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: | 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 425aedc4ac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 57652a5228..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 aa95140971..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0c9b7ff9aa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a4b905b524..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 04d544f30c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 6be477a270..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 36a19cb533..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d6828a8aba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ee39d4257d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 37d8248126..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 2ec7953393..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d985923438..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 349e4a4d35..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 88ee2432f3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4dcf1edb9b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3fb0c6490c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9b98009d81..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e99a610b45..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3788f17e4b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c90b321c91..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f7e8c4e54a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0f1714ad05..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3b0f283d9b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d7a9131ae3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 86e1231142..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e643c970b3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 037c21f721..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 08980edf29..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 412e03bf24..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ece2a3992a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0ad1f01a09..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 92226b0c88..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b8f8444bac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 6a734aa7fb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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-gen/src/model/contracts/scalar.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/model/contracts/scalar.rs index 257a1a85e5..f3da03ecca 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/model/contracts/scalar.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs index 081f4b33b7..bea0d9e44b 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/model/overlay.rs @@ -81,6 +81,8 @@ pub struct OverlayShardFile { #[serde(default)] 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, @@ -505,6 +507,22 @@ pub enum PackedConversionFp8Format { E5m2x2, } +/// Compact admission for packing `bf16x2` into `e4m3x2` or `e5m2x2`. +/// +/// 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 { + 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/cuda-oxide/crates/cuda-intrinsics-gen/src/render/compat.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/compat.rs index ecc48feca7..7275698428 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/compat.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs index 29eabd7073..036c53c8b8 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs +++ b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/dialect/scalar_packed.rs @@ -382,7 +382,8 @@ 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/cuda-oxide/crates/cuda-intrinsics-gen/src/render/families/packed.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/families/packed.rs index f4977f10af..f377c9a597 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/families/packed.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/render/lowering.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/lowering.rs index db9ae260cc..00efa16a52 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/lowering.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/render/tests/mma.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/tests/mma.rs index c9fb824f4f..561f5e421d 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/tests/mma.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/render/tests/scalar_misc.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/render/tests/scalar_misc.rs index b947253880..6c555c0456 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/render/tests/scalar_misc.rs +++ b/cuda-oxide/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/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 b3c0a82a7d..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 @@ -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, }; @@ -16,6 +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, @@ -35,12 +47,107 @@ 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`. +/// +/// 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 { + 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 +541,11 @@ 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 +566,23 @@ 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, 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`. 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") } } } @@ -482,12 +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 { @@ -563,6 +681,64 @@ 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 +845,7 @@ 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 +883,9 @@ 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 +901,7 @@ 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 +966,8 @@ 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 +985,7 @@ 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 +1041,9 @@ 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 +1093,7 @@ 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/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/overlay.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/overlay.rs index 03a02dbea6..4b6206f094 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/overlay.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs index c9bbbdc9ff..4642f19ee8 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/fixtures.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/misc_families.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/misc_families.rs index bbe5c2c982..1fce9ebfdd 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/misc_families.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs index 44707537ae..d10dccace9 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/overlay_schema.rs +++ b/cuda-oxide/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/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 4d7b0dc448..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 @@ -171,6 +171,71 @@ 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_axes_require_the_closed_rn_satfinite_product() { + 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_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] fn packed_atomic_closed_semantics_reject_every_unreviewed_mutation() { let valid = packed_policy("packed_atomic_add_f16x2"); @@ -424,7 +489,7 @@ 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/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/sreg_threadfence.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/sreg_threadfence.rs index f03190029c..959622a23d 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/sreg_threadfence.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tcgen05.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tcgen05.rs index 91aae6815a..48c6f8c1e3 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tcgen05.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs b/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs index c876fea84a..01f1e00f49 100644 --- a/cuda-oxide/crates/cuda-intrinsics-gen/src/resolve/tests/tma_clc_cluster.rs +++ b/cuda-oxide/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/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs b/cuda-oxide/crates/cuda-intrinsics/src/generated/abi_v1.rs index 3eccfe34de..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 aadf572896..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 fe36a8aff8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 1e1509cef6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ead108a430..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8cf09104c6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7020044fa2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 72c7864a4f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 1953d9b216..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 fcb91feaf8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 89d677a028..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c6c7586c8c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3e1feecc57..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 5b15bf968e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 fb9ad47daa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9530f36fbc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9dca2e8bfa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 76f6352d37..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 41849141b1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 643b0122d8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d21b5037bf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f714ace2f2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 26c2b9101e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ca977feaf9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 60066e6a95..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ced2ff35e1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 6987c58ffe..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e6589fd47f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e6dfa4e62d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a71b4e90ce..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 2c5b7977ea..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7519ac479c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8101ef71f2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 942b0d8d20..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b4388a4208..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b05ce7cbf9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d13cbf86fb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a35c18829c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4ea27e9777..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 be7f8fbead..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e56d77b272..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c681aa4b93..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 458a3d4b8f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f1503abbd0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9f808f3507..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e6d33d9513..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 678bbd7350..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 971f157365..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 eacc5c5122..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f0e0a4f733..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0945d8c1f6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7ed4a44ef6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 6f86271313..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8647954b1c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9865a8fa47..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 478197ec78..