[MFMA] Change kWidth parameter semantics

This PR changes kWidth semantics "from elements per instruction" to
"elements per thread per instruction" along k axis.
This commit is contained in:
Aleksandr Efimov
2023-09-11 18:01:57 +00:00
committed by Lixun Zhang
parent 10795d8fd3
commit d80cd2d374
7 changed files with 21 additions and 18 deletions

View File

@@ -389,7 +389,7 @@ bool isMfmaToDotShortcut(RankedTensorType &srcTy, RankedTensorType &dstTy) {
// layout when opIdx == 1.
return mfmaLayout.getWarpsPerCTA()[1] == 1 &&
dotOperandLayout.getOpIdx() == 0 &&
dotOperandLayout.getKWidth() == 8 &&
dotOperandLayout.getKWidth() == 4 &&
dotOperandLayout.getParent() == mfmaLayout &&
mfmaLayout.getIsTransposed() &&
(srcTy.getElementType().isF16() || srcTy.getElementType().isBF16());