Skip to content

Refactor gfx950 A16W16 GEMM to use the universal kernel - #992

Merged
coderfeli merged 15 commits into
mainfrom
xyt/gemm16_950
Aug 11, 2026
Merged

Refactor gfx950 A16W16 GEMM to use the universal kernel#992
coderfeli merged 15 commits into
mainfrom
xyt/gemm16_950

Conversation

@xytpai

@xytpai xytpai commented Aug 10, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Replace the legacy HGEMM split-K and small-M kernels with a unified gfx950 A16W16 GEMM implementation.
  • Support FP16/BF16, NN/NT/TN/TT layouts, bias, FP32 output, slice-K, split-K, and HTI scheduling.
  • Add comprehensive correctness tests and restore HGEMM benchmarks, including BF16 8192³ HTI.
  • Remove obsolete kernels, tests, and documentation references.

Validation

  • gfx950 A16W16 correctness suite
  • Concurrent tuner + split-K stress test
  • GEMM benchmark suite
  • Black and Ruff checks

@xytpai
xytpai requested a review from coderfeli August 10, 2026 08:28
coderfeli
coderfeli previously approved these changes Aug 10, 2026
@coderfeli
coderfeli merged commit 8bd4c53 into main Aug 11, 2026
12 checks passed
@coderfeli
coderfeli deleted the xyt/gemm16_950 branch August 11, 2026 12:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants