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll index c4e0c4ba92..9989a00130 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll index 0a353fd4ae..5cfb0ea15d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll index b180e2a1ce..da536ae700 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll index e6a264bca3..dc58a9dc25 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll index 5bccc23776..3af785f0bd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll index 2a285acd10..68bd22e8f9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll index a3125783ec..7b36c3f64d 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x32bx2_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll index 618adf2bc2..eea2ea9957 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll index 2385efe8fd..ac3c80e5a9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll index e08168d62d..52564ec661 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll index 5e97ab129c..2a3c96178a 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll index ddcfedadf4..cae93a7a62 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll index fe5122d36f..506ea62571 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll index 2d03c1a55f..2dddd7ce99 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll index e5642c1c51..3994c8a4ba 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll index 0ac53dfc0e..15fb8957ff 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll index 4b160f2227..5d05ee095c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll index aa66d99016..f2fae2861f 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll index 3a3791e947..73289f8549 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll index 21a51f9159..acc70601f7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll index 3b1579f29e..2574645ac3 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll index 23b0bd9092..cf8ac59413 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll index 22eba73b99..f7fdb470e7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_16x64b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll index e124078150..b2b72b7be7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll index 234053ecea..51c111b66c 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x128_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll index c2b8849f6a..e905a65ccf 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll index d2328953c6..dfce971dbe 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x16_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll index 611a540a91..a0daa1fa67 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll index 62ea9464d6..27f2dadbcd 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x1_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll index 8e0f1c1a47..65292054b9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll index 1eeb06b6a2..8dc86b0fb2 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x2_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll index bcaa3e3349..04be3d6a60 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll index 9824934037..0ce49b7860 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x32_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll index ba67fdd5de..3bd1c7461b 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll index 247a2227cf..9e8e75fda8 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x4_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll index 740462be72..269ef07736 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll index 6ce4972b5d..d24ba123b9 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x64_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll index deb7d47cfa..25074b9b4e 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_raw.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll index 865b8ccb67..423a7b40f7 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_st_32x32b_x8_unpack16.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll index 1b7b956a0e..a90834a825 100644 --- a/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll +++ b/cuda-oxide/intrinsics/probes/tcgen05_store_wait.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll index 492568717b..0b8a01a198 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_box_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll index 67b7affd7e..576978840f 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll index 95d08cc831..786ce3e147 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_element_type.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll index db514e32a4..983a5c7f45 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_fill_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll index b88aeb493b..6be7af57f0 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_address.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll index f3aa874dc7..69f632e4f0 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_dim.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll index e847ad0cc9..c3528e6c92 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_global_stride.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll index c9d3b1b71a..dcf8a45027 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_interleave_layout.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll index 39964e69ca..b82acad78c 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_rank.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll index f73b3b8e61..e7ca4939ea 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_atomicity.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll index f04918f255..f0faaacf67 100644 --- a/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll +++ b/cuda-oxide/intrinsics/probes/tensormap_replace_swizzle_mode.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_x.ll b/cuda-oxide/intrinsics/probes/thread_idx_x.ll index 16eac4a1c8..861f7b549e 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_x.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_x.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_y.ll b/cuda-oxide/intrinsics/probes/thread_idx_y.ll index ba0ce02788..b9c97e960a 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_y.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_y.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/thread_idx_z.ll b/cuda-oxide/intrinsics/probes/thread_idx_z.ll index 854c169db5..074d5c6627 100644 --- a/cuda-oxide/intrinsics/probes/thread_idx_z.ll +++ b/cuda-oxide/intrinsics/probes/thread_idx_z.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence.ll b/cuda-oxide/intrinsics/probes/threadfence.ll index 881cd4d8d8..81bb0b1da6 100644 --- a/cuda-oxide/intrinsics/probes/threadfence.ll +++ b/cuda-oxide/intrinsics/probes/threadfence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_block.ll b/cuda-oxide/intrinsics/probes/threadfence_block.ll index 18637438f6..2422f916a0 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_block.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_block.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/threadfence_system.ll b/cuda-oxide/intrinsics/probes/threadfence_system.ll index 7dc31e7a53..d5181821c8 100644 --- a/cuda-oxide/intrinsics/probes/threadfence_system.ll +++ b/cuda-oxide/intrinsics/probes/threadfence_system.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll index 1ddd1e6aeb..135eb11317 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll index 83a3246633..9124402f87 100644 --- a/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/total_smem_size.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/trap.ll b/cuda-oxide/intrinsics/probes/trap.ll index 5aa2b7ee63..91eec43597 100644 --- a/cuda-oxide/intrinsics/probes/trap.ll +++ b/cuda-oxide/intrinsics/probes/trap.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/uni_sync.ll b/cuda-oxide/intrinsics/probes/uni_sync.ll index 1cde13d790..b990da9cc0 100644 --- a/cuda-oxide/intrinsics/probes/uni_sync.ll +++ b/cuda-oxide/intrinsics/probes/uni_sync.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll index d26c218655..b36a3105c1 100644 --- a/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll +++ b/cuda-oxide/intrinsics/probes/warpid.lib_nvvm.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll index 3f532a5816..c747459656 100644 --- a/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll +++ b/cuda-oxide/intrinsics/probes/warpid.llvm_nvptx.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll index 50a9a6296b..3584d85f0b 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_commit_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_fence.ll b/cuda-oxide/intrinsics/probes/wgmma_fence.ll index d4195c11ee..a2b40a4c89 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_fence.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_fence.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" diff --git a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll index 294106451e..8ba5bde0b4 100644 --- a/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll +++ b/cuda-oxide/intrinsics/probes/wgmma_wait_group.ll @@ -1,7 +1,7 @@ ; SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. ; SPDX-License-Identifier: Apache-2.0 ; @generated by cuda-intrinsics-gen 0.1.0. DO NOT EDIT. -; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 3f49098fc61d723a7928376c046be496c09479622892fc5bc88f034f643ead7d. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda"