A single GPU tensor-core matrix-multiply-accumulate (MMA) step. More...

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

Inheritance diagram for sdfg::gpu::GpuMmaMatmulNode:

Public Member Functions
	GpuMmaMatmulNode (size_t element_id, const DebugInfo &debug_info, const graph::Vertex vertex, data_flow::DataFlowGraph &parent, const MmaBlockSize &mma_block_size, types::PrimitiveType input_type, types::PrimitiveType acc_type, const data_flow::ImplementationType &impl_type)
const MmaBlockSize &	mma_block_size () const
types::PrimitiveType	input_type () const
types::PrimitiveType	acc_type () const
symbolic::Expression	m () const
	M dimension (rows of A / rows of the accumulator).
symbolic::Expression	n () const
	N dimension (columns of B / columns of the accumulator).
symbolic::Expression	k () const
	K dimension (contraction dimension of A and B for a single MMA step).
void	validate (const Function &function) const override
	Validate the code node's configuration.
symbolic::SymbolSet	symbols () const override
	Get all symbols used in this node.
void	replace (const symbolic::Expression old_expression, const symbolic::Expression new_expression) override
void	replace (const symbolic::ExpressionMapping &replacements) override
std::unique_ptr< data_flow::DataFlowNode >	clone (size_t element_id, const graph::Vertex vertex, data_flow::DataFlowGraph &parent) const override
	Clone this node for graph transformations.
std::string	toStr () const override
	Convert node to string representation.
symbolic::Expression	flop () const override
	Calculate floating point operations for this node.
data_flow::PointerAccessType	pointer_access_type (int input_idx) const override
	Describes what a pointer is used for.
Public Member Functions inherited from sdfg::data_flow::LibraryNode
	LibraryNode (const LibraryNode &data_node)=delete
LibraryNode &	operator= (const LibraryNode &)=delete
virtual	~LibraryNode ()=default
ElementType	type_id () const override
	Returns the type of the element.
const LibraryNodeCode &	code () const
	Get the operation code.
const ImplementationType &	implementation_type () const
	Get the implementation type (const).
ImplementationType &	implementation_type ()
	Get the implementation type (mutable).
void	set_implementation_type (const ImplementationType &impl_type)
bool	side_effect () const override
	Check if this node has side effects.
PointerAccessType	pointer_access_type (const Memlet &edge) const
virtual bool	relocalize_operand (int input_idx, const math::tensor::TensorLayout &packed)
	Repoint the operand at `input_idx` to a locally staged buffer.
virtual bool	can_relocalize_operand (int input_idx, const math::tensor::TensorLayout &packed) const
	Side-effect-free predicate mirroring relocalize_operand: whether the operand at `input_idx` can be repointed at the buffer view `packed`.
virtual std::optional< SubgroupAccess >	operand_access (int input_idx, unsigned subgroup_size) const
	The elements one subgroup of `subgroup_size` lanes reads from the operand at `input_idx`, if the node declares them.
EdgeRemoveOption	can_remove_out_edge (const data_flow::DataFlowGraph &graph, const Memlet *memlet) const override
	Is the edge in question removable, or will this make the node its an output on invalid?
EdgeRemoveOption	can_remove_in_edge (const data_flow::DataFlowGraph &graph, const Memlet *memlet) const override
bool	pointer_use_creates_side_effects (const DataFlowGraph &dataflow, const Function &func)
Public Member Functions inherited from sdfg::data_flow::CodeNode
	CodeNode (const CodeNode &data_node)=delete
CodeNode &	operator= (const CodeNode &)=delete
const std::vector< std::string > &	outputs () const
	Get output connector names (const).
const std::vector< std::string > &	inputs () const
	Get input connector names (const).
std::vector< std::string > &	outputs ()
	Get output connector names (mutable).
std::vector< std::string > &	inputs ()
	Get input connector names (mutable).
const std::string &	output (size_t index) const
	Get output connector name by index.
const std::string &	input (size_t index) const
	Get input connector name by index.
bool	has_constant_input (size_t index) const
	Check if an input is connected to a constant node.
Public Member Functions inherited from sdfg::data_flow::DataFlowNode
	DataFlowNode (const DataFlowNode &data_node)=delete
DataFlowNode &	operator= (const DataFlowNode &)=delete
graph::Vertex	vertex () const
	Get the graph vertex for this node.
const DataFlowGraph &	get_parent () const
	Get the parent dataflow graph (const).
DataFlowGraph &	get_parent ()
	Get the parent dataflow graph (mutable).
virtual bool	update_edge_removed (const std::string &out_conn)
Public Member Functions inherited from sdfg::Element
	Element (size_t element_id, const DebugInfo &debug_info)
virtual	~Element ()=default
size_t	element_id () const
const DebugInfo &	debug_info () const
void	set_debug_info (const DebugInfo &debug_info)
void	add_metadata (const std::string &key, const std::string &value)
void	remove_metadata (const std::string &key)
const std::string &	metadata (const std::string &key) const
const std::string *	metadata_if_exists (const std::string &key) const
const std::unordered_map< std::string, std::string > &	metadata () const
std::string_view	element_type () const
	Returns the type of the element as a human-readable string.