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 60e4b4a4f0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d775d2d9c0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 da31b53b2c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b2477c9d74..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8a1758029b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c136d5c526..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 aa7f15a2bb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0659c6ce7e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 67318df319..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 322035b874..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 1aecf409b0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 5a2af67e32..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0d5606e517..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 764bc401f0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 04c7f5e3a7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ce2d1695bb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f20f4c4491..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 36db24882a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 447579f72e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 129f12b861..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 033064166b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d45478d0ce..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7e1c4672ee..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f44e7d1502..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 90d30ad393..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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/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}" diff --git a/cuda-oxide/crates/cuda-target-spec/src/lib.rs b/cuda-oxide/crates/cuda-target-spec/src/lib.rs index a1ef76a680..62b2f65d99 100644 --- a/cuda-oxide/crates/cuda-target-spec/src/lib.rs +++ b/cuda-oxide/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/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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 616d65ca24..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 1e84152b60..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f475f9c8a1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 373648c5be..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 812c00a924..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b4eed5cbbd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0975f2887d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7ef65fad25..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e8d6750774..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 6bc73ad2ec..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 5643285302..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8d2447c25d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 37a6d2de5f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 6346816724..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 82ebc991f6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 db1dbf890f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 96379c594d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 fe611f1857..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9692546253..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8fef6c41f9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e30a341c2f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f13122c5cd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ff68d7ff74..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0ab9e5d669..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 19a5061577..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ebb671fe9a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 90a3220586..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f59d9a2360..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ea10d67f42..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 082a1c4a04..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 28109e6e97..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b0e5ae390f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8066f000b3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 079d2f4753..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 dbf3e84396..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ade0190400..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 54b62778ae..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e3996216fd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b3af25763b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a04d390784..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9fc04194b7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3b6afc8dde..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 265c0112f4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4ab825e358..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 dca4ce9c0b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 40595ac88b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3178c8f319..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 d6a9cdc83f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b74612cfa4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 531e5c92af..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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. @@ -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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 2f359a9d70..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f66702fb4f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 346a6badbc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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. @@ -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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 269ecd2f67..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a741546fde..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8db6935f06..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 0ae8377b8d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 62df911084..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 86c8b04476..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 042d22fe9a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4ec1b4dc49..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a9533ccfd6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 959c46be61..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 2514ff7c0b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 024eacb481..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b48fe1d28e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ffb9d6def0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c4ab0cc802..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 84bf97d723..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9f0bcb6654..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b0a5c03ee5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 75edecc7e8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c0b0a46333..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ebf3d5ae86..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4f2a4b0dba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 bfe29b8e9a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 fc08e24bd9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4e168a9be0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 bcd1ae8633..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4c4e7a84fc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 2cfdad1ac9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 773f7f2f0c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 dada114a26..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 41d58a16a0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c0c0d792ba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7c515139fa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3539b87869..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 766e88b2f5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 43676af093..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 45ccb2d478..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ea1a765cb8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3e48ff5889..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4e2fa3ace6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 77cf005293..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7c67982816..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4dc9842f60..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 2acc775bc3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 2a3002383a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 98ad159499..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 f33fee4ce7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 fe43a1c911..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 5da7982e82..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 6fb539a5f9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 171dae8c5a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 5b440dda69..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7f967c28c7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9e852b7a6c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 99f9c2491c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8406c947cb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7234063462..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b3a91fd5e6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3e1aa64684..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 c57daa967f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 576e50e4e9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 10bb6fe166..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 e4d3bd35ca..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 89dbae811c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 969c0b61a3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 435dd821d3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 4f53978669..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3555905110..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 75541afe5d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3215b3beee..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 22f41b615a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 1c15e63abb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ab105b7637..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 cd88761a89..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 ba3730d318..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 edb62cf830..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a2a9b9bc70..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a579f8bb6a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3e23e8f32c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 a1469bfddf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 441e699695..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 8e465421a3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 05fd89a47e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 45cfb5e9c1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 3ba049ba8f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 9de55faf9b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 7e15564c79..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b29139f764..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 b84d82d837..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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 5b6f6313f4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// 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/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 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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +// Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. // 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..