sdfg::data_flow Namespace Reference

Classes
class	AccessNode
	Node representing access to a data container in the dataflow graph. More...
class	ConstantNode
	Node representing a constant literal value in the dataflow graph. More...
class	CodeNode
	Base class for computational nodes with inputs and outputs. More...
class	DataFlowGraph
class	DataFlowNode
	Abstract base class for all dataflow graph nodes. More...
struct	SubgroupAccess
	The operand elements one subgroup touches in a single access: for every lane, the operand coordinates it reads, in register order. More...
class	LibraryNode
	Base class for all library nodes in the dataflow graph. More...
class	AtomicScalarOpImpl
	A concrete lowering ("implementation") of an AtomicScalarOpNode. More...
class	AtomicScalarOpNode
	Wrapper around typical atomic operations (add, sub, min, max, swap) on a single scalar value. More...
class	AtomicScalarOpCPUImpl
	CPU (OpenMP) atomic implementation. More...
class	AtomicScalarOpGPUImpl
	Shared base for the GPU implementations (CUDA and ROCm). More...
class	AtomicScalarOpCudaImpl
	CUDA atomic implementation. More...
class	AtomicScalarOpRocmImpl
	ROCm atomic implementation. More...
class	AtomicScalarOpNodeSerializer
class	AtomicScalarOpNodeDispatcher
class	AtomicScalarOpCPUNodeDispatcher
class	AtomicScalarOpGPUNodeDispatcher
class	BarrierLocalNode
class	BarrierLocalNodeSerializer
class	CallNode
class	CallNodeSerializer
class	CallNodeDispatcher
class	InvokeNode
class	InvokeNodeSerializer
class	InvokeNodeDispatcher
class	ConstSource
class	InMemoryConstSource
	Constant data stored inline in SDFG. More...
class	RawFileConstSource
	Constant data stored in external files, using arg_capture_io code to read (raw data, no meta data inside the files themselves, meta data stored here). More...
class	LoadConstNode
class	LoadConstNodeSerializer
class	LoadConstNodeDispatcher
class	MetadataNode
class	MetadataNodeSerializer
class	MetadataDispatcher
class	Memlet
	Edge in the dataflow graph representing data movement. More...
struct	PtrMetaDeleter
class	MemoryAccessPattern
class	ConvexAccessPattern
class	NoAccessPattern
class	TensorLayoutPattern
	A structured access described by a full affine tensor layout (shape, strides and offset in elements). More...
class	PointerAccessMeta
class	PointerReadOnly
	The pointer is only used for reading. More...
class	PointerFullWriteOnly
	The pointer is used solely to write, no reading and over all of the data. More...
class	PointerGenericAccess
	It is unknown what is done with this pointer or it is a mix of reads and writes. More...
class	PointerInvalidate
	Meaning the underlying memory will be deallocated and use of the pointer after this is no longer valid. More...
class	PointerAccessMetaSerializer
class	Tasklet
	Simple computational operation in the dataflow graph. More...

Typedefs
typedef StringEnum	LibraryNodeCode
typedef StringEnum	ImplementationType
	Implementation type identifier for library nodes.
typedef symbolic::MultiExpression	Subset
	Type alias for a subset specification (array of symbolic expressions).
typedef PtrMetaDeleter< PointerAccessMeta >	PtrAccessDeleter
typedef PtrMetaDeleter< MemoryAccessPattern >	AccessPatternsDeleter
typedef std::unique_ptr< PointerAccessMeta, PtrAccessDeleter >	PointerAccessType
typedef std::unique_ptr< MemoryAccessPattern, AccessPatternsDeleter >	MemoryAccessPatternType

Enumerations
enum class	EdgeRemoveOption { NotRemovable , Trivially , RequiresUpdate , RemoveNodeAfter }
enum class	AtomicOpType { Add , Subtract , Min , Max }
enum	MemletType { Computational , Reference , Dereference_Src , Dereference_Dst }
	Type of data movement represented by a memlet. More...
enum	TaskletCode {   assign , fp_neg , fp_add , fp_sub ,   fp_mul , fp_div , fp_rem , fp_fma ,   fp_oeq , fp_one , fp_oge , fp_ogt ,   fp_ole , fp_olt , fp_ord , fp_ueq ,   fp_une , fp_ugt , fp_uge , fp_ult ,   fp_ule , fp_uno , int_add , int_sub ,   int_mul , int_sdiv , int_srem , int_udiv ,   int_urem , int_and , int_or , int_xor ,   int_shl , int_ashr , int_lshr , int_smin ,   int_smax , int_scmp , int_umin , int_umax ,   int_ucmp , int_eq , int_ne , int_sge ,   int_sgt , int_sle , int_slt , int_uge ,   int_ugt , int_ule , int_ult , int_abs }
	Operation codes for tasklet computations. More...

