[MLIR][Vector] Extend elementwise pattern to support unrolling from higher rank to lower rank #162515
+84
−20
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
nbpatel
requested review from
banach-space,
dcaballe,
hanhanW and
nicolasvasilache
as code owners
|
@llvm/pr-subscribers-mlir-vector @llvm/pr-subscribers-mlir Author: Nishant Patel (nbpatel) ChangesFull diff: https://github.com/llvm/llvm-project/pull/162515.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp
index 14639c5f1cdd3..62d65e28e8c2e 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp
@@ -468,23 +468,30 @@ struct UnrollElementwisePattern : public RewritePattern {
auto dstVecType = cast<VectorType>(op->getResult(0).getType());
SmallVector<int64_t> originalSize =
*cast<VectorUnrollOpInterface>(op).getShapeForUnroll();
- // Bail-out if rank(source) != rank(target). The main limitation here is the
- // fact that `ExtractStridedSlice` requires the rank for the input and
- // output to match. If needed, we can relax this later.
- if (originalSize.size() != targetShape->size())
- return rewriter.notifyMatchFailure(
- op, "expected input vector rank to match target shape rank");
+
Location loc = op->getLoc();
+
+ // Handle rank mismatch by adding leading unit dimensions to targetShape
+ SmallVector<int64_t> adjustedTargetShape = *targetShape;
+ SmallVector<int64_t> adjustedOffsets;
+ if (originalSize.size() > targetShape->size()) {
+ // Add leading unit dimensions to targetShape
+ int64_t rankDiff = originalSize.size() - targetShape->size();
+ adjustedTargetShape.insert(adjustedTargetShape.begin(), rankDiff, 1);
+ }
+
// Prepare the result vector.
Value result = arith::ConstantOp::create(rewriter, loc, dstVecType,
rewriter.getZeroAttr(dstVecType));
- SmallVector<int64_t> strides(targetShape->size(), 1);
- VectorType newVecType =
+ SmallVector<int64_t> strides(adjustedTargetShape.size(), 1);
+ VectorType extractVecType =
+ VectorType::get(adjustedTargetShape, dstVecType.getElementType());
+ VectorType computeVecType =
VectorType::get(*targetShape, dstVecType.getElementType());
// Create the unrolled computation.
for (SmallVector<int64_t> offsets :
- StaticTileOffsetRange(originalSize, *targetShape)) {
+ StaticTileOffsetRange(originalSize, adjustedTargetShape)) {
SmallVector<Value> extractOperands;
for (OpOperand &operand : op->getOpOperands()) {
auto vecType = dyn_cast<VectorType>(operand.get().getType());
@@ -492,14 +499,31 @@ struct UnrollElementwisePattern : public RewritePattern {
extractOperands.push_back(operand.get());
continue;
}
- extractOperands.push_back(
- rewriter.createOrFold<vector::ExtractStridedSliceOp>(
- loc, operand.get(), offsets, *targetShape, strides));
+ Value extracted = rewriter.createOrFold<vector::ExtractStridedSliceOp>(
+ loc, operand.get(), offsets, adjustedTargetShape, strides);
+
+ // Reshape to remove leading unit dims if needed
+ if (adjustedTargetShape.size() > targetShape->size()) {
+ extracted = rewriter.createOrFold<vector::ShapeCastOp>(
+ loc, VectorType::get(*targetShape, vecType.getElementType()),
+ extracted);
+ }
+ extractOperands.push_back(extracted);
}
+
Operation *newOp = cloneOpWithOperandsAndTypes(
- rewriter, loc, op, extractOperands, newVecType);
+ rewriter, loc, op, extractOperands, computeVecType);
+
+ Value computeResult = newOp->getResult(0);
+
+ // Reshape back to higher rank if needed for insertion
+ if (adjustedTargetShape.size() > targetShape->size()) {
+ computeResult = rewriter.createOrFold<vector::ShapeCastOp>(
+ loc, extractVecType, computeResult);
+ }
+
result = rewriter.createOrFold<vector::InsertStridedSliceOp>(
- loc, newOp->getResult(0), result, offsets, strides);
+ loc, computeResult, result, offsets, strides);
}
rewriter.replaceOp(op, result);
return success();
diff --git a/mlir/test/Dialect/Vector/vector-unroll-options.mlir b/mlir/test/Dialect/Vector/vector-unroll-options.mlir
index 35db14e0f7f1d..a26e4b0baa05b 100644
--- a/mlir/test/Dialect/Vector/vector-unroll-options.mlir
+++ b/mlir/test/Dialect/Vector/vector-unroll-options.mlir
@@ -188,15 +188,40 @@ func.func @vector_fma(%a: vector<4x4xf32>, %b: vector<4x4xf32>, %c: vector<4x4xf
// CHECK-LABEL: func @vector_fma
// CHECK-COUNT-4: vector.fma %{{.+}}, %{{.+}}, %{{.+}} : vector<2x2xf32>
-// TODO: We should be able to unroll this like the example above - this will require extending UnrollElementwisePattern.
