Namespaces
	detail

Classes
class	LinalgOpToLibraryCallRewrite

struct	ContractionDimensions
	Positions of a Linalg op loops that correspond to different kinds of a contraction dimension. More...

struct	ConvolutionDimensions
	Positions of a Linalg op loops that correspond to different kinds of a convolution dimension. More...

struct	BufferizeToAllocationOptions

struct	LinalgTilingOptions

struct	LinalgTilingAndFusionOptions

struct	LinalgPaddingOptions

struct	LinalgPromotionOptions

struct	SplitReductionOptions
	Split Reduction options. More...

struct	ControlDropUnitDims
	Transformation to drop unit-extent dimensions from `linalg.generic` operations. More...

struct	DropUnitDimsResult

struct	ElementwiseOpFusionResult
	Fuse two `linalg.generic` operations that have a producer-consumer relationship captured through `fusedOperand`. More...

struct	TiledLinalgOp
	Perform standalone tiling of a single LinalgOp by `tileSizes`. More...

struct	PromotionInfo
	Create a new buffer using the `allocationFn` provided. More...

struct	MultiSizeSpecification
	A description of a multi-size tiling comprising tile sizes and numbers of tiles, expressed as Values which may or may not be constant. More...

struct	StaticMultiSizeSpecification

struct	ContinuousTileSizeSpecification

struct	StaticContinuousTileSizeSpecification

struct	ForallReductionTilingResult
	Transformation information returned after reduction tiling. More...

struct	SplitReductionResult
	Apply transformation to split the single linalg op reduction into a parallel and reduction dimension. More...

struct	CollapseResult

struct	LowerPackResult

struct	LowerUnPackOpResult

struct	PackResult
	Struct to hold the result of a `pack` call. More...

struct	PackTransposeResult
	Struct to hold the result of a `packTranspose` call. More...

struct	BlockPackMatmulOptions

struct	DownscaleSizeOneWindowed2DConvolution
	Rewrites 2-D convolution ops with size-1 window dimensions into 1-D convolution ops. More...

struct	DownscaleDepthwiseConv2DNhwcHwcOp
	Rewrites 2-D depthwise convolution ops with size-1 (w, kw) or (h, kh) dimensions into 1-D depthwise convolution ops. More...

struct	DownscaleConv2DOp

struct	LinalgGeneralizationPattern
	Linalg generalization pattern. More...

struct	LinalgSpecializationPattern

struct	CopyVectorizationPattern
	Vectorization pattern for memref::CopyOp. More...

struct	DecomposePadOpPattern
	Rewrite a tensor::PadOp into a sequence of EmptyOp, FillOp and InsertSliceOp. More...

struct	DecomposeOuterUnitDimsPackOpPattern
	Rewrites a tensor::PackOp into a sequence of: More...

struct	DecomposeOuterUnitDimsUnPackOpPattern
	Rewrites a tensor::UnPackOp into a sequence of rank-reduced. More...

struct	LinalgCopyVTRForwardingPattern
	Match and rewrite for the pattern: More...

struct	LinalgCopyVTWForwardingPattern
	Match and rewrite for the pattern: More...

struct	ExtractSliceOfPadTensorSwapPattern
	Rewrite extract_slice(tensor.pad(x)) into tensor.pad(extract_slice(x)). More...

struct	SliceParameters
	A struct containg offsets-sizes-strides arguments of the tiled shape. More...

struct	FusionInfo
	A struct containing the Linalg producer before and after fusion. More...

struct	ProcInfo
	Callback function type used to get processor ID, and number of processors used for distribution for all parallel loops generated. More...

struct	LinalgLoopDistributionOptions
	Options that allow distribution of loops generated in Linalg transforms to processors while generating the loops. More...

struct	RegionMatcher
	A struct containing common matchers over linalg op's region. More...

struct	GenerateLoopNest
	Utility class used to generate nested loops with ranges described by `loopRanges` and loop type described by the `iteratorTypes`. More...

Typedefs
using	TileSizeComputationFunction = std::function< SmallVector< Value, 4 >(OpBuilder &, Operation *)>

using	AllocBufferCallbackFn = std::function< std::optional< Value >(OpBuilder &b, memref::SubViewOp subView, ArrayRef< Value > boundingSubViewSize, DataLayout &layout)>
	Callback function type used to perform the allocation for the promoted `subView`. More...

using	DeallocBufferCallbackFn = std::function< LogicalResult(OpBuilder &b, Value buffer)>
	Callback function type used to deallocate the buffers used to hold the promoted subview. More...

using	CopyCallbackFn = std::function< LogicalResult(OpBuilder &b, Value src, Value dst)>
	Callback function type used to insert copy from original subview to subview of the promoted region for the read operands/subview of promoted region to original subview for the results. More...

using	ControlSplitReductionFn = std::function< SplitReductionOptions(LinalgOp op)>
	Function signature to control reduction splitting. More...

using	LinalgLoops = SmallVector< Operation *, 4 >

using	LoopIndexToRangeIndexMap = DenseMap< int, int >
	Creates a number of ranges equal to the number of non-zero in `tileSizes`. More...

using	ControlBlockPackMatmulFn = std::function< std::optional< BlockPackMatmulOptions >(linalg::LinalgOp)>
	Function type which is used to control matmul packing. More...

using	OptimizeCopyFn = std::function< LogicalResult(RewriterBase &, tensor::PadOp, Value)>

using	ControlFusionFn = std::function< bool(OpOperand *fusedOperand)>
	Function type which is used to control when to stop fusion. More...

using	ControlPropagationFn = std::function< bool(OpOperand *opOperand)>
	Function type which is used to control propagation of tensor.pack/unpack ops. More...

using	GetCollapsableDimensionsFn = std::function< SmallVector< ReassociationIndices >(linalg::LinalgOp)>
	Function type to control generic op dimension collapsing. More...

using	ProcInfoCallBackFn = std::function< SmallVector< ProcInfo >(OpBuilder &b, Location loc, ArrayRef< Range > parallelLoopRanges)>

using	MeshAxis = mesh::MeshAxis

using	ReductionKind = mesh::ReductionKind

using	MeshSharding = mesh::MeshSharding

using	ShardingArray = mesh::ShardingArray

using	MeshOp = mesh::MeshOp

Enumerations
enum class	LinalgTilingLoopType { Loops = 0 , AffineLoops = 1 , ParallelLoops = 2 }
	The type of loops to be generated during tiling. More...

enum class	DistributionMethod { Cyclic = 0 , CyclicNumProcsGeNumIters = 1 , CyclicNumProcsEqNumIters = 2 , None = 3 }
	Scheme used to distribute loops to processors. More...

Functions
void	populateLinalgToStandardConversionPatterns (RewritePatternSet &patterns)
	Populate the given list with patterns that convert from Linalg to Standard. More...

std::string	generateLibraryCallName (Operation *op)
	Returns the name mangled library call name to disambiguate between different overloads at the C level. More...

SmallVector< AffineExpr, 4 >	makeAffineDimExprs (unsigned num, unsigned &startIdx, MLIRContext *context)
	Returns `num` AffineDimExpr dimensions at positions [startIdx, startIdx + num) and increments `startIdx` to `startIdx + num`. More...

AffineMap	extractOrIdentityMap (std::optional< AffineMap > maybeMap, unsigned rank, MLIRContext *context)
	Returns `maybeMap.get()` if `maybeMap` is set, otherwise returns the symbol-less identity map of `rank`. More...

SmallVector< AffineExpr, 4 >	concat (ArrayRef< AffineExpr > a, ArrayRef< AffineExpr > b)
	Return the vector that is the concatenation of `a` and `b`. More...

Value	createOrFoldDimOp (OpBuilder &b, Location loc, Value val, int64_t dim)
	Create one memref::DimOp or tensor::DimOp depending on the type of `val`. More...

OpFoldResult	createFoldedDimOp (OpBuilder &b, Location loc, Value val, int64_t dim)
	Create one memref::DimOp or tensor::DimOp depending on the type of `val`. More...

FailureOr< ContractionDimensions >	inferContractionDims (LinalgOp linalgOp)
	Find at least 2 parallel (m and n) and 1 reduction (k) dimension candidates that form a matmul subcomputation within `linalgOp`. More...

FailureOr< ContractionDimensions >	inferContractionDims (ArrayRef< AffineMap > indexingMaps)

bool	isaContractionOpInterface (LinalgOp linalgOp)
	Checks whether `linalgOp` conforms to ContractionOpInterface. More...

FailureOr< ConvolutionDimensions >	inferConvolutionDims (LinalgOp linalgOp)
	Find at least 1 parallel (output_image) and reduction (filter_loop) dimension candidates that form a convolution subcomputation within `linalgOp`. More...

bool	isaConvolutionOpInterface (LinalgOp linalgOp, bool allowEmptyConvolvedDims=false)
	Checks whether `linalgOp` conforms to ConvolutionOpInterface. More...

bool	isaCopyOpInterface (LinalgOp linalgOp)
	Checks whether `linalgOp` is semantically equivalent to a `linalg.copyOp`. More...

std::optional< SmallVector< int64_t > >	isaBroadcastOpInterface (GenericOp genericOp)
	Checks whether `genericOp` is semantically equivalent to a `linalg.broadcast`. More...

std::optional< SmallVector< int64_t > >	isaTransposeOpInterface (GenericOp genericOp)
	Checks whether `genericOp` is semantically equivalent to a `linalg.transpose`. More...

bool	isaElemwiseSingleUnaryOpInterface (GenericOp genericOp)
	Checks whether a given `genericOp` is semantically equivalent to a single linalgelementwise unary op. More...

bool	isaElemwiseSingleBinaryOpInterface (GenericOp genericOp)
	Checks whether `genericOp` is semantically equivalent to a single linalg elementwise binary op e.g. More...

std::optional< Value >	isaFillOpInterface (GenericOp genericOp)
	Checks whether `genericOp` is semantically equivalent to a `linalg.fill`. More...

void	registerValueBoundsOpInterfaceExternalModels (DialectRegistry &registry)

void	registerTransformDialectExtension (DialectRegistry &registry)

void	registerAllDialectInterfaceImplementations (DialectRegistry &registry)

void	registerBufferizableOpInterfaceExternalModels (DialectRegistry &registry)

void	hoistRedundantVectorTransfers (Operation *root, bool verifyNonZeroTrip=false)
	Hoist vector.transfer_read/vector.transfer_write on buffers pairs out of immediately enclosing scf::ForOp iteratively, if the following conditions are true: More...

void	hoistRedundantVectorBroadcasts (RewriterBase &rewriter, Operation *root)
	Hoist vector.extract/vector.broadcast pairs out of immediately enclosing scf::ForOp iteratively, if the following conditions are met: More...

void	registerMeshShardingInterfaceExternalModels (DialectRegistry &registry)

void	registerRuntimeVerifiableOpInterfaceExternalModels (DialectRegistry &registry)

void	registerSubsetOpInterfaceExternalModels (DialectRegistry &registry)

void	registerTilingInterfaceExternalModels (DialectRegistry &registry)

std::optional< vector::CombiningKind >	getCombinerOpKind (Operation *combinerOp)
	Return vector::CombiningKind for the given op. More...

Value	bufferizeToAllocation (RewriterBase &rewriter, const BufferizeToAllocationOptions &options, tensor::PadOp padOp, Attribute memorySpace={}, Operation *insertionPoint=nullptr)
	Materialize a buffer allocation for the given tensor.pad op and lower the op to linalg.fill/linalg.generic + bufferization.materialize_in_destination. More...

Value	bufferizeToAllocation (RewriterBase &rewriter, const BufferizeToAllocationOptions &options, vector::MaskOp maskOp, Attribute memorySpace={}, Operation *insertionPoint=nullptr)
	Materialize a buffer allocation for the given vector.mask op and bufferize the op, including its region. More...

Value	bufferizeToAllocation (RewriterBase &rewriter, const BufferizeToAllocationOptions &options, bufferization::AllocTensorOp allocTensorOp, Attribute memorySpace={}, Operation *insertionPoint=nullptr)
	Materialize a buffer allocation for the given bufferization.alloc_tensor op and lower the op to memref.alloc + memref.tensor_store. More...

Value	bufferizeToAllocation (RewriterBase &rewriter, const BufferizeToAllocationOptions &options, Operation op, Attribute memorySpace={}, Operation insertionPoint=nullptr)
	Bufferize the given op with tensor semantics and materialize the result in a newly allocated buffer. More...

LogicalResult	linalgOpAnchoredEmptyTensorEliminationStep (RewriterBase &rewriter, Operation *op, bufferization::OneShotAnalysisState &state)
	Try to eliminate tensor::EmptyOps inside `op` that are anchored on a LinalgOp. More...

bool	areElementwiseOpsFusable (OpOperand *fusedOperand)
	Return true if two `linalg.generic` operations with producer/consumer relationship through `fusedOperand` can be fused using elementwise op fusion. More...

LogicalResult	promoteSubviewsPrecondition (Operation *op, LinalgPromotionOptions options)
	Promote memref.subviews feeding linalg-on-buffers operations. More...

LogicalResult	vectorizeOpPrecondition (Operation *op, ArrayRef< int64_t > inputVectorSizes={}, ArrayRef< bool > inputScalableVecDims={}, bool vectorizeNDExtract=false, bool flatten1DDepthwiseConv=false)
	Return success if the operation can be vectorized. More...

FailureOr< DropUnitDimsResult >	dropUnitDims (RewriterBase &rewriter, GenericOp genericOp, const ControlDropUnitDims &options)

FailureOr< ElementwiseOpFusionResult >	fuseElementwiseOps (RewriterBase &rewriter, OpOperand *fusedOperand)

llvm::SmallDenseSet< int >	getPreservedProducerResults (GenericOp producer, GenericOp consumer, OpOperand *fusedOperand)
	Returns a set of indices of the producer's results which would be preserved after the fusion. More...

SmallVector< Value >	peelLoop (RewriterBase &rewriter, Operation *op)
	Try to peel and canonicalize loop `op` and return the new result. More...

void	peelLoops (RewriterBase &rewriter, ArrayRef< scf::ForOp > loops)
	Peel 'loops' and applies affine_min/max bounds simplification on the fly where relevant. More...

LogicalResult	rewriteAsPaddedOp (RewriterBase &rewriter, LinalgOp opToPad, const LinalgPaddingOptions &options, LinalgOp &paddedOp, SmallVector< Value > &replacements, SmallVector< tensor::PadOp > &padOps)
	Pad the iterator dimensions `paddingDimensions` of all `opToPad` operands to a static bounding box. More...

FailureOr< Value >	hoistPaddingOnTensors (RewriterBase &rewriter, tensor::PadOp opToHoist, int64_t numLoops, ArrayRef< int64_t > transposeVector, tensor::PadOp &hoistedOp, SmallVectorImpl< TransposeOp > &transposeOps)
	Mechanically hoist padding operations on tensors by `numLoops` into a new, generally larger tensor. More...

FailureOr< Value >	hoistPaddingOnTensors (tensor::PadOp opToHoist, int64_t numLoops, ArrayRef< int64_t > transposeVector, tensor::PadOp &hoistedOp, SmallVectorImpl< TransposeOp > &transposeOps)
	Calls into `hoistPaddingOnTensors` with a local IRRewriter. More...

FailureOr< LinalgOp >	padAndHoistLinalgOp (RewriterBase &rewriter, LinalgOp linalgOp, const LinalgPaddingOptions &options)
	Apply padding and hoisting to `linalgOp` according to the configuration specified in `options`. More...

std::pair< TilingInterface, TilingInterface >	splitOp (RewriterBase &rewriter, TilingInterface op, unsigned dimension, OpFoldResult splitPoint)
	Split the given `op` into two parts along the given iteration space `dimension` at the specified `splitPoint`, and return the two parts. More...

FailureOr< TiledLinalgOp >	tileLinalgOp (RewriterBase &b, LinalgOp op, const LinalgTilingOptions &options)

FailureOr< GenericOp >	interchangeGenericOp (RewriterBase &rewriter, GenericOp genericOp, ArrayRef< unsigned > interchangeVector)
	Interchange the `iterator_types` and `iterator_maps` dimensions and adapts the index accesses of `op`. More...

FailureOr< GenericOp >	generalizeNamedOp (RewriterBase &rewriter, LinalgOp namedOp)
	Create a GenericOp from the given named operation `namedOp` and replace namedOp. More...

FailureOr< LinalgOp >	specializeGenericOp (RewriterBase &rewriter, GenericOp genericOp)
	Create a namedOp from the given GenericOp and replace the GenericOp. More...

FailureOr< PromotionInfo >	promoteSubviewAsNewBuffer (OpBuilder &b, Location loc, memref::SubViewOp subView, const AllocBufferCallbackFn &allocationFn, DataLayout &layout)

FailureOr< LinalgOp >	promoteSubViews (OpBuilder &b, LinalgOp op, const LinalgPromotionOptions &options)
	Promote the `subViews` into a new buffer allocated at the insertion point `b`. More...

std::optional< Value >	allocateWorkgroupMemory (OpBuilder &builder, memref::SubViewOp subview, ArrayRef< Value > sizeBounds, DataLayout &)
	Allocate the subview in the GPU workgroup memory. More...

LogicalResult	deallocateWorkgroupMemory (OpBuilder &, Value)
	In case of GPU group memory there is no need to deallocate. More...

LogicalResult	copyToWorkgroupMemory (OpBuilder &b, Value src, Value dst)
	Create Memref copy operations and add gpu barrier guards before and after the copy operation to ensure data integrity. More...

std::optional< Value >	allocateGPUPrivateMemory (OpBuilder &builder, memref::SubViewOp subview, ArrayRef< Value > sizeBounds, DataLayout &)
	Allocate the subview in the GPU private memory. More...

LogicalResult	copyToGPUPrivateMemory (OpBuilder &b, Value src, Value dst)
	Normal copy to between src and dst. More...

LogicalResult	deallocateGPUPrivateMemory (OpBuilder &, Value)
	In case of GPU private memory there is no need to deallocate since the memory is freed when going outside of the scope. More...

bool	hasVectorizationImpl (Operation *)
	Return true if there's dedicated logic in the Linalg Vectorizer to vectorize this Op, false otherwise. More...

LogicalResult	vectorize (RewriterBase &rewriter, Operation *op, ArrayRef< int64_t > inputVectorSizes={}, ArrayRef< bool > inputScalableVecDims={}, bool vectorizeNDExtract=false, bool flatten1DDepthwiseConv=false)
	Emit a suitable vector form for an operation. More...

LogicalResult	vectorizeCopy (RewriterBase &builder, memref::CopyOp copyOp)
	Emit a suitable vector form for a Copy op with fully static shape. More...

FailureOr< LinalgLoops >	linalgOpToLoops (RewriterBase &rewriter, LinalgOp linalgOp)
	Emit a loop nest of `scf.for` with the proper body for `linalgOp`. More...

FailureOr< LinalgLoops >	linalgOpToParallelLoops (RewriterBase &rewriter, LinalgOp linalgOp)
	Emit a loop nest of `scf.parallel` with the proper body for `linalgOp`. More...

FailureOr< LinalgLoops >	linalgOpToAffineLoops (RewriterBase &rewriter, LinalgOp linalgOp)
	Emit a loop nest of `affine.for` with the proper body for `linalgOp`. More...

std::tuple< SmallVector< Range, 4 >, LoopIndexToRangeIndexMap >	makeTiledLoopRanges (RewriterBase &b, Location loc, AffineMap map, ArrayRef< OpFoldResult > allShapeSizes, ArrayRef< OpFoldResult > allTileSizes)

FailureOr< MultiSizeSpecification >	computeMultiTileSizes (OpBuilder &builder, LinalgOp op, unsigned dimension, OpFoldResult targetSize, OpFoldResult divisor, bool emitAssertions=true)
	Emits the IR computing the multi-sized tiling specification with two tile sizes not exceeding `targetSize`, each divisible by `sizeDivisor`, such that there exist numbers of tiles with these sizes that fully cover the given iteration space `dimension` of the structured `op`. More...

FailureOr< StaticMultiSizeSpecification >	computeStaticMultiTileSizes (LinalgOp op, unsigned dimension, int64_t targetSize, int64_t divisor)

FailureOr< StaticContinuousTileSizeSpecification >	computeStaticContinuousTileSizes (LinalgOp op, unsigned dimension, unsigned targetSize)

FailureOr< ContinuousTileSizeSpecification >	computeContinuousTileSizes (OpBuilder &builder, TilingInterface op, unsigned dimension, OpFoldResult targetSize, bool emitAssertions)

FailureOr< ForallReductionTilingResult >	tileReductionUsingForall (RewriterBase &b, PartialReductionOpInterface op, ArrayRef< OpFoldResult > numThreads, ArrayRef< OpFoldResult > tileSizes={}, std::optional< ArrayAttr > mapping=std::nullopt)
	Method to tile a reduction to parallel iterations computing partial reductions. More...

void	transformIndexOps (RewriterBase &b, LinalgOp op, SmallVectorImpl< Value > &ivs, const LoopIndexToRangeIndexMap &loopIndexToRangeIndex)
	All indices returned by IndexOp should be invariant with respect to tiling. More...