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 `c243e568` (the stamp in every `ops/generated/` file -header), the dialect holds 576 operations across 42 modules, and they come +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 `c243e568`, 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 @@ -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/abi-v1.toml b/cuda-oxide/intrinsics/abi-v1.toml index 04cefc916a..c82c2cfa98 100644 --- a/cuda-oxide/intrinsics/abi-v1.toml +++ b/cuda-oxide/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/cuda-oxide/intrinsics/catalog.json b/cuda-oxide/intrinsics/catalog.json index 9c57efa92b..52aea22d8d 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: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", @@ -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.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", @@ -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.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", + "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-13.3.73-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-23.1.3-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..8f7e7258b8 --- /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.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" + } + ], + "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.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" + } + ], + "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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. 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.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`. + - 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.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`. + - 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.toml b/cuda-oxide/intrinsics/overlay.toml index 1ba76db28d..540796cc07 100644 --- a/cuda-oxide/intrinsics/overlay.toml +++ b/cuda-oxide/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/cuda-oxide/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml b/cuda-oxide/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml new file mode 100644 index 0000000000..21c092f2a9 --- /dev/null +++ b/cuda-oxide/intrinsics/overlay/packed_conversion_fp8_bf16x2.toml @@ -0,0 +1,20 @@ +# 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 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-23.1.3-aab11ee0" +libnvvm_evidence_profile = "cuda-13.3-libnvvm-13.3.73" +runtime_validation = "unexecuted" +fp8_formats = ["e4m3x2", "e5m2x2"] +relu_variants = true +product_count = 4 diff --git a/cuda-oxide/intrinsics/probes/abs_bf16x2.ll b/cuda-oxide/intrinsics/probes/abs_bf16x2.ll index e4f90dbf4f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d282bddf00..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cfb28da0a0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9e8b4b623c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9c9be96eed..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8bd18c88a6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 15b70e709e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6db9ffcbc0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 969c43674e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aa286f0f62..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b6b6565745..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a03d96f4a1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1c618b8bff..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cdc0104e4b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5af9974339..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 12355cdf98..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 31e4fc10a6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2e8a575e27..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5e098e67bc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6daf152ec6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 731fb3e96a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4922207ad2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1d6bc06f27..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 66bfa0676a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 de5dedd64b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9182470492..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cceab23750..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cb66ad10f4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 97b18cd57a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 96081557da..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 304516396b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cfc169549d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c03df80219..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 08e8e7e515..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4c91be89ba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ad7e518834..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ccaed8aec9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 97fb497976..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 24e0cf60ac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b1d298dd80..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ad845532eb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3d78afddb1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7048c46326..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2066c67b2a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 56b24a2ca6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 69ec9340ab..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fee6f3fa0c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4920f565fa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d3dc26d79b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 418a1c73fa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aec4b27149..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a44a1cfa20..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 38ffb0ce73..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 56dfe3d6b4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aed88470f6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f51209ab7f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3148c69953..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 39fcd3f775..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9384df8463..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 72edce8ff2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d3ee68d839..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b5e1ebc766..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dcb424b2de..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 08cb319243..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0f89613939..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d3007ef56d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 be7a7a08aa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 51b8879cae..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b0f443649b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 69fc747df9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c72d22ab0d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0e3115de83..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6e16ebc12f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 405c6b104e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a1bfe4e31c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 72da0da001..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f9247ae03f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a49fab07b3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 734a7ae722..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a927796f63..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 51a44d497d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5962e7a5a3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b4e0d8cfd5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 42ba75f968..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0c97d6057b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dcf767ab90..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 03600b606f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 899b146871..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1f7da55135..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c990b39f80..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b2266ee055..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8b9b870c67..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 39f9357b2c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 83dd948659..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f726d34eb9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bf08d0169b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7515b63d4c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8b3c1839f9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5c3dedff9e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 96feaf9b06..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 045c4388c4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 35062c7b02..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a529f2b873..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d2977dfc28..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e18e0ecee3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9471b296a5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b7ea799c65..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cd741e34cd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 91d6f48365..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 07e0943450..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ca6edd125f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0aa4d4b10c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0b1a2fe822..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3806624727..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a922f25e39..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0252d6f7e0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1b81e77aeb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d5c2922c22..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1486b3d709..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 41dc2529ea..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ae03b1fceb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9619620a5c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2dd47ef1bc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b2793b839a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f59d0ab3dd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bb49fcd1a4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a6c45925a5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c885e75016..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6a1a53cdaf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9a84e9a095..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d1ed727d67..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 db43ebc6d7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c65c97f089..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 180c30a72c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9e49134245..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6888b3f9fc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2727eb9958..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b26850ff5f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4a303f2f91..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f8d523f2c7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f1104484c8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d25cd2aa46..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c2fdb514b6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a77f118b64..