Harmonize *ScaledMMAAttr operand order and drop MMAFragment
#22465
+401
−437
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
bjacob
requested review from
Groverkss,
IanWood1,
MaheshRavishankar,
Max191,
antiagainst,
benvanik,
hanhanW,
kuhar and
qedawkins
as code owners
|
Very harmonious, this PR restores tranquility. |
bjacob
force-pushed
the
drop-mmafragment-take2
branch
from
079ee69 to
305f0c2
Compare
Signed-off-by: Benoit Jacob <[email protected]>
bjacob
force-pushed
the
drop-mmafragment-take2
branch
from
305f0c2 to
1a009cf
Compare
hanhanW
approved these changes
Oct 30, 2025
Comment on lines
+210
to
+216
| using MMAIntrinsicTy = decltype(mma.getIntrinsic()); | ||
| const bool isScaled = std::is_same<MMAIntrinsicTy, ScaledMMAIntrinsic>::value; | ||
| const bool isLhs = isIntrinsicLhs<MMAIntrinsicTy>(operandIdx); | ||
| const bool isRhs = isIntrinsicRhs<MMAIntrinsicTy>(operandIdx); | ||
| const bool isAcc = isIntrinsicAcc<MMAIntrinsicTy>(operandIdx); | ||
| const bool isLhsScale = isIntrinsicLhsScale<MMAIntrinsicTy>(operandIdx); | ||
| const bool isRhsScale = isIntrinsicRhsScale<MMAIntrinsicTy>(operandIdx); |
There was a problem hiding this comment.
It looks much better than magic numbers to me.
Comment on lines
+1101
to
+1110
| ValueRange inputs = linalgOp->getOperands(); | ||
|
|
||
| SmallVector<Type> eltTypes; | ||
| smmaKind.getElementTypes(eltTypes); | ||
| if (cast<RankedTensorType>(inputs[0].getType()).getElementType() != | ||
| eltTypes[0] || | ||
| cast<RankedTensorType>(inputs[2].getType()).getElementType() != | ||
| eltTypes[2] || | ||
| cast<RankedTensorType>(inputs[4].getType()).getElementType() != | ||
| eltTypes[4]) { | ||
| return failure(); | ||
| for (int i : | ||
| {kScaledMMAOperandLhs, kScaledMMAOperandRhs, kScaledMMAOperandAcc}) { | ||
| if (cast<RankedTensorType>(inputs[i].getType()).getElementType() != | ||
| eltTypes[i]) { | ||
| return failure(); | ||
| } |
There was a problem hiding this comment.
It looks much better to me!
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR does multiple things that were easier done all at once:
*ScaledMMAAttroperand order:ScaledMMAAttrwaslhs, lhs_scale, rhs, rhs_scale, while the operand order ofDataTiledScaledMMAAttrwaslhs, rhs, lhs_scale, rhs_scale.ScaledMMAAttrto match theDataTiledScaledMMAAttrconvention. This propagates to a change of operand order in the enclosinginner_tiledops.MMAFragment:MMAFragment, that had unclear semantics: the enum values were sometimes used as opaque symbolic enums to refer to operand by "role", e.g. "Lhs", and sometimes used as the underlying integer values as operand indices, e.g. "Rhs == 1". This was originally OK as all MMA-like ops had the same 3 operands Lhs, Rhs, Acc. But when ScaledMMAAttr was introduced, that didn't... scale: now the preexisting enum value "Rhs == 1" didn't equal anymore the corresponding operand index under thelhs, lhs_scale, rhs, rhs_scaleconvention (1 != 2) and even regardless of convention, the enum value "Acc ==2" never corresponded to operand index anymore (2 != 4).operandIndex.