FailureOr< SplitReductionResult >	splitReduction (RewriterBase &b, LinalgOp op, const ControlSplitReductionFn &controlSplitReductionFn, bool useAlloc=false)

FailureOr< SplitReductionResult >	splitReductionByScaling (RewriterBase &b, LinalgOp op, const ControlSplitReductionFn &controlSplitReductionFn, bool useAlloc=false)
	Scaling-based implementation of the split reduction transformation. More...

bool	isDimSequencePreserved (AffineMap map, ReassociationIndicesRef dimSequence)
	Return `true` if a given sequence of dimensions are contiguous in the range of the specified indexing map. More...

bool	areDimSequencesPreserved (ArrayRef< AffineMap > maps, ArrayRef< ReassociationIndices > dimSequences)
	Return `true` if all sequences of dimensions specified in `dimSequences` are contiguous in all the ranges of the `maps`. More...

FailureOr< CollapseResult >	collapseOpIterationDims (LinalgOp op, ArrayRef< ReassociationIndices > foldedIterationDims, RewriterBase &rewriter)
	Collapses dimensions of linalg.generic/linalg.copy operation. More...

FailureOr< LowerPackResult >	lowerPack (RewriterBase &rewriter, tensor::PackOp packOp, bool lowerPadLikeWithInsertSlice=true)
	Rewrite pack as pad + reshape + transpose. More...

FailureOr< LowerUnPackOpResult >	lowerUnPack (RewriterBase &rewriter, tensor::UnPackOp unPackOp, bool lowerUnpadLikeWithExtractSlice=true)
	Rewrite pack as empty + transpose + reshape + extract_slice. More...

FailureOr< PackResult >	pack (RewriterBase &rewriter, linalg::LinalgOp linalgOp, ArrayRef< OpFoldResult > packedSizes)
	Implement packing of a single LinalgOp by `packedSizes`. More...

FailureOr< PackTransposeResult >	packTranspose (RewriterBase &rewriter, tensor::PackOp packOp, linalg::LinalgOp linalgOp, tensor::UnPackOp maybeUnPackOp, ArrayRef< int64_t > outerPerm, ArrayRef< int64_t > innerPerm)
	Transpose a single PackOp -> LinalgOp -> UnPackOp chain and return the transposed PackOp -> LinalgOp -> UnPackOp chain after replacements. More...

FailureOr< PackResult >	packMatmulGreedily (RewriterBase &rewriter, LinalgOp linalgOp, ArrayRef< OpFoldResult > mnkPackedSizes, ArrayRef< int64_t > mnkPaddedSizesNextMultipleOf, ArrayRef< int64_t > mnkOrder)
	Pack a LinalgOp by greedily inferring matmul dimensions (m, n, k) where m and n are proper parallel dimensions and k is a proper reduction dimension. More...

FailureOr< PackResult >	blockPackMatmul (RewriterBase &rewriter, linalg::LinalgOp linalgOp, const ControlBlockPackMatmulFn &controlPackMatmul)
	Pack a matmul operation into blocked 4D layout. More...

FailureOr< Operation * >	rewriteInDestinationPassingStyle (RewriterBase &rewriter, tensor::FromElementsOp fromElementsOp)
	Rewrite tensor.from_elements to linalg.generic. More...

FailureOr< Operation * >	rewriteInDestinationPassingStyle (RewriterBase &rewriter, tensor::GenerateOp generateOp)
	Rewrite tensor.generate to linalg.generic. More...

FailureOr< Operation * >	rewriteInDestinationPassingStyle (RewriterBase &rewriter, tensor::PadOp padOp)
	Rewrite tensor.pad to linalg.generic + tensor.insert_slice. More...

FailureOr< std::pair< Operation , Operation > >	rewriteInIm2Col (RewriterBase &rewriter, linalg::Conv2DNhwcHwcfOp convOp)
	Convert linalg.conv_2d_nhwc_hwcf into linalg.generic (for img2col packing) and linalg.matmul. More...

FailureOr< std::pair< Operation , Operation > >	rewriteInIm2Col (RewriterBase &rewriter, linalg::Conv2DNhwcFhwcOp convOp)
	Same as the above but for Fhwc channel orderings in the filter. More...

FailureOr< std::pair< Operation , Operation > >	rewriteInIm2Col (RewriterBase &rewriter, linalg::DepthwiseConv2DNhwcHwcOp convOp)
	Similar to rewriteInIm2Col with linalg::Conv2DNhwcHwcfOp except there is no reduction among the input channels so each convolution can be a matrix-vector product and by transposing both input filter so channels are outer most the computation is a batched matrix-vector product. More...

FailureOr< std::pair< Operation , Operation > >	rewriteInIm2Col (RewriterBase &rewriter, linalg::Conv2DNchwFchwOp convOp)
	Similar to rewriteInIm2Col with linalg::Conv2DNhwcHwcfOp except because the channels are to the left of the image shape dimensions, the position of the contraction dimension in the resulting matmul is reversed. More...

FailureOr< Operation * >	transposeConv2D (RewriterBase &rewriter, linalg::Conv2DNhwcFhwcOp op)
	Convert linalg.conv_2d_nhwc_fhwc(_q) to linalg.conv_2d_nhwc_hwcf(_q) by materializing transpose. More...

FailureOr< Operation * >	transposeConv2D (RewriterBase &rewriter, linalg::Conv2DNhwcFhwcQOp op)

FailureOr< Operation * >	transposeMatmul (RewriterBase &rewriter, linalg::MatmulOp op, bool transposeLHS=true)
	Convert Linalg matmul ops to transposed variants. More...

FailureOr< Operation * >	transposeBatchMatmul (RewriterBase &rewriter, linalg::BatchMatmulOp op, bool transposeLHS=true)
	Pattern to replace. More...

FailureOr< Operation * >	winogradConv2D (RewriterBase &rewriter, linalg::Conv2DNhwcFhwcOp op, int64_t m, int64_t r)
	Convert linalg.conv_2d_nhwc_fhwc to Winograd Conv2D algorithm F(m x m, r x r). More...

FailureOr< Operation * >	decomposeWinogradFilterTransformOp (RewriterBase &rewriter, linalg::WinogradFilterTransformOp op)
	Rewrite linalg.winograd_filter_transform. More...

FailureOr< Operation * >	decomposeWinogradInputTransformOp (RewriterBase &rewriter, linalg::WinogradInputTransformOp op)
	Rewrite linalg.winograd_input_transform. More...

FailureOr< Operation * >	decomposeWinogradOutputTransformOp (RewriterBase &rewriter, linalg::WinogradOutputTransformOp op)
	Rewrite linalg.winograd_output_transform. More...

RewritePatternSet	getLinalgTilingCanonicalizationPatterns (MLIRContext *ctx)
	Canonicalization patterns relevant to apply after tiling patterns. More...

void	populateLinalgTilingCanonicalizationPatterns (RewritePatternSet &patterns)

void	populateLinalgNamedOpsGeneralizationPatterns (RewritePatternSet &patterns)
	Linalg generalization patterns. More...

void	populateLinalgGenericOpsSpecializationPatterns (RewritePatternSet &patterns)
	Populates `patterns` with patterns to convert linalg.generic ops to named ops where possible. More...

void	populateDecomposeConvolutionPatterns (RewritePatternSet &patterns, PatternBenefit benefit=1)
	Linalg decompose convolutions patterns. More...

void	populateDecomposePackUnpackPatterns (RewritePatternSet &patterns)
	Populates patterns to decompose tensor.pack and tensor.unpack Ops into e.g. More...

void	populateDecomposePadPatterns (RewritePatternSet &patterns)
	Populates patterns to decompose tensor.pad into e.g. More...

void	populateConvertConv2DToImg2ColPatterns (RewritePatternSet &patterns)
	Populates patterns to transform linalg.conv_2d_xxx operations into linalg.generic (for img2col packing) and linalg.matmul. More...

void	populateInsertSliceVectorizationPatterns (RewritePatternSet &patterns)
	Populates `patterns` with vectorisation patterns for tensor.insert_slice. More...

void	populatePadOpVectorizationPatterns (RewritePatternSet &patterns, PatternBenefit baseBenefit=1)
	Populates `patterns` with patterns that vectorize tensor.pad. More...

void	populateDecomposeLinalgOpsPattern (RewritePatternSet &patterns, bool removeDeadArgsAndResults=true)
	Populate patterns for splitting a `LinalgOp` with multiple statements within its payload into multiple `GenericOp` that have a single statement. More...

void	populateConvertToDestinationStylePatterns (RewritePatternSet &patterns)
	Populate patterns that convert non-destination-style ops to destination style ops. More...

void	populateConvolutionVectorizationPatterns (RewritePatternSet &patterns, PatternBenefit benefit=1)
	Populate patterns for vectorizing low-D convolution ops. More...

void	populateElementwiseToLinalgConversionPatterns (RewritePatternSet &patterns)
	Populate patterns that convert `ElementwiseMappable` ops to linalg parallel loops. More...

void	populateSparseTensorRewriting (RewritePatternSet &patterns)
	Populate patterns that are only useful in the context of sparse tensors. More...

void	populateElementwiseOpsFusionPatterns (RewritePatternSet &patterns, const ControlFusionFn &controlElementwiseOpFusion)
	Patterns for fusing linalg operation on tensors. More...

void	populateDataLayoutPropagationPatterns (RewritePatternSet &patterns, const ControlPropagationFn &controlPackUnPackPropagation)
	Patterns to bubble up or down data layout ops across other operations. More...

void	populateEraseUnusedOperandsAndResultsPatterns (RewritePatternSet &patterns)
	Pattern to remove dead operands and results of `linalg.generic` operations. More...

void	populateEraseUnnecessaryInputsPatterns (RewritePatternSet &patterns)
	Patterns to promote inputs to outputs and remove unused inputs of `linalg.generic` ops. More...

void	populateCollapseDimensions (RewritePatternSet &patterns, const GetCollapsableDimensionsFn &controlCollapseDimensions)
	Pattern to collapse dimensions in a linalg.generic op. More...

void	populateFoldReshapeOpsByExpansionPatterns (RewritePatternSet &patterns, const ControlFusionFn &controlFoldingReshapes)
	Patterns to fold an expanding (collapsing) tensor_reshape operation with its producer (consumer) generic operation by expanding the dimensionality of the loop in the generic op. More...

void	populateFoldReshapeOpsByCollapsingPatterns (RewritePatternSet &patterns, const ControlFusionFn &controlFoldingReshapes)
	Patterns to fold an expanding tensor.expand_shape operation with its producer generic operation by collapsing the dimensions of the generic op. More...

void	populateConstantFoldLinalgOperations (RewritePatternSet &patterns, const ControlFusionFn &controlFn)
	Patterns to constant fold Linalg operations. More...

void	populateFoldAddIntoDestPatterns (RewritePatternSet &patterns)
	Pattern to replace `linalg.add` when destination passing on a contraction op suffices for achieving the sum. More...

void	populateFuseTensorPadWithProducerLinalgOpPatterns (RewritePatternSet &patterns)
	Pattern to fuse a `tensor.pad` operation with the producer of its source, if the producer is a `linalg` operation with all parallel iterator types. More...

void	populateLinalgNamedOpConversionPatterns (RewritePatternSet &patterns)
	Patterns to convert from one named op to another. More...

void	populateFoldUnitExtentDimsPatterns (RewritePatternSet &patterns, ControlDropUnitDims &options)
	Patterns to fold unit-extent dimensions in operands/results of linalg ops on tensors via reassociative reshape ops. More...

void	populateMoveInitOperandsToInputPattern (RewritePatternSet &patterns)
	A pattern that converts init operands to input operands. More...

void	populateInlineConstantOperandsPatterns (RewritePatternSet &patterns)
	Patterns that are used to inline constant operands into linalg generic ops. More...

void	populateBubbleUpExtractSliceOpPatterns (RewritePatternSet &patterns)
	Patterns that are used to bubble up extract slice op above linalg op. More...

void	populateSwapExtractSliceWithFillPatterns (RewritePatternSet &patterns)
	Adds patterns that waps tensor.extract_slice(linalg.fill(cst, init)) into linalg.fill(cst, tensor.extract_slice(init)). More...

void	populateDecomposeProjectedPermutationPatterns (RewritePatternSet &patterns)
	Add patterns to make explicit broadcasts and transforms in the input operands of a genericOp. More...

void	populateSplitReductionPattern (RewritePatternSet &patterns, const ControlSplitReductionFn &controlSplitReductionFn, bool useAlloc=false)
	Patterns to apply `splitReduction` below. More...

void	populateTransposeMatmulPatterns (RewritePatternSet &patterns, bool transposeLHS=true)
	Patterns to convert Linalg matmul ops to transposed variants. More...

void	populateBlockPackMatmulPatterns (RewritePatternSet &patterns, const ControlBlockPackMatmulFn &controlFn)
	Patterns to block pack Linalg matmul ops. More...

void	populateWinogradConv2DPatterns (RewritePatternSet &patterns, int64_t m, int64_t r)
	Patterns to apply Winograd Conv2D algorithm F(m x m, r x r). More...

void	populateDecomposeWinogradOpsPatterns (RewritePatternSet &patterns)
	Patterns to decompose Winograd operators. More...

void	populateContractionOpRankReducingPatterns (RewritePatternSet &patterns)
	Adds patterns that reduce the rank of named contraction ops that have unit dimensions in the operand(s) by converting to a sequence of `collapse_shape`, `<corresponding linalg named op>`, `expand_shape` (if on tensors). More...

bool	allIndexingsAreProjectedPermutation (LinalgOp op)
	Check if all indexing maps are projected permutations. More...

bool	hasOnlyScalarElementwiseOp (Region &r)
	Detect whether `r` has only ConstantOp, ElementwiseMappable and YieldOp. More...

bool	isElementwise (LinalgOp op)
	Check if a LinalgOp is an element-wise operation. More...

bool	isParallelIterator (utils::IteratorType iteratorType)
	Check if iterator type has "parallel" semantics. More...

bool	isReductionIterator (utils::IteratorType iteratorType)
	Check if iterator type has "reduction" semantics. More...

Value	makeComposedPadHighOp (OpBuilder &b, Location loc, RankedTensorType type, Value source, Value pad, bool nofold)
	Create a tensor::PadOp that pads `source` to the size of the statically sized `type` whose static sizes are assumed to be greater than the dynamic `source` size. More...

GenericOp	makeMemRefCopyOp (OpBuilder &b, Location loc, Value from, Value to)
	Returns GenericOp that copies an n-D memref. More...

std::optional< SmallVector< ReassociationIndices > >	getReassociationMapForFoldingUnitDims (ArrayRef< OpFoldResult > mixedSizes)
	Get the reassociation maps to fold the result of a extract_slice (or source of a insert_slice) operation with given offsets, and sizes to its rank-reduced version. More...

SmallVector< OpFoldResult >	computeTileOffsets (OpBuilder &b, Location loc, ArrayRef< OpFoldResult > ivs, ArrayRef< OpFoldResult > tileSizes)
	Computes tile offsets, given a list of loop `ivs` and `tileSizes`. More...

SmallVector< OpFoldResult >	computeTileSizes (OpBuilder &b, Location loc, ArrayRef< OpFoldResult > tileSizes, ArrayRef< OpFoldResult > sizeBounds)
	Computes tile sizes, given a list of `tileSizes` and dimension sizes (`sizeBounds`). More...

SmallVector< Type >	getTensorOutputTypes (LinalgOp op, ValueRange operands)
	Returns the list of tensor output types produced when the given structured operation `op` is applied to the given `operands`. More...

SmallVector< Value >	insertSlicesBack (OpBuilder &builder, Location loc, LinalgOp op, ValueRange operands, ValueRange results)
	Creates `insert_slice` ops that insert `results` back into larger tensors they were originally extracted from with `extract_slice` before being passed as `operands` to the given structured operation `op` or its clone. More...

SliceParameters	computeSliceParameters (OpBuilder &builder, Location loc, Value valueToTile, ArrayRef< OpFoldResult > tileSizes, AffineMap map, ArrayRef< OpFoldResult > lbs, ArrayRef< OpFoldResult > ubs, ArrayRef< OpFoldResult > subShapeSizes, bool omitPartialTileCheck)
	Computes SliceParameters for a single `valueToTile` assuming that its user is being tiled with the given loop bounds `lbs` and `ubs` and the tile sizes `tileSizes`. More...

SmallVector< std::optional< SliceParameters > >	computeAllSliceParameters (OpBuilder &builder, Location loc, LinalgOp linalgOp, ValueRange valuesToTile, ArrayRef< OpFoldResult > ivs, ArrayRef< OpFoldResult > tileSizes, ArrayRef< OpFoldResult > sizeBounds, bool omitPartialTileCheck)
	Computes SliceParamaters for all `valuesToTile` of the given `linalgOp`, assuming `linalgOp` is being fused into a loop nest. More...

Operation *	makeTiledShape (OpBuilder &builder, Location loc, Value valueToTile, ArrayRef< OpFoldResult > tileSizes, AffineMap map, ArrayRef< OpFoldResult > lbs, ArrayRef< OpFoldResult > ubs, ArrayRef< OpFoldResult > subShapeSizes, bool omitPartialTileCheck)
	Creates an extract_slice/subview op for a single `valueToTile` with `builder`. More...

SmallVector< Value >	makeTiledShapes (OpBuilder &builder, Location loc, LinalgOp linalgOp, ValueRange valuesToTile, ArrayRef< OpFoldResult > ivs, ArrayRef< OpFoldResult > tileSizes, ArrayRef< OpFoldResult > sizeBounds, bool omitPartialTileCheck)
	Creates extract_slice/subview ops for all `valuesToTile` of the given `linalgOp` with `builder`, assuming `linalgOp` is being fused into a loop nest for tiling with the given induction variables `ivs` and tile sizes `tileSizes`. More...

void	offsetIndices (OpBuilder &b, LinalgOp linalgOp, ArrayRef< OpFoldResult > offests)
	Add the specified offsets to any `linalg.index` ops contained in the given `linalgOp`. More...

void	offsetIndices (RewriterBase &b, LinalgOp linalgOp, ArrayRef< OpFoldResult > offests)

FailureOr< FusionInfo >	fuseProducerOfTensor (OpBuilder &b, OpOperand &consumerOpOperand)
	This implements the fusion part of the "tileAndFuse on tensors" transformation and thus requires the `consumerOpOperand` to be a `extract_slice` op (generally obtained by applying the tiling transformation). More...

FailureOr< FusionInfo >	fuseProducerOfTensor (OpBuilder &b, OpResult producerOpResult, OpOperand &consumerOpOperand)
	This implements the fusion part of the "tileAndFuse on tensors" transformation and thus requires the `consumerOpOperand` to be a `extract_slice` op (generally obtained by applying the tiling transformation). More...

void	updateBoundsForCyclicDistribution (OpBuilder &builder, Location loc, Value procId, Value nprocs, Value &lb, Value &ub, Value &step)
	Update the `lb`, `ub` and `step` to get per processor `lb`, `ub` and `step`. More...

template<typename OpTy >
SmallVector< NamedAttribute >	getPrunedAttributeList (OpTy op)
	Returns an attribute list that excludes pre-defined attributes. More...

static bool	hasAllOneValues (DenseIntElementsAttr attr)

static Value	createAdd (Location loc, Value x, Value y, OpBuilder &builder)

static Value	createMul (Location loc, Value x, Value y, Type accType, OpBuilder &builder)

static SmallVector< Value >	unrollIndex (OpBuilder &b, Location loc, Value index, ArrayRef< int64_t > factors)

static Value	getConvolvedIndex (OpBuilder &b, Location loc, Value oIndex, Value fIndex, int64_t stride)

static ReductionKind	getReductionKind (Operation *op)

static std::optional< Operation * >	getCombinerOp (LinalgOp op)

static ReductionKind	getReductionKindOfLinalgOp (LinalgOp op)

static MeshOp	getMesh (Operation *op, ArrayRef< MeshSharding > operandShardings, ArrayRef< MeshSharding > resultShardings, SymbolTableCollection &symbolTable)

static Value	createDestinationPassingStyleInitOperand (LinalgOp op, Value spmdizedOperand, ArrayRef< MeshAxis > reductionMeshAxes, MeshOp meshOp, ImplicitLocOpBuilder &builder)

static SmallVector< Value >	createDestinationPassingStyleInitOperands (LinalgOp op, MeshOp meshOp, ArrayRef< Value > spmdizedOperands, ArrayRef< MeshAxis > reductionMeshAxes, IRMapping &spmdizationMap, ImplicitLocOpBuilder &builder)

static void	createAllReduceForResultWithoutPartialSharding (Value unshardedLinalgOpResult, ArrayRef< MeshAxis > opReductionMeshAxes, MeshSharding resultSharding, ReductionKind reductionKind, IRMapping &spmdizationMap, ImplicitLocOpBuilder &builder)

static void	createAllReduceForResultsWithoutPartialShardings (LinalgOp unshardedOp, ArrayRef< MeshAxis > opReductionMeshAxes, ArrayRef< MeshSharding > resultShardings, IRMapping &spmdizationMap, ImplicitLocOpBuilder &builder)

static void	spmdizeLinalgOpWithShardedReduction (LinalgOp op, ArrayRef< Value > spmdizedOperands, ArrayRef< MeshSharding > operandShardings, ArrayRef< MeshSharding > resultShardings, ArrayRef< utils::IteratorType > loopIteratorTypes, ArrayRef< SmallVector< MeshAxis >> meshAxisAssignmentForLoopIterators, IRMapping &spmdizationMap, SymbolTableCollection &symbolTable, ImplicitLocOpBuilder &builder)

template<typename OpType >
static void	registerOne (MLIRContext *ctx)

template<typename... OpTypes>
static void	registerAll (MLIRContext *ctx)
	Variadic helper function. More...

void	populateTranposeConv2DPatterns (RewritePatternSet &patterns)

static void	generateParallelLoopNest (OpBuilder &b, Location loc, ValueRange lbs, ValueRange ubs, ValueRange steps, ArrayRef< utils::IteratorType > iteratorTypes, ArrayRef< linalg::ProcInfo > procInfo, function_ref< void(OpBuilder &, Location, ValueRange)> bodyBuilderFn, SmallVectorImpl< Value > &ivStorage)
	Generates a loop nest consisting of scf.parallel and scf.for, depending on the `iteratorTypes.` More...

static Operation *	materializeTiledShape (OpBuilder &builder, Location loc, Value valueToTile, const SliceParameters &sliceParams)

Typedef Documentation

◆ AllocBufferCallbackFn

using mlir::linalg::AllocBufferCallbackFn = typedef std::function<std::optional<Value>( OpBuilder &b, memref::SubViewOp subView, ArrayRef<Value> boundingSubViewSize, DataLayout &layout)>

Callback function type used to perform the allocation for the promoted subView.

