CandidateHeuristics contains state and implementations to facilitate making per instruction scheduling decisions; it contains methods used in tryCandidate to decide which instruction to schedule next. More...

#include "Target/AMDGPU/AMDGPUCoExecSchedStrategy.h"

Classes
struct	StallCosts

Public Member Functions
	CandidateHeuristics ()=default
void	initialize (ScheduleDAGMI DAG, const TargetSchedModel SchedModel, const TargetRegisterInfo *TRI)
void	updateForScheduling (SUnit *SU)
	Update the state to reflect that `SU` is going to be scheduled.
HardwareUnitInfo *	getHWUIFromFlavor (AMDGPU::InstructionFlavor Flavor)
	Given a `Flavor` , find the corresponding HardwareUnit.
void	sortHWUIResources ()
	Sort the HardwarUnitInfo vector.
unsigned	getStructuralStallCycles (SchedBoundary &Zone, SUnit *SU)
bool	tryEffectiveStall (GenericSchedulerBase::SchedCandidate &TryCand, GenericSchedulerBase::SchedCandidate &Cand, SchedBoundary &Zone)
bool	tryMemoryPipeline (GenericSchedulerBase::SchedCandidate &TryCand, GenericSchedulerBase::SchedCandidate &Cand, SchedBoundary &Zone)
	Prioritize instructions involved the memory pipeline.
bool	tryCriticalResource (GenericSchedulerBase::SchedCandidate &TryCand, GenericSchedulerBase::SchedCandidate &Cand, SchedBoundary *Zone) const
	Check for critical resource consumption.
bool	tryCriticalResourceDependency (GenericSchedulerBase::SchedCandidate &TryCand, GenericSchedulerBase::SchedCandidate &Cand, SchedBoundary *Zone) const
	Check for dependencies of instructions that use prioritized HardwareUnits.
void	dumpRegionSummary ()

Protected Member Functions
void	collectRegionSummary ()
	Walk over the region and collect characteristics for the various heuristics.
unsigned	getMaxBlockingCycles (const MCSchedClassDesc SC, const MachineInstr MI)
unsigned	getHWUICyclesForSU (SUnit *SU)
	Compute the blocking cycles for the appropriate HardwareUnit given an `SU`.
unsigned	getHWUICyclesForMI (MachineInstr *MI)
	Compute the blocking cycles for the appropriate HardwareUnit given an `MI`.
unsigned	getCarriedLatency (SUnit *SU)
	Estimate the block carried latency from loads for a given `SU`.
StallCosts	getStallCosts (SUnit *SU, SchedBoundary &Zone)

Protected Attributes
ScheduleDAGMI *	DAG
const SIInstrInfo *	SII
const SIRegisterInfo *	SRI
const TargetSchedModel *	SchedModel
SmallVector< HardwareUnitInfo, 8 >	HWUInfo
DenseMap< MachineInstr *, unsigned >	CarriedLatencies

Detailed Description

CandidateHeuristics contains state and implementations to facilitate making per instruction scheduling decisions; it contains methods used in tryCandidate to decide which instruction to schedule next.

Definition at line 214 of file AMDGPUCoExecSchedStrategy.h.

Constructor & Destructor Documentation

◆ CandidateHeuristics()

llvm::CandidateHeuristics::CandidateHeuristics ( )

default

References DAG, SchedModel, and TRI.

Member Function Documentation

◆ collectRegionSummary()

void CandidateHeuristics::collectRegionSummary ( )

protected

Walk over the region and collect characteristics for the various heuristics.

Definition at line 717 of file AMDGPUCoExecSchedStrategy.cpp.

References CarriedLatencies, llvm::AMDGPU::classifyFlavor(), DAG, dumpRegionSummary(), getCarriedLatency(), getHWUICyclesForSU(), HWUInfo, LLVM_DEBUG, MI, SchedModel, and SII.

Referenced by initialize().

◆ dumpRegionSummary()

void CandidateHeuristics::dumpRegionSummary ( )

Definition at line 737 of file AMDGPUCoExecSchedStrategy.cpp.

References DAG, llvm::dbgs(), llvm::AMDGPU::getFlavorName(), llvm::MachineBasicBlock::getNumber(), and HWUInfo.

Referenced by collectRegionSummary().

◆ getCarriedLatency()

unsigned CandidateHeuristics::getCarriedLatency ( SUnit * SU )

protected

Estimate the block carried latency from loads for a given SU.

This is essentially global scheduling info that our local scheduling infrastructure lacks the necessary infrastructure to accurately measure. Thus, this method just attempts to find a reasonable upper bound for carried load latency to avoid long stalls.

Definition at line 638 of file AMDGPUCoExecSchedStrategy.cpp.

