sdfg::rocm Namespace Reference

Namespaces
namespace	blas
namespace	stdlib
namespace	tiles

Classes
class	RocmLibraryNodeRewriter
class	RocmLibraryNodeTransferExtractionVisitor
class	ROCMLanguageExtension
class	ScheduleType_ROCM_Offload
	ROCM schedule type inheriting shared GPU functionality Provides ROCM-specific value() and default block size (64 for wavefront size). More...
class	ScheduleType_ROCM
	ROCM schedule type inheriting shared GPU functionality Provides ROCM-specific value() and default block size (64 for wavefront size). More...
class	ROCMDataOffloadingNode
class	ROCMDataOffloadingNodeDispatcher
class	ROCMDataOffloadingNodeSerializer
class	ROCMMapDispatcher
class	ROCMOffloadDispatcherStrategy
	ROCm/HIP specialization of gpu::GPUOffloadMapDispatcher. More...
class	ROCMReduceDispatcher
	Atomics-based ROCm/HIP code generator for structured_control_flow::Reduce. More...
class	ROCBLASDataTransferExtraction
class	ROCMOffloadTransform
class	ROCMStdlibDataTransferExtraction
class	ROCMTransform
	Offloads a top-level loop to a ROCm/HIP kernel (X grid dimension). More...

Typedefs
typedef passes::VisitorPass< RocmLibraryNodeRewriter >	RocmLibraryNodeRewriterPass
typedef passes::VisitorPass< RocmLibraryNodeTransferExtractionVisitor >	RocmLibraryNodeTransferExtractionPass
using	ROCMDimension = gpu::GPUDimension

Functions
void	register_rocm_plugin (plugins::Context &context)
void	register_rocm_plugin ()
void	rocm_error_checking (codegen::PrettyPrinter &stream, const codegen::LanguageExtension &language_extension, const std::string &status_variable)
bool	do_rocm_error_checking ()
int	rocm_wavefront_size ()
	Wavefront width of the target device, queried once at codegen time.
void	check_rocm_kernel_launch_errors (codegen::PrettyPrinter &stream, const codegen::LanguageExtension &language_extension)
data_flow::LibraryNodeCode	LibraryNodeType_ROCM_Offloading ("ROCMOffloading")

Variables
std::string	ROCM_DEVICE_PREFIX = "__daisy_hip_"
constexpr int	ROCM_WARP_SIZE = 64
data_flow::ImplementationType	ImplementationType_ROCM {"ROCM"}
	Generic ROCM implementation.
data_flow::ImplementationType	ImplementationType_ROCMWithTransfers {"ROCMWithTransfers"}
	ROCM implementation with automatic memory transfers Used for ROCm BLAS, memset, and other ROCm-accelerated library nodes.
data_flow::ImplementationType	ImplementationType_ROCMWithoutTransfers {"ROCMWithoutTransfers"}
	ROCM implementation without memory transfers Used for ROCm BLAS, memset, and other ROCm-accelerated library nodes assuming data is already on GPU.
codegen::TargetType	TargetType_ROCM_Offload {ScheduleType_ROCM_Offload::value()}
codegen::TargetType	TargetType_ROCM {ScheduleType_ROCM::value()}

Typedef Documentation

ROCMDimension

using sdfg::rocm::ROCMDimension = gpu::GPUDimension

RocmLibraryNodeRewriterPass

typedef passes::VisitorPass<RocmLibraryNodeRewriter> sdfg::rocm::RocmLibraryNodeRewriterPass

RocmLibraryNodeTransferExtractionPass

typedef passes::VisitorPass<RocmLibraryNodeTransferExtractionVisitor> sdfg::rocm::RocmLibraryNodeTransferExtractionPass

Function Documentation

check_rocm_kernel_launch_errors()

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

do_rocm_error_checking()

bool sdfg::rocm::do_rocm_error_checking

(

)

LibraryNodeType_ROCM_Offloading()

data_flow::LibraryNodeCode sdfg::rocm::LibraryNodeType_ROCM_Offloading ( "ROCMOffloading" )

inline

register_rocm_plugin() [1/2]

void sdfg::rocm::register_rocm_plugin ( )

inline

Deprecated

use the variant with explicit context

register_rocm_plugin() [2/2]

void sdfg::rocm::register_rocm_plugin

(

plugins::Context &

context

)

rocm_error_checking()

void sdfg::rocm::rocm_error_checking	(	codegen::PrettyPrinter &	stream,
		const codegen::LanguageExtension &	language_extension,
		const std::string &	status_variable )

rocm_wavefront_size()

int sdfg::rocm::rocm_wavefront_size

(

)

Wavefront width of the target device, queried once at codegen time.

RDNA parts (gfx10xx/11xx/12xx) execute wave32 while CDNA/GCN (gfx9xx) execute wave64, so the value that drives warp-shuffle reductions and launch geometry cannot be a compile-time constant. Resolved from the DOCC_ROCM_WAVEFRONT_SIZE override, else the present device via rocminfo, falling back to ROCM_WARP_SIZE. The result is cached.

Variable Documentation

ImplementationType_ROCM

data_flow::ImplementationType sdfg::rocm::ImplementationType_ROCM {"ROCM"}

inline

Generic ROCM implementation.

ImplementationType_ROCMWithoutTransfers

data_flow::ImplementationType sdfg::rocm::ImplementationType_ROCMWithoutTransfers {"ROCMWithoutTransfers"}

inline

ROCM implementation without memory transfers Used for ROCm BLAS, memset, and other ROCm-accelerated library nodes assuming data is already on GPU.

ImplementationType_ROCMWithTransfers

data_flow::ImplementationType sdfg::rocm::ImplementationType_ROCMWithTransfers {"ROCMWithTransfers"}

inline

ROCM implementation with automatic memory transfers Used for ROCm BLAS, memset, and other ROCm-accelerated library nodes.

ROCM_DEVICE_PREFIX

std::string sdfg::rocm::ROCM_DEVICE_PREFIX = "__daisy_hip_"

inline

ROCM_WARP_SIZE

int sdfg::rocm::ROCM_WARP_SIZE = 64

constexpr

TargetType_ROCM

codegen::TargetType sdfg::rocm::TargetType_ROCM {ScheduleType_ROCM::value()}

inline

TargetType_ROCM_Offload

codegen::TargetType sdfg::rocm::TargetType_ROCM_Offload {ScheduleType_ROCM_Offload::value()}

inline