In boundingSubViewsize a best attempt is made to find the smallest constant value for the size of the buffer needed for each dimension. If that is not possible, contains the dynamic size of the subview. The call back should return the buffer to use.

Definition at line 337 of file Transforms.h.

◆ ControlBlockPackMatmulFn

using mlir::linalg::ControlBlockPackMatmulFn = typedef std::function<std::optional<BlockPackMatmulOptions>(linalg::LinalgOp)>

Function type which is used to control matmul packing.

It is expected to return valid packing configuration for each operation. Lack of packing options indicates that no valid configuration could be assigned and the operation will not be packed.

Definition at line 1216 of file Transforms.h.

◆ ControlFusionFn

using mlir::linalg::ControlFusionFn = typedef std::function<bool(OpOperand *fusedOperand)>

Function type which is used to control when to stop fusion.

It is expected that OpOperand is not modified in the callback. The OpOperand is not marked as const to allow callers to use non-const methods.

Definition at line 1769 of file Transforms.h.

◆ ControlPropagationFn

using mlir::linalg::ControlPropagationFn = typedef std::function<bool(OpOperand *opOperand)>

Function type which is used to control propagation of tensor.pack/unpack ops.

Definition at line 1781 of file Transforms.h.

◆ ControlSplitReductionFn

using mlir::linalg::ControlSplitReductionFn = typedef std::function<SplitReductionOptions(LinalgOp op)>

Function signature to control reduction splitting.

This returns SplitReductionOptions.

Definition at line 441 of file Transforms.h.

◆ CopyCallbackFn

using mlir::linalg::CopyCallbackFn = typedef std::function<LogicalResult(OpBuilder &b, Value src, Value dst)>

Callback function type used to insert copy from original subview to subview of the promoted region for the read operands/subview of promoted region to original subview for the results.

The copy has to happen from src to dst.

Definition at line 350 of file Transforms.h.

◆ DeallocBufferCallbackFn

using mlir::linalg::DeallocBufferCallbackFn = typedef std::function<LogicalResult(OpBuilder &b, Value buffer)>

Callback function type used to deallocate the buffers used to hold the promoted subview.

Definition at line 343 of file Transforms.h.

◆ GetCollapsableDimensionsFn

using mlir::linalg::GetCollapsableDimensionsFn = typedef std::function<SmallVector<ReassociationIndices>(linalg::LinalgOp)>

Function type to control generic op dimension collapsing.

It is expected to return an array of ReassociationIndices representing dimensions that should be merged.

Definition at line 1799 of file Transforms.h.

◆ LinalgLoops

using mlir::linalg::LinalgLoops = typedef SmallVector<Operation *, 4>

Definition at line 469 of file Transforms.h.

◆ LoopIndexToRangeIndexMap

using mlir::linalg::LoopIndexToRangeIndexMap = typedef DenseMap<int, int>

Creates a number of ranges equal to the number of non-zero in tileSizes.

One for each loop of the LinalgOp that is tiled. The tileSizes argument has one entry per surrounding loop. It uses zero as the convention that a particular loop is not tiled. This convention simplifies implementations by avoiding affine map manipulations. The returned ranges correspond to the loop ranges, in the proper order, that are tiled and for which new loops will be created. Also the function returns a map from loop indices of the LinalgOp to the corresponding non-empty range indices of newly created loops.

Definition at line 809 of file Transforms.h.

◆ MeshAxis

using mlir::linalg::MeshAxis = typedef mesh::MeshAxis

Definition at line 44 of file MeshShardingInterfaceImpl.cpp.

◆ MeshOp

using mlir::linalg::MeshOp = typedef mesh::MeshOp

Definition at line 48 of file MeshShardingInterfaceImpl.cpp.

◆ MeshSharding

using mlir::linalg::MeshSharding = typedef mesh::MeshSharding

Definition at line 46 of file MeshShardingInterfaceImpl.cpp.

◆ OptimizeCopyFn

using mlir::linalg::OptimizeCopyFn = typedef std::function<LogicalResult(RewriterBase &, tensor::PadOp, Value)>

Definition at line 1503 of file Transforms.h.

◆ ProcInfoCallBackFn

using mlir::linalg::ProcInfoCallBackFn = typedef std::function<SmallVector<ProcInfo>( OpBuilder &b, Location loc, ArrayRef<Range> parallelLoopRanges)>

Definition at line 293 of file Utils.h.

◆ ReductionKind

using mlir::linalg::ReductionKind = typedef mesh::ReductionKind

Definition at line 45 of file MeshShardingInterfaceImpl.cpp.

◆ ShardingArray

using mlir::linalg::ShardingArray = typedef mesh::ShardingArray

Definition at line 47 of file MeshShardingInterfaceImpl.cpp.

◆ TileSizeComputationFunction

using mlir::linalg::TileSizeComputationFunction = typedef std::function<SmallVector<Value, 4>(OpBuilder &, Operation *)>

Definition at line 185 of file Transforms.h.

Enumeration Type Documentation

◆ DistributionMethod

enum mlir::linalg::DistributionMethod

strong

Scheme used to distribute loops to processors.

Enumerator
Cyclic	Cyclic distribution where no assumption is made about the dynamic relationship between number of processors and number of iterations of the distributed loop. Distributes the following loop scf.parallel (iv) = (lb) to (ub) step (step) to scf.parallel(iv)= (lb + procId * step) to (ub) step (step * nprocs)
CyclicNumProcsGeNumIters	Cyclic distribution where the number of processors can be assumed to be more than or equal to the number of iterations of the distributed loop. In such cases, a simple in-bounds check is enough (instead of materializing a loop). Distributes the following loop scf.parallel (iv) = (lb) to (ub) step (step) to iv = lb + procId * step cond = arith.cmpi "slt", iv, ub scf.if cond { ... }
CyclicNumProcsEqNumIters	Cyclic distribution where the number of processors can be assumed to be equal to the number of iterations of the distributed loop. In such cases, no bounds check is needed. Distributes the following loop scf.parallel (iv) = (lb) to (ub) step (step) to iv = lb + procId * step
None	No Distribution.

Definition at line 238 of file Utils.h.

◆ LinalgTilingLoopType

enum mlir::linalg::LinalgTilingLoopType

strong

The type of loops to be generated during tiling.

Enumerator
Loops
AffineLoops
ParallelLoops

Definition at line 97 of file Utils.h.

Function Documentation

◆ allIndexingsAreProjectedPermutation()

bool mlir::linalg::allIndexingsAreProjectedPermutation ( LinalgOp op )

Check if all indexing maps are projected permutations.

Definition at line 149 of file Utils.cpp.

Referenced by isElementwise(), and vectorizeLinalgOpPrecondition().

◆ allocateGPUPrivateMemory()

std::optional< Value > mlir::linalg::allocateGPUPrivateMemory	(	OpBuilder &	builder,
		memref::SubViewOp	subview,
		ArrayRef< Value >	sizeBounds,
		DataLayout &
	)

Allocate the subview in the GPU private memory.

Definition at line 495 of file Promotion.cpp.

References allocateSubviewGPUMemoryInAddressSpace().

◆ allocateWorkgroupMemory()

std::optional< Value > mlir::linalg::allocateWorkgroupMemory	(	OpBuilder &	builder,
		memref::SubViewOp	subview,
		ArrayRef< Value >	sizeBounds,
		DataLayout &
	)

Allocate the subview in the GPU workgroup memory.

Definition at line 470 of file Promotion.cpp.

References allocateSubviewGPUMemoryInAddressSpace().

◆ areDimSequencesPreserved()

bool mlir::linalg::areDimSequencesPreserved	(	ArrayRef< AffineMap >	maps,
		ArrayRef< ReassociationIndices >	dimSequences
	)

Return true if all sequences of dimensions specified in dimSequences are contiguous in all the ranges of the maps.

Definition at line 1204 of file ElementwiseOpFusion.cpp.

References isDimSequencePreserved().

◆ areElementwiseOpsFusable()

bool mlir::linalg::areElementwiseOpsFusable ( OpOperand * fusedOperand )

Return true if two linalg.generic operations with producer/consumer relationship through fusedOperand can be fused using elementwise op fusion.

Conditions for elementwise fusion of generic operations.

Definition at line 137 of file ElementwiseOpFusion.cpp.

References mlir::IROperand< DerivedT, IRValueT >::get(), mlir::Value::getDefiningOp(), getIndexingMapOfProducerOperandsInCoordinatesOfFusedOp(), mlir::AffineMap::getNumResults(), mlir::detail::IROperandBase::getOwner(), mlir::Value::getType(), and mlir::AffineMap::isPermutation().

Referenced by fuseElementwiseOps().

◆ blockPackMatmul()

FailureOr< PackResult > mlir::linalg::blockPackMatmul	(	RewriterBase &	rewriter,
		linalg::LinalgOp	linalgOp,
		const ControlBlockPackMatmulFn &	controlPackMatmul
	)

Pack a matmul operation into blocked 4D layout.

Relayout a matmul operation into blocked layout with two levels of subdivision:

major 2D blocks - outer dimensions, consist of minor blocks
minor 2D blocks - inner dimensions, consist of scalar elements

A 2D matmul MxNxK gets reshaped into blocked 4D representation as: [MB][NB][mb][nb] += [MB][KB][mb][kb] * [NB][KB][nb][kb] where the (MB, NB, KB) dimensions represent the major blocks, and the (mb, nb, kb) are the minor blocks of their respective original 2D dimensions (M, N, K).

Depending on the initial operands' data layout and the specified packing options, the major blocks dimensions might get transposed e.g., [MB][KB] -> [KB][MB]. The minor blocks can also be transposed e.g., [mb][kb] -> [kb][mb]. Any present batch dimensions remain unchanged. The final result is unpacked back to the original shape.

Return failure if no valid packing options are provided.

Definition at line 139 of file BlockPackMatmul.cpp.

References mlir::getAsOpFoldResult(), mlir::Builder::getI64ArrayAttr(), inferContractionDims(), mlir::RewriterBase::notifyMatchFailure(), options, packMatmulGreedily(), mlir::OpBuilder::setInsertionPointAfter(), transposePackedMatmul(), and validateFullTilesOnDims().

◆ bufferizeToAllocation() [1/4]

Value mlir::linalg::bufferizeToAllocation	(	RewriterBase &	rewriter,
		const BufferizeToAllocationOptions &	options,
		bufferization::AllocTensorOp	allocTensorOp,
		Attribute	memorySpace = `{}`,
		Operation *	insertionPoint = `nullptr`
	)

Materialize a buffer allocation for the given bufferization.alloc_tensor op and lower the op to memref.alloc + memref.tensor_store.

In addition to rewriting the IR, this function returns the newly allocated buffer. The insertionPoint parameter can be used to specify a custom insertion point for the buffer allocation.

Definition at line 322 of file ConvertToDestinationStyle.cpp.

References mlir::OpBuilder::create(), createAllocationForTensor(), options, mlir::RewriterBase::replaceOp(), and mlir::OpBuilder::setInsertionPoint().

◆ bufferizeToAllocation() [2/4]

Value mlir::linalg::bufferizeToAllocation	(	RewriterBase &	rewriter,
		const BufferizeToAllocationOptions &	options,
		Operation *	op,
		Attribute	memorySpace = `{}`,
		Operation *	insertionPoint = `nullptr`
	)

Bufferize the given op with tensor semantics and materialize the result in a newly allocated buffer.

Only bufferizable ops that bufferize to a memory write or have an aliasing OpOperand (and do not themselves bufferize to an allocation) are supported. They are bufferized using their BufferizableOpInterface implementation.

Selected ops that bufferize to an allocation (or need special handling) are also supported:

tensor.pad
vector.mask

This function returns the newly allocated buffer. The insertionPoint parameter can be used to specify a custom insertion point for the buffer allocation.

Definition at line 470 of file ConvertToDestinationStyle.cpp.

References bufferizeToAllocation(), and options.

◆ bufferizeToAllocation() [3/4]

Value mlir::linalg::bufferizeToAllocation	(	RewriterBase &	rewriter,
		const BufferizeToAllocationOptions &	options,
		tensor::PadOp	padOp,
		Attribute	memorySpace = `{}`,
		Operation *	insertionPoint = `nullptr`
	)

Materialize a buffer allocation for the given tensor.pad op and lower the op to linalg.fill/linalg.generic + bufferization.materialize_in_destination.

E.g.:

%0 = tensor.pad low[l] high[h] t ...

is lowered to:

alloc = memref.alloc linalg.fill ... outs(alloc) subview = memref.subview alloc [l] [...] [1] bufferization.materialize_in_destination t in subview %0 = bufferization.to_tensor alloc restrict writable

In addition to rewriting the IR as shown above, this function returns the newly allocated buffer. The insertionPoint parameter can be used to specify a custom insertion point for the buffer allocation.

Referenced by bufferizeToAllocation().

◆ bufferizeToAllocation() [4/4]

Value mlir::linalg::bufferizeToAllocation	(	RewriterBase &	rewriter,
		const BufferizeToAllocationOptions &	options,
		vector::MaskOp	maskOp,
		Attribute	memorySpace = `{}`,
		Operation *	insertionPoint = `nullptr`
	)

Materialize a buffer allocation for the given vector.mask op and bufferize the op, including its region.

E.g.:

%0 = vector.mask { vector.transfer_write v, t : vector<16xf32>, tensor<?xf32> } : vector<16xi1> -> tensor<?xf32>

is lowered to:

alloc = memref.alloc bufferization.materialize_in_destination t in subview vector.mask { vector.transfer_write arg0, alloc : vector<16xf32>, memref<?xf32> } : vector<16xi1> %0 = bufferization.to_tensor alloc restrict writable

In addition to rewriting the IR as shown above, this function returns the newly allocated buffer. The insertionPoint parameter can be used to specify a custom insertion point for the buffer allocation.

Definition at line 260 of file ConvertToDestinationStyle.cpp.

References bufferizeToAllocation(), mlir::RewriterBase::eraseOp(), mlir::RewriterBase::modifyOpInPlace(), options, and mlir::OpBuilder::setInsertionPoint().

◆ collapseOpIterationDims()

FailureOr< CollapseResult > mlir::linalg::collapseOpIterationDims	(	LinalgOp	op,
		ArrayRef< ReassociationIndices >	foldedIterationDims,
		RewriterBase &	rewriter
	)

Collapses dimensions of linalg.generic/linalg.copy operation.

Implementation of fusion with reshape operation by collapsing dimensions.

A precondition to calling this method is that for each list in foldedIterationDim, the sequence of dimensions is contiguous in domains of all indexing_maps of the linalgOp. This can be checked using areDimSequencePreserved method. When valid, the method also collapses the operands of the op. Returns replacement values of the results of the original linalgOp by inserting reshapes to get back values of compatible types.

Definition at line 1674 of file ElementwiseOpFusion.cpp.

References mlir::OpBuilder::create(), createCollapsedOp(), mlir::detail::enumerate(), generateCollapsedIndexingRegion(), mlir::get(), getOperandReassociation(), mlir::Value::getType(), mlir::getValueOrCreateConstantIndexOp(), mlir::m_ConstantInt(), mlir::matchPattern(), mlir::RewriterBase::notifyMatchFailure(), mlir::Range::offset, mlir::OpBuilder::setInsertionPoint(), mlir::Range::size, and mlir::Range::stride.

◆ computeAllSliceParameters()

SmallVector< std::optional< SliceParameters > > mlir::linalg::computeAllSliceParameters	(	OpBuilder &	builder,
		Location	loc,
		LinalgOp	linalgOp,
		ValueRange	valuesToTile,
		ArrayRef< OpFoldResult >	ivs,
		ArrayRef< OpFoldResult >	tileSizes,
		ArrayRef< OpFoldResult >	sizeBounds,
		bool	omitPartialTileCheck
	)

Computes SliceParamaters for all valuesToTile of the given linalgOp, assuming linalgOp is being fused into a loop nest.

Calls computeSliceParameters for every individual value.

Note that a constant zero in tileSizes means no tiling at that implicit loop. The number of non-zero values in tileSizes should be equal to the number of values in ivs.

Some of the valuesToTile won't be affected by tiling. For these values, std::nullopt will be returned.

Definition at line 743 of file Utils.cpp.

References computeSliceParameters(), computeTileOffsets(), computeTileSizes(), and isTiled().

Referenced by makeTiledShapes().

◆ computeContinuousTileSizes()

FailureOr< ContinuousTileSizeSpecification > mlir::linalg::computeContinuousTileSizes	(	OpBuilder &	builder,
		TilingInterface	op,
		unsigned	dimension,
		OpFoldResult	targetSize,
		bool	emitAssertions
	)

Definition at line 162 of file Tiling.cpp.

References mlir::OpBuilder::createOrFold(), emitIsPositiveIndexAssertion(), mlir::AffineExpr::floorDiv(), mlir::Builder::getAffineSymbolExpr(), mlir::getConstantIntValue(), mlir::ImplicitLocOpBuilder::getLoc(), mlir::getValueOrCreateConstantIndexOp(), mlir::affine::makeComposedAffineApply(), mlir::affine::makeComposedFoldedAffineApply(), mlir::linalg::detail::ContinuousTileSizeSpecificationBase< T >::tileSizes, and mlir::linalg::detail::ContinuousTileSizeSpecificationBase< T >::tripCounts.

◆ computeMultiTileSizes()

FailureOr< MultiSizeSpecification > mlir::linalg::computeMultiTileSizes	(	OpBuilder &	builder,
		LinalgOp	op,
		unsigned	dimension,
		OpFoldResult	targetSize,
		OpFoldResult	divisor,
		bool	emitAssertions = `true`
	)

Emits the IR computing the multi-sized tiling specification with two tile sizes not exceeding targetSize, each divisible by sizeDivisor, such that there exist numbers of tiles with these sizes that fully cover the given iteration space dimension of the structured op.

The computation is as follows:

b = originalTripCount floordiv sizeDivisor t = (targetSize + sizeDivisor - 1) floordiv sizeDivisor d = (b + t - 1) floordiv t s = (b floordiv d) * sizeDivisor v = b % d u = d - v

where the tile sizes are s and s + sizeDivisor, and the numbers of the corresponding tiles are u and v, respectively. Alternatively,

s * u + (s + sizeDivisor) * v == original size, where s mod sizeDivisor = 0.

Expects all values to be positive. In some cases with the target tile size sufficiently close to the dimension shape and non-unit divisor, it is impossible to compute such sizes. If emitAssertion is set, also emit the assertion that size computation succeeded.

Returns the specification consisting of both tile values and the number of tiles of each size.

Definition at line 268 of file Tiling.cpp.

References mlir::ImplicitLocOpBuilder::create(), emitIsPositiveIndexAssertion(), mlir::AffineExpr::floorDiv(), mlir::Builder::getAffineSymbolExpr(), mlir::ImplicitLocOpBuilder::getLoc(), mlir::Builder::getStringAttr(), mlir::getValueOrCreateConstantIndexOp(), mlir::linalg::detail::MultiSizeSpecificationBase< T >::lowTileSize, mlir::affine::makeComposedAffineApply(), and mlir::affine::makeComposedFoldedMultiResultAffineApply().

◆ computeSliceParameters()

SliceParameters mlir::linalg::computeSliceParameters	(	OpBuilder &	builder,
		Location	loc,
		Value	valueToTile,
		ArrayRef< OpFoldResult >	tileSizes,
		AffineMap	map,
		ArrayRef< OpFoldResult >	lbs,
		ArrayRef< OpFoldResult >	ubs,
		ArrayRef< OpFoldResult >	subShapeSizes,
		bool	omitPartialTileCheck
	)

Computes SliceParameters for a single valueToTile assuming that its user is being tiled with the given loop bounds lbs and ubs and the tile sizes tileSizes.

omitPartialTileCheck controls whether to omit the partial/boundary tile condition check in cases where we statically know that it is unnecessary.

Definition at line 567 of file Utils.cpp.

References mlir::Value::getType(), mlir::linalg::SliceParameters::offsets, mlir::linalg::SliceParameters::sizes, and mlir::linalg::SliceParameters::strides.

Referenced by computeAllSliceParameters(), and makeTiledShape().

