Rewrite use of tensor::PadOp result in TransferWriteOp. More...
Inheritance diagram for PadOpVectorizationWithTransferWritePattern:
Public Member Functions | |
| LogicalResult | rewriteUser (PatternRewriter &rewriter, tensor::PadOp padOp, vector::TransferWriteOp xferOp) const override |
| bool | hasSameTensorSize (Value beforePadding, tensor::ExtractSliceOp afterTrimming) const |
Check if beforePadding and afterTrimming have the same tensor size, i.e., same dimensions. More... | |
| Public Member Functions inherited from VectorizePadOpUserPattern< vector::TransferWriteOp > | |
| LogicalResult | matchAndRewrite (tensor::PadOp padOp, PatternRewriter &rewriter) const final |
| Public Member Functions inherited from mlir::OpRewritePattern< tensor::PadOp > | |
| OpRewritePattern (MLIRContext *context, PatternBenefit benefit=1, ArrayRef< StringRef > generatedNames={}) | |
| Patterns must specify the root operation name they match against, and can also specify the benefit of the pattern matching and a list of generated ops. More... | |
| Public Member Functions inherited from mlir::detail::OpOrInterfaceRewritePatternBase< SourceOp > | |
| void | rewrite (Operation *op, PatternRewriter &rewriter) const final |
| Wrappers around the RewritePattern methods that pass the derived op type. More... | |
| LogicalResult | match (Operation *op) const final |
| Attempt to match against code rooted at the specified operation, which is the same operation code as getRootKind(). More... | |
| LogicalResult | matchAndRewrite (Operation *op, PatternRewriter &rewriter) const final |
| Attempt to match against code rooted at the specified operation, which is the same operation code as getRootKind(). More... | |
| virtual void | rewrite (SourceOp op, PatternRewriter &rewriter) const |
| Rewrite and Match methods that operate on the SourceOp type. More... | |
| virtual LogicalResult | match (SourceOp op) const |
| virtual LogicalResult | matchAndRewrite (SourceOp op, PatternRewriter &rewriter) const |
| Public Member Functions inherited from mlir::RewritePattern | |
| virtual | ~RewritePattern ()=default |
| Public Member Functions inherited from mlir::Pattern | |
| ArrayRef< OperationName > | getGeneratedOps () const |
| Return a list of operations that may be generated when rewriting an operation instance with this pattern. More... | |
| std::optional< OperationName > | getRootKind () const |
| Return the root node that this pattern matches. More... | |
| std::optional< TypeID > | getRootInterfaceID () const |
| Return the interface ID used to match the root operation of this pattern. More... | |
| std::optional< TypeID > | getRootTraitID () const |
| Return the trait ID used to match the root operation of this pattern. More... | |
| PatternBenefit | getBenefit () const |
| Return the benefit (the inverse of "cost") of matching this pattern. More... | |
| bool | hasBoundedRewriteRecursion () const |
| Returns true if this pattern is known to result in recursive application, i.e. More... | |
| MLIRContext * | getContext () const |
| Return the MLIRContext used to create this pattern. More... | |
| StringRef | getDebugName () const |
| Return a readable name for this pattern. More... | |
| void | setDebugName (StringRef name) |
| Set the human readable debug name used for this pattern. More... | |
| ArrayRef< StringRef > | getDebugLabels () const |
| Return the set of debug labels attached to this pattern. More... | |
| void | addDebugLabels (ArrayRef< StringRef > labels) |
| Add the provided debug labels to this pattern. More... | |
| void | addDebugLabels (StringRef label) |
Additional Inherited Members | |
| Static Public Member Functions inherited from mlir::RewritePattern | |
| template<typename T , typename... Args> | |
| static std::unique_ptr< T > | create (Args &&...args) |
This method provides a convenient interface for creating and initializing derived rewrite patterns of the given type T. More... | |
| Protected Member Functions inherited from VectorizePadOpUserPattern< vector::TransferWriteOp > | |
| virtual LogicalResult | rewriteUser (PatternRewriter &rewriter, tensor::PadOp padOp, vector::TransferWriteOp op) const=0 |
