Code

[js/webgpu] optimize MatmulNBits#21747

Code

Merged

guschmue merged 11 commits intomicrosoft:mainmicrosoft/onnxruntime:mainfrom qjia7:opt_matmulnbitsqjia7/onnxruntime:opt_matmulnbitsCopy head branch name to clipboardAug 23, 2024

+124-187

Commits

Commits on Aug 16, 2024

• opt matmulnbits

  qjia7

  committed
  79a4ac1

  Copy full SHA for 79a4ac1

• add outputNumber > 1

  qjia7

  committed
  6bd5417

  Copy full SHA for 6bd5417

• clean code

  qjia7

  committed
  d28ac0a

  Copy full SHA for d28ac0a

• formant and add missing shaderCache hints

  qjia7

  committed
  dab4542

  Copy full SHA for dab4542

Commits on Aug 19, 2024

• use global_idx

  qjia7

  committed
  e86150a

  Copy full SHA for e86150a

• tune outputNumber

  qjia7

  committed
  e007210

  Copy full SHA for e007210

• add limitations

  qjia7

  committed
  2bf70ef

  Copy full SHA for 2bf70ef

Commits on Aug 21, 2024

• Merge branch 'main' into opt_matmulnbits

  qjia7

  committed
  f68a2da

  Copy full SHA for f68a2da

Commits on Aug 22, 2024

• fix workgroupSize to reduce shader recompilation

  qjia7

  committed
  cbacc4a

  Copy full SHA for cbacc4a

• support zeroPoints as input

  qjia7

  committed
  cb775ed

  Copy full SHA for cb775ed

• replace the old algorithm

  qjia7

  committed
  2ab2c4e

  Copy full SHA for 2ab2c4e