[mlir][tensor] Add shape inference support for tensor.concat op.
#140168
Conversation
|
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 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. |
|
@llvm/pr-subscribers-mlir @llvm/pr-subscribers-mlir-tensor Author: Aaron St George (AaronStGeorge) Changesdescription
example // Second operand dim 1 has dynamic shape constrained by dim 1 of first
// operand.
%2 = tensor.concat dim(0) %0, %1: (tensor<?x12xi32>, tensor<?x?xi32>) ->tensor<?x12xi32>becomes: %cast = tensor.cast %1 : tensor<?x?xi32> to tensor<?x12xi32>
%2 = tensor.concat dim(0) %0, %cast : (tensor<?x12xi32>, tensor<?x12xi32>) -> tensor<?x12xi32>Full diff: https://github.com/llvm/llvm-project/pull/140168.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
index 815806f06b472..633e502db5a3a 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
@@ -330,8 +330,9 @@ bool mlir::tensor::canFoldIntoConsumerOp(CastOp castOp) {
/// Determines whether the tensor::CastOp casts to a more static version of the
/// source tensor. This is useful to fold into a producing op and implement
-/// canonicaliation patterns with the `tensor.cast` op as the root, but producer
-/// being from different dialects. Returns true when all conditions are met:
+/// canonicalization patterns with the `tensor.cast` op as the root, but
+/// producer being from different dialects. Returns true when all conditions are
+/// met:
/// 1. source and result and ranked tensors with same element type and rank.
/// 2. the result type has more static information than the source.
///
@@ -773,11 +774,118 @@ struct SingleInputConcatOp : public OpRewritePattern<ConcatOp> {
return success();
}
};
+
+/// Propagate static shapes into the operands of a `tensor.concat`.
+///
+/// `tensor.concat` requires every operand to match on all dimensions except the
+/// concatenation dimension. If one operand is already static in those
+/// dimensions, the other operands may safely be refined to that same static
+/// shape.
+///
+/// Example:
+///
+/// ```mlir
+/// // Second operand dim 1 has dynamic shape constrained by dim 1 of first
+/// // operand.
+/// %2 = tensor.concat dim(0) %0, %1: (tensor<?x12xi32>, tensor<?x?xi32>) ->
+/// tensor<?x12xi32>
+/// ```
+/// ->
+/// ```mlir
+/// %cast = tensor.cast %1 : tensor<?x?xi32> to tensor<?x12xi32>
+/// %2 = tensor.concat dim(0) %0, %cast :
+/// (tensor<?x12xi32>, tensor<?x12xi32>) -> tensor<?x12xi32>
+/// ```
+struct InferConcatOperandTypes : public OpRewritePattern<ConcatOp> {
+ using OpRewritePattern<ConcatOp>::OpRewritePattern;
+
+ LogicalResult matchAndRewrite(ConcatOp concatOp,
+ PatternRewriter &rewriter) const override {
+ auto operandTensorTypes =
+ llvm::map_range(concatOp->getOperandTypes(), [](Type type) {
+ return llvm::cast<RankedTensorType>(type);
+ });
+
+ int64_t dim = concatOp.getDim();
+ ArrayRef<int64_t> inferredResultShape =
+ concatOp.inferResultType(dim, concatOp->getOperandTypes()).getShape();
+
+ // Find operands for which a more static shape can be inferred.
+ SmallVector<std::tuple<size_t, RankedTensorType>> refinedTypes;
+ for (auto [operandIdx, operandType] : llvm::enumerate(operandTensorTypes)) {
+ // Compute inferred type for operand.
+ SmallVector<int64_t> inferredOperandShape(inferredResultShape);
+ inferredOperandShape[dim] = operandType.getDimSize(dim);
+ auto inferredOperandType = RankedTensorType::get(
+ inferredOperandShape, operandType.getElementType());
+
+ // Check if inferred type is more static.
+ if (!preservesStaticInformation(inferredOperandType, operandType)) {
+ refinedTypes.push_back({operandIdx, inferredOperandType});
+ }
+ }
+
+ if (refinedTypes.empty()) {
+ return failure();
+ }
+
+ // Use refined types for operands, insert casts for original type.
+ SmallVector<Value> newOperands = concatOp.getOperands();
+ for (auto [operandIdx, refinedType] : refinedTypes) {
+ newOperands[operandIdx] = rewriter.create<CastOp>(
+ concatOp->getLoc(), refinedType, concatOp.getOperand(operandIdx));
+ }
+ rewriter.replaceOpWithNewOp<ConcatOp>(concatOp, concatOp.getResultType(),
+ dim, newOperands);
+
+ return success();
+ }
+};
+
+// Ensure `tensor.concat`'s result type is at least as static as can be inferred
+// from its operand types.
+///
+/// Example:
+/// ```mlir
+/// %2 = tensor.concat dim(0) %0, %1: (tensor<?x12xi32>, tensor<?x12xi32>) ->
+/// tensor<?x?xi32>
+/// ```
+/// ->
+/// ```mlir
+/// %2 = tensor.concat dim(0) %0, %cast : (tensor<?x12xi32>, tensor<?x12xi32>)
+/// -> tensor<?x12xi32> %cast = tensor.cast %2 : tensor<?x12xi32> to
+/// tensor<?x?xi32>
+/// ```
+struct InferConcatResultType : public OpRewritePattern<ConcatOp> {
+ using OpRewritePattern<ConcatOp>::OpRewritePattern;
+
+ LogicalResult matchAndRewrite(ConcatOp concatOp,
+ PatternRewriter &rewriter) const override {
+ int64_t dim = concatOp.getDim();
+ RankedTensorType inferredResultType =
+ concatOp.inferResultType(dim, concatOp->getOperandTypes());
+
+ // The result type should be at least as static as inferred result type.
