[mlir] Extract RHS rows once when lowering vector.contract to dot #130130

abulavin · 2025-03-06T16:10:49Z

The vector.contract op on two matrices A and B will be lowered to individual dot products of each row and column of A and B respectively. The existing lowering will extract each column of B for each row of A, which leads to multiple values in the IR representing the same columns of B.

This PR makes changes to the ContractOpToDotLowering to make sure that the columns of B are only ever extracted once, so then the SSA values representing the extracted columns are then re-used in the IR for later dot products.

I have updated the existing vector-contract-to-dot-transforms test.

github-actions · 2025-03-06T16:11:08Z

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

llvmbot · 2025-03-06T16:11:44Z

@llvm/pr-subscribers-mlir

Author: Artemiy Bulavin (abulavin)

Changes

The vector.contract op on two matrices A and B will be lowered to individual dot products of each row and column of A and B respectively. The existing lowering will extract each column of B for each column of A, which leads to multiple values in the IR representing the same columns of B.

This PR makes changes to the ContractOpToDotLowering to make sure that the columns of B are only ever extracted once, and then that the SSA values representing the extracted columns are then re-used in the IR for later dot products.

I have update the existing vector-contract-to-dot-transforms test.

Full diff: https://github.com/llvm/llvm-project/pull/130130.diff

2 Files Affected:

(modified) mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp (+11-3)
(modified) mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir (+2-4)

diff --git a/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp b/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
index 21261478f0648..88227ecb6c675 100644
--- a/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
@@ -758,12 +758,20 @@ FailureOr<Value> ContractionOpToDotLowering::matchAndRewriteMaskableOp(
   Value res = rewriter.create<arith::ConstantOp>(loc, dstType,
                                                  rewriter.getZeroAttr(dstType));
   bool isInt = isa<IntegerType>(dstType.getElementType());
+  llvm::SmallVector<Value> extractedCols;
+  extractedCols.reserve(dstColumns);
   for (unsigned r = 0; r < dstRows; ++r) {
     Value a = rewriter.create<vector::ExtractOp>(op.getLoc(), lhs, r);
     for (unsigned c = 0; c < dstColumns; ++c) {
-      Value b = rank == 1
-                    ? rhs
-                    : rewriter.create<vector::ExtractOp>(op.getLoc(), rhs, c);
+      if (r == 0) {
+        // We only need to extract the rows of the RHS once
+        // and then re-use them later.
+        Value b = rank == 1
+                      ? rhs
+                      : rewriter.create<vector::ExtractOp>(op.getLoc(), rhs, c);
+        extractedCols.push_back(b);
+      }
+      Value b = extractedCols[c];
       Value m = createMul(op.getLoc(), a, b, isInt, rewriter);
       Value reduced = rewriter.create<vector::ReductionOp>(
           op.getLoc(), vector::CombiningKind::ADD, m);
diff --git a/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir b/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir
index 0ba185bb84760..f77027c3c2d95 100644
--- a/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir
+++ b/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir
@@ -169,13 +169,11 @@ func.func @extract_contract3(%arg0: vector<3xf32>,
 // CHECK:    %[[T21:.*]] = vector.insert %[[T20]], %[[T11]] [0, 1] : f32 into vector<2x2xf32>
 //
 // CHECK:    %[[T23:.*]] = vector.extract %[[A]][1] : vector<2xf32> from vector<2x2xf32>
-// CHECK:    %[[T24:.*]] = vector.extract %[[Bt]][0] : vector<2xf32> from vector<2x2xf32>
-// CHECK:    %[[T32:.*]] = arith.mulf %[[T23]], %[[T24]] : vector<2xf32>
+// CHECK:    %[[T32:.*]] = arith.mulf %[[T23]], %[[T2]] : vector<2xf32>
 // CHECK:    %[[T33:.*]] = vector.reduction <add>, %[[T32]] : vector<2xf32> into f32
 // CHECK:    %[[T34:.*]] = vector.insert %[[T33]], %[[T21]] [1, 0] : f32 into vector<2x2xf32>
 //
-// CHECK:    %[[T40:.*]] = vector.extract %[[Bt]][1] : vector<2xf32> from vector<2x2xf32>
-// CHECK:    %[[T41:.*]] = arith.mulf %[[T23]], %[[T40]] : vector<2xf32>
+// CHECK:    %[[T41:.*]] = arith.mulf %[[T23]], %[[T12]] : vector<2xf32>
 // CHECK:    %[[T42:.*]] = vector.reduction <add>, %[[T41]] : vector<2xf32> into f32
 // CHECK:    %[[T43:.*]] = vector.insert %[[T42]], %[[T34]] [1, 1] : f32 into vector<2x2xf32>
 //

llvmbot · 2025-03-06T16:11:44Z

@llvm/pr-subscribers-mlir-vector

Author: Artemiy Bulavin (abulavin)

Changes

The vector.contract op on two matrices A and B will be lowered to individual dot products of each row and column of A and B respectively. The existing lowering will extract each column of B for each column of A, which leads to multiple values in the IR representing the same columns of B.