Functions
constexpr const char *	atomic_op_type_to_string (AtomicOpType type)
constexpr AtomicOpType	atomic_op_type_from_string (std::string_view str)
LibraryNodeCode	LibraryNodeType_Call ("Call")
LibraryNodeCode	LibraryNodeType_Invoke ("Invoke")
LibraryNodeCode	LibraryNodeType_LoadConst ("LoadConst")
LibraryNodeCode	LibraryNodeType_Metadata ("Metadata")
constexpr size_t	arity (TaskletCode c)
	Get the number of inputs for a tasklet operation.
constexpr bool	is_unsigned (TaskletCode c)
	Check if a tasklet operation is unsigned.
constexpr bool	is_integer (TaskletCode c)
	Check if a tasklet operation is an integer operation.
constexpr bool	is_floating_point (TaskletCode c)
	Check if a tasklet operation is a floating-point operation.

Variables
ImplementationType	ImplementationType_NONE {""}
	Default implementation type indicating no specific implementation.
LibraryNodeCode	LibraryNodeType_AtomicScalarOp {"AtomicScalarOp"}
LibraryNodeCode	LibraryNodeType_BarrierLocal {"barrier_local"}

Typedef Documentation

AccessPatternsDeleter

typedef PtrMetaDeleter<MemoryAccessPattern> sdfg::data_flow::AccessPatternsDeleter

ImplementationType

typedef StringEnum sdfg::data_flow::ImplementationType

Implementation type identifier for library nodes.

Determines how a library node should be implemented during code generation. Each library node can specify an implementation type that indicates:

• Whether is only has a single dispatcher or expand the node into primitive operations (NONE)
• Which library implementation to use (BLAS, CUBLAS, etc.)

LibraryNodeCode

typedef StringEnum sdfg::data_flow::LibraryNodeCode

MemoryAccessPatternType

typedef std::unique_ptr<MemoryAccessPattern, AccessPatternsDeleter> sdfg::data_flow::MemoryAccessPatternType

PointerAccessType

typedef std::unique_ptr<PointerAccessMeta, PtrAccessDeleter> sdfg::data_flow::PointerAccessType

PtrAccessDeleter

typedef PtrMetaDeleter<PointerAccessMeta> sdfg::data_flow::PtrAccessDeleter

Subset

typedef std::vector< symbolic::Expression > sdfg::data_flow::Subset

Type alias for a subset specification (array of symbolic expressions).

Vector of symbolic expressions describing data access pattern.

A subset specifies which portion of a data container is accessed. Each dimension is represented by a symbolic expression. Examples:

• Empty vector: Full scalar or container access
• {i}: Single element at index i
• {i, j}: 2D array element at [i][j]

Enumeration Type Documentation

AtomicOpType

enum class sdfg::data_flow::AtomicOpType

strong

Enumerator
Add
Subtract
Min
Max

EdgeRemoveOption

enum class sdfg::data_flow::EdgeRemoveOption

strong

Enumerator
NotRemovable	You must not remove this edge unless you can guarantee the entire node will be removed.
Trivially	You may remove the edge without further action.
RequiresUpdate	It can be removed, but the node requires a call to update itself to match the removed edge.
RemoveNodeAfter	This is the only relevant edge of this node. If it will be removed, the entire node must be removed as well, ignoring any potential side-effect flags of the node

MemletType

enum sdfg::data_flow::MemletType

Type of data movement represented by a memlet.

Enumerator
Computational	Standard data transfer for computation.
Reference	Creates a reference/pointer to data.
Dereference_Src	Dereferences pointer at source.
Dereference_Dst	Dereferences pointer at destination.

TaskletCode

enum sdfg::data_flow::TaskletCode

Operation codes for tasklet computations.

Defines all primitive operations that tasklets can perform, including arithmetic, logical, comparison, and type conversion operations.