Static Public Attributes
static constexpr auto	Y_INPUT_IDX = 0
static constexpr auto	A_INPUT_IDX = 1
static constexpr auto	B_INPUT_IDX = 2
Static Public Attributes inherited from sdfg::data_flow::CodeNode
static constexpr ElementType	TypeGroup = ElementType::Tasklet \| ElementType::LibraryNode
	Bitmask of all concrete ElementTypes that are CodeNodes.
Static Public Attributes inherited from sdfg::data_flow::DataFlowNode
static constexpr ElementType	TypeGroup
	Bitmask of all concrete ElementTypes that are DataFlowNodes.

Additional Inherited Members
Static Public Member Functions inherited from sdfg::data_flow::LibraryNode
static bool	classof (const Element &element)
Static Public Member Functions inherited from sdfg::data_flow::CodeNode
static bool	classof (const Element &element)
	LLVM-style RTTI predicate: true if `element` is a CodeNode.
Static Public Member Functions inherited from sdfg::data_flow::DataFlowNode
static bool	classof (const Element &element)
	LLVM-style RTTI predicate: true if `element` is a DataFlowNode.
Protected Member Functions inherited from sdfg::data_flow::LibraryNode
	LibraryNode (size_t element_id, const DebugInfo &debug_info, const graph::Vertex vertex, DataFlowGraph &parent, const LibraryNodeCode &code, const std::vector< std::string > &outputs, const std::vector< std::string > &inputs, const bool side_effect, const ImplementationType &implementation_type)
	Protected constructor for library nodes.
Protected Member Functions inherited from sdfg::data_flow::CodeNode
	CodeNode (size_t element_id, const DebugInfo &debug_info, const graph::Vertex vertex, DataFlowGraph &parent, const std::vector< std::string > &outputs, const std::vector< std::string > &inputs)
	Protected constructor for code nodes.
Protected Member Functions inherited from sdfg::data_flow::DataFlowNode
	DataFlowNode (size_t element_id, const DebugInfo &debug_info, const graph::Vertex vertex, DataFlowGraph &parent)
	Protected constructor for dataflow nodes.
Protected Attributes inherited from sdfg::data_flow::LibraryNode
LibraryNodeCode	code_
	Operation identifier (e.g., "GEMM", "Add").
bool	side_effect_
	Whether this node has side effects.
ImplementationType	implementation_type_
	How to implement this node.
Protected Attributes inherited from sdfg::data_flow::CodeNode
std::vector< std::string >	outputs_
	Names of output connectors.
std::vector< std::string >	inputs_
	Names of input connectors.
Protected Attributes inherited from sdfg::Element
size_t	element_id_
DebugInfo	debug_info_
std::unordered_map< std::string, std::string >	metadata_

Detailed Description

A single GPU tensor-core matrix-multiply-accumulate (MMA) step.

Unlike math::tensor::MatMulNode, this is a code-generation leaf node dedicated to MMA usage. It performs one Y += A * B MMA on mma fragments A and B into the accumulator behind its Y input, which is expected to be a preinitialized accumulator fragment (see GpuMmaFillNode).

It has no K loop (iterating K is done externally), does not add C, and does not store the result back. Y is an input treated like a pointer to the accumulator storage. So

  Decl(fragAcc)
  GpuMmaFillNode(fragAcc, 0.0, acc_type)
  for (int k = 0; k < K; k += mma_k_size) {
    Decl(fragA, fragB)
    GpuMmaLoad(fragB, A+offset, ...layout)
    GpuMmaLoad(fragA, B+offset, ...layout)
    GpuMmaNode(fragA, fragB, fragAcc){mma_block_dims, input_type, acc_type}
  }
  // Load tile of C, add onto it, store fragAcc to ptr & write back to C (can load C into another frag add as frags
or write acc to ptr and add manually there)

would achieve the same as an entire MMA Matmul before

Currently reuses TensorLayouts for convenience, but is very strict. They need to have immediate shape, shape must be divisible by mma_block sizes and strides must follow gemm-semantics (there can be one stride that is customizable, the other must be 1 to represent either col-major or row-major with leading dimension). Types must be supported by the underlying hardware