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a171fba27f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1538221ea6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d3832a0185..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9d7a79214c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2b8f449c46..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ea00c76c5b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b259dd25a2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d485f438b6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2f33dd8e47..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6238356da1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a9cf6db63d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d14f710cef..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 299203b666..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8d20bef888..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e3ea390a0b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7c785b4521..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5f0fadf127..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cfa609bfc4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d107ebada2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9a8167a5fa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fd68791d7b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4d7856699e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e002dcbaf2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c7dd69139f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 73ea5503d5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 378d524a1d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9ed6a52162..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e285d27077..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0f8eb28414..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0addb10c1a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 61f4b8c68b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 99de175278..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d49656c4e0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 26ce456bc6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 347c99ea5c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e68eafcffa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aa9ccb605e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5feca63319..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 daab8d78d9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9e8688dea8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 920123763b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6e4b6697e5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2c100257bb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 72f2068054..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 798279c8b3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 new file mode 100644 index 0000000000..081544c233 --- /dev/null +++ b/cuda-oxide/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: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. + +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/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e4m3x2_f16x2.ll index d90d9e288a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2df0cc6f42..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 new file mode 100644 index 0000000000..0bb0c24789 --- /dev/null +++ b/cuda-oxide/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: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. + +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/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll b/cuda-oxide/intrinsics/probes/cvt_rn_satfinite_e5m2x2_f16x2.ll index fc62e942db..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 694cc51b18..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 new file mode 100644 index 0000000000..0b467a2c90 --- /dev/null +++ b/cuda-oxide/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: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. + +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/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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 45cdaeb7a7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 new file mode 100644 index 0000000000..299f50b531 --- /dev/null +++ b/cuda-oxide/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: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. +; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. + +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/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..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cac2a150ff..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fa711fd95f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a1ddc82b97..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 764d2a6e71..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a5c282a4c9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6a946aed71..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 000b7054c0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f7a12ec46c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 733dda6360..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5b6c24128c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f8f39abb08..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e8f91ad578..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 09f00888e4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0ea9c58769..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 228cc0aa87..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bd90372579..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 985d40669e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 264c51acac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2a63516b60..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2afa9efd1d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e0e3aa902a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a0efcae75c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 656c119c87..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c876a0c3b8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f7c3e2f850..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ac0329ea6f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9a6bbb84c9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 693c323247..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 89000848ba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b0ae0bf16e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3ceb7bacc1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 56c243c066..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2084941712..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c26aa21244..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 57c4d39162..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 af884de851..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dc1d7a4a5b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b0d7997a38..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7306e10b93..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ad98279bac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 54f2db264f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fd539cbdcd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 07cdb52a2a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1af7bc56ea..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a1e3f909a1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b2b7da6b0a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d02c4efd94..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d63f549758..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dbdb54654f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c63b505f08..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a48d2cb019..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0227a87cf5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3006ca5867..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4178195aff..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1dab638241..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 88f236546c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e6ce5730d2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 69c4856a3f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1f06311c75..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 85fc23c5e9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bd78e7fec2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 172f5a198f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 15c7ff5a84..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d5d90f504e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f6a0afb978..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fe95c6605a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aa8de7458c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7f8f6bcdb5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7675704ac7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 abd1aa265f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c2c322df2a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 64bc22b5ab..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e938739361..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 39143c43a8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 795d4c65bd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4d0f422f23..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 22b01235a1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ed8b184bc3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 03b8c41659..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ef0b109944..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bb61385a0f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 341ecde82b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c562b17969..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b90c665711..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2cbb8fa061..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d46c40c165..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 00341c087b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7cfe56e7bc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a2b3c2796c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4bf2bb73f3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3810a36af7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 20799aed61..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f29b87e61f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 53a7542517..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 924858d48a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 189d4d8182..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2ba6ad2898..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6056d52646..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7eeffa64bd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c64347b04d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0333251475..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5f2ccf3356..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1819f1a48a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fb06db92da..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d610e0ceba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4ddddf4817..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e6ea8ec6e2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 216a433048..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3f3b32a4b7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4c4f562852..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5437865344..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 38c1e31d97..