sdfg::passes::scheduler::CUDAOffloadScheduler Class Reference

#include <sdfg/passes/scheduler/cuda_offload_scheduler.h>

Inheritance diagram for sdfg::passes::scheduler::CUDAOffloadScheduler:

Public Member Functions
SchedulerAction	find (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager, structured_control_flow::StructuredLoop &loop, bool offload_unknown_sizes=false) override
	Phase 1: Determine the action for a loop during loop discovery.
SchedulerAction	find (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager, structured_control_flow::While &loop, bool offload_unknown_sizes=false) override
bool	can_apply_schedule (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager, structured_control_flow::StructuredLoop &loop, bool offload_unknown_sizes=false) override
	Check if the scheduling transform can be applied to a loop.
void	apply_schedule (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager, structured_control_flow::StructuredLoop &loop, bool offload_unknown_sizes=false) override
	Apply the scheduling transform to a single loop.
void	pre_schedule (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager, std::vector< structured_control_flow::StructuredLoop * > &applicable_loops) override
	Pre-scheduling phase: collapse and cleanup on the applicable loops.
bool	post_schedule (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager, std::vector< structured_control_flow::StructuredLoop * > &scheduled_loops) override
	Post-scheduling phase: additional transforms on the scheduled loops.
std::unordered_set< ScheduleTypeCategory >	compatible_types () override
Public Member Functions inherited from sdfg::passes::scheduler::LoopScheduler
virtual	~LoopScheduler ()=default
virtual void	set_report (PassReportConsumer *report)
virtual void	set_recorder (transformations::Recorder *recorder)
	Attach an optional Recorder that captures each scheduling transform applied via apply_schedule as a replayable descriptor.

Static Public Member Functions
static std::string	target ()
static symbolic::Integer	get_parallel_size (structured_control_flow::StructuredLoop &loop)

Additional Inherited Members
Protected Attributes inherited from sdfg::passes::scheduler::LoopScheduler
PassReportConsumer *	report_ = nullptr
transformations::Recorder *	recorder_ = nullptr

Member Function Documentation

apply_schedule()

void sdfg::passes::scheduler::CUDAOffloadScheduler::apply_schedule ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager, structured_control_flow::StructuredLoop & loop, bool offload_unknown_sizes = false )

overridevirtual

Apply the scheduling transform to a single loop.

Called only on loops that passed can_apply_schedule().

Implements sdfg::passes::scheduler::LoopScheduler.

can_apply_schedule()

bool sdfg::passes::scheduler::CUDAOffloadScheduler::can_apply_schedule ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager, structured_control_flow::StructuredLoop & loop, bool offload_unknown_sizes = false )

overridevirtual

Check if the scheduling transform can be applied to a loop.

Called after pre_schedule. Used to filter the applicable loops before applying.

Implements sdfg::passes::scheduler::LoopScheduler.

compatible_types()

std::unordered_set< ScheduleTypeCategory > sdfg::passes::scheduler::CUDAOffloadScheduler::compatible_types ( )

overridevirtual

Implements sdfg::passes::scheduler::LoopScheduler.

find() [1/2]

SchedulerAction sdfg::passes::scheduler::CUDAOffloadScheduler::find ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager, structured_control_flow::StructuredLoop & loop, bool offload_unknown_sizes = false )

overridevirtual

Phase 1: Determine the action for a loop during loop discovery.

Returns NEXT if the loop should be scheduled (added to the applicable set), CHILDREN if the scheduler should descend into child loops instead, or SKIP if the loop already carries a non-None schedule and must be left untouched (neither scheduled nor traversed into).

Implements sdfg::passes::scheduler::LoopScheduler.

find() [2/2]

SchedulerAction sdfg::passes::scheduler::CUDAOffloadScheduler::find ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager, structured_control_flow::While & loop, bool offload_unknown_sizes = false )

overridevirtual

Implements sdfg::passes::scheduler::LoopScheduler.

get_parallel_size()

symbolic::Integer sdfg::passes::scheduler::CUDAOffloadScheduler::get_parallel_size ( structured_control_flow::StructuredLoop & loop )

static

post_schedule()

bool sdfg::passes::scheduler::CUDAOffloadScheduler::post_schedule ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager, std::vector< structured_control_flow::StructuredLoop * > & scheduled_loops )

overridevirtual

Post-scheduling phase: additional transforms on the scheduled loops.

Called once per target after apply_schedule, even if no loops were scheduled (scheduled_loops may be empty). GPU schedulers use this for loop reordering, nested parallelization, and library node transfer extraction.

Default implementation is a no-op.

Returns

true if the SDFG was modified

Reimplemented from sdfg::passes::scheduler::LoopScheduler.

pre_schedule()

void sdfg::passes::scheduler::CUDAOffloadScheduler::pre_schedule ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager, std::vector< structured_control_flow::StructuredLoop * > & applicable_loops )

overridevirtual

Pre-scheduling phase: collapse and cleanup on the applicable loops.

Called after loop discovery and before can_apply_schedule/apply_schedule. Implementations may collapse loop nests, run symbol propagation, dead data/CFG elimination, etc. The vector is modified in-place to reflect any pointer changes (e.g. after collapse).

Default implementation is a no-op.

Reimplemented from sdfg::passes::scheduler::LoopScheduler.

target()

std::string sdfg::passes::scheduler::CUDAOffloadScheduler::target ( )

inlinestatic

---

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

• opt/include/sdfg/passes/scheduler/cuda_offload_scheduler.h