This PR makes changes to the ContractOpToDotLowering to make sure that the columns of B are only ever extracted once, and then that the SSA values representing the extracted columns are then re-used in the IR for later dot products.

I have update the existing vector-contract-to-dot-transforms test.

Full diff: https://github.com/llvm/llvm-project/pull/130130.diff

2 Files Affected:

(modified) mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp (+11-3)
(modified) mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir (+2-4)

diff --git a/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp b/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
index 21261478f0648..88227ecb6c675 100644
--- a/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
@@ -758,12 +758,20 @@ FailureOr<Value> ContractionOpToDotLowering::matchAndRewriteMaskableOp(
   Value res = rewriter.create<arith::ConstantOp>(loc, dstType,
                                                  rewriter.getZeroAttr(dstType));
   bool isInt = isa<IntegerType>(dstType.getElementType());
+  llvm::SmallVector<Value> extractedCols;
+  extractedCols.reserve(dstColumns);
   for (unsigned r = 0; r < dstRows; ++r) {
     Value a = rewriter.create<vector::ExtractOp>(op.getLoc(), lhs, r);
     for (unsigned c = 0; c < dstColumns; ++c) {
-      Value b = rank == 1
-                    ? rhs
-                    : rewriter.create<vector::ExtractOp>(op.getLoc(), rhs, c);
+      if (r == 0) {
+        // We only need to extract the rows of the RHS once
+        // and then re-use them later.
+        Value b = rank == 1
+                      ? rhs
+                      : rewriter.create<vector::ExtractOp>(op.getLoc(), rhs, c);
+        extractedCols.push_back(b);
+      }
+      Value b = extractedCols[c];
       Value m = createMul(op.getLoc(), a, b, isInt, rewriter);
       Value reduced = rewriter.create<vector::ReductionOp>(
           op.getLoc(), vector::CombiningKind::ADD, m);
diff --git a/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir b/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir
index 0ba185bb84760..f77027c3c2d95 100644
--- a/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir
+++ b/mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir
@@ -169,13 +169,11 @@ func.func @extract_contract3(%arg0: vector<3xf32>,
 // CHECK:    %[[T21:.*]] = vector.insert %[[T20]], %[[T11]] [0, 1] : f32 into vector<2x2xf32>
 //
 // CHECK:    %[[T23:.*]] = vector.extract %[[A]][1] : vector<2xf32> from vector<2x2xf32>
-// CHECK:    %[[T24:.*]] = vector.extract %[[Bt]][0] : vector<2xf32> from vector<2x2xf32>
-// CHECK:    %[[T32:.*]] = arith.mulf %[[T23]], %[[T24]] : vector<2xf32>
+// CHECK:    %[[T32:.*]] = arith.mulf %[[T23]], %[[T2]] : vector<2xf32>
 // CHECK:    %[[T33:.*]] = vector.reduction <add>, %[[T32]] : vector<2xf32> into f32
 // CHECK:    %[[T34:.*]] = vector.insert %[[T33]], %[[T21]] [1, 0] : f32 into vector<2x2xf32>
 //
-// CHECK:    %[[T40:.*]] = vector.extract %[[Bt]][1] : vector<2xf32> from vector<2x2xf32>
-// CHECK:    %[[T41:.*]] = arith.mulf %[[T23]], %[[T40]] : vector<2xf32>
+// CHECK:    %[[T41:.*]] = arith.mulf %[[T23]], %[[T12]] : vector<2xf32>
 // CHECK:    %[[T42:.*]] = vector.reduction <add>, %[[T41]] : vector<2xf32> into f32
 // CHECK:    %[[T43:.*]] = vector.insert %[[T42]], %[[T34]] [1, 1] : f32 into vector<2x2xf32>
 //

banach-space

Thanks, mostly just nits.

mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp

mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir

banach-space

I love the updates in the test, thank you!

