Code

Subgroupmatmul memory optimzations#23758

Code

Draft

sushraja-msft wants to merge 14 commits intomainmicrosoft/onnxruntime:mainfrom user/sushraja/subgroupMatrix_load_optimizationmicrosoft/onnxruntime:user/sushraja/subgroupMatrix_load_optimizationCopy head branch name to clipboard

+285-4

Commits

Commits on Feb 19, 2025

• Upgrade Dawn

  sushraja-msft

  committed
  7e2b8db

  Copy full SHA for 7e2b8db

• Code Builds

  sushraja-msft

  committed
  91cda62

  Copy full SHA for 91cda62

• Runs but garbage outputs

  sushraja-msft

  committed
  9f838ff

  Copy full SHA for 9f838ff

• Add restriction to use subgroup matmul for prefill only

  sushraja-msft

  committed
  aaeb58b

  Copy full SHA for aaeb58b

• First round bug fixes

  sushraja-msft

  committed
  2d406da

  Copy full SHA for 2d406da

• Remove safeMatrixStore

  sushraja-msft

  committed
  a3102cd

  Copy full SHA for a3102cd

• Add FP32 support

  sushraja-msft

  committed
  2866c26

  Copy full SHA for 2866c26

• Add support for compute precision

  sushraja-msft

  committed
  acc3ab6

  Copy full SHA for acc3ab6

• Restrict to accuracy level 4

  sushraja-msft

  committed
  2ad480c

  Copy full SHA for 2ad480c

• Lintrunner

  sushraja-msft

  committed
  8d7ae12

  Copy full SHA for 8d7ae12

• Revert change to .gitignore

  sushraja-msft

  committed
  c5856df

  Copy full SHA for c5856df

• Fix mac build break

  sushraja-msft

  committed
  aa6e8c0

  Copy full SHA for aa6e8c0

• add comments

  sushraja-msft

  committed
  09e30be

  Copy full SHA for 09e30be

• Subgroupmatmul memory optimzations

  sushraja-msft

  committed
  91b7eb6

  Copy full SHA for 91b7eb6