Namespaces
	amd

	index_lowering

Classes
class	OffloadingTranslationAttrTrait
	This class indicates that the attribute associated with this trait is a GPU offloading translation attribute. More...

class	TargetOptions
	This class serves as an opaque interface for passing options to the `TargetAttrInterface` methods. More...

struct	KernelDim3
	Utility class for the GPU dialect to represent triples of `Value`s accessible through `.x`, `.y`, and `.z` similarly to CUDA notation. More...

class	AsyncTokenType

struct	MMAMatrixStorageType
	MMAMatrixType storage and uniquing. More...

class	MMAMatrixType
	MMAMatrix represents a matrix held by a subgroup for matrix-matrix multiply accumulate operations. More...

class	SparseDnTensorHandleType

class	SparseSpMatHandleType

class	SparseSpGEMMOpHandleType

struct	GPUToNVVMPipelineOptions
	Options for the gpu to nvvm pipeline. More...

struct	WarpDistributionPattern

Enumerations
enum class	SparseHandleKind { SpMat , DnTensor , SpGEMMOp }

Functions
void	registerConvertGpuToLLVMInterface (DialectRegistry &registry)
	Registers the `ConvertToLLVMOpInterface` interface on the `gpu::GPUModuleOP` operation. More...

void	addAsyncDependency (Operation *op, Value token)

void	registerValueBoundsOpInterfaceExternalModels (DialectRegistry &registry)

void	buildLowerToNVVMPassPipeline (OpPassManager &pm, const GPUToNVVMPipelineOptions &options)
	Adds the GPU to NVVM pipeline to the given pass manager. More...

void	registerGPUToNVVMPipeline ()
	Register all pipeleines for the `gpu` dialect. More...

void	registerTransformDialectExtension (DialectRegistry &registry)

void	registerBufferDeallocationOpInterfaceExternalModels (DialectRegistry &registry)

StringRef	getMappingAttrName ()
	Name of the mapping attribute produced by loop mappers. More...

LogicalResult	setMappingAttr (scf::ParallelOp ploopOp, ArrayRef< ParallelLoopDimMappingAttr > mapping)
	Sets the mapping attribute of a scf.parallel operation. More...

LogicalResult	transformGpuModulesToBinaries (Operation *op, OffloadingLLVMTranslationAttrInterface handler=nullptr, const gpu::TargetOptions &options={})
	Searches for all GPU modules in `op` and transforms them into GPU binary operations. More...

vector::CombiningKind	convertReductionKind (gpu::AllReduceOperation mode)
	Returns the matching vector combining kind. More...

void	registerOffloadingLLVMTranslationInterfaceExternalModels (mlir::DialectRegistry &registry)
	Registers the offloading LLVM translation interfaces for `gpu.select_object`. More...

static MappingLevel &	operator++ (MappingLevel &mappingLevel)
	Bounded increment on MappingLevel. More...

static Processor	getHardwareIdForMapping (MappingLevel level, int dimension)
	Computed the hardware id to use for a given mapping level. More...

static void	mapParallelOp (ParallelOp parallelOp, MappingLevel mappingLevel=MapGrid)
	Add mapping information to the given parallel loop. More...

Variables
static constexpr int	kNumHardwareIds = 3

Enumeration Type Documentation

◆ SparseHandleKind

enum mlir::gpu::SparseHandleKind

strong

Enumerator
SpMat
DnTensor
SpGEMMOp

Definition at line 176 of file GPUDialect.h.

Function Documentation

◆ addAsyncDependency()

void mlir::gpu::addAsyncDependency	(	Operation *	op,
		Value	token
	)

Definition at line 669 of file GPUDialect.cpp.

References mlir::Operation::getContext(), mlir::Builder::getDenseI32ArrayAttr(), mlir::OpTrait::AttrSizedOperandSegments< ConcreteType >::getOperandSegmentSizeAttr(), mlir::Operation::insertOperands(), and mlir::Operation::setAttr().

◆ buildLowerToNVVMPassPipeline()

void mlir::gpu::buildLowerToNVVMPassPipeline	(	OpPassManager &	pm,
		const GPUToNVVMPipelineOptions &	options
	)

Adds the GPU to NVVM pipeline to the given pass manager.

Transforms main dialects into NVVM targets. Begins with GPU code regions, then handles host code.

Definition at line 107 of file GPUToNVVMPipeline.cpp.

References options.

Referenced by registerGPUToNVVMPipeline().

◆ convertReductionKind()

vector::CombiningKind mlir::gpu::convertReductionKind ( gpu::AllReduceOperation mode )

Returns the matching vector combining kind.

Definition at line 18 of file Utils.cpp.

References MAP_CASE, and MINUI.

◆ getHardwareIdForMapping()

static Processor mlir::gpu::getHardwareIdForMapping	(	MappingLevel	level,
		int	dimension
	)

static

Computed the hardware id to use for a given mapping level.

Will assign x,y and z hardware ids for the first 3 dimensions and use sequential after. TODO: Make this use x for the inner-most loop that is distributed to map to x, the next innermost to y and the next innermost to z.

Definition at line 74 of file ParallelLoopMapper.cpp.

References kNumHardwareIds.

Referenced by mapParallelOp().