Constructor & Destructor Documentation

◆ GpuMmaMatmulNode()

sdfg::gpu::GpuMmaMatmulNode::GpuMmaMatmulNode	(	size_t	element_id,
		const DebugInfo &	debug_info,
		const graph::Vertex	vertex,
		data_flow::DataFlowGraph &	parent,
		const MmaBlockSize &	mma_block_size,
		types::PrimitiveType	input_type,
		types::PrimitiveType	acc_type,
		const data_flow::ImplementationType &	impl_type )

Member Function Documentation

◆ acc_type()

types::PrimitiveType sdfg::gpu::GpuMmaMatmulNode::acc_type ( ) const

inline

◆ clone()

std::unique_ptr< data_flow::DataFlowNode > sdfg::gpu::GpuMmaMatmulNode::clone	(	size_t	element_id,
		const graph::Vertex	vertex,
		data_flow::DataFlowGraph &	parent ) const

overridevirtual

Clone this node for graph transformations.

Parameters

element_id	New element identifier for the clone
vertex	New graph vertex for the clone
parent	Parent graph for the clone

Returns: Unique pointer to the cloned node

Pure virtual function that must be implemented by derived classes to support graph transformations and optimizations.

Implements sdfg::data_flow::DataFlowNode.

◆ flop()

symbolic::Expression sdfg::gpu::GpuMmaMatmulNode::flop ( ) const

overridevirtual

Calculate floating point operations for this node.

Returns: Symbolic expression for FLOP count (or null)

Reimplemented from sdfg::data_flow::LibraryNode.

◆ input_type()

types::PrimitiveType sdfg::gpu::GpuMmaMatmulNode::input_type ( ) const

inline

◆ k()

symbolic::Expression sdfg::gpu::GpuMmaMatmulNode::k ( ) const

inline

K dimension (contraction dimension of A and B for a single MMA step).

◆ m()

symbolic::Expression sdfg::gpu::GpuMmaMatmulNode::m ( ) const

inline

M dimension (rows of A / rows of the accumulator).

◆ mma_block_size()

const MmaBlockSize & sdfg::gpu::GpuMmaMatmulNode::mma_block_size ( ) const

inline

◆ n()

symbolic::Expression sdfg::gpu::GpuMmaMatmulNode::n ( ) const

inline

N dimension (columns of B / columns of the accumulator).

◆ pointer_access_type()

data_flow::PointerAccessType sdfg::gpu::GpuMmaMatmulNode::pointer_access_type ( int input_idx ) const

overridevirtual

Describes what a pointer is used for.

Parameters

input_idx index of input that is a pointer.

Returns: Invalid if not asked about a pointer input

Reimplemented from sdfg::data_flow::LibraryNode.

◆ replace() [1/2]

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

overridevirtual

Implements sdfg::Element.

◆ replace() [2/2]

void sdfg::gpu::GpuMmaMatmulNode::replace ( const symbolic::ExpressionMapping & replacements )

overridevirtual

Reimplemented from sdfg::Element.

◆ symbols()

symbolic::SymbolSet sdfg::gpu::GpuMmaMatmulNode::symbols ( ) const

overridevirtual

Get all symbols used in this node.

Returns: Set of symbolic expressions used by this node

Implements sdfg::data_flow::LibraryNode.

◆ toStr()

std::string sdfg::gpu::GpuMmaMatmulNode::toStr ( ) const

overridevirtual

Convert node to string representation.

Returns: String representation of the node

Reimplemented from sdfg::data_flow::LibraryNode.

◆ validate()

void sdfg::gpu::GpuMmaMatmulNode::validate ( const Function & function ) const

overridevirtual

Validate the code node's configuration.

Parameters

function Function context for validation

Ensures that:

No two access nodes connect to the same data for inputs/outputs

Reimplemented from sdfg::data_flow::CodeNode.

Member Data Documentation

◆ A_INPUT_IDX

auto sdfg::gpu::GpuMmaMatmulNode::A_INPUT_IDX = 1

staticconstexpr

◆ B_INPUT_IDX

auto sdfg::gpu::GpuMmaMatmulNode::B_INPUT_IDX = 2

staticconstexpr

◆ Y_INPUT_IDX

auto sdfg::gpu::GpuMmaMatmulNode::Y_INPUT_IDX = 0

staticconstexpr

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

opt/include/sdfg/targets/gpu/gpu_mma_matmul_node.h

Public Member Functions

Static Public Attributes

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ GpuMmaMatmulNode()

Member Function Documentation

◆ acc_type()

◆ clone()

◆ flop()

◆ input_type()

◆ k()

◆ m()

◆ mma_block_size()

◆ n()

◆ pointer_access_type()

◆ replace() [1/2]

◆ replace() [2/2]

◆ symbols()

◆ toStr()

◆ validate()

Member Data Documentation

◆ A_INPUT_IDX

◆ B_INPUT_IDX

◆ Y_INPUT_IDX