References assert(), BlockCarriedLatency, llvm::AMDGPU::classifyFlavor(), DAG, llvm::AMDGPU::DS, llvm::AMDGPU::Fence, getHWUICyclesForMI(), llvm::SUnit::getInstr(), I, llvm::SlotIndex::isEarlierInstr(), llvm::Latency, MBB, MI, and SII.

Referenced by collectRegionSummary().

◆ getHWUICyclesForMI()

unsigned CandidateHeuristics::getHWUICyclesForMI ( MachineInstr * MI )

protected

Compute the blocking cycles for the appropriate HardwareUnit given an MI.

Definition at line 597 of file AMDGPUCoExecSchedStrategy.cpp.

References assert(), getMaxBlockingCycles(), MI, and SchedModel.

Referenced by getCarriedLatency().

◆ getHWUICyclesForSU()

unsigned CandidateHeuristics::getHWUICyclesForSU ( SUnit * SU )

protected

Compute the blocking cycles for the appropriate HardwareUnit given an SU.

Definition at line 589 of file AMDGPUCoExecSchedStrategy.cpp.

References assert(), DAG, llvm::SUnit::getInstr(), getMaxBlockingCycles(), MI, SchedModel, and SII.

Referenced by collectRegionSummary(), getStallCosts(), and updateForScheduling().

◆ getHWUIFromFlavor()

HardwareUnitInfo * CandidateHeuristics::getHWUIFromFlavor ( AMDGPU::InstructionFlavor Flavor )

Given a Flavor , find the corresponding HardwareUnit.

Returns: the mapped HardwareUnit.

Definition at line 563 of file AMDGPUCoExecSchedStrategy.cpp.

References HWUInfo.

Referenced by getStallCosts(), and updateForScheduling().

◆ getMaxBlockingCycles()

unsigned CandidateHeuristics::getMaxBlockingCycles	(	const MCSchedClassDesc *	SC,
		const MachineInstr *	MI )

protected

Returns: the maximum blocking cycles according to the SchedModel for a given MCSchedClassDesc SC.

Definition at line 572 of file AMDGPUCoExecSchedStrategy.cpp.

References MI, SchedModel, and SII.

Referenced by getHWUICyclesForMI(), and getHWUICyclesForSU().

◆ getStallCosts()

CandidateHeuristics::StallCosts CandidateHeuristics::getStallCosts	(	SUnit *	SU,
		SchedBoundary &	Zone )

protected

Definition at line 810 of file AMDGPUCoExecSchedStrategy.cpp.

References llvm::CandidateHeuristics::StallCosts::Buffer, llvm::CandidateHeuristics::StallCosts::Carried, CarriedLatencies, llvm::AMDGPU::classifyFlavor(), DAG, llvm::AMDGPU::DS, llvm::CandidateHeuristics::StallCosts::Effective, llvm::AMDGPU::Fence, llvm::CandidateHeuristics::StallCosts::Fence, llvm::HardwareUnitInfo::getBufferAvailableCycle(), llvm::HardwareUnitInfo::getBufferSize(), llvm::SchedBoundary::getCurrCycle(), getHWUICyclesForSU(), getHWUIFromFlavor(), llvm::SUnit::getInstr(), llvm::HardwareUnitInfo::getLastScheduledSU(), llvm::SchedBoundary::getLatencyStallCycles(), getStructuralStallCycles(), llvm::SchedBoundary::isTop(), llvm::CandidateHeuristics::StallCosts::Latency, llvm::CandidateHeuristics::StallCosts::Ready, llvm::CandidateHeuristics::StallCosts::Structural, and llvm::SUnit::TopReadyCycle.

Referenced by tryEffectiveStall(), and tryMemoryPipeline().

◆ getStructuralStallCycles()

unsigned CandidateHeuristics::getStructuralStallCycles	(	SchedBoundary &	Zone,
		SUnit *	SU )

Definition at line 775 of file AMDGPUCoExecSchedStrategy.cpp.

References assert(), DAG, llvm::SchedBoundary::getCurrCycle(), llvm::SUnit::getInstr(), llvm::SchedBoundary::getNextResourceCycle(), llvm::SUnit::hasReservedResource, llvm::SchedBoundary::HazardRec, llvm::SchedBoundary::isTop(), llvm::make_range(), MI, SchedModel, and llvm::AMDGPU::Stall.

Referenced by getStallCosts().

◆ initialize()

void CandidateHeuristics::initialize	(	ScheduleDAGMI *	DAG,
		const TargetSchedModel *	SchedModel,
		const TargetRegisterInfo *	TRI )

Definition at line 609 of file AMDGPUCoExecSchedStrategy.cpp.

