Add vec4 gemm! - #7
Closed
haixuanTao wants to merge 2 commits into
Closed
haixuanTao wants to merge 2 commits into
haixuanTao wants to merge 2 commits into
Conversation
Adds vortx::linalg::{Activation (tanh + tanh_backward), Adam} and their shaders,
the GPU building blocks for MLP training (used by nexus RL demos / zealot-rl).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two vec4 changes to the tiled GEMM, both verified bit-exact (0.0) vs the scalar gemm_tiled: - gemm_tiled inner loop now does 4-wide vec4 FMAs (acc: [Vec4;4], bvec.mul_add(Vec4::splat(a), acc)) instead of 16 scalar MACs. - new gemm_tiled_vec4 shader + Gemm::dispatch_tiled_vec4 host: loads the A/B tiles from global memory with 128-bit vec4 transactions (host reinterprets the buffers as glamx::Vec4). Assumes contiguous row-major, single batch, and tile-filling dims (M%64, N%64, K%16); the host falls back to scalar gemm_tiled for transposed/odd-dim cases. Pulls in glamx. Measured on an RTX 5090: the compute-FMA vec4 is ~+12% (rust-gpu -> SPIR-V -> NVIDIA does not auto-vectorize the inner loop the way Metal does); the vec4 *global* loads are ~+0% because a tiled GEMM already amortizes global memory (load once into smem, reuse), so it is not bandwidth-bound. Lesson: for a tiled GEMM, vec4 the compute, not the global loads. Note: the one-line glamx Cargo.toml dependency is shared with the ELU and PPO-kernels branches; whichever lands first, the others need a trivial rebase of that line. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Member
|
This is being merged as part of #11 |
sebcrozet
added a commit
that referenced
this pull request
Aug 27, 2026
* feat: move inferi shaders into an optional ml module for vortx * chore: cargo fmt * feat(ml): add tanh activation + Adam optimizer GPU ops Replaces #5 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * refactor(ml): move the tanh-backward + Adam kernels into vortx::ml and drop the duplicated tanh forward Completes #5 * feat(ml): add ELU activation GPU ops (fwd/backward/vec4) Replaces #6 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * refactor(ml): keep only the ELU backward pass, the forward is already UnaryOp::Elu Completes #6 * feat(ml): add PPO loss-gradient GPU kernels Replaces #5 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * refactor(ml): wire the PPO kernels to vortx::ml and use StepRng for uniform control flow Completes #5 * perf(linalg): vec4 GEMM (compute-FMA inner loop + vec4 global-load variant) Replaces #7 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * perf(linalg): keep the vec4 FMA inner loop, drop the unmeasured vec4 global-load GEMM variant Completes #7 * feat(ml): gpu_ppo_stage_batch, building the PPO minibatch on device * chore(ml): comment cleanups * chore: CI fixes --------- Co-authored-by: Haixuan Xavier Tao <tao.xavier@outlook.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two vec4 changes to the tiled GEMM, both verified bit-exact (0.0) vs the
scalar gemm_tiled:
bvec.mul_add(Vec4::splat(a), acc)) instead of 16 scalar MACs.
A/B tiles from global memory with 128-bit vec4 transactions (host
reinterprets the buffers as glamx::Vec4). Assumes contiguous row-major,
single batch, and tile-filling dims (M%64, N%64, K%16); the host falls
back to scalar gemm_tiled for transposed/odd-dim cases.
Pulls in glamx. Measured on an RTX 5090: the compute-FMA vec4 is ~+12%
(rust-gpu -> SPIR-V -> NVIDIA does not auto-vectorize the inner loop the
way Metal does); the vec4 global loads are ~+0% because a tiled GEMM
already amortizes global memory (load once into smem, reuse), so it is
not bandwidth-bound. Lesson: for a tiled GEMM, vec4 the compute, not the
global loads.