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f823d1f56d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2c7a254706..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d20f9b8100..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3150cdcf51..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 17ee6027fd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c0c2733487..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b076a21e01..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cfc9acc3e1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e4e65e322a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 77be96d793..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7150683ad3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d12d32818b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 eb099b1f93..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9ce110e2e6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 99431d33f4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5b041b2ded..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3fce4e7b67..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5ccb70e93a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d1b271ce91..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0663c3c250..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6c218589e8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 289a6f955e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3fa48796ac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9239b9fe52..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0da393f4ff..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f91129746f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4656697104..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 974357c43b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 437daaa6bb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5d73b53155..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ab4af618e8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4f558e905c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bd8265fe52..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5e8e6ed0d7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 db5370ba75..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6a2e6c31c9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 03e42f432b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a73ef96512..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f053059555..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 04aefd53a3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ca09146c29..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1b0d581793..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aff83d50be..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4e35deeb04..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 65fe02b980..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3cf7ee744f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 910ac349d8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2c7fae458f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0e21dc951d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 df497af274..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8374649481..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 128985b3ea..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3bbd2641ca..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f57634c72f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a944cfb4dc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f295381bb6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3d9104e413..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b06fa76805..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 00d289910f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b583273e3c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aa8217ccf9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 53ceb1f1a5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 037e2349a9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 153ba0e6d6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d2e8fa9463..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 27fd6d5ccc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c533c9544b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 20513a48e4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1a5d56cda6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 16d01c1151..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1607ff0343..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e65f385040..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 35b19c4fd1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f87889e445..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 653e8bf18d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c08d69fe08..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ab32ac384a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2f9c14a0d2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 211614bc5d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c4e6e964e5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3b46362412..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9e9415875f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0c5a76f1d3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9486508752..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1ea6db4179..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f343c80ca5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f147489e24..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7776405699..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8e3187734e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2f60560537..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0107ddc2be..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 348ca5a36b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 044208ada0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 da44fee12a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cb86d24427..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 10859f722d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0ba1467481..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 92879daa3a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 07c76884b2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3fa61efac6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a5032ec5cb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 70397939ac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 39b5948174..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f6f7f63e30..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3038cd2516..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e7ed2d289f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7db1a4bb35..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4f85a99466..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d2104e7fb4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8ca6cb55a3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 db9953f57d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 87a760c4ff..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dfd197e064..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 735b58026a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 eaa10620d8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e44cce86e8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8fbc3419a0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e3f1714a86..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 76c3c32edf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3c8be80062..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 eee0fafb8f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fd52e2f349..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3baa32de0e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 add7cc4475..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9e7c95d969..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4b7122373b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c4ec2080fb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 630f56bf62..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fecfeab6b8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1a0bc3f45b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bc46a827c3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 951a05b0f6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 96c95b15a5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b0b2bcc4a1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 52db2b582e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 84c6daf029..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 022877ae35..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 96c70b0d9e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 97fe74ef02..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 69f1fabc6e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 de067d5495..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ed9138286c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1c775c4e24..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4cc5ed515c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6f432aa964..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5141ec0786..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ffccc1412f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b4d7276108..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1221b46de6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 779b69c37b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b51a697a76..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 767763cb26..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 92c8be40b4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 980aa082e0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2101d232ac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 46d5d57412..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4bec7816f6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c6f7ee644a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 38c8822a13..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 823495e8ab..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 00aa64998e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fd74bfbf88..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 751d6de387..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e62f38b979..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9bb91ff4be..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6ac77efe23..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a29c45e41e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8b6865a36d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f1a47b1431..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c141a5ed75..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bcd74f3058..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 15c81cafb8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1090808214..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d8a60527d0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9a10044a3c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c9839d36d0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0ad35af407..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b3433362fc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d009f0a87d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7f33262008..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c3de14683a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 db4ba8e839..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e220edb8f2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7900e8e9e6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f879eaf258..