sdfg::gpu::GPUOffloadBaseDispatcher Class Reference

Shared scaffolding for the GPU offload dispatchers (Map and Reduce). More...

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

Inheritance diagram for sdfg::gpu::GPUOffloadBaseDispatcher:

Public Member Functions
void	dispatch_node (codegen::PrettyPrinter &main_stream, codegen::PrettyPrinter &globals_stream, codegen::CodeSnippetFactory &library_snippet_factory) override
codegen::InstrumentationInfo	instrumentation_info () const override
	Bad design.
Public Member Functions inherited from sdfg::codegen::NodeDispatcher
	NodeDispatcher (LanguageExtension &language_extension, StructuredSDFG &sdfg, analysis::AnalysisManager &analysis_manager, structured_control_flow::ControlFlowNode &node, InstrumentationPlan &instrumentation_plan, ArgCapturePlan &arg_capture_plan)
virtual	~NodeDispatcher ()=default
InstrumentationInfo	instrumentation_info_for_group () const
virtual void	dispatch (PrettyPrinter &main_stream, PrettyPrinter &globals_stream, CodeSnippetFactory &library_snippet_factory)

Protected Member Functions
	GPUOffloadBaseDispatcher (codegen::LanguageExtension &language_extension, StructuredSDFG &sdfg, analysis::AnalysisManager &analysis_manager, structured_control_flow::StructuredLoop &node, codegen::InstrumentationPlan &instrumentation_plan, codegen::ArgCapturePlan &arg_capture_plan, std::unique_ptr< GPUOffloadDispatcherStrategy > strategy)
bool	is_outermost_map (analysis::AnalysisManager &analysis_manager)
void	dispatch_header (codegen::PrettyPrinter &globals_stream, const std::string &kernel_name, std::vector< std::string > &arguments_declaration)
void	dispatch_kernel_preamble (codegen::PrettyPrinter &library_stream, const std::string &kernel_name, std::vector< std::string > &arguments_declaration)
void	emit_lib_dependency_includes (codegen::PrettyPrinter &kernel_header_stream, codegen::NestedCodeSnippetFactory &nested_snippet_factory)
virtual void	validate_before_dispatch (analysis::AnalysisManager &analysis_manager)
	Node-specific preconditions checked before any code is emitted (default: none).
virtual void	dispatch_kernel_body (codegen::NestedCodeSnippetFactory &kernel_snippet_factory, codegen::PrettyPrinter &kernel_source_stream, codegen::PrettyPrinter &kernel_header_stream, codegen::LanguageExtension &kernel_lang_ext, symbolic::Symbol indvar, std::vector< std::string > &scope_variables, symbolic::Expression &num_iterations)=0
Protected Member Functions inherited from sdfg::codegen::NodeDispatcher
virtual bool	begin_node (PrettyPrinter &stream)
virtual void	end_node (PrettyPrinter &stream, bool has_declaration)

Protected Attributes
structured_control_flow::StructuredLoop &	node_
std::unique_ptr< GPUOffloadDispatcherStrategy >	strategy_
long	launch_bounds_ = 0
Protected Attributes inherited from sdfg::codegen::NodeDispatcher
LanguageExtension &	language_extension_
StructuredSDFG &	sdfg_
InstrumentationPlan &	instrumentation_plan_
ArgCapturePlan &	arg_capture_plan_
analysis::AnalysisManager &	analysis_manager_

Detailed Description

Shared scaffolding for the GPU offload dispatchers (Map and Reduce).

Owns the code common to every offloaded kernel: collecting kernel arguments and scope variables, computing the launch grid/block sizes, emitting the kernel declaration, opening the kernel translation unit + its include header, and calling the (device-)target launch. The node-specific body and the target-specific launch/language policy are supplied by subclasses.

The target specific parts are handled by the DispatcherStrategy across Map and Reduce kernels

Constructor & Destructor Documentation

GPUOffloadBaseDispatcher()

sdfg::gpu::GPUOffloadBaseDispatcher::GPUOffloadBaseDispatcher ( codegen::LanguageExtension & language_extension, StructuredSDFG & sdfg, analysis::AnalysisManager & analysis_manager, structured_control_flow::StructuredLoop & node, codegen::InstrumentationPlan & instrumentation_plan, codegen::ArgCapturePlan & arg_capture_plan, std::unique_ptr< GPUOffloadDispatcherStrategy > strategy )

protected

Member Function Documentation

dispatch_header()

void sdfg::gpu::GPUOffloadBaseDispatcher::dispatch_header ( codegen::PrettyPrinter & globals_stream, const std::string & kernel_name, std::vector< std::string > & arguments_declaration )

protected

dispatch_kernel_body()

virtual void sdfg::gpu::GPUOffloadBaseDispatcher::dispatch_kernel_body ( codegen::NestedCodeSnippetFactory & kernel_snippet_factory, codegen::PrettyPrinter & kernel_source_stream, codegen::PrettyPrinter & kernel_header_stream, codegen::LanguageExtension & kernel_lang_ext, symbolic::Symbol indvar, std::vector< std::string > & scope_variables, symbolic::Expression & num_iterations )

protectedpure virtual

Implemented in sdfg::gpu::GPUOffloadMapDispatcher, and sdfg::gpu::GPUOffloadReduceDispatcher.

dispatch_kernel_preamble()

void sdfg::gpu::GPUOffloadBaseDispatcher::dispatch_kernel_preamble ( codegen::PrettyPrinter & library_stream, const std::string & kernel_name, std::vector< std::string > & arguments_declaration )

protected

dispatch_node()

void sdfg::gpu::GPUOffloadBaseDispatcher::dispatch_node ( codegen::PrettyPrinter & main_stream, codegen::PrettyPrinter & globals_stream, codegen::CodeSnippetFactory & library_snippet_factory )

overridevirtual

Implements sdfg::codegen::NodeDispatcher.

emit_lib_dependency_includes()

void sdfg::gpu::GPUOffloadBaseDispatcher::emit_lib_dependency_includes ( codegen::PrettyPrinter & kernel_header_stream, codegen::NestedCodeSnippetFactory & nested_snippet_factory )

protected

instrumentation_info()

codegen::InstrumentationInfo sdfg::gpu::GPUOffloadBaseDispatcher::instrumentation_info ( ) const

overridevirtual

Bad design.

We already have fields in this class and have to bind an instance to a single node. Just set a field to sth. of the instrumentation plan where the node can register more details. But this is a breaking change

Reimplemented from sdfg::codegen::NodeDispatcher.

is_outermost_map()

bool sdfg::gpu::GPUOffloadBaseDispatcher::is_outermost_map ( analysis::AnalysisManager & analysis_manager )

protected

validate_before_dispatch()

virtual void sdfg::gpu::GPUOffloadBaseDispatcher::validate_before_dispatch ( analysis::AnalysisManager & analysis_manager )

inlineprotectedvirtual

Node-specific preconditions checked before any code is emitted (default: none).

Reimplemented in sdfg::gpu::GPUOffloadReduceDispatcher.

Member Data Documentation

launch_bounds_

long sdfg::gpu::GPUOffloadBaseDispatcher::launch_bounds_ = 0

protected

node_

structured_control_flow::StructuredLoop& sdfg::gpu::GPUOffloadBaseDispatcher::node_

protected

strategy_

std::unique_ptr<GPUOffloadDispatcherStrategy> sdfg::gpu::GPUOffloadBaseDispatcher::strategy_

protected

---

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

• opt/include/sdfg/targets/gpu/gpu_offload_base_dispatcher.h