| Protected Member Functions inherited from mlir::RewritePattern | |
| Pattern (StringRef rootName, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
Inherit the base constructors from Pattern. More... | |
| Pattern (MatchAnyOpTypeTag tag, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
Inherit the base constructors from Pattern. More... | |
| Pattern (MatchInterfaceOpTypeTag tag, TypeID interfaceID, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
Inherit the base constructors from Pattern. More... | |
| Pattern (MatchTraitOpTypeTag tag, TypeID traitID, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
Inherit the base constructors from Pattern. More... | |
| Protected Member Functions inherited from mlir::Pattern | |
| Pattern (StringRef rootName, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
| Construct a pattern with a certain benefit that matches the operation with the given root name. More... | |
| Pattern (MatchAnyOpTypeTag tag, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
| Construct a pattern that may match any operation type. More... | |
| Pattern (MatchInterfaceOpTypeTag tag, TypeID interfaceID, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
Construct a pattern that may match any operation that implements the interface defined by the provided interfaceID. More... | |
| Pattern (MatchTraitOpTypeTag tag, TypeID traitID, PatternBenefit benefit, MLIRContext *context, ArrayRef< StringRef > generatedNames={}) | |
Construct a pattern that may match any operation that implements the trait defined by the provided traitID. More... | |
| void | setHasBoundedRewriteRecursion (bool hasBoundedRecursionArg=true) |
| Set the flag detailing if this pattern has bounded rewrite recursion or not. More... | |
Detailed Description
Rewrite use of tensor::PadOp result in TransferWriteOp.
This pattern rewrites TransferWriteOps that write to a padded tensor value, where the same amount of padding is immediately removed again after the write. In such cases, the TransferWriteOp can write to the non-padded tensor value and apply out-of-bounds masking. E.g.:
%0 = tensor.extract_slice ...[...] [%s0, %s1] [1, 1]
: tensor<...> to tensor<?x?xf32>
%1 = tensor.pad %0 ... : tensor<?x?xf32> to tensor<17x5xf32>
%2 = vector.transfer_write %vec, %1[...]
: vector<17x5xf32>, tensor<17x5xf32>
%r = tensor.extract_slice %2[0, 0] [%s0, %s1] [1, 1]
: tensor<17x5xf32> to tensor<?x?xf32>
is rewritten to:
%0 = tensor.extract_slice ...[...] [%s0, %s1] [1, 1]
: tensor<...> to tensor<?x?xf32>
%r = vector.transfer_write %vec, %0[...] : vector<17x5xf32>,
tensor<?x?xf32>
Note: It is important that the ExtractSliceOp r resizes the result of the TransferWriteOp to the same size as the input of the TensorPadOp (or an even smaller size). Otherwise, r's new (dynamic) dimensions would differ from r's old dimensions.
This rewrite is possible if:
- Low padding is static 0.
xferOphas exactly one use, which is an ExtractSliceOp. This ExtractSliceOp trims the same amount of padding that was added beforehand.- Single, scalar padding value.
Definition at line 2391 of file Vectorization.cpp.
Member Function Documentation
◆ hasSameTensorSize()
|
inline |
Check if beforePadding and afterTrimming have the same tensor size, i.e., same dimensions.
Dimensions may be static, dynamic or mix of both. In case of dynamic dimensions, this function tries to infer the (static) tensor size by looking at the defining op and utilizing op-specific knowledge.
This is a conservative analysis. In case equal tensor sizes cannot be proven statically, this analysis returns false even though the tensor sizes may turn out to be equal at runtime.
Definition at line 2445 of file Vectorization.cpp.
References mlir::Value::getDefiningOp(), mlir::Value::getType(), and mlir::isEqualConstantIntOrValue().
◆ rewriteUser()
|
inlineoverride |
Definition at line 2396 of file Vectorization.cpp.
References mlir::Builder::getBoolArrayAttr(), mlir::RewriterBase::replaceOp(), mlir::RewriterBase::replaceOpWithNewOp(), and mlir::OpBuilder::setInsertionPoint().
The documentation for this struct was generated from the following file:
- lib/Dialect/Linalg/Transforms/Vectorization.cpp