-func.func @negative_vector_fma_3d(%a: vector<3x2x2xf32>) -> vector<3x2x2xf32>{
+func.func @vector_fma_3d(%a: vector<3x2x2xf32>) -> vector<3x2x2xf32>{
%0 = vector.fma %a, %a, %a : vector<3x2x2xf32>
return %0 : vector<3x2x2xf32>
}
-// CHECK-LABEL: func @negative_vector_fma_3d
-// CHECK-NOT: vector.extract_strided_slice
-// CHECK: %[[R0:.*]] = vector.fma %{{.+}} : vector<3x2x2xf32>
-// CHECK: return
+// CHECK-LABEL: func @vector_fma_3d
+// CHECK: %[[CST:.*]] = arith.constant dense<0.000000e+00> : vector<3x2x2xf32>
+// CHECK: %[[E0:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [0, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S0:.*]] = vector.shape_cast %[[E0]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E1:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [0, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S1:.*]] = vector.shape_cast %[[E1]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E2:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [0, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S2:.*]] = vector.shape_cast %[[E2]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[FMA0:.*]] = vector.fma %[[S0]], %[[S1]], %[[S2]] : vector<2x2xf32>
+// CHECK: %[[SC0:.*]] = vector.shape_cast %[[FMA0]] : vector<2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[I0:.*]] = vector.insert_strided_slice %[[SC0]], %[[CST]] {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x2x2xf32> into vector<3x2x2xf32>
+// CHECK: %[[E3:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [1, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S3:.*]] = vector.shape_cast %[[E3]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E4:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [1, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S4:.*]] = vector.shape_cast %[[E4]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E5:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [1, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S5:.*]] = vector.shape_cast %[[E5]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[FMA1:.*]] = vector.fma %[[S3]], %[[S4]], %[[S5]] : vector<2x2xf32>
+// CHECK: %[[SC1:.*]] = vector.shape_cast %[[FMA1]] : vector<2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[I1:.*]] = vector.insert_strided_slice %[[SC1]], %[[I0]] {offsets = [1, 0, 0], strides = [1, 1, 1]} : vector<1x2x2xf32> into vector<3x2x2xf32>
+// CHECK: %[[E6:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [2, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S6:.*]] = vector.shape_cast %[[E6]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E7:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [2, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S7:.*]] = vector.shape_cast %[[E7]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E8:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [2, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<3x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S8:.*]] = vector.shape_cast %[[E8]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[FMA2:.*]] = vector.fma %[[S6]], %[[S7]], %[[S8]] : vector<2x2xf32>
+// CHECK: %[[SC2:.*]] = vector.shape_cast %[[FMA2]] : vector<2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[I2:.*]] = vector.insert_strided_slice %[[SC2]], %[[I1]] {offsets = [2, 0, 0], strides = [1, 1, 1]} : vector<1x2x2xf32> into vector<3x2x2xf32>
+// CHECK: return %[[I2]] : vector<3x2x2xf32>
func.func @vector_multi_reduction(%v : vector<4x6xf32>, %acc: vector<4xf32>) -> vector<4xf32> {
%0 = vector.multi_reduction #vector.kind<add>, %v, %acc [1] : vector<4x6xf32> to vector<4xf32>
@@ -440,3 +465,26 @@ func.func @vector_step() -> vector<32xindex> {
// CHECK: %[[ADD3:.*]] = arith.addi %[[STEP]], %[[CST]] : vector<8xindex>
// CHECK: %[[INS3:.*]] = vector.insert_strided_slice %[[ADD3]], %[[INS2]] {offsets = [24], strides = [1]} : vector<8xindex> into vector<32xindex>
// CHECK: return %[[INS3]] : vector<32xindex>
+
+
+func.func @elementwise(%v1: vector<2x2x2xf32>, %v2: vector<2x2x2xf32>) -> vector<2x2x2xf32> {
+ %0 = arith.addf %v1, %v2 : vector<2x2x2xf32>
+ return %0 : vector<2x2x2xf32>
+}
+// CHECK-LABEL: func @elementwise
+// CHECK: %[[CST:.*]] = arith.constant dense<0.000000e+00> : vector<2x2x2xf32>
+// CHECK: %[[E0:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [0, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S0:.*]] = vector.shape_cast %[[E0]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E1:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [0, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S1:.*]] = vector.shape_cast %[[E1]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[ADD0:.*]] = arith.addf %[[S0]], %[[S1]] : vector<2x2xf32>
+// CHECK: %[[SC0:.*]] = vector.shape_cast %[[ADD0]] : vector<2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[I0:.*]] = vector.insert_strided_slice %[[SC0]], %[[CST]] {offsets = [0, 0, 0], strides = [1, 1, 1]} : vector<1x2x2xf32> into vector<2x2x2xf32>
+// CHECK: %[[E2:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [1, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S2:.*]] = vector.shape_cast %[[E2]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[E3:.*]] = vector.extract_strided_slice %{{.*}} {offsets = [1, 0, 0], sizes = [1, 2, 2], strides = [1, 1, 1]} : vector<2x2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[S3:.*]] = vector.shape_cast %[[E3]] : vector<1x2x2xf32> to vector<2x2xf32>
+// CHECK: %[[ADD1:.*]] = arith.addf %[[S2]], %[[S3]] : vector<2x2xf32>
+// CHECK: %[[SC1:.*]] = vector.shape_cast %[[ADD1]] : vector<2x2xf32> to vector<1x2x2xf32>
+// CHECK: %[[I1:.*]] = vector.insert_strided_slice %[[SC1]], %[[I0]] {offsets = [1, 0, 0], strides = [1, 1, 1]} : vector<1x2x2xf32> into vector<2x2x2xf32>
+// CHECK: return %[[I1]] : vector<2x2x2xf32>
|
|
@newling please take a look |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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 enhances the elementwise unrolling pattern to support higher rank to lower rank unroll. The approach is to add leading unit dims to lower rank targetShape to match the rank of original vector (because ExtractStridedSlice requires same rank to extractSlices), extract slice, reshape to targetShape's rank and perform the operation.