◆ computeStaticContinuousTileSizes()

FailureOr< StaticContinuousTileSizeSpecification > mlir::linalg::computeStaticContinuousTileSizes	(	LinalgOp	op,
		unsigned	dimension,
		unsigned	targetSize
	)

Definition at line 111 of file Tiling.cpp.

References mlir::linalg::detail::ContinuousTileSizeSpecificationBase< T >::tileSizes, and mlir::linalg::detail::ContinuousTileSizeSpecificationBase< T >::tripCounts.

◆ computeStaticMultiTileSizes()

FailureOr< StaticMultiSizeSpecification > mlir::linalg::computeStaticMultiTileSizes	(	LinalgOp	op,
		unsigned	dimension,
		int64_t	targetSize,
		int64_t	divisor
	)

Definition at line 242 of file Tiling.cpp.

References mlir::linalg::detail::MultiSizeSpecificationBase< T >::highTileSize, mlir::linalg::detail::MultiSizeSpecificationBase< T >::highTripCount, mlir::linalg::detail::MultiSizeSpecificationBase< T >::lowTileSize, and mlir::linalg::detail::MultiSizeSpecificationBase< T >::lowTripCount.

◆ computeTileOffsets()

SmallVector< OpFoldResult > mlir::linalg::computeTileOffsets	(	OpBuilder &	b,
		Location	loc,
		ArrayRef< OpFoldResult >	ivs,
		ArrayRef< OpFoldResult >	tileSizes
	)

Computes tile offsets, given a list of loop ivs and tileSizes.

In case a tile size is zero (i.e., no tiling), the corresponding offset is also zero.

Definition at line 675 of file Utils.cpp.

References mlir::Builder::getIndexAttr(), isTiled(), and mlir::isZeroIndex().

Referenced by computeAllSliceParameters().

◆ computeTileSizes()

SmallVector< OpFoldResult > mlir::linalg::computeTileSizes	(	OpBuilder &	b,
		Location	loc,
		ArrayRef< OpFoldResult >	tileSizes,
		ArrayRef< OpFoldResult >	sizeBounds
	)

Computes tile sizes, given a list of tileSizes and dimension sizes (sizeBounds).

In case a tile size is zero (i.e., no tiling), the corresponding result size is the corresponding value from sizeBounds. Note: The returned tile sizes are closed intervals.

Definition at line 689 of file Utils.cpp.

References mlir::getAffineDimExpr(), mlir::Builder::getContext(), isTiled(), mlir::isZeroIndex(), and mlir::affine::makeComposedFoldedAffineApply().

Referenced by computeAllSliceParameters().

◆ concat()

SmallVector< AffineExpr, 4 > mlir::linalg::concat	(	ArrayRef< AffineExpr >	a,
		ArrayRef< AffineExpr >	b
	)

Return the vector that is the concatenation of a and b.

Definition at line 2299 of file LinalgOps.cpp.

Referenced by mlir::presburger::Simplex::makeProduct().

◆ copyToGPUPrivateMemory()

LogicalResult mlir::linalg::copyToGPUPrivateMemory	(	OpBuilder &	b,
		Value	src,
		Value	dst
	)

Normal copy to between src and dst.

Definition at line 503 of file Promotion.cpp.

References mlir::OpBuilder::create(), and mlir::Value::getLoc().

◆ copyToWorkgroupMemory()

LogicalResult mlir::linalg::copyToWorkgroupMemory	(	OpBuilder &	b,
		Value	src,
		Value	dst
	)

Create Memref copy operations and add gpu barrier guards before and after the copy operation to ensure data integrity.

Definition at line 486 of file Promotion.cpp.

References mlir::OpBuilder::create(), and mlir::Value::getLoc().

◆ createAdd()

static Value mlir::linalg::createAdd	(	Location	loc,
		Value	x,
		Value	y,
		OpBuilder &	builder
	)

static

Definition at line 32 of file ConvertConv2DToImg2Col.cpp.

References mlir::OpBuilder::create(), and mlir::Value::getType().

Referenced by rewriteInIm2Col().

◆ createAllReduceForResultsWithoutPartialShardings()

static void mlir::linalg::createAllReduceForResultsWithoutPartialShardings	(	LinalgOp	unshardedOp,
		ArrayRef< MeshAxis >	opReductionMeshAxes,
		ArrayRef< MeshSharding >	resultShardings,
		IRMapping &	spmdizationMap,
		ImplicitLocOpBuilder &	builder
	)

static

Definition at line 206 of file MeshShardingInterfaceImpl.cpp.

References createAllReduceForResultWithoutPartialSharding(), and getReductionKindOfLinalgOp().

Referenced by spmdizeLinalgOpWithShardedReduction().

◆ createAllReduceForResultWithoutPartialSharding()

static void mlir::linalg::createAllReduceForResultWithoutPartialSharding	(	Value	unshardedLinalgOpResult,
		ArrayRef< MeshAxis >	opReductionMeshAxes,
		MeshSharding	resultSharding,
		ReductionKind	reductionKind,
		IRMapping &	spmdizationMap,
		ImplicitLocOpBuilder &	builder
	)

static

Definition at line 185 of file MeshShardingInterfaceImpl.cpp.

References mlir::ImplicitLocOpBuilder::create(), mlir::mesh::MeshSharding::getMesh(), mlir::mesh::MeshSharding::getPartialAxes(), mlir::IRMapping::lookup(), and mlir::IRMapping::map().

Referenced by createAllReduceForResultsWithoutPartialShardings().

◆ createDestinationPassingStyleInitOperand()

static Value mlir::linalg::createDestinationPassingStyleInitOperand	(	LinalgOp	op,
		Value	spmdizedOperand,
		ArrayRef< MeshAxis >	reductionMeshAxes,
		MeshOp	meshOp,
		ImplicitLocOpBuilder &	builder
	)

static

Definition at line 131 of file MeshShardingInterfaceImpl.cpp.

References mlir::ImplicitLocOpBuilder::create(), mlir::mesh::createProcessLinearIndex(), mlir::Builder::getI1Type(), mlir::ImplicitLocOpBuilder::getLoc(), mlir::tensor::getMixedSizes(), mlir::Value::getType(), and mlir::OpBuilder::setInsertionPointToEnd().

Referenced by createDestinationPassingStyleInitOperands().

◆ createDestinationPassingStyleInitOperands()

static SmallVector<Value> mlir::linalg::createDestinationPassingStyleInitOperands	(	LinalgOp	op,
		MeshOp	meshOp,
		ArrayRef< Value >	spmdizedOperands,
		ArrayRef< MeshAxis >	reductionMeshAxes,
		IRMapping &	spmdizationMap,
		ImplicitLocOpBuilder &	builder
	)

static

Definition at line 169 of file MeshShardingInterfaceImpl.cpp.

References createDestinationPassingStyleInitOperand(), and mlir::IRMapping::lookup().

Referenced by spmdizeLinalgOpWithShardedReduction().

◆ createFoldedDimOp()

OpFoldResult mlir::linalg::createFoldedDimOp	(	OpBuilder &	b,
		Location	loc,
		Value	val,
		int64_t	dim
	)

Create one memref::DimOp or tensor::DimOp depending on the type of val.

This is a polymorphic convenience function to abstract away the rank and concrete type of val. Asserts that val is a memref or tensor type.

Definition at line 105 of file LinalgOps.cpp.

References createOrFoldDimOp(), mlir::Builder::getIndexAttr(), and mlir::Value::getType().

Referenced by fuse().

◆ createMul()

static Value mlir::linalg::createMul	(	Location	loc,
		Value	x,
		Value	y,
		Type	accType,
		OpBuilder &	builder
	)

static

Definition at line 40 of file ConvertConv2DToImg2Col.cpp.

References mlir::convertScalarToDtype(), and mlir::OpBuilder::create().

Referenced by rewriteInIm2Col().

◆ createOrFoldDimOp()

Value mlir::linalg::createOrFoldDimOp	(	OpBuilder &	b,
		Location	loc,
		Value	val,
		int64_t	dim
	)

Create one memref::DimOp or tensor::DimOp depending on the type of val.

This is a polymorphic convenience function to abstract away the rank and concrete type of val. Asserts that val is a memref or tensor type.

Definition at line 96 of file LinalgOps.cpp.

References mlir::OpBuilder::createOrFold(), and mlir::Value::getType().

Referenced by concatSizesFromInputs(), createFoldedDimOp(), and mlir::sparse_tensor::sizesFromSrc().

◆ deallocateGPUPrivateMemory()

LogicalResult mlir::linalg::deallocateGPUPrivateMemory	(	OpBuilder &	,
		Value
	)

In case of GPU private memory there is no need to deallocate since the memory is freed when going outside of the scope.

Definition at line 511 of file Promotion.cpp.

◆ deallocateWorkgroupMemory()

LogicalResult mlir::linalg::deallocateWorkgroupMemory	(	OpBuilder &	,
		Value
	)

In case of GPU group memory there is no need to deallocate.

Definition at line 479 of file Promotion.cpp.

◆ decomposeWinogradFilterTransformOp()

FailureOr< Operation * > mlir::linalg::decomposeWinogradFilterTransformOp	(	RewriterBase &	rewriter,
		linalg::WinogradFilterTransformOp	op
	)

Rewrite linalg.winograd_filter_transform.

The data layout of the filter is FHWC. The transformation matrix is 2-dimension. We need to extract H x W from FHWC first. We generate 2 levels of loops to iterate on F and C. After the rewriting, we get

scf.for f = lo_f to hi_f step 1 scf.for c = lo_c to hi_c step 1 extracted = extract filter<h x w> from filter<f x h x w x c> ret = linalg.matmul G, extracted ret = linalg.matmul ret, GT inserted = insert ret into filter<h x w x c x f>

Definition at line 1191 of file WinogradConv2D.cpp.

◆ decomposeWinogradInputTransformOp()

FailureOr< Operation * > mlir::linalg::decomposeWinogradInputTransformOp	(	RewriterBase &	rewriter,
		linalg::WinogradInputTransformOp	op
	)

Rewrite linalg.winograd_input_transform.

The data layout of the input is NHWC. The transformation matrix is 2-dimension. We need to extract H x W from NHWC first. We generate 4 levels of loops to iterate on N, C, tileH, and tileW. After the rewriting, we get

scf.for h = 0 to tileH step 1 scf.for w = 0 to tileW step 1 scf.for n = 0 to N step 1 scf.for c = 0 to C step 1 extracted = extract extracted<alphaH x alphaW> from input<N x H x W x C> at [n, (h x m), (w x m), c] ret = linalg.matmul BT, extracted ret = linalg.matmul ret, B inserted = insert ret<alphaH x alphaW> into output<alphaH x alphaW x tileH x tileW x N x C> at [0, 0, h, w, n, c]

Definition at line 1197 of file WinogradConv2D.cpp.

◆ decomposeWinogradOutputTransformOp()

FailureOr< Operation * > mlir::linalg::decomposeWinogradOutputTransformOp	(	RewriterBase &	rewriter,
		linalg::WinogradOutputTransformOp	op
	)

Rewrite linalg.winograd_output_transform.

The data layout of the output is HWNF. The transformation matrix is 2-dimension. We need to extract H x W from HWNF first. We generate 4 levels of loops to iterate on N, F, tileH, and tileW. After the transformation, we get

scf.for h = 0 to tileH step 1 scf.for w = 0 to tileW step 1 scf.for n = 0 to N step 1 scf.for f = 0 to F step 1 extracted = extract extracted<alphaH x alphaW> from input<alphaH x alphaW x tileH x tileW x N x F> at [0, 0, h, w, n, f] ret = linalg.matmul AT, extracted ret = linalg.matmul ret, A inserted = insert ret<alphaH x alphaW> into output<N x H x W x F> at [n, (h x m), (w x m), f]

Definition at line 1203 of file WinogradConv2D.cpp.

◆ dropUnitDims()

FailureOr< DropUnitDimsResult > mlir::linalg::dropUnitDims	(	RewriterBase &	rewriter,
		GenericOp	genericOp,
		const ControlDropUnitDims &	options
	)

Definition at line 386 of file DropUnitDims.cpp.

References collapseValue(), mlir::concatAffineMaps(), mlir::OpBuilder::create(), dropUnitExtentFromOperandMetadata(), mlir::detail::enumerate(), expandValue(), mlir::getAffineConstantExpr(), mlir::getAffineDimExpr(), mlir::Builder::getContext(), mlir::AffineMap::getResults(), mlir::getType(), mlir::RewriterBase::inlineRegionBefore(), mlir::inversePermutation(), mlir::RewriterBase::notifyMatchFailure(), options, mlir::AffineMap::replaceDimsAndSymbols(), and replaceUnitDimIndexOps().

◆ extractOrIdentityMap()

AffineMap mlir::linalg::extractOrIdentityMap	(	std::optional< AffineMap >	maybeMap,
		unsigned	rank,
		MLIRContext *	context
	)

Returns maybeMap.get() if maybeMap is set, otherwise returns the symbol-less identity map of rank.

Definition at line 2279 of file LinalgOps.cpp.

References mlir::AffineMap::get(), and mlir::AffineMap::getMultiDimIdentityMap().

◆ fuseElementwiseOps()

FailureOr< mlir::linalg::ElementwiseOpFusionResult > mlir::linalg::fuseElementwiseOps	(	RewriterBase &	rewriter,
		OpOperand *	fusedOperand
	)

Find the results of the producer that have uses outside of the consumer, after the fusion.

Definition at line 337 of file ElementwiseOpFusion.cpp.

References areElementwiseOpsFusable(), mlir::AffineMap::compose(), mlir::OpBuilder::create(), mlir::detail::enumerate(), mlir::RewriterBase::eraseOp(), mlir::linalg::ElementwiseOpFusionResult::fusedOp, generateFusedElementwiseOpRegion(), mlir::IROperand< DerivedT, IRValueT >::get(), mlir::Builder::getAffineMapArrayAttr(), getIndexingMapOfProducerOperandsInCoordinatesOfFusedOp(), mlir::detail::IROperandBase::getOwner(), getPreservedProducerResults(), mlir::inversePermutation(), mlir::RewriterBase::notifyMatchFailure(), and mlir::linalg::ElementwiseOpFusionResult::replacements.

◆ fuseProducerOfTensor() [1/2]

FailureOr< FusionInfo > mlir::linalg::fuseProducerOfTensor	(	OpBuilder &	b,
		OpOperand &	consumerOpOperand
	)

This implements the fusion part of the "tileAndFuse on tensors" transformation and thus requires the consumerOpOperand to be a extract_slice op (generally obtained by applying the tiling transformation).

Definition at line 227 of file Fusion.cpp.

References mlir::IROperand< DerivedT, IRValueT >::get(), and getProducerOfTensor().

◆ fuseProducerOfTensor() [2/2]

FailureOr< FusionInfo > mlir::linalg::fuseProducerOfTensor	(	OpBuilder &	b,
		OpResult	producerOpResult,
		OpOperand &	consumerOpOperand
	)

This implements the fusion part of the "tileAndFuse on tensors" transformation and thus requires the consumerOpOperand to be a extract_slice op (generally obtained by applying the tiling transformation).

Assumes producerOfTensor is a Linalg op that produces consumerOpOperand.

Definition at line 239 of file Fusion.cpp.

References mlir::OpBuilder::create(), fuse(), mlir::IROperand< DerivedT, IRValueT >::get(), mlir::Value::getDefiningOp(), mlir::detail::IROperandBase::getOwner(), mlir::OpResult::getOwner(), mlir::Value::getParentBlock(), mlir::OpResult::getResultNumber(), mlir::Value::getType(), mlir::IROperand< DerivedT, IRValueT >::set(), and mlir::OpBuilder::setInsertionPoint().

◆ generalizeNamedOp()

FailureOr< GenericOp > mlir::linalg::generalizeNamedOp	(	RewriterBase &	rewriter,
		LinalgOp	namedOp
	)

Create a GenericOp from the given named operation namedOp and replace namedOp.

Return failure if namedOp is a GenericOp or misses a region builder.

Definition at line 53 of file Generalization.cpp.

References mlir::OpBuilder::create(), generalizeNamedOpPrecondition(), mlir::RewriterBase::inlineRegionBefore(), mlir::RewriterBase::notifyMatchFailure(), and mlir::RewriterBase::replaceOp().

Referenced by packMatmulGreedily(), and mlir::linalg::LinalgGeneralizationPattern::returningMatchAndRewrite().

◆ generateLibraryCallName()

std::string mlir::linalg::generateLibraryCallName ( Operation * op )

Returns the name mangled library call name to disambiguate between different overloads at the C level.

The name mangling scheme is basic and uses MLIR type names:

form a string which is the concatenation of the linalg op name with all the operand type names, separate by underscores;
drop the linalg. prefix, and the <, >, ? symbols from the type. Assumes op is a LinalgOp.

Examples:

linalg.fill(f, A) : f32, memref<f32> name mangles into linalg_fill_f32_viewf32
linalg.dot A, B, C : (memref<?xf32, stride_specification>, memref<?xf32, stride_specification>, memref<f32>) name mangles into linalg_dot_viewxf32_viewxf32_viewf32
linalg.matmul(...) : memref<?x?xf32, stride_specification>, memref<?x?xf32, stride_specification>, memref<?x?xf32, stride_specification> name mangles into linalg_matmul_viewxxf32_viewxxf32_viewxxf32

Definition at line 2340 of file LinalgOps.cpp.

References appendMangledType(), mlir::Operation::getAttrs(), mlir::Operation::getName(), mlir::Operation::getOperandTypes(), and mlir::OperationName::getStringRef().

◆ generateParallelLoopNest()

static void mlir::linalg::generateParallelLoopNest	(	OpBuilder &	b,
		Location	loc,
		ValueRange	lbs,
		ValueRange	ubs,
		ValueRange	steps,
		ArrayRef< utils::IteratorType >	iteratorTypes,
		ArrayRef< linalg::ProcInfo >	procInfo,
		function_ref< void(OpBuilder &, Location, ValueRange)>	bodyBuilderFn,
		SmallVectorImpl< Value > &	ivStorage
	)

static

Generates a loop nest consisting of scf.parallel and scf.for, depending on the iteratorTypes.

Consecutive parallel loops create a single scf.parallel operation; each sequential loop creates a new scf.for operation. The body of the innermost loop is populated by bodyBuilderFn that accepts a range of induction variables for all loops. ivStorage is used to store the partial list of induction variables.

Definition at line 372 of file Utils.cpp.

References mlir::ArithBuilder::_and(), mlir::scf::buildLoopNest(), mlir::OpBuilder::create(), mlir::linalg::ProcInfo::distributionMethod, isParallelIterator(), None, and mlir::ArithBuilder::slt().

Referenced by mlir::linalg::GenerateLoopNest< LoopTy >::doit().

◆ getCombinerOp()

static std::optional<Operation *> mlir::linalg::getCombinerOp ( LinalgOp op )

static

Definition at line 81 of file MeshShardingInterfaceImpl.cpp.

References mlir::matchReduction().

Referenced by getReductionKindOfLinalgOp().

◆ getCombinerOpKind()

std::optional< vector::CombiningKind > mlir::linalg::getCombinerOpKind ( Operation * combinerOp )

Return vector::CombiningKind for the given op.

Definition at line 544 of file Vectorization.cpp.

References MINUI.

Referenced by buildMultiDimReduce(), and reductionPreconditions().

◆ getConvolvedIndex()

static Value mlir::linalg::getConvolvedIndex	(	OpBuilder &	b,
		Location	loc,
		Value	oIndex,
		Value	fIndex,
		int64_t	stride
	)

static

Definition at line 70 of file ConvertConv2DToImg2Col.cpp.

References mlir::bindSymbols(), mlir::AffineMap::get(), mlir::Builder::getContext(), and mlir::affine::makeComposedAffineApply().

Referenced by rewriteInIm2Col().

◆ getLinalgTilingCanonicalizationPatterns()

RewritePatternSet mlir::linalg::getLinalgTilingCanonicalizationPatterns ( MLIRContext * ctx )

Canonicalization patterns relevant to apply after tiling patterns.

These are applied automatically by the tiling pass but need to be applied manually when tiling is called programmatically.

Definition at line 854 of file Tiling.cpp.

References mlir::patterns, and populateLinalgTilingCanonicalizationPatterns().

◆ getMesh()

static MeshOp mlir::linalg::getMesh	(	Operation *	op,
		ArrayRef< MeshSharding >	operandShardings,
		ArrayRef< MeshSharding >	resultShardings,
		SymbolTableCollection &	symbolTable
	)

static

Definition at line 105 of file MeshShardingInterfaceImpl.cpp.

References mlir::mesh::getMesh().

Referenced by spmdizeLinalgOpWithShardedReduction().

◆ getPreservedProducerResults()

llvm::SmallDenseSet< int > mlir::linalg::getPreservedProducerResults	(	GenericOp	producer,
		GenericOp	consumer,
		OpOperand *	fusedOperand
	)

Returns a set of indices of the producer's results which would be preserved after the fusion.

There is a chance that the implementation of the transformation does not agree with the result of this method. This function gives a prediction based on an optimized fusion.

Definition at line 110 of file ElementwiseOpFusion.cpp.

References mlir::detail::enumerate(), and isOpOperandCanBeDroppedAfterFusedLinalgs().

Referenced by fuseElementwiseOps().