Enumerator
assign	Assignment/copy operation (can be cast if types differ).
fp_neg	Floating-point negation.
fp_add	Floating-point addition.
fp_sub	Floating-point subtraction.
fp_mul	Floating-point multiplication.
fp_div	Floating-point division.
fp_rem	Floating-point remainder (fmod semantics).
fp_fma	Floating-point fused multiply-add.
fp_oeq	Floating-point ordered equal.
fp_one	Floating-point ordered not equal.
fp_oge	Floating-point ordered greater or equal.
fp_ogt	Floating-point ordered greater than.
fp_ole	Floating-point ordered less or equal.
fp_olt	Floating-point ordered less than.
fp_ord	Floating-point ordered (neither operand is NaN).
fp_ueq	Floating-point unordered equal.
fp_une	Floating-point unordered not equal.
fp_ugt	Floating-point unordered greater than.
fp_uge	Floating-point unordered greater or equal.
fp_ult	Floating-point unordered less than.
fp_ule	Floating-point unordered less or equal.
fp_uno	Floating-point unordered (at least one operand is NaN).
int_add	Integer addition.
int_sub	Integer subtraction.
int_mul	Integer multiplication.
int_sdiv	Signed integer division.
int_srem	Signed integer remainder.
int_udiv	Unsigned integer division.
int_urem	Unsigned integer remainder.
int_and	Bitwise AND.
int_or	Bitwise OR.
int_xor	Bitwise XOR.
int_shl	Shift left.
int_ashr	Arithmetic shift right (sign-extend).
int_lshr	Logical shift right (zero-extend).
int_smin	Signed minimum.
int_smax	Signed maximum.
int_scmp	Signed comparison (-1, 0, 1).
int_umin	Unsigned minimum.
int_umax	Unsigned maximum.
int_ucmp	Unsigned comparison (-1, 0, 1).
int_eq	Integer equal.
int_ne	Integer not equal.
int_sge	Signed greater or equal.
int_sgt	Signed greater than.
int_sle	Signed less or equal.
int_slt	Signed less than.
int_uge	Unsigned greater or equal.
int_ugt	Unsigned greater than.
int_ule	Unsigned less or equal.
int_ult	Unsigned less than.
int_abs	Integer absolute value.

Function Documentation

arity()

size_t sdfg::data_flow::arity ( TaskletCode c )

constexpr

Get the number of inputs for a tasklet operation.

Parameters

c	TaskletCode operation

Returns

Number of inputs required (arity)

Exceptions

InvalidSDFGException

if code is invalid

atomic_op_type_from_string()

AtomicOpType sdfg::data_flow::atomic_op_type_from_string ( std::string_view str )

constexpr

atomic_op_type_to_string()

const char * sdfg::data_flow::atomic_op_type_to_string ( AtomicOpType type )

constexpr

is_floating_point()

bool sdfg::data_flow::is_floating_point ( TaskletCode c )

constexpr

Check if a tasklet operation is a floating-point operation.

Parameters

c	TaskletCode operation

Returns

True if the operation works on floating-point types

is_integer()

bool sdfg::data_flow::is_integer ( TaskletCode c )

constexpr

Check if a tasklet operation is an integer operation.

Parameters

c	TaskletCode operation

Returns

True if the operation works on integer types

is_unsigned()

bool sdfg::data_flow::is_unsigned ( TaskletCode c )

constexpr

Check if a tasklet operation is unsigned.

Parameters

c	TaskletCode operation

Returns

True if the operation treats operands as unsigned

LibraryNodeType_Call()

LibraryNodeCode sdfg::data_flow::LibraryNodeType_Call ( "Call" )

inline

LibraryNodeType_Invoke()

LibraryNodeCode sdfg::data_flow::LibraryNodeType_Invoke ( "Invoke" )

inline

LibraryNodeType_LoadConst()

LibraryNodeCode sdfg::data_flow::LibraryNodeType_LoadConst ( "LoadConst" )

inline

LibraryNodeType_Metadata()

LibraryNodeCode sdfg::data_flow::LibraryNodeType_Metadata ( "Metadata" )

inline

Variable Documentation

ImplementationType_NONE

ImplementationType sdfg::data_flow::ImplementationType_NONE {""}

inline

Default implementation type indicating no specific implementation.

When a library node has ImplementationType_NONE, it may be code-generated using a single dispatcher or is expanded into more primitive operations during the expansion pass.

LibraryNodeType_AtomicScalarOp

LibraryNodeCode sdfg::data_flow::LibraryNodeType_AtomicScalarOp {"AtomicScalarOp"}

inline

LibraryNodeType_BarrierLocal

LibraryNodeCode sdfg::data_flow::LibraryNodeType_BarrierLocal {"barrier_local"}

inline