Offloads a top-level loop to a CUDA kernel (X grid dimension). More...

#include <sdfg/transformations/offloading/cuda_transform.h>

Inheritance diagram for sdfg::cuda::CUDATransform:

Public Member Functions
	CUDATransform (structured_control_flow::StructuredLoop &loop, int block_size=32, bool allow_dynamic_sizes=false)
std::string	name () const override
	Get the name of this transformation.
bool	can_be_applied (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) override
	Check if this transformation can be safely applied.
void	to_json (nlohmann::json &j) const override
	Serialize this transformation to JSON.
Public Member Functions inherited from sdfg::transformations::OffloadTransform
	OffloadTransform (structured_control_flow::StructuredLoop &loop, bool allow_dynamic_sizes=false)
bool	can_be_applied (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) override
	Check if this transformation can be safely applied.
void	apply (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) override
	Apply this transformation to the SDFG.
Public Member Functions inherited from sdfg::transformations::Transformation
virtual	~Transformation ()=default
virtual void	set_report (PassReportConsumer *report)
	Set the report consumer for transformation feedback.
virtual bool	try_apply (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager)
	Attempt to apply the transformation if possible.

Static Public Member Functions
static CUDATransform	from_json (builder::StructuredSDFGBuilder &builder, const nlohmann::json &desc)

Protected Member Functions
types::StorageType	local_device_storage_type () override
types::StorageType	global_device_storage_type (symbolic::Expression arg_size) override
ScheduleType	transformed_schedule_type () override
std::string	copy_prefix () override
void	add_device_buffer (builder::StructuredSDFGBuilder &builder, std::string host_arg_name, std::string device_arg_name, symbolic::Expression arg_size) override
void	allocate_device_arg (builder::StructuredSDFGBuilder &builder, Block &alloc_block, std::string host_arg_name, std::string device_arg_name, symbolic::Expression arg_size, symbolic::Expression page_size) override
void	deallocate_device_arg (builder::StructuredSDFGBuilder &builder, Block &dealloc_block, std::string device_arg_name, symbolic::Expression arg_size, symbolic::Expression page_size) override
void	copy_to_device (builder::StructuredSDFGBuilder &builder, std::string host_arg_name, std::string device_arg_name, symbolic::Expression size, symbolic::Expression page_size, Block &copy_block) override
void	copy_to_device_with_allocation (builder::StructuredSDFGBuilder &builder, std::string host_arg_name, std::string device_arg_name, symbolic::Expression size, symbolic::Expression page_size, Block &copy_block) override
void	copy_from_device (builder::StructuredSDFGBuilder &builder, Block &copy_out_block, std::string host_arg_name, std::string device_arg_name, symbolic::Expression size, symbolic::Expression page_size) override
void	copy_from_device_with_free (builder::StructuredSDFGBuilder &builder, Block &copy_out_block, std::string host_arg_name, std::string device_arg_name, symbolic::Expression size, symbolic::Expression page_size) override
void	setup_device (builder::StructuredSDFGBuilder &builder, Block &global_alloc_block) override
void	teardown_device (builder::StructuredSDFGBuilder &builder, Block &global_alloc_block) override
Protected Member Functions inherited from sdfg::transformations::OffloadTransform
template<class Container>
void	allocate_locals_on_device_stack (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager, Container &locals)
void	handle_device_setup_and_teardown (builder::StructuredSDFGBuilder &builder, const std::map< std::string, analysis::RegionArgument > &arguments, const std::unordered_map< std::string, symbolic::Expression > &argument_sizes, std::string prefix)
void	update_loop_containers (const std::map< std::string, analysis::RegionArgument > &arguments, std::string prefix)

Additional Inherited Members
Protected Attributes inherited from sdfg::transformations::OffloadTransform
structured_control_flow::StructuredLoop &	loop_
bool	allow_dynamic_sizes_ = false
bool	skip_unneeded_d2h_ = true
Protected Attributes inherited from sdfg::transformations::Transformation
PassReportConsumer *	report_ = nullptr

Detailed Description

Offloads a top-level loop to a CUDA kernel (X grid dimension).

This transformation does not perform blocking or tiling on its own. It expects the scheduler to have already identified a suitable loop for GPU offloading. The transformation assigns the loop to the CUDA X grid dimension and handles data transfers between host and device.

The resulting grid X-dimension is validated against CUDA hardware limits (2^31 - 1 blocks). If the grid would exceed this limit, the transformation is rejected (can_be_applied returns false).