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7afaf17ef0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6b40349352..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 699d72600e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 df4a91ccd7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7af2a65bb3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7b171cbe43..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7eb4d7daae..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c6017a7c15..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 051cc6c569..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b7f1b1db9e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b206c7932c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ac519e39fc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d4f89e99bb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 209a3d9989..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a681419efb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5c96b8193e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cecdf1f30f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7c350cf31b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0bebe9da87..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 980f0166b2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0efe91ff83..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6e70e7d1ec..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e9890ce30d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9ee23211c7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 14d8201f7e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 19343fb6de..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8826984e26..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3344055106..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a33954d5ee..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 934c12dff5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b021c14fac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e64504fff5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7caff5c91c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d030101330..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 473c2e79de..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3a432cfb15..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 adddb22d13..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4a8d18d3da..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 054ea5e279..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9bf0b3be33..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 50bc3245d2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6e4ffb7eb5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e2160e822c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dd72002d02..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b1d3623e4d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2ad3a82c38..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 530fd956cb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 53b117b1c6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7b524c3780..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 598c261e2d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b69f73456e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d3892576bf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2e369de46c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 96d692c1ac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 87cb08ca39..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cfe144b4f5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 07e3b09849..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 54beec52dd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3ae0c3ab34..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 33b8a3e468..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3d7fdf7538..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c476d84a31..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7114cc098e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ee3ac32bed..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5cc5b3a4f4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 97dda1e812..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f05b0a4760..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 00a0005305..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f60dcffd1b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7352440cd0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 64747e411f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 99f94d3258..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5cebce211f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1a2729aa84..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4ffcd5c1b4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 74262a9bd4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c798abd261..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f685ca86ab..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6d046a2fa9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b0be097754..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1f9c965c2d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4240c1d1fa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3011a63159..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 82307523e8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 35d18671bf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d70ff12bf7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 83026e728b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1598d12615..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4332a7521b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 684f3a2e70..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 21bb6d0c6d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c54884e450..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cb2ddfa7aa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 01817551d1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 525df8de25..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b8fd3dde7e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 31643145aa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4c4533956b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b284e897b7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f0f1dbd4a4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 be5dde36bc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 85056d3498..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cbba60c4ba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 15c849298e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8c42464754..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b1050defb6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d88fe3e3eb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0fd3ba60e1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4d1c1830ae..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b3dab246b3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c9e1a911e9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 021358fd1d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 70b7321527..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4914fd926b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6fd7141112..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6e85ddd82a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 02a23a1169..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 618904c371..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2d9b3994b7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 456c9a0483..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c5f468014c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9fc3d9b536..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 30466576ba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e56812c604..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 83c991c331..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5ba1ecb190..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c3fbfa9f94..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d32068231a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 da1a18a093..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4e2a6b84ec..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 60d12b5bba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a0b26fc7e0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7f1f38bee5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e51fcbf0e9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 78bffdbec0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 eb254e0845..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 961ec8b463..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cccadb49d9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 13c28b8e69..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 87433ab492..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 377d250f54..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4274cd36df..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2f26e87cc5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 78a15d17ca..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3b5459eae9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 080082879d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2973d38341..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5fd43812e4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 eb3ff37c98..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 25662838f7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e4554753f0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 86de360fe1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 017859b219..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e1e2970290..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ede1412f6d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6e33033b3a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ed5694e626..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b034f065e4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3f4d201382..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bca458a7c3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3a2cc89340..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2173d4db76..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4a9fdedf9a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 20e4d53b8a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ddbe82e11f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6aff5c6f88..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0421f0ffc6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1d59d7a529..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1a9cce88d0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 17d4263041..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9c8a310db5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 491a194d31..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1d1a5e1c84..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d4e56c6b32..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 254cf0452f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a03a94023e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f235abf757..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0e68928c1e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 10b9927d32..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 204e98b791..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 726b7ea310..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 092afdea4a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7a4daaeeb6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e0e2b089d4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1e194d6780..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 513893e676..