◆ getPrunedAttributeList()

template<typename OpTy >

SmallVector<NamedAttribute> mlir::linalg::getPrunedAttributeList ( OpTy op )

Returns an attribute list that excludes pre-defined attributes.

Definition at line 364 of file Utils.h.

Referenced by matchAndReplaceDepthwiseConv().

◆ getReassociationMapForFoldingUnitDims()

std::optional< SmallVector< ReassociationIndices > > mlir::linalg::getReassociationMapForFoldingUnitDims ( ArrayRef< OpFoldResult > mixedSizes )

Get the reassociation maps to fold the result of a extract_slice (or source of a insert_slice) operation with given offsets, and sizes to its rank-reduced version.

This is only done for the cases where the size is 1 and offset is 0. Strictly speaking the offset 0 is not required in general, but non-zero offsets are not handled by SPIR-V backend at this point (and potentially cannot be handled).

Definition at line 852 of file Utils.cpp.

References mlir::detail::enumerate().

◆ getReductionKind()

static ReductionKind mlir::linalg::getReductionKind ( Operation * op )

static

Definition at line 51 of file MeshShardingInterfaceImpl.cpp.

Referenced by getReductionKindOfLinalgOp().

◆ getReductionKindOfLinalgOp()

static ReductionKind mlir::linalg::getReductionKindOfLinalgOp ( LinalgOp op )

static

Definition at line 91 of file MeshShardingInterfaceImpl.cpp.

References getCombinerOp(), and getReductionKind().

Referenced by createAllReduceForResultsWithoutPartialShardings().

◆ getTensorOutputTypes()

SmallVector< Type > mlir::linalg::getTensorOutputTypes	(	LinalgOp	op,
		ValueRange	operands
	)

Returns the list of tensor output types produced when the given structured operation op is applied to the given operands.

Note that operands are not necessarily the actual operands of op.

Definition at line 705 of file Utils.cpp.

Referenced by tileLinalgOpImpl().

◆ hasAllOneValues()

static bool mlir::linalg::hasAllOneValues ( DenseIntElementsAttr attr )

static

Definition at line 27 of file ConvertConv2DToImg2Col.cpp.

Referenced by rewriteInIm2Col().

◆ hasOnlyScalarElementwiseOp()

bool mlir::linalg::hasOnlyScalarElementwiseOp ( Region & r )

Detect whether r has only ConstantOp, ElementwiseMappable and YieldOp.

Definition at line 155 of file Utils.cpp.

Referenced by isElementwise().

◆ hasVectorizationImpl()

bool mlir::linalg::hasVectorizationImpl ( Operation * op )

Return true if there's dedicated logic in the Linalg Vectorizer to vectorize this Op, false otherwise.

Note that this helper merely implements a very high level check and that the vectorizer also requires various additional pre-conditions to be met for it to work (these are checked by the vectorizer itself).

Definition at line 2165 of file Vectorization.cpp.

Referenced by vectorizeOpPrecondition().

◆ hoistPaddingOnTensors() [1/2]

FailureOr< Value > mlir::linalg::hoistPaddingOnTensors	(	RewriterBase &	rewriter,
		tensor::PadOp	opToHoist,
		int64_t	numLoops,
		ArrayRef< int64_t >	transposeVector,
		tensor::PadOp &	hoistedOp,
		SmallVectorImpl< TransposeOp > &	transposeOps
	)

Mechanically hoist padding operations on tensors by numLoops into a new, generally larger tensor.

This achieves packing of multiple padding ops into a larger tensor. On success, opToHoist is replaced by the cloned version in the packing loop so the caller can continue reasoning about the padding operation. If transposeVector is non-empty, hoist padding introduces a TransposeOp to transpose the padded tensor before inserting it into the packed tensor. A transposeVector can change the storage order of the padded tensor but does not change the order of the pack or compute loops.

TODO: In the future, we should consider rewriting as a tensor.pack after hoisting since this abstraction is now available.

Example in pseudo-mlir:

If hoistPaddingOnTensors is called with nLoops = 2 on the following IR.

scf.for (%i, %j, %k)
  %st0 = tensor.extract_slice f(%i, %k) : ... to tensor<?x?xf32>
  %0 = tensor.pad %st0 low[0, 0] high[...] {
  ^bb0( ... ):
    linalg.yield %pad
  } : tensor<?x?xf32> to tensor<4x8xf32>
  compute(%0)

IR resembling the following is produced:

scf.for (%i) {
  %packed_init = tensor.empty range(%j) : tensor<?x4x8xf32>
  %packed = scf.for (%k) iter_args(%p : %packed_init) {
    %st0 = tensor.extract_slice f(%i, %k) : ... to tensor<?x?xf32>
    %0 = tensor.pad %st0 low[0, 0] high[...] {
    ^bb0( ... ):
      linalg.yield %pad
    } : tensor<?x?xf32> to tensor<4x8xf32>
    %1 = tensor.insert_slice %0 ...
        : tensor<4x8xf32> to tensor<?x4x8xf32>
    scf.yield %1: tensor<?x4x8xf32>
  } -> tensor<?x4x8xf32>
  scf.for (%j, %k) {
    %st0 = tensor.extract_slice %packed [%k, 0, 0][1, 4, 8][1, 1, 1] :
             tensor<?x4x8xf32> to tensor<4x8xf32>
    compute(%st0)
  }
}

Construct the packing loop nest.

Definition at line 938 of file HoistPadding.cpp.

References buildPackingLoopNestImpl(), mlir::tensor::computeTransposedType(), mlir::OpBuilder::create(), DBGS, mlir::Value::getDefiningOp(), mlir::Operation::getParentOfType(), mlir::Operation::getResult(), replaceByPackingResult(), and mlir::OpBuilder::setInsertionPointAfter().

Referenced by hoistPaddingOnTensors(), and padAndHoistLinalgOp().

◆ hoistPaddingOnTensors() [2/2]

FailureOr< Value > mlir::linalg::hoistPaddingOnTensors	(	tensor::PadOp	opToHoist,
		int64_t	numLoops,
		ArrayRef< int64_t >	transposeVector,
		tensor::PadOp &	hoistedOp,
		SmallVectorImpl< TransposeOp > &	transposeOps
	)

Calls into hoistPaddingOnTensors with a local IRRewriter.

Definition at line 1002 of file HoistPadding.cpp.

References hoistPaddingOnTensors().

◆ hoistRedundantVectorBroadcasts()

void mlir::linalg::hoistRedundantVectorBroadcasts	(	RewriterBase &	rewriter,
		Operation *	root
	)

Hoist vector.extract/vector.broadcast pairs out of immediately enclosing scf::ForOp iteratively, if the following conditions are met:

The vector.extract operation is applied on an iter_argument, and no other operator is using this argument in the body of the loop.
The position of the vector.extract is either a static value, or defined outside of the loop.
The vector.broadcast operation is yielded by the loop. To improve hoisting opportunities, call the moveLoopInvariantCode helper function on the candidate loop above which to hoist.

Definition at line 97 of file Hoisting.cpp.

References mlir::WalkResult::advance(), broadcast(), mlir::changed, DBGS, mlir::IROperand< DerivedT, IRValueT >::get(), mlir::Value::hasOneUse(), mlir::WalkResult::interrupt(), mlir::RewriterBase::modifyOpInPlace(), mlir::moveLoopInvariantCode(), mlir::RewriterBase::moveOpAfter(), mlir::RewriterBase::replaceAllUsesWith(), replaceWithDifferentYield(), and mlir::Operation::walk().

◆ hoistRedundantVectorTransfers()

void mlir::linalg::hoistRedundantVectorTransfers	(	Operation *	root,
		bool	verifyNonZeroTrip = `false`
	)

Hoist vector.transfer_read/vector.transfer_write on buffers pairs out of immediately enclosing scf::ForOp iteratively, if the following conditions are true:

The two ops access the same memref with the same indices.
All operands are invariant under the enclosing scf::ForOp.
No uses of the memref either dominate the transfer_read or are dominated by the transfer_write (i.e. no aliasing between the write and the read across the loop)
The source operands for vector.transfer_{read|write} do not originate from Ops implementing ViewLikeOpInterface (to reduce the risk of aliasing).
If verifyNonZeroTrip is true, then the lower bound of the loop must be statically smaller than the upper bound of the loop, guaranteeing that the loop body will execute at least once. To improve hoisting opportunities, call the moveLoopInvariantCode helper function on the candidate loop above which to hoist. Hoisting the transfers results in scf::ForOp yielding the value that originally transited through memory.

TODO: To further improve hoisting opportunities, fold aliasing memref operations into respective vector.transfer{read|write} operations and avoid using ops implementing ViewLikeOpInterface as the source for transfer Ops.

WARNING: This hoisting does not model parallelism and is generally incorrect when used on distributed loops with memref semantics! NOTE: Setting verifyNonZeroTrip = true makes this more stable for distributed loops with memref semantics, but there could still be some issues when loops are executed a different number of times for different threads.

Definition at line 202 of file Hoisting.cpp.

References mlir::WalkResult::advance(), mlir::changed, mlir::ValueBoundsConstraintSet::computeConstantBound(), DBGS, mlir::getForwardSlice(), mlir::WalkResult::interrupt(), mlir::vector::isDisjointTransferSet(), mlir::presburger::LB, mlir::moveLoopInvariantCode(), noAliasingUseInLoop(), mlir::DominanceInfo::properlyDominates(), mlir::presburger::UB, and mlir::Operation::walk().

◆ inferContractionDims() [1/2]

FailureOr< ContractionDimensions > mlir::linalg::inferContractionDims ( ArrayRef< AffineMap > indexingMaps )

Definition at line 457 of file LinalgInterfaces.cpp.

References inferContractionDimsImpl(), and inferIteratorsFromOutMap().

◆ inferContractionDims() [2/2]

FailureOr< ContractionDimensions > mlir::linalg::inferContractionDims ( LinalgOp linalgOp )

Find at least 2 parallel (m and n) and 1 reduction (k) dimension candidates that form a matmul subcomputation within linalgOp.

These dimensions are such that:

The m dimension is involved in an outer-product along LHS (i.e. it is a permutation on RES and LHS and does not appear in RHS).
The n dimension is involved in an outer-product along RHS (i.e. it is a permutation on RES and RHS and does not appear in LHS).
The k dimension appears as a permutation on LHS and RHS.
m, n and k appear only once in any given indexing.
Optional batch dimensions that appear in all operands are captured. This allows e.g. detecting that some contraction is embedded within linalgOp with some orthogonal heuristic. When multiple dimension occurrences exist that match batch, m, n, or k, indices are returned in sorted order. Returns a failure if any of m, n or k is empty.

Definition at line 449 of file LinalgInterfaces.cpp.

References inferContractionDimsImpl().

Referenced by blockPackMatmul(), packMatmulGreedily(), and validateFullTilesOnDims().

◆ inferConvolutionDims()

FailureOr< ConvolutionDimensions > mlir::linalg::inferConvolutionDims ( LinalgOp linalgOp )

Find at least 1 parallel (output_image) and reduction (filter_loop) dimension candidates that form a convolution subcomputation within linalgOp.

The LHS is assumed to be the convolution input while the RHS is assumed as the filter. These dimensions are such that:

Optional batch dimensions that appear in the input and filter.
The output_image dimension is involved in a cross-correlation along LHS (i.e. it is a permutation on RES and LHS and has an associated filter_loop in RHS).
Optional output_channel dimension is involved in an outer-product along RHS (i.e. it is a permutation on RES and RHS and does not appear in LHS).
Optional input_channel dimension appears as a permutation on LHS and RHS.
The filter_loop dimension appears as a permutation on the RHS and represents the shape of the kernel cross-correlated along a corresponding output_image dim.
The input_channel dimension appears as a permutation on LHS and RHS.
All dimensions appear only once in any given indexing map. This allows e.g. detecting that some convolution is embedded within linalgOp with some orthogonal heuristic. When multiple dimension occurrences exist that match any classification indices are returned in sorted order. Returns a failure if output_image (and implicitly filter_loop) is empty.

Definition at line 818 of file LinalgInterfaces.cpp.

References inferConvolutionDimsImpl().

◆ insertSlicesBack()

SmallVector< Value > mlir::linalg::insertSlicesBack	(	OpBuilder &	builder,
		Location	loc,
		LinalgOp	op,
		ValueRange	operands,
		ValueRange	results
	)

Creates insert_slice ops that insert results back into larger tensors they were originally extracted from with extract_slice before being passed as operands to the given structured operation op or its clone.

Note that operands are not necessarily the actual operands of op, the operation serves only as metadata container for operand types and positions.

Definition at line 714 of file Utils.cpp.

References mlir::OpBuilder::create(), and mlir::Value::getDefiningOp().

Referenced by tileLinalgOpImpl().

◆ interchangeGenericOp()

FailureOr< GenericOp > mlir::linalg::interchangeGenericOp	(	RewriterBase &	rewriter,
		GenericOp	genericOp,
		ArrayRef< unsigned >	interchangeVector
	)

Interchange the iterator_types and iterator_maps dimensions and adapts the index accesses of op.

This is an in-place transformation controlled by interchangeVector. An empty vector is interpreted as the identity permutation and the transformation returns early.

E.g. the permutation (i,j,k) -> (j,k,i) is expressed with interchangeVector = [1,2,0]. All values in interchangeVector must be integers, in the range 0..op.rank without duplications (i.e. [1,1,2] is an invalid permutation).

Return failure if the permutation is not valid.

Definition at line 50 of file Interchange.cpp.

References mlir::RewriterBase::finalizeOpModification(), mlir::AffineMap::getPermutationMap(), interchangeGenericOpPrecondition(), mlir::inversePermutation(), mlir::RewriterBase::notifyMatchFailure(), and mlir::RewriterBase::startOpModification().

Referenced by packMatmulGreedily().

◆ isaBroadcastOpInterface()

std::optional< SmallVector< int64_t > > mlir::linalg::isaBroadcastOpInterface ( GenericOp genericOp )

Checks whether genericOp is semantically equivalent to a linalg.broadcast.

Returns broadcast dimensions if true.

Definition at line 103 of file LinalgInterfaces.cpp.

Referenced by specializeGenericOp().

◆ isaContractionOpInterface()

bool mlir::linalg::isaContractionOpInterface ( LinalgOp linalgOp )

Checks whether linalgOp conforms to ContractionOpInterface.

Definition at line 530 of file LinalgInterfaces.cpp.

References mlir::linalg::detail::isContractionInterfaceImpl(), and mlir::linalg::detail::Success.

Referenced by FoldAddIntoDest::matchAndRewrite(), and specializeGenericOp().

◆ isaConvolutionOpInterface()

bool mlir::linalg::isaConvolutionOpInterface	(	LinalgOp	linalgOp,
		bool	allowEmptyConvolvedDims = `false`
	)

Checks whether linalgOp conforms to ConvolutionOpInterface.

By default, we require the linalgOp to have non-empty convolved dims (implicitly non-empty output_image and filter_loop). Users can loosen the constraint by setting allowEmptyConvolvedDims to true

Definition at line 1007 of file LinalgInterfaces.cpp.

References mlir::linalg::detail::isConvolutionInterfaceImpl(), and mlir::linalg::detail::Success.

◆ isaCopyOpInterface()

bool mlir::linalg::isaCopyOpInterface ( LinalgOp linalgOp )

Checks whether linalgOp is semantically equivalent to a linalg.copyOp.

Definition at line 65 of file LinalgInterfaces.cpp.

Referenced by specializeGenericOp().

◆ isaElemwiseSingleBinaryOpInterface()

bool mlir::linalg::isaElemwiseSingleBinaryOpInterface ( GenericOp genericOp )

Checks whether genericOp is semantically equivalent to a single linalg elementwise binary op e.g.

linalg.sub.

Referenced by specializeGenericOp().

◆ isaElemwiseSingleUnaryOpInterface()

bool mlir::linalg::isaElemwiseSingleUnaryOpInterface ( GenericOp genericOp )

Checks whether a given genericOp is semantically equivalent to a single linalgelementwise unary op.

e.g. linalg.exp. A linalg.generic body could be a series of unary elementwise ops e.g. exp(neg(x)), such as formed by linalg op fusion. Here we restrict it to detecting cases where body is is a single computation op.

Referenced by specializeGenericOp().

◆ isaFillOpInterface()

std::optional< Value > mlir::linalg::isaFillOpInterface ( GenericOp genericOp )

Checks whether genericOp is semantically equivalent to a linalg.fill.

Returns the scalar fill value if true.

Definition at line 82 of file LinalgInterfaces.cpp.

References mlir::IROperand< DerivedT, IRValueT >::get().

Referenced by specializeGenericOp().

◆ isaTransposeOpInterface()

std::optional< SmallVector< int64_t > > mlir::linalg::isaTransposeOpInterface ( GenericOp genericOp )

Checks whether genericOp is semantically equivalent to a linalg.transpose.

Returns permuted dimensions if true.

Definition at line 153 of file LinalgInterfaces.cpp.

Referenced by specializeGenericOp().

◆ isDimSequencePreserved()

bool mlir::linalg::isDimSequencePreserved	(	AffineMap	indexingMap,
		ReassociationIndicesRef	dimSequence
	)

Return true if a given sequence of dimensions are contiguous in the range of the specified indexing map.

For a given dimSequence, check if the sequence is conserved in the indexingMap.

indexingMap is expected to be a projected permutation. Non-existence of the sequence returns true as well.

Definition at line 1163 of file ElementwiseOpFusion.cpp.

References mlir::detail::enumerate(), mlir::AffineMap::getNumResults(), mlir::AffineMap::getResult(), mlir::AffineMap::getResults(), and mlir::AffineMap::isProjectedPermutation().

Referenced by areDimSequencesPreserved().

◆ isElementwise()

bool mlir::linalg::isElementwise ( LinalgOp op )

Check if a LinalgOp is an element-wise operation.

Definition at line 169 of file Utils.cpp.

References allIndexingsAreProjectedPermutation(), and hasOnlyScalarElementwiseOp().

Referenced by vectorizeDynamicLinalgOpPrecondition(), vectorizeLinalgOpPrecondition(), and vectorizeScalableVectorPrecondition().

◆ isParallelIterator()

bool mlir::linalg::isParallelIterator ( utils::IteratorType iteratorType )

Check if iterator type has "parallel" semantics.

Definition at line 184 of file Utils.cpp.

Referenced by generateParallelLoopNest(), getCollapsableIterationSpaceDims(), and tileLinalgOpImpl().

◆ isReductionIterator()

bool mlir::linalg::isReductionIterator ( utils::IteratorType iteratorType )

Check if iterator type has "reduction" semantics.

Definition at line 188 of file Utils.cpp.

Referenced by getCollapsableIterationSpaceDims(), getDimsToReduce(), hasReductionIterator(), mlir::sparse_tensor::CodegenEnv::isAdmissibleTensorExp(), and reductionPreconditions().

◆ linalgOpAnchoredEmptyTensorEliminationStep()

LogicalResult mlir::linalg::linalgOpAnchoredEmptyTensorEliminationStep	(	RewriterBase &	rewriter,
		Operation *	op,
		bufferization::OneShotAnalysisState &	state
	)

Try to eliminate tensor::EmptyOps inside op that are anchored on a LinalgOp.

This transforms looks for LinalgOps that have an unused output operand and an input operand that is rooted in a tensor::EmptyOp. The tensor::EmptyOp uses are replaced with the output operand and the two operands of the LinalgOp are swapped.

Example: %0 = tensor.empty() %1 = linalg.matmul ins(...) outs(%0) %2 = linalg.generic ins(%1) outs(dest) { ^bb0(in: f32, out: f32): // out not used }

The IR is transformed as follows: %0 = tensor.empty() %1 = linalg.matmul ins(...) outs(dest) %2 = linalg.generic ins(%0) outs(%1) { ^bb0(in: f32, out: f32): // Use out instead of in }

The "ins" operand has no uses inside the body of the LinalgOp and can be folded away with existing cleanup patterns. Afterwards, the tensor::EmptyOp can also fold away.

Definition at line 40 of file EliminateEmptyTensors.cpp.

◆ linalgOpToAffineLoops()

FailureOr< LinalgLoops > mlir::linalg::linalgOpToAffineLoops	(	RewriterBase &	rewriter,
		LinalgOp	linalgOp
	)

Emit a loop nest of affine.for with the proper body for linalgOp.

Emits a loop nest of affine.for with the proper body for linalgOp.

Definition at line 363 of file Loops.cpp.

◆ linalgOpToLoops()

FailureOr< LinalgLoops > mlir::linalg::linalgOpToLoops	(	RewriterBase &	rewriter,
		LinalgOp	linalgOp
	)

Emit a loop nest of scf.for with the proper body for linalgOp.

Emits a loop nest of scf.for with the proper body for linalgOp.

Definition at line 368 of file Loops.cpp.

◆ linalgOpToParallelLoops()

FailureOr< LinalgLoops > mlir::linalg::linalgOpToParallelLoops	(	RewriterBase &	rewriter,
		LinalgOp	linalgOp
	)

Emit a loop nest of scf.parallel with the proper body for linalgOp.

Emits a loop nest of scf.parallel with the proper body for linalgOp.

Definition at line 375 of file Loops.cpp.

