A structured access described by a full affine tensor layout (shape, strides and offset in elements). More...

#include <sdfg/data_flow/pointer_metadata.h>

Inheritance diagram for sdfg::data_flow::TensorLayoutPattern:

Public Member Functions
	TensorLayoutPattern (math::tensor::TensorLayout layout, bool not_sparse=false)
bool	empty () const override
	There are no elements in this pattern = no accesses of this type.
bool	every_element_accessed () const override
	Every element that is part of this pattern is actually accessed.
const math::tensor::TensorLayout *	layout () const override
	The structured affine layout (shape/strides/offset, in elements) backing this pattern, or nullptr when the pattern carries no explicit layout (a flat convex span or no access).
void	replace (const symbolic::Expression old_expression, const symbolic::Expression new_expression) override
void	replace (const symbolic::ExpressionMapping &replacements) override
MemoryAccessPatternType	clone () const override
void	serialize_to_json (nlohmann::json &entry) override
Public Member Functions inherited from sdfg::data_flow::MemoryAccessPattern
virtual	~MemoryAccessPattern ()=default
MemoryAccessPatternType	ref () const

Static Public Member Functions
static MemoryAccessPatternType	create (math::tensor::TensorLayout layout, bool not_sparse=false)

Detailed Description

A structured access described by a full affine tensor layout (shape, strides and offset in elements).

Unlike ConvexAccessPattern, which only bounds a flat convex span, this preserves the operand's real multi-dimensional shape so a consumer can reconstruct exactly which elements are touched (e.g. a padded GEMM/MatMul operand).

Constructor & Destructor Documentation

◆ TensorLayoutPattern()

sdfg::data_flow::TensorLayoutPattern::TensorLayoutPattern	(	math::tensor::TensorLayout	layout,
		bool	not_sparse = false )

inline

Member Function Documentation

◆ clone()

MemoryAccessPatternType sdfg::data_flow::TensorLayoutPattern::clone ( ) const

overridevirtual

Implements sdfg::data_flow::MemoryAccessPattern.

◆ create()

MemoryAccessPatternType sdfg::data_flow::TensorLayoutPattern::create	(	math::tensor::TensorLayout	layout,
		bool	not_sparse = false )

static

◆ empty()

bool sdfg::data_flow::TensorLayoutPattern::empty ( ) const

inlineoverridevirtual

There are no elements in this pattern = no accesses of this type.

Implements sdfg::data_flow::MemoryAccessPattern.

◆ every_element_accessed()

bool sdfg::data_flow::TensorLayoutPattern::every_element_accessed ( ) const

inlineoverridevirtual

Every element that is part of this pattern is actually accessed.

If false, this is just an approximation that may contain elements in between the bounds that are not guaranteed to be accessed Ex. indicating write of a[0] and a[9] as ConvexPattern(10) is approximate. Writes a[0], a[1], .. [9] would be every element being accessed. Bounds should never be violated regardless. We may in the future

Implements sdfg::data_flow::MemoryAccessPattern.

◆ layout()

const math::tensor::TensorLayout * sdfg::data_flow::TensorLayoutPattern::layout ( ) const

inlineoverridevirtual

The structured affine layout (shape/strides/offset, in elements) backing this pattern, or nullptr when the pattern carries no explicit layout (a flat convex span or no access).

Only TensorLayoutPattern provides one.

Reimplemented from sdfg::data_flow::MemoryAccessPattern.