LGTM % nits

mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp

github-actions · 2025-03-11T11:25:44Z

✅ With the latest revision this PR passed the C/C++ code formatter.

abulavin · 2025-03-12T14:32:28Z

@banach-space Hi again, I believe I've addressed all your comments, so if there's nothing more then could this please be merged for me? I can see that the Check code formatting check failed after my last commit to remove some trailing whitespace that got caught by this check. The most recent check failure actually has nothing to do with code formatting and looks like an infra issue. Can we just ignore this or is there something I can do?

banach-space · 2025-03-12T15:02:46Z

@banach-space Hi again, I believe I've addressed all your comments, so if there's nothing more then could this please be merged for me? I can see that the Check code formatting check failed after my last commit to remove some trailing whitespace that got caught by this check. The most recent check failure actually has nothing to do with code formatting and looks like an infra issue. Can we just ignore this or is there something I can do?

Thanks for the ping! Sorry, I didn't realise those failures were bogus. Yes, I think that it is safe to land this.

EDIT Let me try re-running the formatter job first.

github-actions · 2025-03-12T17:17:11Z

@abulavin Congratulations on having your first Pull Request (PR) merged into the LLVM Project!

Your changes will be combined with recent changes from other authors, then tested by our build bots. If there is a problem with a build, you may receive a report in an email or a comment on this PR.

Please check whether problems have been caused by your change specifically, as the builds can include changes from many authors. It is not uncommon for your change to be included in a build that fails due to someone else's changes, or infrastructure issues.

How to do this, and the rest of the post-merge process, is covered in detail here.

If your change does cause a problem, it may be reverted, or you can revert it yourself. This is a normal part of LLVM development. You can fix your changes and open a new PR to merge them again.

If you don't get any reports, no action is required from you. Your changes are working as expected, well done!

…vm#130130) The `vector.contract` op on two matrices A and B will be lowered to individual dot products of each row and column of A and B respectively. The existing lowering will extract each column of B for each row of A, which leads to multiple values in the IR representing the same columns of B. This PR makes changes to the `ContractOpToDotLowering` to make sure that the columns of B are only ever extracted once, so then the SSA values representing the extracted columns are then re-used in the IR for later dot products. I have updated the existing vector-contract-to-dot-transforms test.

Extract RHS rows once when lowering vector.contract to dot

8e4544f

abulavin requested review from banach-space, dcaballe, hanhanW and nicolasvasilache as code owners March 6, 2025 16:10

llvmbot added mlir:vectorops mlir mlir:vector labels Mar 6, 2025

abulavin changed the title ~~Extract RHS rows once when lowering vector.contract to dot~~ [mlir] Extract RHS rows once when lowering vector.contract to dot Mar 6, 2025

banach-space reviewed Mar 6, 2025

View reviewed changes

mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp Outdated Show resolved Hide resolved

mlir/test/Dialect/Vector/vector-contract-to-dot-transforms.mlir Outdated Show resolved Hide resolved

abulavin added 2 commits March 6, 2025 20:07

fixup: Refactor variable names in vector contract to dot transform test

ce9c720

fixup: refer to columns instead of rows in reuse comment

4018f23

banach-space approved these changes Mar 10, 2025

View reviewed changes

mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp Outdated Show resolved Hide resolved

mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp Outdated Show resolved Hide resolved

fixup: Rephrase comment and rename column/row variables

4c74d73

fixup: remove trailing whitespace

86853bf

banach-space merged commit fc127ff into llvm:main Mar 12, 2025
11 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[mlir] Extract RHS rows once when lowering vector.contract to dot #130130

[mlir] Extract RHS rows once when lowering vector.contract to dot #130130

abulavin commented Mar 6, 2025 •

edited

Loading

github-actions bot commented Mar 6, 2025

llvmbot commented Mar 6, 2025

llvmbot commented Mar 6, 2025

banach-space left a comment

banach-space left a comment

github-actions bot commented Mar 11, 2025 •

edited

Loading

abulavin commented Mar 12, 2025

banach-space commented Mar 12, 2025 •

edited

Loading

github-actions bot commented Mar 12, 2025

[mlir] Extract RHS rows once when lowering vector.contract to dot #130130

[mlir] Extract RHS rows once when lowering vector.contract to dot #130130

Conversation

abulavin commented Mar 6, 2025 • edited Loading

github-actions bot commented Mar 6, 2025

llvmbot commented Mar 6, 2025

llvmbot commented Mar 6, 2025

banach-space left a comment

Choose a reason for hiding this comment

banach-space left a comment

Choose a reason for hiding this comment

github-actions bot commented Mar 11, 2025 • edited Loading

abulavin commented Mar 12, 2025

banach-space commented Mar 12, 2025 • edited Loading

github-actions bot commented Mar 12, 2025

abulavin commented Mar 6, 2025 •

edited

Loading

github-actions bot commented Mar 11, 2025 •

edited

Loading

banach-space commented Mar 12, 2025 •

edited

Loading