◆ lowerPack()

FailureOr< LowerPackResult > mlir::linalg::lowerPack	(	RewriterBase &	rewriter,
		tensor::PackOp	packOp,
		bool	lowerPadLikeWithInsertSlice = `true`
	)

◆ lowerUnPack()

FailureOr< LowerUnPackOpResult > mlir::linalg::lowerUnPack	(	RewriterBase &	rewriter,
		tensor::UnPackOp	unPackOp,
		bool	lowerUnpadLikeWithExtractSlice = `true`
	)

Rewrite pack as empty + transpose + reshape + extract_slice.

Definition at line 356 of file Transforms.cpp.

References mlir::applyPermutationToVector(), mlir::OpBuilder::create(), DBGS, DBGSNL, mlir::Builder::getIndexAttr(), mlir::tensor::getMixedSizes(), mlir::tensor::getUnPackInverseSrcPerm(), mlir::RewriterBase::replaceOp(), mlir::OpBuilder::setInsertionPoint(), and mlir::RankedTensorType::Builder::setShape().

◆ makeAffineDimExprs()

SmallVector< AffineExpr, 4 > mlir::linalg::makeAffineDimExprs	(	unsigned	num,
		unsigned &	startIdx,
		MLIRContext *	context
	)

Returns num AffineDimExpr dimensions at positions [startIdx, startIdx + num) and increments startIdx to startIdx + num.

Definition at line 2290 of file LinalgOps.cpp.

References mlir::getAffineDimExpr().

◆ makeComposedPadHighOp()

Value mlir::linalg::makeComposedPadHighOp	(	OpBuilder &	b,
		Location	loc,
		RankedTensorType	type,
		Value	source,
		Value	pad,
		bool	nofold
	)

Create a tensor::PadOp that pads source to the size of the statically sized type whose static sizes are assumed to be greater than the dynamic source size.

The padding introduces trailing pad values until the target size is met. If source is defined by one or more LinalgOps that have been padded with the same value and sizes, return their padded result instead of creating a tensor::PadOp.

Example:

%0 = tensor.extract_slice %arg0 [%iv0, %iv1] [%sz0, %sz1]
%1 = tensor.pad %0 low[0, 0] high[...] { tensor.yield %cst }
%2 = linalg.matmul ins(...) outs(%1)
%3 = tensor.extract_slice %2 [0, 0] [%sz0, %sz1]

makeComposedPadHighOp(source=%3, pad=cst) returns %2 makeComposedPadHighOp(source=%3, pad=other_cst) returns %4

%4 = tensor.pad %3 low[0, 0] high[...] { tensor.yield %other_cst }

Definition at line 192 of file Utils.cpp.

References mlir::tensor::createPadHighOp(), mlir::Value::getDefiningOp(), mlir::OpResult::getResultNumber(), mlir::m_Constant(), and mlir::matchPattern().

Referenced by padOperandToSmallestStaticBoundingBox().

◆ makeMemRefCopyOp()

GenericOp mlir::linalg::makeMemRefCopyOp	(	OpBuilder &	b,
		Location	loc,
		Value	from,
		Value	to
	)

Returns GenericOp that copies an n-D memref.

Unlike the current implementation of memref::CopyOp, this op can further tile, lower to loops or vectorize.

Definition at line 252 of file Utils.cpp.

References mlir::OpBuilder::create(), mlir::Builder::getContext(), mlir::AffineMap::getMultiDimIdentityMap(), and mlir::Value::getType().

◆ makeTiledLoopRanges()

std::tuple< SmallVector< Range, 4 >, LoopIndexToRangeIndexMap > mlir::linalg::makeTiledLoopRanges	(	RewriterBase &	b,
		Location	loc,
		AffineMap	map,
		ArrayRef< OpFoldResult >	allShapeSizes,
		ArrayRef< OpFoldResult >	allTileSizes
	)

Definition at line 49 of file Tiling.cpp.

References mlir::getConstantIntValue(), mlir::Builder::getIndexAttr(), mlir::AffineMap::getNumResults(), and mlir::affine::makeComposedFoldedMultiResultAffineApply().

Referenced by tileLinalgOpImpl().

◆ makeTiledShape()

Operation * mlir::linalg::makeTiledShape	(	OpBuilder &	builder,
		Location	loc,
		Value	valueToTile,
		ArrayRef< OpFoldResult >	tileSizes,
		AffineMap	map,
		ArrayRef< OpFoldResult >	lbs,
		ArrayRef< OpFoldResult >	ubs,
		ArrayRef< OpFoldResult >	subShapeSizes,
		bool	omitPartialTileCheck
	)

Creates an extract_slice/subview op for a single valueToTile with builder.

This new operation extracts a tile of valueToTile, starting at offsets lbs and with sizes subShapeSizes. omitPartialTileCheck controls whether to omit the partial/boundary tile condition check in cases where we statically know that it is unnecessary.

Definition at line 554 of file Utils.cpp.

References computeSliceParameters(), and materializeTiledShape().

◆ makeTiledShapes()

SmallVector< Value > mlir::linalg::makeTiledShapes	(	OpBuilder &	builder,
		Location	loc,
		LinalgOp	linalgOp,
		ValueRange	valuesToTile,
		ArrayRef< OpFoldResult >	ivs,
		ArrayRef< OpFoldResult >	tileSizes,
		ArrayRef< OpFoldResult >	sizeBounds,
		bool	omitPartialTileCheck
	)

Creates extract_slice/subview ops for all valuesToTile of the given linalgOp with builder, assuming linalgOp is being fused into a loop nest for tiling with the given induction variables ivs and tile sizes tileSizes.

sizeBounds are the iteration space bounds for all the implicit loops in linalgOp. omitPartialTileCheck controls whether to omit the partial/boundary tile condition check in cases where we statically know that it is unnecessary.

Note that a constant zero in tileSizes means no tiling at that implicit loop. The number of non-zero values in tileSizes should be equal to the number of values in ivs.

Definition at line 794 of file Utils.cpp.

References computeAllSliceParameters(), and materializeTiledShape().

Referenced by fuse(), and tileLinalgOpImpl().

◆ materializeTiledShape()

static Operation* mlir::linalg::materializeTiledShape	(	OpBuilder &	builder,
		Location	loc,
		Value	valueToTile,
		const SliceParameters &	sliceParams
	)

static

Definition at line 533 of file Utils.cpp.

References mlir::OpBuilder::create(), mlir::Value::getType(), mlir::linalg::SliceParameters::offsets, mlir::linalg::SliceParameters::sizes, and mlir::linalg::SliceParameters::strides.

Referenced by makeTiledShape(), and makeTiledShapes().

◆ offsetIndices() [1/2]

void mlir::linalg::offsetIndices	(	OpBuilder &	b,
		LinalgOp	linalgOp,
		ArrayRef< OpFoldResult >	offests
	)

Add the specified offsets to any linalg.index ops contained in the given linalgOp.

The offsets are provided in the same order as iteration space dimensions. Null offests are assumed to be zero.

Definition at line 816 of file Utils.cpp.

Referenced by fuse(), and transformIndexOps().

◆ offsetIndices() [2/2]

void mlir::linalg::offsetIndices	(	RewriterBase &	b,
		LinalgOp	linalgOp,
		ArrayRef< OpFoldResult >	offests
	)

Definition at line 822 of file Utils.cpp.

References mlir::bindDims(), mlir::Builder::getContext(), mlir::Value::getDefiningOp(), mlir::detail::IROperandBase::getOwner(), mlir::getValueOrCreateConstantIndexOp(), mlir::affine::makeComposedFoldedAffineApply(), mlir::RewriterBase::replaceUsesWithIf(), and mlir::OpBuilder::setInsertionPointAfter().

◆ pack()

FailureOr< PackResult > mlir::linalg::pack	(	RewriterBase &	rewriter,
		linalg::LinalgOp	linalgOp,
		ArrayRef< OpFoldResult >	packedSizes
	)

Implement packing of a single LinalgOp by packedSizes.

Implement packing of a single LinalgOp by performing packing by packedSizes.

There must be one packedSizes entry per linalgOp iterator. Return the packed Linalg op on success, failure otherwise.

Definition at line 479 of file Transforms.cpp.

References mlir::OpBuilder::create(), DBGS, DBGSNL, mlir::getConstantIntValue(), mlir::getElementTypeOrSelf(), mlir::Operation::getRegion(), mlir::Value::getType(), mlir::ValueRange::getTypes(), mlir::Builder::getZeroAttr(), mlir::RewriterBase::notifyMatchFailure(), packLinalgMetadataOnce(), mlir::RewriterBase::replaceOp(), mlir::Region::takeBody(), and mlir::tile().

◆ packMatmulGreedily()

FailureOr< PackResult > mlir::linalg::packMatmulGreedily	(	RewriterBase &	rewriter,
		LinalgOp	linalgOp,
		ArrayRef< OpFoldResult >	mnkPackedSizes,
		ArrayRef< int64_t >	mnkPaddedSizesNextMultipleOf,
		ArrayRef< int64_t >	mnkOrder
	)

Pack a LinalgOp by greedily inferring matmul dimensions (m, n, k) where m and n are proper parallel dimensions and k is a proper reduction dimension.

Packing occurs by rewriting the op as a linalg.generic and calling linalg::pack by mnkPackedSizes. The order of the packed dimensions is customizable: the mnkOrder is a permutation of {0, 1, 2} to reorder {m, n, k} into one of the 8 possible forms. The outer dimensions of the operands are not permuted at this time, this is left for future work.

Definition at line 768 of file Transforms.cpp.

References mlir::computePermutationVector(), DBGS, DBGSNL, generalizeNamedOp(), inferContractionDims(), interchangeGenericOp(), mlir::isPermutationVector(), and mlir::RewriterBase::notifyMatchFailure().

Referenced by blockPackMatmul().

◆ packTranspose()

FailureOr< PackTransposeResult > mlir::linalg::packTranspose	(	RewriterBase &	rewriter,
		tensor::PackOp	packOp,
		linalg::LinalgOp	linalgOp,
		tensor::UnPackOp	maybeUnPackOp,
		ArrayRef< int64_t >	outerPerm,
		ArrayRef< int64_t >	innerPerm
	)

Transpose a single PackOp -> LinalgOp -> UnPackOp chain and return the transposed PackOp -> LinalgOp -> UnPackOp chain after replacements.

Return failure if either:

the packOp does not have the linalgOp as its unique use.
the maybeUnPackOp, if specified must be a consumer of the result tied to the unique packOp use.
outerPerm (resp. innerPerm) must be valid permutations of packOp.getOuterDimsPerm (resp. packOp.getInnerDimsPerm) or empty.

Definition at line 677 of file Transforms.cpp.

References mlir::OpOperand::getOperandNumber(), mlir::detail::IROperandBase::getOwner(), mlir::isPermutationVector(), mlir::RewriterBase::notifyMatchFailure(), mlir::RewriterBase::replaceOp(), mlir::OpBuilder::setInsertionPoint(), and transposeOneLinalgOperandAndReplace().

Referenced by transposePackedMatmul().

◆ padAndHoistLinalgOp()

FailureOr< LinalgOp > mlir::linalg::padAndHoistLinalgOp	(	RewriterBase &	rewriter,
		LinalgOp	linalgOp,
		const LinalgPaddingOptions &	options
	)

Apply padding and hoisting to linalgOp according to the configuration specified in options.

Definition at line 265 of file Padding.cpp.

References mlir::detail::enumerate(), mlir::IROperand< DerivedT, IRValueT >::get(), mlir::Value::getDefiningOp(), hoistPaddingOnTensors(), mlir::linalg::LinalgPaddingOptions::None, mlir::RewriterBase::notifyMatchFailure(), options, mlir::RewriterBase::replaceOp(), and rewriteAsPaddedOp().

◆ peelLoop()

SmallVector< Value > mlir::linalg::peelLoop	(	RewriterBase &	rewriter,
		Operation *	op
	)

Try to peel and canonicalize loop op and return the new result.

Also applies affine_min/max bounds simplification on the fly where relevant.

Definition at line 59 of file Transforms.cpp.

References mlir::Operation::getResults(), and mlir::scf::peelForLoopAndSimplifyBounds().

Referenced by peelLoops().

◆ peelLoops()

void mlir::linalg::peelLoops	(	RewriterBase &	rewriter,
		ArrayRef< scf::ForOp >	loops
	)

Peel 'loops' and applies affine_min/max bounds simplification on the fly where relevant.

Definition at line 75 of file Transforms.cpp.

References peelLoop().

◆ populateBlockPackMatmulPatterns()

void mlir::linalg::populateBlockPackMatmulPatterns	(	RewritePatternSet &	patterns,
		const ControlBlockPackMatmulFn &	controlFn
	)

Patterns to block pack Linalg matmul ops.

Definition at line 310 of file BlockPackMatmul.cpp.

References mlir::patterns.

◆ populateBubbleUpExtractSliceOpPatterns()

void mlir::linalg::populateBubbleUpExtractSliceOpPatterns ( RewritePatternSet & patterns )

Patterns that are used to bubble up extract slice op above linalg op.

Definition at line 134 of file BubbleUpExtractSlice.cpp.

References mlir::patterns.

◆ populateCollapseDimensions()

void mlir::linalg::populateCollapseDimensions	(	RewritePatternSet &	patterns,
		const GetCollapsableDimensionsFn &	controlCollapseDimensions
	)

Pattern to collapse dimensions in a linalg.generic op.

This will collapse tensor operands when needed and expand back the result tensors.

Definition at line 2154 of file ElementwiseOpFusion.cpp.

References mlir::patterns.

◆ populateConstantFoldLinalgOperations()

void mlir::linalg::populateConstantFoldLinalgOperations	(	RewritePatternSet &	patterns,
		const ControlFusionFn &	controlFn
	)

Patterns to constant fold Linalg operations.

Definition at line 306 of file ConstantFold.cpp.

References mlir::patterns.

◆ populateContractionOpRankReducingPatterns()

void mlir::linalg::populateContractionOpRankReducingPatterns ( RewritePatternSet & patterns )

Adds patterns that reduce the rank of named contraction ops that have unit dimensions in the operand(s) by converting to a sequence of collapse_shape, <corresponding linalg named op>, expand_shape (if on tensors).

For example a linalg.batch_matmul with unit batch size will convert to linalg.matmul and a linalg.matvec with with unit spatial dim in lhs will convert to a linalg.dot.

Definition at line 1068 of file DropUnitDims.cpp.

References mlir::patterns.

◆ populateConvertConv2DToImg2ColPatterns()

void mlir::linalg::populateConvertConv2DToImg2ColPatterns ( RewritePatternSet & patterns )

Populates patterns to transform linalg.conv_2d_xxx operations into linalg.generic (for img2col packing) and linalg.matmul.

See also: rewriteInIm2Col for more details.

Definition at line 687 of file ConvertConv2DToImg2Col.cpp.

References mlir::patterns.

◆ populateConvertToDestinationStylePatterns()

void mlir::linalg::populateConvertToDestinationStylePatterns ( RewritePatternSet & patterns )

Populate patterns that convert non-destination-style ops to destination style ops.

Definition at line 604 of file ConvertToDestinationStyle.cpp.

References mlir::patterns.

◆ populateConvolutionVectorizationPatterns()

void mlir::linalg::populateConvolutionVectorizationPatterns	(	RewritePatternSet &	patterns,
		PatternBenefit	benefit = `1`
	)

Populate patterns for vectorizing low-D convolution ops.

This is a step in progressive lowering for convolution ops, it assume high-D convolution ops were decomposed previously.

Definition at line 3892 of file Vectorization.cpp.

References mlir::patterns.

◆ populateDataLayoutPropagationPatterns()

void mlir::linalg::populateDataLayoutPropagationPatterns	(	RewritePatternSet &	patterns,
		const ControlPropagationFn &	controlPackUnPackPropagation
	)

Patterns to bubble up or down data layout ops across other operations.

Definition at line 1206 of file DataLayoutPropagation.cpp.

References mlir::patterns.

◆ populateDecomposeConvolutionPatterns()

void mlir::linalg::populateDecomposeConvolutionPatterns	(	RewritePatternSet &	patterns,
		PatternBenefit	benefit = `1`
	)

Linalg decompose convolutions patterns.

Populates patterns to decompose high-D convolution ops into low-D ones. This is a step in progressive lowering for convolution ops, afterwards we can vectorize the low-D convolution ops.

Definition at line 1637 of file Transforms.cpp.

References mlir::patterns.

◆ populateDecomposeLinalgOpsPattern()

void mlir::linalg::populateDecomposeLinalgOpsPattern	(	RewritePatternSet &	patterns,
		bool	removeDeadArgsAndResults = `true`
	)

Populate patterns for splitting a LinalgOp with multiple statements within its payload into multiple GenericOp that have a single statement.

The option removeDeadArgsAndResults adds patterns to remove dead arguments and results from the generated decomposed ops. This is default true since the core decomposition patterns relies on these clean up patterns. It is set to false only for testing purposes.

Definition at line 384 of file DecomposeLinalgOps.cpp.

References mlir::patterns, and populateEraseUnusedOperandsAndResultsPatterns().

◆ populateDecomposePackUnpackPatterns()

void mlir::linalg::populateDecomposePackUnpackPatterns ( RewritePatternSet & patterns )

Populates patterns to decompose tensor.pack and tensor.unpack Ops into e.g.

tensor.pad, linalg.transpose, tensor.{insert|extract}_slice. Require all outer dims to be unit.

Definition at line 1658 of file Transforms.cpp.

References mlir::patterns.

◆ populateDecomposePadPatterns()

void mlir::linalg::populateDecomposePadPatterns ( RewritePatternSet & patterns )

Populates patterns to decompose tensor.pad into e.g.

tensor.empty, linalg.fill, tensor.insert_slice.

Definition at line 1663 of file Transforms.cpp.

References mlir::patterns.

◆ populateDecomposeProjectedPermutationPatterns()

void mlir::linalg::populateDecomposeProjectedPermutationPatterns ( RewritePatternSet & patterns )

Add patterns to make explicit broadcasts and transforms in the input operands of a genericOp.

Definition at line 246 of file DecomposeGenericByUnfoldingPermutation.cpp.

References mlir::patterns.

◆ populateDecomposeWinogradOpsPatterns()

void mlir::linalg::populateDecomposeWinogradOpsPatterns ( RewritePatternSet & patterns )

Patterns to decompose Winograd operators.

Definition at line 1215 of file WinogradConv2D.cpp.

References mlir::patterns.

◆ populateElementwiseOpsFusionPatterns()

void mlir::linalg::populateElementwiseOpsFusionPatterns	(	RewritePatternSet &	patterns,
		const ControlFusionFn &	controlElementwiseOpFusion
	)

Patterns for fusing linalg operation on tensors.

Pattern to fuse linalg.generic -> linalg.generic operations when both operations are fusable elementwise operations.

Definition at line 2143 of file ElementwiseOpFusion.cpp.

References mlir::patterns, and populateEraseUnusedOperandsAndResultsPatterns().

◆ populateElementwiseToLinalgConversionPatterns()

void mlir::linalg::populateElementwiseToLinalgConversionPatterns ( RewritePatternSet & patterns )

Populate patterns that convert ElementwiseMappable ops to linalg parallel loops.

Definition at line 115 of file ElementwiseToLinalg.cpp.

References mlir::patterns.

◆ populateEraseUnnecessaryInputsPatterns()

void mlir::linalg::populateEraseUnnecessaryInputsPatterns ( RewritePatternSet & patterns )

Patterns to promote inputs to outputs and remove unused inputs of linalg.generic ops.

Definition at line 428 of file EraseUnusedOperandsAndResults.cpp.

References mlir::patterns.

◆ populateEraseUnusedOperandsAndResultsPatterns()

void mlir::linalg::populateEraseUnusedOperandsAndResultsPatterns ( RewritePatternSet & patterns )

Pattern to remove dead operands and results of linalg.generic operations.

This is effectively DCE for a linalg op.

Definition at line 421 of file EraseUnusedOperandsAndResults.cpp.

References mlir::patterns.

Referenced by populateDecomposeLinalgOpsPattern(), and populateElementwiseOpsFusionPatterns().

◆ populateFoldAddIntoDestPatterns()

void mlir::linalg::populateFoldAddIntoDestPatterns ( RewritePatternSet & patterns )

Pattern to replace linalg.add when destination passing on a contraction op suffices for achieving the sum.

Definition at line 147 of file FoldAddIntoDest.cpp.

References mlir::patterns.

◆ populateFoldReshapeOpsByCollapsingPatterns()

void mlir::linalg::populateFoldReshapeOpsByCollapsingPatterns	(	RewritePatternSet &	patterns,
		const ControlFusionFn &	controlFoldingReshapes
	)

Patterns to fold an expanding tensor.expand_shape operation with its producer generic operation by collapsing the dimensions of the generic op.

Definition at line 2134 of file ElementwiseOpFusion.cpp.

References mlir::patterns.

◆ populateFoldReshapeOpsByExpansionPatterns()

void mlir::linalg::populateFoldReshapeOpsByExpansionPatterns	(	RewritePatternSet &	patterns,
		const ControlFusionFn &	controlFoldingReshapes
	)

Patterns to fold an expanding (collapsing) tensor_reshape operation with its producer (consumer) generic operation by expanding the dimensionality of the loop in the generic op.

Definition at line 2123 of file ElementwiseOpFusion.cpp.

References mlir::patterns.