+ if (preservesStaticInformation(inferredResultType,
+ concatOp.getResultType())) {
+ return failure();
+ }
+
+ auto newConcatOp = rewriter.create<ConcatOp>(
+ concatOp->getLoc(), inferredResultType, dim, concatOp->getOperands());
+ rewriter.replaceOpWithNewOp<CastOp>(concatOp, concatOp.getResultType(),
+ newConcatOp);
+
+ return llvm::success();
+ }
+};
} // namespace
void ConcatOp::getCanonicalizationPatterns(RewritePatternSet &results,
MLIRContext *context) {
- results.add<SingleInputConcatOp>(context);
+ results
+ .add<SingleInputConcatOp, InferConcatOperandTypes, InferConcatResultType>(
+ context);
}
//===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/Tensor/canonicalize.mlir b/mlir/test/Dialect/Tensor/canonicalize.mlir
index 85bf6fba52aa4..cdcd7f305d2d9 100644
--- a/mlir/test/Dialect/Tensor/canonicalize.mlir
+++ b/mlir/test/Dialect/Tensor/canonicalize.mlir
@@ -136,6 +136,32 @@ func.func @fold_concat(%arg0: tensor<1x2x?xi32>) -> (tensor<1x2x3xi32>, tensor<1
// -----
+// CHECK-LABEL: infer_concat_operand_types
+// CHECK-SAME: %[[ARG0:.+]]: tensor<?x12xi32>
+// CHECK-SAME: %[[ARG1:.+]]: tensor<?x?xi32>
+func.func @infer_concat_operand_types(%arg0: tensor<?x12xi32>, %arg1: tensor<?x?xi32>) -> (tensor<?x12xi32>) {
+ // CHECK-NEXT: %[[CAST:.+]] = tensor.cast %[[ARG1]] : tensor<?x?xi32> to tensor<?x12xi32>
+ %0 = tensor.concat dim(0) %arg0, %arg1: (tensor<?x12xi32>, tensor<?x?xi32>) -> tensor<?x12xi32>
+ // CHECK-NEXT: %[[CONCAT:.+]] = tensor.concat dim(0) %[[ARG0]], %[[CAST]] : (tensor<?x12xi32>, tensor<?x12xi32>) -> tensor<?x12xi32>
+ return %0 : tensor<?x12xi32>
+ // CHECK-NEXT: return %[[CONCAT]] : tensor<?x12xi32>
+}
+
+// -----
+
+// CHECK-LABEL: infer_concat_return_type
+// CHECK-SAME: %[[ARG0:.+]]: tensor<5x12xi32>
+// CHECK-SAME: %[[ARG1:.+]]: tensor<?x12xi32>
+func.func @infer_concat_return_type(%arg0: tensor<5x12xi32>, %arg1: tensor<?x12xi32>) -> (tensor<?x?xi32>) {
+ %0 = tensor.concat dim(0) %arg0, %arg1: (tensor<5x12xi32>, tensor<?x12xi32>) -> tensor<?x?xi32>
+ // CHECK-NEXT: %[[CONCAT:.+]] = tensor.concat dim(0) %[[ARG0]], %[[ARG1]] : (tensor<5x12xi32>, tensor<?x12xi32>) -> tensor<?x12xi32>
+ // CHECK-NEXT: %[[CAST:.+]] = tensor.cast %[[CONCAT]] : tensor<?x12xi32> to tensor<?x?xi32>
+ return %0 : tensor<?x?xi32>
+ // CHECK-NEXT: return %[[CAST]] : tensor<?x?xi32>
+}
+
+// -----
+
// CHECK-LABEL: func @fold_extract
func.func @fold_extract(%arg0 : index) -> (f32, f16, f16, i32, complex<f32>) {
%const_0 = arith.constant 0 : index
|
27233af to
db8aaff
Compare
tensor.concat cast propagationtensor.concat op.
| rewriter.create<CastOp>(concatOp->getLoc(), inferredOperandType, | ||
| concatOp.getOperand(operandIdx)); | ||
| rewriter.modifyOpInPlace( | ||
| concatOp, [=, operandIdx = (size_t)operandIdx] { |
There was a problem hiding this comment.
operandIdx = (size_t)operandIdx is to avoid "structured bindings are a C++20 extension" error.
There was a problem hiding this comment.
Yeah its a bit annoying to have to do this. Is there a reason for the cast?
There was a problem hiding this comment.
Nope, I'll remove that.
Adds canonicalization patterns which propagate inferred static shapes to `tensor.concat` operands and result types. Static is propagated to other canonicalization patterns through casts.
Co-authored-by: Ian Wood <ianwood2024@u.northwestern.edu>
6b32885 to
b0d1405
Compare
| rewriter.create<CastOp>(concatOp->getLoc(), inferredOperandType, | ||
| concatOp.getOperand(operandIdx)); | ||
| rewriter.modifyOpInPlace( | ||
| concatOp, [=, operandIdx = (size_t)operandIdx] { |
There was a problem hiding this comment.
Yeah its a bit annoying to have to do this. Is there a reason for the cast?
|
@AaronStGeorge 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! |
description
tensor.concatrequires operands and the result to match on all dimensions except the concatenation dimension. If one operand is already static in those dimensions, the other operands and result type may safely be refined to that same static shape. This PR adds canonicalization patterns to refinetensor.concattypes and propagate static shapes to other canonicalization patterns through casts.example
becomes: