sdfg::gpu::GpuMmaSupport Struct Reference

#include <sdfg/targets/gpu/gpu_mma_fragment.h>

Inheritance diagram for sdfg::gpu::GpuMmaSupport:

Public Member Functions
virtual	~GpuMmaSupport ()=default
virtual types::PrimitiveType	get_accumulator_type (types::PrimitiveType input_type, types::PrimitiveType output_type, types::PrimitiveType desired_acc_type) const =0
types::PrimitiveType	get_accumulator_type (types::PrimitiveType input_type, types::PrimitiveType output_type) const
	GpuMmaSupport ()=default
virtual bool	is_valid_block_size (const MmaBlockSize &block_size, types::PrimitiveType input_type, types::PrimitiveType acc_type) const =0
virtual data_flow::ImplementationType	get_mma_impl_type (const GpuMmaTiling &tiling) const =0
virtual bool	supported_types (types::PrimitiveType input_type, types::PrimitiveType output_type) const =0
virtual std::optional< GpuMmaTiling >	get_mma_tiling (const symbolic::MultiExpression &res_shape, types::PrimitiveType input_type, types::PrimitiveType acc_type, const MmaBlockSize *block_size_hint, const BlockCounts *wave_tile_hint) const =0
virtual std::optional< GpuMmaTiling >	get_ragged_mma_tiling (const symbolic::MultiExpression &padded_shape, types::PrimitiveType input_type, types::PrimitiveType output_type, types::PrimitiveType acc_type, const MmaBlockSize *block_size_hint, const BlockCounts *wave_tile_hint) const
	Tiling for a ragged problem: padded_shape is a constant upper bound of each extent that need not be a multiple of the MMA block.
virtual void	set_mma_fragment_storage_type (types::StorageType &storage_type, const GpuMmaTiling &tiling, MmaFragmentType type, MmaFragmentLayout layout) const =0
virtual std::vector< MmaBlockSize >	get_supported_block_sizes (types::PrimitiveType input_type, types::PrimitiveType acc_type) const =0

Static Public Member Functions
static int	get_integer_block_count (const symbolic::Expression &size, uint16_t block_size)
static int	get_ceil_block_count (const symbolic::Expression &size, uint16_t block_size)
	ceil(size / block_size) for a positive integer size, otherwise 0.

Static Protected Member Functions
static int	get_storage_type_arg_as_int (const types::StorageType &storage, int idx)

Constructor & Destructor Documentation

~GpuMmaSupport()

virtual sdfg::gpu::GpuMmaSupport::~GpuMmaSupport ( )

virtualdefault

GpuMmaSupport()

sdfg::gpu::GpuMmaSupport::GpuMmaSupport ( )

default

Member Function Documentation

get_accumulator_type() [1/2]

types::PrimitiveType sdfg::gpu::GpuMmaSupport::get_accumulator_type ( types::PrimitiveType input_type, types::PrimitiveType output_type ) const

inline

get_accumulator_type() [2/2]

virtual types::PrimitiveType sdfg::gpu::GpuMmaSupport::get_accumulator_type ( types::PrimitiveType input_type, types::PrimitiveType output_type, types::PrimitiveType desired_acc_type ) const

pure virtual

Implemented in sdfg::gpu::cuda::CudaMmaSupport, and sdfg::gpu::rocm::RocmMmaSupport.

get_ceil_block_count()

int sdfg::gpu::GpuMmaSupport::get_ceil_block_count ( const symbolic::Expression & size, uint16_t block_size )

static

ceil(size / block_size) for a positive integer size, otherwise 0.

get_integer_block_count()

int sdfg::gpu::GpuMmaSupport::get_integer_block_count ( const symbolic::Expression & size, uint16_t block_size )

static

get_mma_impl_type()

virtual data_flow::ImplementationType sdfg::gpu::GpuMmaSupport::get_mma_impl_type ( const GpuMmaTiling & tiling ) const

pure virtual

Implemented in sdfg::gpu::cuda::CudaMmaSupport, and sdfg::gpu::rocm::RocmMmaSupport.

get_mma_tiling()

virtual std::optional< GpuMmaTiling > sdfg::gpu::GpuMmaSupport::get_mma_tiling ( const symbolic::MultiExpression & res_shape, types::PrimitiveType input_type, types::PrimitiveType acc_type, const MmaBlockSize * block_size_hint, const BlockCounts * wave_tile_hint ) const

pure virtual

Implemented in sdfg::gpu::cuda::CudaMmaSupport, and sdfg::gpu::rocm::RocmMmaSupport.

get_ragged_mma_tiling()

virtual std::optional< GpuMmaTiling > sdfg::gpu::GpuMmaSupport::get_ragged_mma_tiling ( const symbolic::MultiExpression & padded_shape, types::PrimitiveType input_type, types::PrimitiveType output_type, types::PrimitiveType acc_type, const MmaBlockSize * block_size_hint, const BlockCounts * wave_tile_hint ) const

inlinevirtual

Tiling for a ragged problem: padded_shape is a constant upper bound of each extent that need not be a multiple of the MMA block.

Fragments crossing the exact extent are clipped (zero-filled loads, masked stores; see GpuMmaFromMemoryLayout::rows/cols). Targets without masked fragment transfers return nullopt.

Reimplemented in sdfg::gpu::rocm::RocmMmaSupport.

get_storage_type_arg_as_int()

int sdfg::gpu::GpuMmaSupport::get_storage_type_arg_as_int ( const types::StorageType & storage, int idx )

staticprotected

get_supported_block_sizes()

virtual std::vector< MmaBlockSize > sdfg::gpu::GpuMmaSupport::get_supported_block_sizes ( types::PrimitiveType input_type, types::PrimitiveType acc_type ) const

pure virtual

Implemented in sdfg::gpu::cuda::CudaMmaSupport, and sdfg::gpu::rocm::RocmMmaSupport.

is_valid_block_size()

virtual bool sdfg::gpu::GpuMmaSupport::is_valid_block_size ( const MmaBlockSize & block_size, types::PrimitiveType input_type, types::PrimitiveType acc_type ) const

pure virtual

Implemented in sdfg::gpu::cuda::CudaMmaSupport, and sdfg::gpu::rocm::RocmMmaSupport.

set_mma_fragment_storage_type()

virtual void sdfg::gpu::GpuMmaSupport::set_mma_fragment_storage_type ( types::StorageType & storage_type, const GpuMmaTiling & tiling, MmaFragmentType type, MmaFragmentLayout layout ) const

pure virtual

Implemented in sdfg::gpu::cuda::CudaMmaSupport, and sdfg::gpu::rocm::RocmMmaSupport.

supported_types()

virtual bool sdfg::gpu::GpuMmaSupport::supported_types ( types::PrimitiveType input_type, types::PrimitiveType output_type ) const

pure virtual

Implemented in sdfg::gpu::cuda::CudaMmaSupport, and sdfg::gpu::rocm::RocmMmaSupport.

---

The documentation for this struct was generated from the following file:

• opt/include/sdfg/targets/gpu/gpu_mma_fragment.h