◆ populateFoldUnitExtentDimsPatterns()

void mlir::linalg::populateFoldUnitExtentDimsPatterns	(	RewritePatternSet &	patterns,
		linalg::ControlDropUnitDims &	options
	)

Patterns to fold unit-extent dimensions in operands/results of linalg ops on tensors via reassociative reshape ops.

Definition at line 799 of file DropUnitDims.cpp.

References mlir::linalg::ControlDropUnitDims::ExtractInsertSlice, options, mlir::patterns, populateFoldUnitExtentDimsViaReshapesPatterns(), and populateFoldUnitExtentDimsViaSlicesPatterns().

◆ populateFuseTensorPadWithProducerLinalgOpPatterns()

void mlir::linalg::populateFuseTensorPadWithProducerLinalgOpPatterns ( RewritePatternSet & patterns )

Pattern to fuse a tensor.pad operation with the producer of its source, if the producer is a linalg operation with all parallel iterator types.

Definition at line 121 of file FusePadOpWithLinalgProducer.cpp.

References mlir::patterns.

◆ populateInlineConstantOperandsPatterns()

void mlir::linalg::populateInlineConstantOperandsPatterns ( RewritePatternSet & patterns )

Patterns that are used to inline constant operands into linalg generic ops.

Definition at line 98 of file InlineScalarOperands.cpp.

References mlir::patterns.

◆ populateInsertSliceVectorizationPatterns()

void mlir::linalg::populateInsertSliceVectorizationPatterns ( RewritePatternSet & patterns )

Populates patterns with vectorisation patterns for tensor.insert_slice.

TODO: Avoid having a dedicated populate{} for one pattern. Instead, either expand or merge with other populate{}.

Definition at line 2781 of file Vectorization.cpp.

References mlir::patterns.

◆ populateLinalgGenericOpsSpecializationPatterns()

void mlir::linalg::populateLinalgGenericOpsSpecializationPatterns ( RewritePatternSet & patterns )

Populates patterns with patterns to convert linalg.generic ops to named ops where possible.

A linalg.generic can represent wide range and complex computations for which equivalent linalg named op may not exist e.g. linalg.generic that takes a tensor and computes a polynomial such as: p(x) = an*x^n + ... + a1x + a0 There is no equivalent named op to convert to. Many such cases exist.

Definition at line 356 of file Specialize.cpp.

References mlir::patterns.

◆ populateLinalgNamedOpConversionPatterns()

void mlir::linalg::populateLinalgNamedOpConversionPatterns ( RewritePatternSet & patterns )

Patterns to convert from one named op to another.

These can be seen as canonicalizations of named ops into another named op.

Definition at line 161 of file NamedOpConversions.cpp.

References mlir::patterns.

◆ populateLinalgNamedOpsGeneralizationPatterns()

void mlir::linalg::populateLinalgNamedOpsGeneralizationPatterns ( RewritePatternSet & patterns )

Linalg generalization patterns.

Populates patterns with patterns to convert spec-generated named ops to linalg.generic ops.

Definition at line 95 of file Generalization.cpp.

References mlir::patterns.

◆ populateLinalgTilingCanonicalizationPatterns()

void mlir::linalg::populateLinalgTilingCanonicalizationPatterns ( RewritePatternSet & patterns )

Definition at line 860 of file Tiling.cpp.

References mlir::patterns.

Referenced by getLinalgTilingCanonicalizationPatterns().

◆ populateLinalgToStandardConversionPatterns()

void mlir::linalg::populateLinalgToStandardConversionPatterns ( RewritePatternSet & patterns )

Populate the given list with patterns that convert from Linalg to Standard.

Definition at line 127 of file LinalgToStandard.cpp.

References mlir::patterns.

◆ populateMoveInitOperandsToInputPattern()

void mlir::linalg::populateMoveInitOperandsToInputPattern ( RewritePatternSet & patterns )

A pattern that converts init operands to input operands.

Definition at line 811 of file DropUnitDims.cpp.

References mlir::patterns.

◆ populatePadOpVectorizationPatterns()

void mlir::linalg::populatePadOpVectorizationPatterns	(	RewritePatternSet &	patterns,
		PatternBenefit	baseBenefit = `1`
	)

Populates patterns with patterns that vectorize tensor.pad.

These patterns are meant to apply in a complementary fashion. Benefits are used to encode a certain ordering of pattern application. To avoid scattering magic constants throughout the code base, the patterns must be added with this function. baseBenefit can be used to offset the benefit of all tensor::PadOp vectorization patterns by a certain value.

Definition at line 2786 of file Vectorization.cpp.

References mlir::PatternBenefit::getBenefit(), and mlir::patterns.

◆ populateSparseTensorRewriting()

void mlir::linalg::populateSparseTensorRewriting ( RewritePatternSet & patterns )

Populate patterns that are only useful in the context of sparse tensors.

◆ populateSplitReductionPattern()

void mlir::linalg::populateSplitReductionPattern	(	RewritePatternSet &	patterns,
		const ControlSplitReductionFn &	controlSplitReductionFn,
		bool	useAlloc = `false`
	)

Patterns to apply splitReduction below.

Definition at line 448 of file SplitReduction.cpp.

References mlir::patterns.

◆ populateSwapExtractSliceWithFillPatterns()

void mlir::linalg::populateSwapExtractSliceWithFillPatterns ( RewritePatternSet & patterns )

Adds patterns that waps tensor.extract_slice(linalg.fill(cst, init)) into linalg.fill(cst, tensor.extract_slice(init)).

Definition at line 38 of file SwapExtractSliceWithFillPatterns.cpp.

References mlir::patterns.

◆ populateTranposeConv2DPatterns()

void mlir::linalg::populateTranposeConv2DPatterns ( RewritePatternSet & patterns )

Definition at line 141 of file TransposeConv2D.cpp.

References mlir::patterns.

◆ populateTransposeMatmulPatterns()

void mlir::linalg::populateTransposeMatmulPatterns	(	RewritePatternSet &	patterns,
		bool	transposeLHS = `true`
	)

Patterns to convert Linalg matmul ops to transposed variants.

Definition at line 164 of file TransposeMatmul.cpp.

References mlir::patterns.

◆ populateWinogradConv2DPatterns()

void mlir::linalg::populateWinogradConv2DPatterns	(	RewritePatternSet &	patterns,
		int64_t	m,
		int64_t	r
	)

Patterns to apply Winograd Conv2D algorithm F(m x m, r x r).

Definition at line 1208 of file WinogradConv2D.cpp.

◆ promoteSubviewAsNewBuffer()

FailureOr< PromotionInfo > mlir::linalg::promoteSubviewAsNewBuffer	(	OpBuilder &	b,
		Location	loc,
		memref::SubViewOp	subView,
		const AllocBufferCallbackFn &	allocationFn,
		DataLayout &	layout
	)

Definition at line 238 of file Promotion.cpp.

References mlir::OpBuilder::create(), mlir::OpBuilder::createOrFold(), mlir::detail::enumerate(), mlir::Builder::getIndexAttr(), and mlir::getValueOrCreateConstantIndexOp().

Referenced by promoteSubViews().

◆ promoteSubViews()

FailureOr< LinalgOp > mlir::linalg::promoteSubViews	(	OpBuilder &	b,
		LinalgOp	op,
		const LinalgPromotionOptions &	options
	)

Promote the subViews into a new buffer allocated at the insertion point b.

Promotion occurs in 3 steps:

Create a new buffer for a full tile (i.e. not clipped at the boundary).
Take a full view on the buffer.
Take a partial slice of the full view in step 2. and copy into it.

Return the modified linalg op (the modification happens in place) as well as all the copy ops created.

Definition at line 421 of file Promotion.cpp.

References options, and promoteSubViews().

◆ promoteSubviewsPrecondition()

LogicalResult mlir::linalg::promoteSubviewsPrecondition	(	Operation *	op,
		LinalgPromotionOptions	options
	)

Promote memref.subviews feeding linalg-on-buffers operations.

Definition at line 399 of file Promotion.cpp.

References options.

◆ registerAll()

template<typename... OpTypes>

static void mlir::linalg::registerAll ( MLIRContext * ctx )

static

Variadic helper function.

Definition at line 344 of file MeshShardingInterfaceImpl.cpp.

Referenced by registerMeshShardingInterfaceExternalModels().

◆ registerAllDialectInterfaceImplementations()

void mlir::linalg::registerAllDialectInterfaceImplementations ( DialectRegistry & registry )

Definition at line 17 of file AllInterfaces.cpp.

References mlir::arith::registerBufferizableOpInterfaceExternalModels(), registerMeshShardingInterfaceExternalModels(), registerSubsetOpInterfaceExternalModels(), registerTilingInterfaceExternalModels(), and mlir::affine::registerValueBoundsOpInterfaceExternalModels().

Referenced by mlir::registerAllDialects().

◆ registerBufferizableOpInterfaceExternalModels()

void mlir::linalg::registerBufferizableOpInterfaceExternalModels ( DialectRegistry & registry )

Definition at line 196 of file BufferizableOpInterfaceImpl.cpp.

References mlir::DialectRegistry::addExtension().

◆ registerMeshShardingInterfaceExternalModels()

void mlir::linalg::registerMeshShardingInterfaceExternalModels ( DialectRegistry & registry )

Definition at line 348 of file MeshShardingInterfaceImpl.cpp.

References mlir::DialectRegistry::addExtension(), mlir::MLIRContext::appendDialectRegistry(), mlir::DialectRegistry::getDialectNames(), mlir::MLIRContext::getOrLoadDialect(), mlir::DialectRegistry::insert(), and registerAll().

Referenced by registerAllDialectInterfaceImplementations().

◆ registerOne()

template<typename OpType >

static void mlir::linalg::registerOne ( MLIRContext * ctx )

static

Definition at line 338 of file MeshShardingInterfaceImpl.cpp.

◆ registerRuntimeVerifiableOpInterfaceExternalModels()

void mlir::linalg::registerRuntimeVerifiableOpInterfaceExternalModels ( DialectRegistry & registry )

Definition at line 122 of file RuntimeOpVerification.cpp.

References mlir::DialectRegistry::addExtension(), and mlir::MLIRContext::loadDialect().

Referenced by mlir::registerAllDialects().

◆ registerSubsetOpInterfaceExternalModels()

void mlir::linalg::registerSubsetOpInterfaceExternalModels ( DialectRegistry & registry )

Definition at line 75 of file SubsetInsertionOpInterfaceImpl.cpp.

References mlir::DialectRegistry::addExtension().

Referenced by registerAllDialectInterfaceImplementations().

◆ registerTilingInterfaceExternalModels()

void mlir::linalg::registerTilingInterfaceExternalModels ( DialectRegistry & registry )

Definition at line 522 of file TilingInterfaceImpl.cpp.

References mlir::DialectRegistry::addExtension(), and registerAll().

Referenced by registerAllDialectInterfaceImplementations().

◆ registerTransformDialectExtension()

void mlir::nvgpu::registerTransformDialectExtension ( DialectRegistry & registry )

Definition at line 60 of file DialectExtension.cpp.

References mlir::DialectRegistry::addExtensions().

Referenced by mlir::registerAllExtensions().

◆ registerValueBoundsOpInterfaceExternalModels()

void mlir::linalg::registerValueBoundsOpInterfaceExternalModels ( DialectRegistry & registry )

Definition at line 54 of file ValueBoundsOpInterfaceImpl.cpp.

References mlir::DialectRegistry::addExtension().

◆ rewriteAsPaddedOp()

LogicalResult mlir::linalg::rewriteAsPaddedOp	(	RewriterBase &	rewriter,
		LinalgOp	opToPad,
		const LinalgPaddingOptions &	options,
		LinalgOp &	paddedOp,
		SmallVector< Value > &	replacements,
		SmallVector< tensor::PadOp > &	padOps
	)

Pad the iterator dimensions paddingDimensions of all opToPad operands to a static bounding box.

The original opToPad is cloned and operates on the padded tensors.

"options.padToMultipleOf" indicates that each padding dimension should be padded to the specified multiple.
Use "options.paddingValues" and "options.nofoldFlags" to set padding value and nofold attribute of the created tensor::PadOps, respectively.
The unpadded results (extracted slice of the cloned operation) are returned via replacements.
The tensor::PadOps are returned via padOps.
"options.copyBackOp" specifies the op type for copying back the unpadded result to the original destination tensor.

Definition at line 153 of file Padding.cpp.

References mlir::clone(), mlir::OpBuilder::create(), DBGS, mlir::detail::enumerate(), mlir::get(), mlir::getElementTypeOrSelf(), mlir::Builder::getIndexAttr(), mlir::Value::getType(), mlir::ValueRange::getTypes(), mlir::Builder::getZeroAttr(), mlir::linalg::LinalgPaddingOptions::LinalgCopy, mlir::linalg::LinalgPaddingOptions::None, mlir::RewriterBase::notifyMatchFailure(), options, padOperandToSmallestStaticBoundingBox(), mlir::reifyResultShapes(), and mlir::OpBuilder::setInsertionPointAfter().

Referenced by padAndHoistLinalgOp().

◆ rewriteInDestinationPassingStyle() [1/3]

FailureOr< Operation * > mlir::linalg::rewriteInDestinationPassingStyle	(	RewriterBase &	rewriter,
		tensor::FromElementsOp	fromElementsOp
	)

Rewrite tensor.from_elements to linalg.generic.

Lower tensor.from_elements to a sequence of chained tensor.insert.

Definition at line 344 of file ConvertToDestinationStyle.cpp.

References mlir::OpBuilder::create(), createInserts(), mlir::Value::getDefiningOp(), mlir::RewriterBase::replaceOp(), and mlir::RewriterBase::replaceOpWithNewOp().

◆ rewriteInDestinationPassingStyle() [2/3]

FailureOr< Operation * > mlir::linalg::rewriteInDestinationPassingStyle	(	RewriterBase &	rewriter,
		tensor::GenerateOp	generateOp
	)

Rewrite tensor.generate to linalg.generic.

Lower tensor.generate to linalg.generic.

Definition at line 382 of file ConvertToDestinationStyle.cpp.

References mlir::OpBuilder::create(), mlir::OpBuilder::createBlock(), mlir::Builder::getMultiDimIdentityMap(), mlir::RewriterBase::mergeBlocks(), mlir::RewriterBase::replaceOp(), mlir::RewriterBase::replaceOpWithNewOp(), and mlir::OpBuilder::setInsertionPointToStart().

◆ rewriteInDestinationPassingStyle() [3/3]

FailureOr< Operation * > mlir::linalg::rewriteInDestinationPassingStyle	(	RewriterBase &	rewriter,
		tensor::PadOp	padOp
	)

Rewrite tensor.pad to linalg.generic + tensor.insert_slice.

Lower tensor.pad to linalg.generic + tensor.insert_slice.

Definition at line 423 of file ConvertToDestinationStyle.cpp.

References mlir::OpBuilder::create(), mlir::Builder::getIndexAttr(), mlir::tensor::getMixedSizes(), mlir::isZeroIndex(), movePaddingToFillOrGenericOp(), mlir::RewriterBase::notifyMatchFailure(), mlir::reifyResultShapes(), mlir::RewriterBase::replaceOpWithNewOp(), and mlir::OpBuilder::setInsertionPointAfter().

◆ rewriteInIm2Col() [1/4]

FailureOr< std::pair< Operation , Operation > > mlir::linalg::rewriteInIm2Col	(	RewriterBase &	rewriter,
		linalg::Conv2DNchwFchwOp	convOp
	)

Similar to rewriteInIm2Col with linalg::Conv2DNhwcHwcfOp except because the channels are to the left of the image shape dimensions, the position of the contraction dimension in the resulting matmul is reversed.

This swaps the LHS and RHS of the matmul when compared with nhwc (i.e. (D, C x Kh x Kw) * (C x Kh x Kw, Ho x Wo))

Definition at line 365 of file ConvertConv2DToImg2Col.cpp.

References mlir::bindDims(), mlir::OpBuilder::create(), createAdd(), createMul(), mlir::AffineMap::get(), mlir::get(), mlir::Builder::getContext(), getConvolvedIndex(), mlir::Value::getLoc(), mlir::AffineMap::getMultiDimIdentityMap(), mlir::Value::getType(), mlir::getType(), hasAllOneValues(), mlir::RewriterBase::notifyMatchFailure(), mlir::RewriterBase::replaceOp(), and unrollIndex().

◆ rewriteInIm2Col() [2/4]

FailureOr< std::pair< Operation , Operation > > mlir::linalg::rewriteInIm2Col	(	RewriterBase &	rewriter,
		linalg::Conv2DNhwcFhwcOp	convOp
	)

Same as the above but for Fhwc channel orderings in the filter.

In this case the matrix multiplication is actually a row-wise dot-product rather than a row-column dot-product. This is to avoid transposing the filter matrix which would be required for a regular matrix multiplication to produce the correct output dimensions.

Definition at line 498 of file ConvertConv2DToImg2Col.cpp.

References mlir::bindDims(), mlir::OpBuilder::create(), createAdd(), createMul(), mlir::AffineMap::get(), mlir::get(), mlir::Builder::getContext(), getConvolvedIndex(), mlir::AffineMap::getMultiDimIdentityMap(), mlir::Value::getType(), mlir::getType(), hasAllOneValues(), mlir::RewriterBase::notifyMatchFailure(), mlir::RewriterBase::replaceOp(), and unrollIndex().

◆ rewriteInIm2Col() [3/4]

FailureOr< std::pair< Operation , Operation > > mlir::linalg::rewriteInIm2Col	(	RewriterBase &	rewriter,
		linalg::Conv2DNhwcHwcfOp	convOp
	)

Convert linalg.conv_2d_nhwc_hwcf into linalg.generic (for img2col packing) and linalg.matmul.

A convolution operation can be written as a matrix-matrix multiplication by unfolding the cross-correlation between input and filter and explicitly copy overlapped sliding window inputs.

Consider 2D input X with single channel input and output and 2x2 filter W: [x(0, 0) , x(0, 1) , ..., x(0, n) ] [x(1, 0) , x(1, 1) , ..., x(1, n) ] [. , . ,. , . ] [w(0, 0), w(0, 1)] [. , . , . , . ] (conv) [w(1, 0), w(1, 1)] [. , . , ., . ] [x(n-1, 0), x(n-1, 1), ..., x(n-1, n-1)]

The packed input data (img2col) is a matrix with |rows| = output spatial size, |columns| = filter spatial size. To compute the output Y(i, j) we need to calculate the dot product between filter window at input X(x, y)) and the filter which will look like the following where r.h.s is the img2col matrix and l.h.s is the flattened filter:

[x(0,0), x(0,1), x(1,0), x(1,1)] [x(0,1), x(1,1), x(0,2), x(1,2)] (matmul) [w(0,0), w(0,1), w(1,0), w(1,1)] [x(0,1), x(1,1), x(0,2), x(1,2)] [ . , . , . , . ]

In general for 2D case with (N, H, W, C) input and (Kh, Kw, C, D) filter and output (N, Ho, Wo, D) the convolution is the following matrix-matrix multiplication (Ho x Wo, Kh x Kw x C) * (Kh x Kw x C, D) for each input in the N input. For the case where N > 1 its a batched matrix-matrix multiplication.

On success, return both the operation that produces the img2col tensor and the final operation of the sequence that replaces the original convolution.

Definition at line 79 of file ConvertConv2DToImg2Col.cpp.

References mlir::bindDims(), mlir::OpBuilder::create(), createAdd(), createMul(), mlir::AffineMap::get(), mlir::get(), mlir::Builder::getContext(), getConvolvedIndex(), mlir::AffineMap::getMultiDimIdentityMap(), mlir::Value::getType(), mlir::getType(), hasAllOneValues(), mlir::RewriterBase::notifyMatchFailure(), mlir::RewriterBase::replaceOp(), and unrollIndex().

◆ rewriteInIm2Col() [4/4]

FailureOr< std::pair< Operation , Operation > > mlir::linalg::rewriteInIm2Col	(	RewriterBase &	rewriter,
		linalg::DepthwiseConv2DNhwcHwcOp	convOp
	)

Similar to rewriteInIm2Col with linalg::Conv2DNhwcHwcfOp except there is no reduction among the input channels so each convolution can be a matrix-vector product and by transposing both input filter so channels are outer most the computation is a batched matrix-vector product.

Definition at line 214 of file ConvertConv2DToImg2Col.cpp.

References mlir::bindDims(), mlir::OpBuilder::create(), mlir::AffineMap::get(), mlir::get(), mlir::Builder::getAffineConstantExpr(), mlir::Builder::getAffineDimExpr(), mlir::Builder::getContext(), mlir::AffineMap::getMultiDimIdentityMap(), mlir::Operation::getResult(), mlir::Value::getType(), hasAllOneValues(), mlir::inversePermutation(), mlir::RewriterBase::notifyMatchFailure(), and mlir::RewriterBase::replaceOp().

◆ specializeGenericOp()

FailureOr< LinalgOp > mlir::linalg::specializeGenericOp	(	RewriterBase &	rewriter,
		GenericOp	genericOp
	)

Create a namedOp from the given GenericOp and replace the GenericOp.

Currently we can specialize only trivial linalg copy operations.

Definition at line 260 of file Specialize.cpp.

References areBinOpsSwapped(), isaBroadcastOpInterface(), isaContractionOpInterface(), isaCopyOpInterface(), isaElemwiseSingleBinaryOpInterface(), isaElemwiseSingleUnaryOpInterface(), isaFillOpInterface(), isaTransposeOpInterface(), REPLACE_BINARY_OP, REPLACE_UNARY_OP, and mlir::RewriterBase::replaceOpWithNewOp().

