sdfg::rocm::ROCMOffloadDispatcherStrategy Class Reference

ROCm/HIP specialization of gpu::GPUOffloadMapDispatcher. More...

#include <sdfg/targets/rocm/rocm_offload_dispatcher_strategy.h>

Inheritance diagram for sdfg::rocm::ROCMOffloadDispatcherStrategy:

Public Member Functions
void	dispatch_kernel_call (codegen::PrettyPrinter &main_stream, const std::string &kernel_name, codegen::LanguageExtension &host_language_ext, symbolic::Expression &num_blocks_x, symbolic::Expression &num_blocks_y, symbolic::Expression &num_blocks_z, symbolic::Expression &block_size_x, symbolic::Expression &block_size_y, symbolic::Expression &block_size_z, std::vector< std::string > &arguments_device) override
void	emit_target_header_declarations (codegen::PrettyPrinter &kernel_header_stream) override
	Target-specific includes/declarations emitted into the kernel include header (device fp16/bf16 intrinsics, lib-dependency includes, ...).
void	emit_target_global_includes (codegen::CodeSnippetFactory &library_snippet_factory) override
	ROCMOffloadDispatcherStrategy ()
void	dispatch_kernel_launch_error_check (codegen::PrettyPrinter &stream, const codegen::LanguageExtension &language_extension)
int	get_warp_size () const override
bool	is_device_pointer_storage (const types::StorageType &storage) const override
std::string	kernel_file_extension () const override
	File extension for the kernel translation unit ("cu"/"rocm.cpp").
std::string	kernel_header_file_extension () const override
	File extension for the kernel's include header ("cu.h"/"rocm.h").
std::string	warp_shuffle_xor (const std::string &value, const std::string &lane_mask) const override
	Cross-lane XOR butterfly shuffle of value by lane_mask (CUDA/HIP __shfl_xor_sync).
codegen::TargetType	get_instrumentation_kernel_target_type () const override
bool	is_matching_kernel_language_extension (const StructuredSDFG &sdfg, const StructuredLoop &loop, codegen::LanguageExtension &language_extension) override
std::unique_ptr< codegen::LanguageExtension >	create_kernel_language_extension (StructuredSDFG &sdfg, const StructuredLoop &loop) override
Public Member Functions inherited from sdfg::gpu::GPUOffloadDispatcherStrategy
virtual	~GPUOffloadDispatcherStrategy ()=default
void	require_matching_kernel_language_extension (const StructuredSDFG &sdfg, const StructuredLoop &loop, codegen::LanguageExtension &language_extension)

Static Public Attributes
static constexpr const char *	KERNEL_SNIPPET_FILE_EXT = "rocm.cpp"
static constexpr const char *	KERNEL_SNIPPET_HEADER_EXT = "rocm.h"

Detailed Description

ROCm/HIP specialization of gpu::GPUOffloadMapDispatcher.

Supplies the HIP-specific policy for the shared coverage-loop offload lowering: the device language extension, the hipLaunchKernelGGL kernel launch and launch-error check.

Constructor & Destructor Documentation

ROCMOffloadDispatcherStrategy()

sdfg::rocm::ROCMOffloadDispatcherStrategy::ROCMOffloadDispatcherStrategy

(

)

Member Function Documentation

create_kernel_language_extension()

std::unique_ptr< codegen::LanguageExtension > sdfg::rocm::ROCMOffloadDispatcherStrategy::create_kernel_language_extension ( StructuredSDFG & sdfg, const StructuredLoop & loop )

overridevirtual

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

dispatch_kernel_call()

void sdfg::rocm::ROCMOffloadDispatcherStrategy::dispatch_kernel_call ( codegen::PrettyPrinter & main_stream, const std::string & kernel_name, codegen::LanguageExtension & host_language_ext, symbolic::Expression & num_blocks_x, symbolic::Expression & num_blocks_y, symbolic::Expression & num_blocks_z, symbolic::Expression & block_size_x, symbolic::Expression & block_size_y, symbolic::Expression & block_size_z, std::vector< std::string > & arguments_device )

overridevirtual

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

dispatch_kernel_launch_error_check()

void sdfg::rocm::ROCMOffloadDispatcherStrategy::dispatch_kernel_launch_error_check	(	codegen::PrettyPrinter &	stream,
		const codegen::LanguageExtension &	language_extension )

emit_target_global_includes()

void sdfg::rocm::ROCMOffloadDispatcherStrategy::emit_target_global_includes ( codegen::CodeSnippetFactory & library_snippet_factory )

overridevirtual

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

emit_target_header_declarations()

void sdfg::rocm::ROCMOffloadDispatcherStrategy::emit_target_header_declarations ( codegen::PrettyPrinter & kernel_header_stream )

overridevirtual

Target-specific includes/declarations emitted into the kernel include header (device fp16/bf16 intrinsics, lib-dependency includes, ...).

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

get_instrumentation_kernel_target_type()

codegen::TargetType sdfg::rocm::ROCMOffloadDispatcherStrategy::get_instrumentation_kernel_target_type ( ) const

overridevirtual

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

get_warp_size()

int sdfg::rocm::ROCMOffloadDispatcherStrategy::get_warp_size ( ) const

overridevirtual

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

is_device_pointer_storage()

bool sdfg::rocm::ROCMOffloadDispatcherStrategy::is_device_pointer_storage ( const types::StorageType & storage ) const

overridevirtual

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

is_matching_kernel_language_extension()

bool sdfg::rocm::ROCMOffloadDispatcherStrategy::is_matching_kernel_language_extension ( const StructuredSDFG & sdfg, const StructuredLoop & loop, codegen::LanguageExtension & language_extension )

overridevirtual

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

kernel_file_extension()

std::string sdfg::rocm::ROCMOffloadDispatcherStrategy::kernel_file_extension ( ) const

overridevirtual

File extension for the kernel translation unit ("cu"/"rocm.cpp").

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

kernel_header_file_extension()

std::string sdfg::rocm::ROCMOffloadDispatcherStrategy::kernel_header_file_extension ( ) const

overridevirtual

File extension for the kernel's include header ("cu.h"/"rocm.h").

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

warp_shuffle_xor()

std::string sdfg::rocm::ROCMOffloadDispatcherStrategy::warp_shuffle_xor ( const std::string & value, const std::string & lane_mask ) const

overridevirtual

Cross-lane XOR butterfly shuffle of value by lane_mask (CUDA/HIP __shfl_xor_sync).

Implements sdfg::gpu::GPUOffloadDispatcherStrategy.

Member Data Documentation

KERNEL_SNIPPET_FILE_EXT

const char* sdfg::rocm::ROCMOffloadDispatcherStrategy::KERNEL_SNIPPET_FILE_EXT = "rocm.cpp"

staticconstexpr

KERNEL_SNIPPET_HEADER_EXT

const char* sdfg::rocm::ROCMOffloadDispatcherStrategy::KERNEL_SNIPPET_HEADER_EXT = "rocm.h"

staticconstexpr

---

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

• opt/include/sdfg/targets/rocm/rocm_offload_dispatcher_strategy.h