Atomics-based ROCm/HIP code generator for structured_control_flow::Reduce. More...

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

Inheritance diagram for sdfg::rocm::ROCMReduceDispatcher:

Public Member Functions
	ROCMReduceDispatcher (codegen::LanguageExtension &language_extension, StructuredSDFG &sdfg, analysis::AnalysisManager &analysis_manager, structured_control_flow::Reduce &node, codegen::InstrumentationPlan &instrumentation_plan, codegen::ArgCapturePlan &arg_capture_plan)
Public Member Functions inherited from sdfg::gpu::GPUReduceDispatcher
	GPUReduceDispatcher (codegen::LanguageExtension &language_extension, StructuredSDFG &sdfg, analysis::AnalysisManager &analysis_manager, structured_control_flow::Reduce &node, codegen::InstrumentationPlan &instrumentation_plan, codegen::ArgCapturePlan &arg_capture_plan)
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
std::string	schedule_value () const override
	Schedule value string identifying this backend's GPU schedule ("CUDA"/"ROCM").
codegen::TargetType	target_type () const override
	Target type used for instrumentation info.
std::unique_ptr< codegen::LanguageExtension >	create_device_language_extension () const override
	Create the backend device-code language extension (CUDA/HIP).
bool	is_device_pointer_storage (const types::StorageType &storage) const override
	True if `storage` designates a device-resident generic pointer for this backend.
std::string	kernel_file_extension () const override
	File extension for the kernel translation unit ("cu"/"rocm.cpp").
void	emit_kernel_includes (codegen::CodeSnippetFactory &library_snippet_factory) const override
	Emit the backend include directives required by the generated kernel.
void	emit_library_preamble (codegen::PrettyPrinter &library_stream) const override
	Emit any extra preamble at the top of the kernel translation unit.
void	emit_kernel_call (codegen::PrettyPrinter &main_stream, const std::string &kernel_name, symbolic::Expression &num_blocks, symbolic::Expression &block_size, std::vector< std::string > &arguments_device) override
	Emit the host-side kernel launch.
Protected Member Functions inherited from sdfg::gpu::GPUReduceDispatcher
bool	is_nested_in_gpu_kernel ()
	True if this Reduce is nested inside an enclosing GPU kernel (an ancestor Map with this backend's GPU schedule).
void	dispatch_header (codegen::PrettyPrinter &stream, const std::string &kernel_name, std::vector< std::string > &arguments_declaration)
void	dispatch_reduce_core (codegen::CodeSnippetFactory &library_snippet_factory, codegen::PrettyPrinter &library_stream, GPUDimension dimension, bool declare_scope_variables, std::vector< std::string > &scope_variables)
	Emit the privatize + grid-stride reduce loop + atomic merge into the given (device) stream, parallelized on `dimension`.
void	dispatch_atomic_merge (codegen::PrettyPrinter &library_stream, codegen::CodeSnippetFactory &library_snippet_factory, const structured_control_flow::ReductionInfo &reduction)
	Emit the device-side helper(s) and the atomic merge statement that combines a thread's private partial into the global accumulator.
Protected Member Functions inherited from sdfg::codegen::NodeDispatcher
virtual bool	begin_node (PrettyPrinter &stream)
virtual void	end_node (PrettyPrinter &stream, bool has_declaration)

Additional Inherited Members
Protected Attributes inherited from sdfg::gpu::GPUReduceDispatcher
structured_control_flow::Reduce &	node_
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

Atomics-based ROCm/HIP code generator for structured_control_flow::Reduce.

Thin ROCm specialization of gpu::GPUReduceDispatcher: the whole lowering (grid-stride reduce, privatization, atomic merge) lives in the shared base; this class only supplies the ROCm-specific policy (language extension, device-pointer storage, kernel launch syntax and includes).

Constructor & Destructor Documentation

◆ ROCMReduceDispatcher()

sdfg::rocm::ROCMReduceDispatcher::ROCMReduceDispatcher	(	codegen::LanguageExtension &	language_extension,
		StructuredSDFG &	sdfg,
		analysis::AnalysisManager &	analysis_manager,
		structured_control_flow::Reduce &	node,
		codegen::InstrumentationPlan &	instrumentation_plan,
		codegen::ArgCapturePlan &	arg_capture_plan )

Member Function Documentation

◆ create_device_language_extension()

std::unique_ptr< codegen::LanguageExtension > sdfg::rocm::ROCMReduceDispatcher::create_device_language_extension ( ) const

overrideprotectedvirtual

Create the backend device-code language extension (CUDA/HIP).

Implements sdfg::gpu::GPUReduceDispatcher.

◆ emit_kernel_call()

void sdfg::rocm::ROCMReduceDispatcher::emit_kernel_call	(	codegen::PrettyPrinter &	main_stream,
		const std::string &	kernel_name,
		symbolic::Expression &	num_blocks,
		symbolic::Expression &	block_size,
		std::vector< std::string > &	arguments_device )

overrideprotectedvirtual

Emit the host-side kernel launch.

Implements sdfg::gpu::GPUReduceDispatcher.

◆ emit_kernel_includes()

void sdfg::rocm::ROCMReduceDispatcher::emit_kernel_includes ( codegen::CodeSnippetFactory & library_snippet_factory ) const

overrideprotectedvirtual

Emit the backend include directives required by the generated kernel.

Implements sdfg::gpu::GPUReduceDispatcher.

◆ emit_library_preamble()

void sdfg::rocm::ROCMReduceDispatcher::emit_library_preamble ( codegen::PrettyPrinter & library_stream ) const

overrideprotectedvirtual

Emit any extra preamble at the top of the kernel translation unit.

Implements sdfg::gpu::GPUReduceDispatcher.

◆ is_device_pointer_storage()

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

overrideprotectedvirtual

True if storage designates a device-resident generic pointer for this backend.

Implements sdfg::gpu::GPUReduceDispatcher.

◆ kernel_file_extension()

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

overrideprotectedvirtual

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

Implements sdfg::gpu::GPUReduceDispatcher.

◆ schedule_value()

std::string sdfg::rocm::ROCMReduceDispatcher::schedule_value ( ) const

overrideprotectedvirtual

Schedule value string identifying this backend's GPU schedule ("CUDA"/"ROCM").

Implements sdfg::gpu::GPUReduceDispatcher.

◆ target_type()

codegen::TargetType sdfg::rocm::ROCMReduceDispatcher::target_type ( ) const

overrideprotectedvirtual

Target type used for instrumentation info.

Implements sdfg::gpu::GPUReduceDispatcher.

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

opt/include/sdfg/targets/rocm/rocm_reduce_dispatcher.h

Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ ROCMReduceDispatcher()

Member Function Documentation

◆ create_device_language_extension()

◆ emit_kernel_call()

◆ emit_kernel_includes()

◆ emit_library_preamble()

◆ is_device_pointer_storage()

◆ kernel_file_extension()

◆ schedule_value()

◆ target_type()