◆ getMappingAttrName()

StringRef mlir::gpu::getMappingAttrName ( )

Name of the mapping attribute produced by loop mappers.

Definition at line 31 of file ParallelLoopMapper.cpp.

Referenced by mlir::configureParallelLoopToGPULegality(), mapParallelOp(), and processParallelLoop().

◆ mapParallelOp()

static void mlir::gpu::mapParallelOp	(	ParallelOp	parallelOp,
		MappingLevel	mappingLevel = `MapGrid`
	)

static

Add mapping information to the given parallel loop.

Do not add mapping information if the loop already has it. Also, don't start a mapping at a nested loop.

Definition at line 110 of file ParallelLoopMapper.cpp.

References mlir::Builder::getAttr(), mlir::Builder::getDimIdentityMap(), getHardwareIdForMapping(), getMappingAttrName(), and setMappingAttr().

◆ operator++()

static MappingLevel& mlir::gpu::operator++ ( MappingLevel & mappingLevel )

static

Bounded increment on MappingLevel.

Increments to the next level unless Sequential was already reached.

Definition at line 61 of file ParallelLoopMapper.cpp.

◆ registerBufferDeallocationOpInterfaceExternalModels()

void mlir::gpu::registerBufferDeallocationOpInterfaceExternalModels ( DialectRegistry & registry )

Definition at line 32 of file BufferDeallocationOpInterfaceImpl.cpp.

References mlir::DialectRegistry::addExtension().

Referenced by mlir::registerAllDialects().

◆ registerConvertGpuToLLVMInterface()

void mlir::gpu::registerConvertGpuToLLVMInterface ( DialectRegistry & registry )

Registers the ConvertToLLVMOpInterface interface on the gpu::GPUModuleOP operation.

Definition at line 1858 of file GPUToLLVMConversion.cpp.

References mlir::DialectRegistry::addExtension().

Referenced by mlir::registerAllExtensions().

◆ registerGPUToNVVMPipeline()

void mlir::gpu::registerGPUToNVVMPipeline ( )

Register all pipeleines for the gpu dialect.

Definition at line 119 of file GPUToNVVMPipeline.cpp.

References buildLowerToNVVMPassPipeline().

Referenced by mlir::registerAllPasses().

◆ registerOffloadingLLVMTranslationInterfaceExternalModels()

void mlir::gpu::registerOffloadingLLVMTranslationInterfaceExternalModels ( mlir::DialectRegistry & registry )

Registers the offloading LLVM translation interfaces for gpu.select_object.

Definition at line 468 of file SelectObjectAttr.cpp.

References mlir::DialectRegistry::addExtension().

Referenced by mlir::registerAllGPUToLLVMIRTranslations(), and mlir::registerAllToLLVMIRTranslations().

◆ registerTransformDialectExtension()

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

Definition at line 938 of file GPUTransformOps.cpp.

References mlir::DialectRegistry::addExtensions().

Referenced by mlir::registerAllExtensions().

◆ registerValueBoundsOpInterfaceExternalModels()

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

Definition at line 93 of file ValueBoundsOpInterfaceImpl.cpp.

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

Referenced by mlir::registerAllDialects().

◆ setMappingAttr()

LogicalResult mlir::gpu::setMappingAttr	(	scf::ParallelOp	ploopOp,
		ArrayRef< ParallelLoopDimMappingAttr >	mapping
	)

Sets the mapping attribute of a scf.parallel operation.

Verifies that the mapping passed is valid.

the number of DimMapperAttr provided is same as the number of loops of the ploopOp.
the mapping does not map multiple loops to the same processor.

Referenced by mapParallelOp().

◆ transformGpuModulesToBinaries()

LogicalResult mlir::gpu::transformGpuModulesToBinaries	(	Operation *	op,
		OffloadingLLVMTranslationAttrInterface	handler = `nullptr`,
		const gpu::TargetOptions &	options = `{}`
	)

Searches for all GPU modules in op and transforms them into GPU binary operations.

The resulting gpu.binary has handler as its offloading handler attribute.

Definition at line 125 of file ModuleToBinary.cpp.

References mlir::Operation::getRegions().

Variable Documentation

◆ kNumHardwareIds

constexpr int mlir::gpu::kNumHardwareIds = 3

staticconstexpr

Definition at line 57 of file ParallelLoopMapper.cpp.

Referenced by getHardwareIdForMapping().

Namespaces

Classes

Enumerations

Functions

Variables

Enumeration Type Documentation

◆ SparseHandleKind

Function Documentation

◆ addAsyncDependency()

◆ buildLowerToNVVMPassPipeline()

◆ convertReductionKind()

◆ getHardwareIdForMapping()

◆ getMappingAttrName()

◆ mapParallelOp()

◆ operator++()

◆ registerBufferDeallocationOpInterfaceExternalModels()

◆ registerConvertGpuToLLVMInterface()

◆ registerGPUToNVVMPipeline()

◆ registerOffloadingLLVMTranslationInterfaceExternalModels()

◆ registerTransformDialectExtension()

◆ registerValueBoundsOpInterfaceExternalModels()

◆ setMappingAttr()

◆ transformGpuModulesToBinaries()

Variable Documentation

◆ kNumHardwareIds