Referenced by mlir::linalg::LinalgSpecializationPattern::returningMatchAndRewrite().

◆ splitOp()

std::pair< TilingInterface, TilingInterface > mlir::linalg::splitOp	(	RewriterBase &	rewriter,
		TilingInterface	op,
		unsigned	dimension,
		OpFoldResult	splitPoint
	)

Split the given op into two parts along the given iteration space dimension at the specified splitPoint, and return the two parts.

If the second part is statically known to be empty, do not create it and return nullptr instead. Error state is signalled by returning a pair of nullptrs.

For example, the following op:

linalg.matmul ins(%0, %1 : tensor<128x32xf32>, tensor<32x64xf32>) outs(%2 : tensor<128x64xf32>)

split along the first dimension at position 42 will result in:

%3 = tensor.extract_slice %0[0, 0][42, 32][1, 1] %4 = tensor.extract_slice %2[0, 0][42, 64][1, 1] %5 = linalg.matmul ins(%3, %1 : tensor<42x32xf32>, tensor<32x64xf32>) outs(%5 : tensor<42x64xf32>) %6 = tensor.insert_slice %5 into %2[0, 0][42, 64][1, 1]

%7 = tensor.extract_slice %0[42, 0][86, 32][1, 1] %8 = tensor.extract_slice %6[42, 0][86, 64][1, 1] %9 = linalg.matmul ins(%7, %1 : tensor<86x32xf32>, tensor<32x64xf32>) outs(%8 : tensor<86x64xf32>) tensor.insert_slice %5 into %6[42, 0][86, 64][1, 1]

Note that there is no simplification other than constant propagation applied to slice extraction and insertion.

Definition at line 67 of file Split.cpp.

References mlir::bindDims(), createSplitPart(), mlir::Builder::getContext(), mlir::tensor::getOrCreateDestinations(), mlir::AffineMap::inferFromExprList(), mlir::affine::makeComposedFoldedAffineApply(), mlir::affine::makeComposedFoldedAffineMin(), mlir::RewriterBase::modifyOpInPlace(), mlir::Range::offset, mlir::RewriterBase::replaceOp(), and mlir::Range::size.

◆ splitReduction()

FailureOr< SplitReductionResult > mlir::linalg::splitReduction	(	RewriterBase &	b,
		LinalgOp	op,
		const ControlSplitReductionFn &	controlSplitReductionFn,
		bool	useAlloc = `false`
	)

◆ splitReductionByScaling()

FailureOr< SplitReductionResult > mlir::linalg::splitReductionByScaling	(	RewriterBase &	b,
		LinalgOp	op,
		const ControlSplitReductionFn &	controlSplitReductionFn,
		bool	useAlloc = `false`
	)

Scaling-based implementation of the split reduction transformation.

Core rewrite implementation.

Instead of introducing an ExpandShapeOp, this rewrites a reduction dimension k into k * scale + kk.

Example: ``` %0 = linalg.matmul ins(A, B: tensor<16x256xf32>, tensor<256x32xf32>) outs(C: tensor<16x32xf32>) -> tensor<16x32xf32> ```

Is transformed to:

``` #map0 = affine_map<(d0, d1, d2, d3) -> (d0, d2 * 4 + d3)> #map1 = affine_map<(d0, d1, d2, d3) -> (d2 * 4 + d3, d1)> #map2 = affine_map<(d0, d1, d2, d3) -> (d2, d3)> #map3 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)> #map4 = affine_map<(d0, d1, d2) -> (d0, d1, d2)> #map5 = affine_map<(d0, d1, d2) -> (d0, d1)> %0 = tensor.empty [16, 32, 64] : tensor<16x32x64xf32> cst = arith.constant 0.000000e+00 : f32 %1 = linalg.fill ins(cst : f32) outs(%0 : tensor<16x32x64xf32>) -> tensor<16x32x64xf32> %2 = tensor.empty [64, 4] : tensor<64x4xi1>

%3 = linalg.generic {indexing_maps = [#map0, #map1, #map2, #map3], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(A, B, %2 : tensor<16x256xf32>, tensor<256x32xf32>, tensor<64x4xi1>) outs(%1 : tensor<16x32x64xf32>) { ^bb0(arg3: f32, arg4: f32, arg5: i1, arg6: f32): %5 = arith.mulf arg3, arg4 : f32 %6 = arith.addf arg6, %5 : f32 linalg.yield %6 : f32 } -> tensor<16x32x64xf32>

%4 = linalg.generic {indexing_maps = [#map4, #map5], iterator_types = ["parallel", "parallel", "reduction"]} */ // ins(%3 : tensor<16x32x64xf32>) /** outs(C : tensor<16x32xf32>) { ^bb0(arg3: f32, arg4: f32): %5 = arith.addf arg3, arg4 : f32 linalg.yield %5 : f32 } -> tensor<16x32xf32>

return %4 : tensor<16x32xf32> ```

Definition at line 241 of file SplitReduction.cpp.

References mlir::Operation::clone(), mlir::OpBuilder::create(), mlir::tensor::createDynamicDimValues(), mlir::AffineMap::dropResult(), mlir::AffineMap::get(), mlir::getAffineDimExpr(), mlir::Value::getDefiningOp(), mlir::Builder::getIntegerType(), mlir::Builder::getMultiDimIdentityMap(), mlir::arith::getNeutralElement(), mlir::Operation::getResult(), mlir::Value::getType(), mlir::ValueRange::getTypes(), mlir::linalg::SplitReductionOptions::index, mlir::RewriterBase::inlineRegionBefore(), mlir::linalg::SplitReductionOptions::innerParallel, mlir::RankedTensorType::Builder::insertDim(), insertParallelDim(), mlir::matchReduction(), mlir::RewriterBase::notifyMatchFailure(), mlir::linalg::SplitReductionOptions::ratio, mlir::RewriterBase::replaceOp(), scaleReductionDim(), mlir::OpBuilder::setInsertionPoint(), and mlir::Operation::setOperand().

◆ spmdizeLinalgOpWithShardedReduction()

static void mlir::linalg::spmdizeLinalgOpWithShardedReduction	(	LinalgOp	op,
		ArrayRef< Value >	spmdizedOperands,
		ArrayRef< MeshSharding >	operandShardings,
		ArrayRef< MeshSharding >	resultShardings,
		ArrayRef< utils::IteratorType >	loopIteratorTypes,
		ArrayRef< SmallVector< MeshAxis >>	meshAxisAssignmentForLoopIterators,
		IRMapping &	spmdizationMap,
		SymbolTableCollection &	symbolTable,
		ImplicitLocOpBuilder &	builder
	)

static

Definition at line 219 of file MeshShardingInterfaceImpl.cpp.

References createAllReduceForResultsWithoutPartialShardings(), createDestinationPassingStyleInitOperands(), getMesh(), mlir::mesh::getReductionMeshAxes(), mlir::IRMapping::lookup(), mlir::IRMapping::map(), and mlir::mesh::spmdizeTriviallyShardableOperation().

◆ tileLinalgOp()

FailureOr< TiledLinalgOp > mlir::linalg::tileLinalgOp	(	RewriterBase &	b,
		LinalgOp	op,
		const LinalgTilingOptions &	options
	)

Definition at line 820 of file Tiling.cpp.

References options.

◆ tileReductionUsingForall()

FailureOr< linalg::ForallReductionTilingResult > mlir::linalg::tileReductionUsingForall	(	RewriterBase &	b,
		PartialReductionOpInterface	op,
		ArrayRef< OpFoldResult >	numThreads,
		ArrayRef< OpFoldResult >	tileSizes = `{}`,
		std::optional< ArrayAttr >	mapping = `std::nullopt`
	)

Method to tile a reduction to parallel iterations computing partial reductions.

After the loop all the partial reduction are merged into a final reduction. For example for the following sequence

%0 = linalg.generic %in ["parallel", "reduction"]

: tensor<7x9xf32> -> tensor<7xf32>

into:

%0 = linalg.fill ... : tensor<7x4xf32>
%1 = scf.forall (%iv) in (%c4) shared_outs(%arg0 = %0)
  -> (tensor<7x4xf32>) {
  %2 = tensor.extract_slice %arg3 : tensor<7x4xf32> to tensor<7xf32>
  %3 = tensor.extract_slice %in : tensor<7x9xf32> -> tensor<7x?xf32>
  %4 = linalg.generic %2, %3 ["parallel", "reduction"]
    : tensor<7x?xf32> -> tensor<7xf32>
  %5 = tensor.insert_slice %3, %arg0[0, %iv] : tensor<7x4xf32>
}
%6 = linalg.generic %1 ["parallel", "reduction"]
  : tensor<7x4xf32> -> tensor<7xf32>

Definition at line 594 of file Tiling.cpp.

References calculateTileOffsetsAndSizes(), mlir::OpBuilder::clone(), mlir::OpBuilder::create(), mlir::Operation::emitError(), mlir::RewriterBase::eraseOp(), mlir::getAsOpFoldResult(), mlir::Builder::getIndexAttr(), mlir::tensor::getOrCreateDestinations(), mlir::getValueOrCreateConstantIndexOp(), mlir::linalg::ForallReductionTilingResult::initialValues, mlir::isConstantIntValue(), mlir::linalg::ForallReductionTilingResult::loops, mlir::affine::mapLoopToProcessorIds(), mlir::linalg::ForallReductionTilingResult::mergeOps, mlir::RewriterBase::modifyOpInPlace(), mlir::RewriterBase::notifyMatchFailure(), options, mlir::linalg::ForallReductionTilingResult::parallelTiledOps, mlir::RewriterBase::replaceOp(), mlir::OpBuilder::setInsertionPoint(), mlir::OpBuilder::setInsertionPointAfter(), and mlir::OpBuilder::setInsertionPointToEnd().

◆ transformIndexOps()

void mlir::linalg::transformIndexOps	(	RewriterBase &	b,
		LinalgOp	op,
		SmallVectorImpl< Value > &	ivs,
		const LoopIndexToRangeIndexMap &	loopIndexToRangeIndex
	)

All indices returned by IndexOp should be invariant with respect to tiling.

Therefore, if an operation is tiled, we have to transform the indices accordingly, i.e. offset them by the values of the corresponding induction variables that are captured implicitly in the body of the op.

Example. linalg.generic before tiling:

#id_2d = (i, j) -> (i, j) #pointwise_2d_trait = { indexing_maps = [#id_2d, #id_2d], iterator_types = ["parallel", "parallel"] } linalg.generic #pointwise_2d_trait operand, result { ^bb0(operand_in: f32, result_in: f32): i = linalg.index 0 : index j = linalg.index 1 : index <some operations that use i, j> }: memref<50x100xf32>, memref<50x100xf32>

After tiling pass with tiles sizes 10 and 25:

#strided = (i, j)[s0, s1, s2] -> (i * s1 + s0 + j * s2)

c1 = arith.constant 1 : index c0 = arith.constant 0 : index c25 = arith.constant 25 : index c10 = arith.constant 10 : index operand_dim_0 = dim operand, 0 : memref<50x100xf32> operand_dim_1 = dim operand, 1 : memref<50x100xf32> scf.for k = c0 to operand_dim_0 step c10 { scf.for l = c0 to operand_dim_1 step c25 { %4 = memref.subview operand[k, l][c10, c25][c1, c1] : memref<50x100xf32> to memref<?x?xf32, #strided> %5 = memref.subview result[k, l][c10, c25][c1, c1] : memref<50x100xf32> to memref<?x?xf32, #strided> linalg.generic pointwise_2d_trait %4, %5 { ^bb0(operand_in: f32, result_in: f32): i = linalg.index 0 : index j = linalg.index 1 : index // Indices k and l are implicitly captured in the body. transformed_i = arith.addi i, k : index // index i is offset by k transformed_j = arith.addi j, l : index // index j is offset by l // Every use of i, j is replaced with transformed_i, transformed_j <some operations that use transformed_i, transformed_j> }: memref<?x?xf32, #strided>, memref<?x?xf32, #strided> } }

TODO: Investigate whether mixing implicit and explicit indices does not lead to losing information.

Definition at line 78 of file Tiling.cpp.

References mlir::detail::enumerate(), mlir::getAsOpFoldResult(), and offsetIndices().

Referenced by tileLinalgOpImpl().

◆ transposeBatchMatmul()

FailureOr< Operation * > mlir::linalg::transposeBatchMatmul	(	RewriterBase &	rewriter,
		linalg::BatchMatmulOp	batchMatmulOp,
		bool	transposeLHS = `true`
	)

Pattern to replace.

linalg.batch_matmul(a, b)

with

linalg.batch_matmul_transpose_a(linalg.transpose(a), b)

Only the non-batch dimensions are transposed. By default the LHS is transposed. Set transposeLHS=false to transpose RHS instead.

Definition at line 88 of file TransposeMatmul.cpp.

References mlir::OpBuilder::create(), mlir::Value::getType(), mlir::bufferization::hasTensorSemantics(), mlir::RewriterBase::notifyMatchFailure(), and mlir::RewriterBase::replaceOp().

◆ transposeConv2D() [1/2]

FailureOr< Operation * > mlir::linalg::transposeConv2D	(	RewriterBase &	rewriter,
		linalg::Conv2DNhwcFhwcOp	op
	)

Convert linalg.conv_2d_nhwc_fhwc(_q) to linalg.conv_2d_nhwc_hwcf(_q) by materializing transpose.

Definition at line 127 of file TransposeConv2D.cpp.

◆ transposeConv2D() [2/2]

FailureOr< Operation * > mlir::linalg::transposeConv2D	(	RewriterBase &	rewriter,
		linalg::Conv2DNhwcFhwcQOp	op
	)

Definition at line 134 of file TransposeConv2D.cpp.

◆ transposeMatmul()

FailureOr< Operation * > mlir::linalg::transposeMatmul	(	RewriterBase &	rewriter,
		linalg::MatmulOp	matmulOp,
		bool	transposeLHS = `true`
	)

Convert Linalg matmul ops to transposed variants.

Pattern to replace.

linalg.matmul(a, b)

with

linalg.matmul_transpose_a(linalg.transpose(a), b)

By default the LHS is transposed. Set transposeLHS=false to transpose RHS instead.

Definition at line 31 of file TransposeMatmul.cpp.

References mlir::OpBuilder::create(), mlir::Value::getType(), mlir::bufferization::hasTensorSemantics(), mlir::RewriterBase::notifyMatchFailure(), and mlir::RewriterBase::replaceOp().

◆ unrollIndex()

static SmallVector<Value> mlir::linalg::unrollIndex	(	OpBuilder &	b,
		Location	loc,
		Value	index,
		ArrayRef< int64_t >	factors
	)

static

Definition at line 55 of file ConvertConv2DToImg2Col.cpp.

References mlir::OpBuilder::create(), mlir::affine::delinearizeIndex(), and mlir::Builder::getIndexAttr().

Referenced by rewriteInIm2Col().

◆ updateBoundsForCyclicDistribution()

void mlir::linalg::updateBoundsForCyclicDistribution	(	OpBuilder &	builder,
		Location	loc,
		Value	procId,
		Value	nprocs,
		Value &	lb,
		Value &	ub,
		Value &	step
	)

Update the lb, ub and step to get per processor lb, ub and step.

Definition at line 351 of file Utils.cpp.

References mlir::bindDims(), mlir::getAffineSymbolExpr(), mlir::Builder::getContext(), and mlir::affine::makeComposedAffineApply().

Referenced by mlir::linalg::GenerateLoopNest< LoopTy >::doit().

◆ vectorize()

LogicalResult mlir::linalg::vectorize	(	RewriterBase &	rewriter,
		Operation *	op,
		ArrayRef< int64_t >	inputVectorSizes = `{}`,
		ArrayRef< bool >	inputScalableVecDims = `{}`,
		bool	vectorizeNDExtract = `false`,
		bool	flatten1DDepthwiseConv = `false`
	)

Emit a suitable vector form for an operation.

If provided, inputVectorSizes are used to vectorize this operation. inputVectorSizes must match the rank of the iteration space of the operation and the sizes must be smaller or equal than their counterpart interation space sizes, if static. inputVectorShapes also allows the vectorization of operations with dynamic shapes.

If provided, inputVectorSizes are used to vectorize this operation. inputVectorSizes must match the rank of the iteration space of the operation and the input vector sizes must be greater than or equal to their counterpart iteration space sizes, if static. inputVectorShapes also allows the vectorization of operations with dynamic shapes.

Definition at line 2176 of file Vectorization.cpp.

References LDBG, and vectorizeOpPrecondition().

◆ vectorizeCopy()

LogicalResult mlir::linalg::vectorizeCopy	(	RewriterBase &	builder,
		memref::CopyOp	copyOp
	)

Emit a suitable vector form for a Copy op with fully static shape.

Definition at line 2266 of file Vectorization.cpp.

References mlir::OpBuilder::create(), mlir::get(), mlir::getElementTypeOrSelf(), mlir::Builder::getMultiDimIdentityMap(), mlir::Operation::getResults(), mlir::sparse_tensor::detail::readValue(), and mlir::RewriterBase::replaceOp().

Referenced by mlir::linalg::CopyVectorizationPattern::matchAndRewrite().

◆ vectorizeOpPrecondition()

LogicalResult mlir::linalg::vectorizeOpPrecondition	(	Operation *	op,
		ArrayRef< int64_t >	inputVectorSizes = `{}`,
		ArrayRef< bool >	inputScalableVecDims = `{}`,
		bool	vectorizeNDExtract = `false`,
		bool	flatten1DDepthwiseConv = `false`
	)

Return success if the operation can be vectorized.

Definition at line 2120 of file Vectorization.cpp.

References hasVectorizationImpl(), vectorizeLinalgOpPrecondition(), vectorizePackOpPrecondition(), vectorizePadOpPrecondition(), vectorizeScalableVectorPrecondition(), and vectorizeUnPackOpPrecondition().

Referenced by vectorize().

◆ winogradConv2D()

FailureOr< Operation * > mlir::linalg::winogradConv2D	(	RewriterBase &	rewriter,
		linalg::Conv2DNhwcFhwcOp	op,
		int64_t	m,
		int64_t	r
	)

Convert linalg.conv_2d_nhwc_fhwc to Winograd Conv2D algorithm F(m x m, r x r).

m is the dimension size of output and r is the dimension size of filter.

Definition at line 1184 of file WinogradConv2D.cpp.

Namespaces

Classes

Typedefs

Enumerations

Functions

Typedef Documentation

◆ AllocBufferCallbackFn

◆ ControlBlockPackMatmulFn

◆ ControlFusionFn

◆ ControlPropagationFn

◆ ControlSplitReductionFn

◆ CopyCallbackFn

◆ DeallocBufferCallbackFn

◆ GetCollapsableDimensionsFn

◆ LinalgLoops

◆ LoopIndexToRangeIndexMap

◆ MeshAxis

◆ MeshOp

◆ MeshSharding

◆ OptimizeCopyFn

◆ ProcInfoCallBackFn

◆ ReductionKind

◆ ShardingArray

◆ TileSizeComputationFunction

Enumeration Type Documentation

◆ DistributionMethod

◆ LinalgTilingLoopType

Function Documentation

◆ allIndexingsAreProjectedPermutation()

◆ allocateGPUPrivateMemory()

◆ allocateWorkgroupMemory()

◆ areDimSequencesPreserved()

◆ areElementwiseOpsFusable()

◆ blockPackMatmul()

◆ bufferizeToAllocation() [1/4]

◆ bufferizeToAllocation() [2/4]

◆ bufferizeToAllocation() [3/4]

◆ bufferizeToAllocation() [4/4]

◆ collapseOpIterationDims()

◆ computeAllSliceParameters()

◆ computeContinuousTileSizes()

◆ computeMultiTileSizes()

◆ computeSliceParameters()

◆ computeStaticContinuousTileSizes()

◆ computeStaticMultiTileSizes()

◆ computeTileOffsets()

◆ computeTileSizes()

◆ concat()

◆ copyToGPUPrivateMemory()

◆ copyToWorkgroupMemory()

◆ createAdd()

◆ createAllReduceForResultsWithoutPartialShardings()

◆ createAllReduceForResultWithoutPartialSharding()

◆ createDestinationPassingStyleInitOperand()

◆ createDestinationPassingStyleInitOperands()

◆ createFoldedDimOp()

◆ createMul()

◆ createOrFoldDimOp()

◆ deallocateGPUPrivateMemory()

◆ deallocateWorkgroupMemory()

◆ decomposeWinogradFilterTransformOp()

◆ decomposeWinogradInputTransformOp()

◆ decomposeWinogradOutputTransformOp()

◆ dropUnitDims()

◆ extractOrIdentityMap()

◆ fuseElementwiseOps()

◆ fuseProducerOfTensor() [1/2]

◆ fuseProducerOfTensor() [2/2]

◆ generalizeNamedOp()

◆ generateLibraryCallName()

◆ generateParallelLoopNest()

◆ getCombinerOp()

◆ getCombinerOpKind()

◆ getConvolvedIndex()

◆ getLinalgTilingCanonicalizationPatterns()

◆ getMesh()

◆ getPreservedProducerResults()

◆ getPrunedAttributeList()

◆ getReassociationMapForFoldingUnitDims()

◆ getReductionKind()