Constructor & Destructor Documentation

◆ CUDATransform()

sdfg::cuda::CUDATransform::CUDATransform	(	structured_control_flow::StructuredLoop &	loop,
		int	block_size = 32,
		bool	allow_dynamic_sizes = false )

inlineexplicit

Member Function Documentation

◆ add_device_buffer()

void sdfg::cuda::CUDATransform::add_device_buffer	(	builder::StructuredSDFGBuilder &	builder,
		std::string	host_arg_name,
		std::string	device_arg_name,
		symbolic::Expression	arg_size )

overrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ allocate_device_arg()

void sdfg::cuda::CUDATransform::allocate_device_arg	(	builder::StructuredSDFGBuilder &	builder,
		Block &	alloc_block,
		std::string	host_arg_name,
		std::string	device_arg_name,
		symbolic::Expression	arg_size,
		symbolic::Expression	page_size )

overrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ can_be_applied()

bool sdfg::cuda::CUDATransform::can_be_applied	(	builder::StructuredSDFGBuilder &	builder,
		analysis::AnalysisManager &	analysis_manager )

overridevirtual

Check if this transformation can be safely applied.

Parameters

builder	The SDFG builder
analysis_manager	The analysis manager providing necessary analyses

Returns: true if all preconditions are met

Implements sdfg::transformations::Transformation.

◆ copy_from_device()

void sdfg::cuda::CUDATransform::copy_from_device	(	builder::StructuredSDFGBuilder &	builder,
		Block &	copy_out_block,
		std::string	host_arg_name,
		std::string	device_arg_name,
		symbolic::Expression	size,
		symbolic::Expression	page_size )

overrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ copy_from_device_with_free()

void sdfg::cuda::CUDATransform::copy_from_device_with_free	(	builder::StructuredSDFGBuilder &	builder,
		Block &	copy_out_block,
		std::string	host_arg_name,
		std::string	device_arg_name,
		symbolic::Expression	size,
		symbolic::Expression	page_size )

overrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ copy_prefix()

std::string sdfg::cuda::CUDATransform::copy_prefix ( )

inlineoverrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ copy_to_device()

void sdfg::cuda::CUDATransform::copy_to_device	(	builder::StructuredSDFGBuilder &	builder,
		std::string	host_arg_name,
		std::string	device_arg_name,
		symbolic::Expression	size,
		symbolic::Expression	page_size,
		Block &	copy_block )

overrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ copy_to_device_with_allocation()

void sdfg::cuda::CUDATransform::copy_to_device_with_allocation	(	builder::StructuredSDFGBuilder &	builder,
		std::string	host_arg_name,
		std::string	device_arg_name,
		symbolic::Expression	size,
		symbolic::Expression	page_size,
		Block &	copy_block )

overrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ deallocate_device_arg()

void sdfg::cuda::CUDATransform::deallocate_device_arg	(	builder::StructuredSDFGBuilder &	builder,
		Block &	dealloc_block,
		std::string	device_arg_name,
		symbolic::Expression	arg_size,
		symbolic::Expression	page_size )

overrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ from_json()

CUDATransform sdfg::cuda::CUDATransform::from_json	(	builder::StructuredSDFGBuilder &	builder,
		const nlohmann::json &	desc )

static

◆ global_device_storage_type()

types::StorageType sdfg::cuda::CUDATransform::global_device_storage_type ( symbolic::Expression arg_size )

inlineoverrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ local_device_storage_type()

types::StorageType sdfg::cuda::CUDATransform::local_device_storage_type ( )

inlineoverrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ name()

std::string sdfg::cuda::CUDATransform::name ( ) const

overridevirtual

Get the name of this transformation.

Returns: The transformation name (e.g., "LoopInterchange")

Implements sdfg::transformations::Transformation.

◆ setup_device()

void sdfg::cuda::CUDATransform::setup_device	(	builder::StructuredSDFGBuilder &	builder,
		Block &	global_alloc_block )

inlineoverrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ teardown_device()

void sdfg::cuda::CUDATransform::teardown_device	(	builder::StructuredSDFGBuilder &	builder,
		Block &	global_alloc_block )

inlineoverrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

◆ to_json()

void sdfg::cuda::CUDATransform::to_json ( nlohmann::json & j ) const

overridevirtual

Serialize this transformation to JSON.

Parameters

j	JSON object to populate with transformation description

Exceptions

std::logic_error if not implemented

Implements sdfg::transformations::Transformation.

◆ transformed_schedule_type()

ScheduleType sdfg::cuda::CUDATransform::transformed_schedule_type ( )

inlineoverrideprotectedvirtual

Implements sdfg::transformations::OffloadTransform.

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

opt/include/sdfg/transformations/offloading/cuda_transform.h

Public Member Functions

Static Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ CUDATransform()

Member Function Documentation

◆ add_device_buffer()

◆ allocate_device_arg()

◆ can_be_applied()

◆ copy_from_device()

◆ copy_from_device_with_free()

◆ copy_prefix()

◆ copy_to_device()

◆ copy_to_device_with_allocation()

◆ deallocate_device_arg()

◆ from_json()

◆ global_device_storage_type()

◆ local_device_storage_type()

◆ name()

◆ setup_device()

◆ teardown_device()

◆ to_json()

◆ transformed_schedule_type()