Skip to content

[MLIR][Vector] Add support for inner-parallel masked multi-reductions #126722

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Feb 14, 2025
Merged

[MLIR][Vector] Add support for inner-parallel masked multi-reductions #126722

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 21 additions & 9 deletions mlir/lib/Dialect/Vector/Transforms/LowerVectorMultiReduction.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -308,12 +308,6 @@ struct TwoDimMultiReductionToElementWise

LogicalResult matchAndRewrite(vector::MultiDimReductionOp multiReductionOp,
PatternRewriter &rewriter) const override {
auto maskableOp =
cast<vector::MaskableOpInterface>(multiReductionOp.getOperation());
if (maskableOp.isMasked())
// TODO: Support masking.
return failure();

auto srcRank = multiReductionOp.getSourceVectorType().getRank();
// Rank-2 ["parallel", "reduce"] or bail.
if (srcRank != 2)
Expand All @@ -330,15 +324,33 @@ struct TwoDimMultiReductionToElementWise
if (!elementType.isIntOrIndexOrFloat())
return failure();

OpBuilder::InsertionGuard guard(rewriter);
auto maskableOp =
cast<vector::MaskableOpInterface>(multiReductionOp.getOperation());
Operation *rootOp;
Value mask = nullptr;
if (maskableOp.isMasked()) {
rewriter.setInsertionPoint(maskableOp.getMaskingOp());
rootOp = maskableOp.getMaskingOp();
mask = maskableOp.getMaskingOp().getMask();
} else {
rootOp = multiReductionOp;
}

Value result = multiReductionOp.getAcc();
for (int64_t i = 0; i < srcShape[0]; i++) {
auto operand = rewriter.create<vector::ExtractOp>(
loc, multiReductionOp.getSource(), i);
result = makeArithReduction(rewriter, loc, multiReductionOp.getKind(),
operand, result);
Value extractMask = nullptr;
if (mask) {
extractMask = rewriter.create<vector::ExtractOp>(loc, mask, i);
}
result =
makeArithReduction(rewriter, loc, multiReductionOp.getKind(), operand,
result, /*fastmath=*/nullptr, extractMask);
}

rewriter.replaceOp(multiReductionOp, result);
rewriter.replaceOp(rootOp, result);
return success();
}
};
Expand Down
20 changes: 20 additions & 0 deletions mlir/test/Dialect/Vector/vector-multi-reduction-pass-lowering.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,3 +41,23 @@ func.func @vector_multi_reduction_parallel_middle(%arg0: vector<3x4x5xf32>, %acc
// ALL-SAME: %[[INPUT:.+]]: vector<3x4x5xf32>, %[[ACC:.+]]: vector<4xf32>
// INNER-REDUCTION: vector.transpose %[[INPUT]], [1, 0, 2] : vector<3x4x5xf32> to vector<4x3x5xf32>
// INNER-PARALLEL: vector.transpose %[[INPUT]], [0, 2, 1] : vector<3x4x5xf32> to vector<3x5x4xf32>

// -----

func.func @vector_multi_reduction_masked(%arg0: vector<2x4xf32>, %acc: vector<2xf32>, %mask: vector<2x4xi1>) -> vector<2xf32> {
%0 = vector.mask %mask { vector.multi_reduction <mul>, %arg0, %acc [1] : vector<2x4xf32> to vector<2xf32> } : vector<2x4xi1> -> vector<2xf32>
return %0 : vector<2xf32>
}

// ALL-LABEL: func @vector_multi_reduction_masked
// ALL-SAME: %[[INPUT:.+]]: vector<2x4xf32>, %[[ACC:.+]]: vector<2xf32>, %[[MASK:.+]]: vector<2x4xi1>
// INNER-REDUCTION: %[[INNERVEC:.+]] = vector.extract %[[INPUT]][0] : vector<4xf32> from vector<2x4xf32>
// INNER-REDUCTION: %[[INNERACC:.+]] = vector.extract %[[ACC]][0] : f32 from vector<2xf32>
// INNER-REDUCTION: %[[INNERMASK:.+]] = vector.extract %[[MASK]][0] : vector<4xi1> from vector<2x4xi1>
// INNER-REDUCTION: vector.mask %[[INNERMASK]] { vector.reduction <mul>, %[[INNERVEC]], %[[INNERACC]] : vector<4xf32> into f32 } : vector<4xi1> -> f32
// INNER-PARALLEL: %[[TPMASK:.+]] = vector.transpose %[[MASK]], [1, 0] : vector<2x4xi1> to vector<4x2xi1>
// INNER-PARALLEL: %[[TPINPUT:.+]] = vector.transpose %[[INPUT]], [1, 0] : vector<2x4xf32> to vector<4x2xf32>
// INNER-PARALLEL: %[[INNERVEC:.+]] = vector.extract %[[TPINPUT]][0] : vector<2xf32> from vector<4x2xf32>
// INNER-PARALLEL: %[[INNERMASK:.+]] = vector.extract %[[TPMASK]][0] : vector<2xi1> from vector<4x2xi1>
// INNER-PARALLEL: %[[REDUCED:.+]] = arith.mulf %[[INNERVEC]], %[[ACC]] : vector<2xf32>
// INNER-PARALLEL: arith.select %[[INNERMASK]], %[[REDUCED]], %[[ACC]] : vector<2xi1>, vector<2xf32>