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 86b0f26e72..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c936466501..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d6b9062a8a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e43e931133..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 98b013e2c0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a0b25e06c2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b5932f240f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1edbbaaef0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a8f00b6bd3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9fa2605615..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ef3c013ae5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dc2600aad3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ee5840d2d9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7d0749f632..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 69cffffa8a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b9f83e5cb8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ee1c99f101..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1111cab0e8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0ec238e0da..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ff9af628e0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dc36ea55be..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e7d238a19f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a1218e54b3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bbb2e73e77..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 48b9949bd5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bbcd5d8d54..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aae7467217..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8830088f2e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cb9e316c01..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ad7ad7aaef..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e16fed6685..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e52fc61d12..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 818212973b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e0357d9a35..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f1cf96b1e4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d534303ff7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d116a4b0fd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4e6231b9b1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 91617991d7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 42337793ae..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 211d1bb57f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ce89a31458..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0c2e753bb3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b8c214ab64..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 71ed83b2fc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 26e5f081cb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fbc23a8b97..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 119729a5fb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 602124e3a7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 26e1140870..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e9770680f2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0ca86b7845..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f96e6ddf2f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ffe5fd9168..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ee0fc0d5a4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 457e80576a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ed161fd7e1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c03c50cf02..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e4d7228376..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4f48c804f1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 828ad30b00..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d2cba3359d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e1953f3cfa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 db2087224f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 02a95a332e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6b0e005a30..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c627065725..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e6f72c28ae..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2a7a9a7a70..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c14e1bc459..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bb43fd0b6d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 13be404b92..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 37030a5272..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b25cc907c1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 46b9532127..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 56238c93d7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e46f7410da..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4d5c293ef1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 597b0dcc44..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 66cae7bf2d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f30d3a602f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e4f66f1f98..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2ffca499c0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ae698906d1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a0f3e67b69..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 84f69a4c2b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2d18411bf8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e582bd4225..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e36ad71f3d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d77884c98c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 08308e002c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 abe907ac33..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e6498b0c18..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8b1fbcef6e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 32b43dd8c2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5db2851f97..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4f7d355392..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8970e79012..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d77114a98d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 686258e982..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a402a2c3f7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 683b1be50a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bfa844406b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5f0d7e4493..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b823ac413b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dfb302cecb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 779a70d70e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cb0f8096d4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2e30ccc923..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f92286e1f9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 59af03b61b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 df4b9c0626..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a56fb45429..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 31a987f4e0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1b5002737d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b2ff7cd266..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5a92491c7c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 40c0f29f78..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 12f23ad13d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 09e05d6ef3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 431b81da7a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 09774f7173..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e6128ec64b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a729bc01d9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e63edfab81..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dca98ba632..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3f05d3a6e2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c09e87a586..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8d4d634cbb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 822e77604e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1d376316dc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 272e8ef0e1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bf72fbf488..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5b4bd94c8d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7837962693..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0c8ef2cfd3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a78f8aa6d6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a0b7e81098..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 02a4d9562c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 14d660a9cc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8656971789..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7bd6d95497..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d646da8df8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 82a1b27449..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 69ca6f3776..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 82d525ada3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 21adeb6aec..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f335cec4d1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 34170acc77..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6c4b3c8e84..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 463b96a822..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 7eb0ca8654..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5ae9bd0f15..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 586087f6a9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fed4990a1b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 647f31d933..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 650cea233f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3c1b0e92fd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 910354cb3f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 eca0f558cb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e9bbbe0942..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0231c8cee7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 18d52449a3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 52c3931b6d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5c14240d1c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d5780d2b58..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e473742565..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4446a354ea..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b45bcd8f64..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bfd93db9ea..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 61e93d7b86..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b970b32f99..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5e6ec4657f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f869e87f30..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 197489f82c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 070622fe59..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 03fb61e7fb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4a51274d3b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 90cb0d5fe4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2a94f9813d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4e2ab19a27..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b034aee05b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5fe0a8efed..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 83cac7b1e1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 71aa4fc632..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 43fac69148..