sdfg::gpu::GpuMmaFromMemoryLayout Struct Reference

Describes how a single MMA fragment maps to memory relative to a base pointer. More...

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

Public Member Functions
bool	clipped () const
std::string	toStr () const
void	collect_symbols (symbolic::SymbolSet &syms) const
void	replace (const symbolic::Expression old_expression, const symbolic::Expression new_expression)
void	replace (const symbolic::ExpressionMapping &replacements)
std::optional< math::tensor::TensorLayout >	to_tensor_layout (const MmaBlockSize &block_size, MmaFragmentType frag) const
	The accessed tile; a clipped fragment reports its clipped (in-bounds) extents min(block, rows/cols).
bool	matches_access_shape (const symbolic::MultiExpression &shape, const MmaBlockSize &block_size, MmaFragmentType frag) const
	Whether shape is this fragment's block shape or its clipped shape (see to_tensor_layout).

Static Public Member Functions
static std::optional< GpuMmaFromMemoryLayout >	from_tensor_layout (const math::tensor::TensorLayout &layout)

Public Attributes
symbolic::Expression	offset
	element offset added to the base pointer
symbolic::Expression	ldstride
	leading-dimension stride
MmaFragmentLayout	layout
	memory layout of the tile
symbolic::Expression	rows = SymEngine::null
	Clip of a ragged fragment: only logical rows r < rows and columns c < cols (from the fragment origin) are in bounds; loads zero-fill the rest, stores skip it.
symbolic::Expression	cols = SymEngine::null

Detailed Description

Describes how a single MMA fragment maps to memory relative to a base pointer.

Shared by the fragment load/store nodes. The tile it describes must follow GEMM semantics: one customizable leading-dimension stride, the other dimension contiguous (row- or column-major).

Member Function Documentation

clipped()

bool sdfg::gpu::GpuMmaFromMemoryLayout::clipped ( ) const

inline

collect_symbols()

void sdfg::gpu::GpuMmaFromMemoryLayout::collect_symbols

(

symbolic::SymbolSet &

syms

)

const

from_tensor_layout()

std::optional< GpuMmaFromMemoryLayout > sdfg::gpu::GpuMmaFromMemoryLayout::from_tensor_layout ( const math::tensor::TensorLayout & layout )

static

matches_access_shape()

bool sdfg::gpu::GpuMmaFromMemoryLayout::matches_access_shape	(	const symbolic::MultiExpression &	shape,
		const MmaBlockSize &	block_size,
		MmaFragmentType	frag ) const

Whether shape is this fragment's block shape or its clipped shape (see to_tensor_layout).

replace() [1/2]

void sdfg::gpu::GpuMmaFromMemoryLayout::replace	(	const symbolic::Expression	old_expression,
		const symbolic::Expression	new_expression )

replace() [2/2]

void sdfg::gpu::GpuMmaFromMemoryLayout::replace

(

const symbolic::ExpressionMapping &

replacements

)

to_tensor_layout()

std::optional< math::tensor::TensorLayout > sdfg::gpu::GpuMmaFromMemoryLayout::to_tensor_layout ( const MmaBlockSize & block_size, MmaFragmentType frag ) const

nodiscard

The accessed tile; a clipped fragment reports its clipped (in-bounds) extents min(block, rows/cols).

toStr()

std::string sdfg::gpu::GpuMmaFromMemoryLayout::toStr

(

)

const

Member Data Documentation

cols

symbolic::Expression sdfg::gpu::GpuMmaFromMemoryLayout::cols = SymEngine::null

layout

MmaFragmentLayout sdfg::gpu::GpuMmaFromMemoryLayout::layout

memory layout of the tile

ldstride

symbolic::Expression sdfg::gpu::GpuMmaFromMemoryLayout::ldstride

leading-dimension stride

offset

symbolic::Expression sdfg::gpu::GpuMmaFromMemoryLayout::offset

element offset added to the base pointer

rows

symbolic::Expression sdfg::gpu::GpuMmaFromMemoryLayout::rows = SymEngine::null

Clip of a ragged fragment: only logical rows r < rows and columns c < cols (from the fragment origin) are in bounds; loads zero-fill the rest, stores skip it.

Null = the whole block is in bounds. Values above the block extent or below zero are allowed (fully inside / fully outside).

---

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

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