Code

DP4AMatMul perf refinements#23539

Code

Merged

guschmue merged 2 commits intomainmicrosoft/onnxruntime:mainfrom user/sushraja/dp4matmul_finetunesmicrosoft/onnxruntime:user/sushraja/dp4matmul_finetunesCopy head branch name to clipboardJan 31, 2025

+107-63

Commits

Commits on Jan 30, 2025

• Implement memory layout optimizations. Move away from array for lane_output which leads to it being put on registers.

  sushraja-msft

  committed
  d613a0f

  Copy full SHA for d613a0f

Commits on Jan 31, 2025

• Unroll nvidia half of the loop

  sushraja-msft

  committed
  4a3799d

  Copy full SHA for 4a3799d