References assert(), collectRegionSummary(), DAG, llvm::AMDGPU::DefaultBufferSizes::DS, llvm::AMDGPU::DS, HWUInfo, I, llvm::AMDGPU::MultiCycleVALU, llvm::AMDGPU::NUM_FLAVORS, SchedModel, SII, SRI, llvm::AMDGPU::TRANS, TRI, and llvm::AMDGPU::WMMA.

◆ sortHWUIResources()

void CandidateHeuristics::sortHWUIResources ( )

Sort the HardwarUnitInfo vector.

After sorting, the HWUI that are highest priority are first. Priority is determined by maximizing coexecution and keeping the critical HardwareUnit busy.

Definition at line 754 of file AMDGPUCoExecSchedStrategy.cpp.

References A(), B(), HWUInfo, and llvm::sort().

◆ tryCriticalResource()

bool CandidateHeuristics::tryCriticalResource	(	GenericSchedulerBase::SchedCandidate &	TryCand,
		GenericSchedulerBase::SchedCandidate &	Cand,
		SchedBoundary *	Zone ) const

Check for critical resource consumption.

Prefer the candidate that uses the most prioritized HardwareUnit. If both candidates use the same HarwareUnit, prefer the candidate with higher priority on that HardwareUnit.

Definition at line 1022 of file AMDGPUCoExecSchedStrategy.cpp.

References llvm::HardwareUnitInfo::contains(), llvm::AMDGPU::DS, llvm::HardwareUnitInfo::getHigherPriority(), llvm::HardwareUnitInfo::getType(), HWUInfo, I, llvm::GenericSchedulerBase::SchedCandidate::Reason, llvm::GenericSchedulerBase::RegCritical, llvm::GenericSchedulerBase::SchedCandidate::SU, and tryCriticalResourceDependency().

◆ tryCriticalResourceDependency()

bool CandidateHeuristics::tryCriticalResourceDependency	(	GenericSchedulerBase::SchedCandidate &	TryCand,
		GenericSchedulerBase::SchedCandidate &	Cand,
		SchedBoundary *	Zone ) const

Check for dependencies of instructions that use prioritized HardwareUnits.

Prefer the candidate that is a dependency of an instruction that uses the most prioritized HardwareUnit. If both candidates enable the same HardwareUnit, prefer the candidate that enables the higher priority instruction on that HardwareUnit.

Definition at line 935 of file AMDGPUCoExecSchedStrategy.cpp.

References llvm::AMDGPU::classifyFlavor(), DAG, llvm::AMDGPU::DS, llvm::Enabled, llvm::SUnit::getHeight(), llvm::SUnit::getInstr(), llvm::HardwareUnitInfo::getNextTargetSU(), llvm::HardwareUnitInfo::getType(), HWUInfo, I, llvm::GenericSchedulerBase::SchedCandidate::Reason, llvm::GenericSchedulerBase::RegCritical, SII, llvm::GenericSchedulerBase::SchedCandidate::SU, and llvm::AMDGPU::WMMA.

Referenced by tryCriticalResource().

◆ tryEffectiveStall()

bool CandidateHeuristics::tryEffectiveStall	(	GenericSchedulerBase::SchedCandidate &	TryCand,
		GenericSchedulerBase::SchedCandidate &	Cand,
		SchedBoundary &	Zone )

Definition at line 873 of file AMDGPUCoExecSchedStrategy.cpp.

References llvm::CandidateHeuristics::StallCosts::Effective, getStallCosts(), llvm::SchedBoundary::isTop(), LLVM_DEBUG, llvm::GenericSchedulerBase::Stall, llvm::GenericSchedulerBase::SchedCandidate::SU, and llvm::tryLess().

◆ tryMemoryPipeline()

bool CandidateHeuristics::tryMemoryPipeline	(	GenericSchedulerBase::SchedCandidate &	TryCand,
		GenericSchedulerBase::SchedCandidate &	Cand,
		SchedBoundary &	Zone )

Prioritize instructions involved the memory pipeline.

Currently we don't have any modelling of pipelined loads, so we control the layout of the pipeline per iteration by giving the user some control over the stalls (e.g. between s_barrier_signal and s_barrier_wait) and scheduling the pipeline instructions as soon as they are ready.

TODO – add better modelling and heuristics for pipelining based scheduling.

Definition at line 899 of file AMDGPUCoExecSchedStrategy.cpp.

References llvm::AMDGPU::classifyFlavor(), llvm::AMDGPU::DMA, llvm::CandidateHeuristics::StallCosts::Effective, llvm::AMDGPU::Fence, llvm::SUnit::getInstr(), getStallCosts(), llvm::GenericSchedulerBase::SchedCandidate::Reason, llvm::GenericSchedulerBase::RegCritical, SII, and llvm::GenericSchedulerBase::SchedCandidate::SU.