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 22f0043a56..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 261adbdbc9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e5f9e24f90..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 02be0e8cad..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e74d6957ac..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 371d6f78b0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 71c6cc8313..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 722bbb9938..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 68d013a0dc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d9db4b69fd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 578be1d4dd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 730e54dbfb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 71c61e2deb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 183b1767e6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bb25fecc3c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e527fb2524..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 15dcf1f873..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f59cd79143..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8947077001..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 23f0103b48..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e2003259a6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 83f8e8f89c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 01c6816cfb..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1f54053c1c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8c24485dba..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4165c98728..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dcf3eb1268..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bd0569970f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 830163dda2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 41bf253cc1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fa2725a12d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 89ca9a137d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f290ab329c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f1bbfeef13..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1f4b9e89b4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e46f0ca74e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a7d62aaf59..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c94cd41efc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 dbf59bafdd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6dead57bca..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9ea197a58a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 74df38e7c9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d3fd44cf7b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2c1773e38b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e975159ebf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 54fbc8fca1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a160caea4e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fa6921722a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8de63e138d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2f264a9b12..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0f32015fb1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 35f7bc7167..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9117cdfd9e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5facb4658a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 748f593bc6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e3a2ddf422..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6bb3038653..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aedbde911b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c9211ebf93..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ff4d689ed2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6119c1b78a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3587b55966..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3d1513573e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 db6f697ace..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 94ffd5f6aa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e9445ec605..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e6d9cdf552..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 13650f29e6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9881f97c0a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 981f53b0f8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e80ecc14d9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a341605a9e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 61a4985f62..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 405b42e46a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b6cd43c4ed..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 edf0f0f845..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ba656768e3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 325dcf58ec..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 06939a001b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fe2ce75b77..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3f7835672d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9ce2e04316..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e0399b1b80..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ea382a3e4c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 eb14712d1a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cef55806f6..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cce1d76678..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ae71ecb02d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1309cf9a45..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 561ad2b226..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4a7a422acf..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 102e728e71..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 77e2d52b11..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2d6e832e41..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8f76bfac3b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8ed6bcc5aa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c3d4ae1412..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 88b8146e7f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e7c2de1788..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 734f56972d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9368684e3e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cd978dbea9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b452b20c0d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 291d882495..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 334d590622..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a1cc78d36b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3fc331675c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 351a96ea42..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 33179a9019..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c9e00cdf2e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8969ab99f5..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 13ae0c1322..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5615fdc594..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ccc1986b6d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 76ba355871..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 93ce5bfb45..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1d98697b7d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6ec802a078..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 54408310a9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 fce5aab34f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 579190ce50..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 945c41b3a1..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6c3d0cbc2b..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ebe4861d94..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cc93a7369e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8d9474b696..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 f5a16d9133..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 693fc0b853..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 2c73baa378..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5d0cf33455..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4f3dbb60ef..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 57ce4c0886..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5e85f3a09d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 9b709f422e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5a9c9373e0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bee9bffa8c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8c77ac107a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c8cc9d74ed..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 333f2b5205..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ac0554d50e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d9eb449d59..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 610f0af819..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 33a2ee4102..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 548bd59e8d..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 baad23fe6a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 98d7bbbb7e..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0ec73445c8..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5a3eb0984c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 bca6fc8d35..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b811beb1f0..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 43185ffd6a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d0f0f31dd9..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 cb489ffa7c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 b65e7f27aa..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 02e4af589f..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 8c289fca96..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 1e1f8ae263..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 aee07673e4..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 ce6ff316b2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e335b2b467..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 afb08c6673..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 d63b34a14c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 406589b9ea..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 df4ebf993a..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 04e6cb9ae7..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3e40aeed62..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 c75307a800..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 826e9530a3..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 4d92137b25..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5707a1d084..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 6da7e92ab2..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 a33eef123c..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 3106842b64..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 5a050dcd77..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 e9bf779030..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 0f230c57dd..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; 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 25ac4f95cc..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: c243e568fc16c9f907c9fd87fbacbe9f823043a9badd8633e43e9faa87adbd10. +; Catalog schema/version: 46/0.1.0; intrinsic ABI: v1; catalog SHA-256: 31b5238818203ac4c636984d416ea7791c5861feb0ad30d1f5fc693ebefb4399. ; LLVM source: aab11ee0ea5ea9a1d462ad0bc1fae9d9bb643019. target triple = "nvptx64-nvidia-cuda" 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 ]] \