#include <gpu_dyn_inst.hh>

Inheritance diagram for GPUDynInst:

Public Types
typedef std::unordered_map< Addr, std::vector< int > >	StatusVector

Public Member Functions
	GPUDynInst (ComputeUnit _cu, Wavefront _wf, GPUStaticInst *static_inst, uint64_t instSeqNum)

	~GPUDynInst ()

void	execute (GPUDynInstPtr gpuDynInst)

int	numSrcRegOperands ()

int	numDstRegOperands ()

int	numDstVecOperands ()

int	numSrcVecOperands ()

int	numSrcVecDWORDs ()

int	numDstVecDWORDs ()

int	numOpdDWORDs (int operandIdx)

int	getNumOperands ()

bool	isVectorRegister (int operandIdx)

bool	isScalarRegister (int operandIdx)

int	getRegisterIndex (int operandIdx, GPUDynInstPtr gpuDynInst)

int	getOperandSize (int operandIdx)

bool	isDstOperand (int operandIdx)

bool	isSrcOperand (int operandIdx)

bool	hasDestinationSgpr () const

bool	hasSourceSgpr () const

bool	hasDestinationVgpr () const

bool	hasSourceVgpr () const

bool	hasSgprRawDependence (GPUDynInstPtr s)

bool	hasVgprRawDependence (GPUDynInstPtr s)

bool	isOpcode (const std::string &opcodeStr) const

bool	isOpcode (const std::string &opcodeStr, const std::string &extStr) const

bool	srcIsVgpr (int index) const

const std::string &	disassemble () const

InstSeqNum	seqNum () const

Enums::StorageClassType	executedAs ()

void	initiateAcc (GPUDynInstPtr gpuDynInst)

void	completeAcc (GPUDynInstPtr gpuDynInst)

void	updateStats ()

GPUStaticInst *	staticInstruction ()

TheGpuISA::ScalarRegU32	srcLiteral () const

bool	isALU () const
	accessor methods for the attributes of the underlying GPU static instruction More...

bool	isBranch () const

bool	isCondBranch () const

bool	isNop () const

bool	isReturn () const

bool	isEndOfKernel () const

bool	isKernelLaunch () const

bool	isSDWAInst () const

bool	isDPPInst () const

bool	isUnconditionalJump () const

bool	isSpecialOp () const

bool	isWaitcnt () const

bool	isBarrier () const

bool	isMemSync () const

bool	isMemRef () const

bool	isFlat () const

bool	isLoad () const

bool	isStore () const

bool	isAtomic () const

bool	isAtomicNoRet () const

bool	isAtomicRet () const

bool	isScalar () const

bool	isVector () const

bool	readsSCC () const

bool	writesSCC () const

bool	readsVCC () const

bool	writesVCC () const

bool	readsEXEC () const

bool	writesEXEC () const

bool	readsMode () const

bool	writesMode () const

bool	ignoreExec () const

bool	readsFlatScratch () const

bool	writesFlatScratch () const

bool	readsExecMask () const

bool	writesExecMask () const

bool	isAtomicAnd () const

bool	isAtomicOr () const

bool	isAtomicXor () const

bool	isAtomicCAS () const

bool	isAtomicExch () const

bool	isAtomicAdd () const

bool	isAtomicSub () const

bool	isAtomicInc () const

bool	isAtomicDec () const

bool	isAtomicMax () const

bool	isAtomicMin () const

bool	isArgLoad () const

bool	isGlobalMem () const

bool	isLocalMem () const

bool	isArgSeg () const

bool	isGlobalSeg () const

bool	isGroupSeg () const

bool	isKernArgSeg () const

bool	isPrivateSeg () const

bool	isReadOnlySeg () const

bool	isSpillSeg () const

bool	isGloballyCoherent () const

bool	isSystemCoherent () const

bool	isF16 () const

bool	isF32 () const

bool	isF64 () const

bool	isFMA () const

bool	isMAC () const

bool	isMAD () const

void	doApertureCheck (const VectorMask &mask)

void	resolveFlatSegment (const VectorMask &mask)

template<typename c0 >
AtomicOpFunctorPtr	makeAtomicOpFunctor (c0 reg0, c0 reg1)

void	setRequestFlags (RequestPtr req) const

void	resetEntireStatusVector ()

void	resetStatusVector (int lane)

void	setStatusVector (int lane, int newVal)

void	decrementStatusVector (int lane)

int	getLaneStatus (int lane) const

bool	allLanesZero () const

std::string	printStatusVector () const

Tick	getAccessTime () const

void	setAccessTime (Tick currentTime)

void	profileRoundTripTime (Tick currentTime, int hopId)

std::vector< Tick >	getRoundTripTime () const

void	profileLineAddressTime (Addr addr, Tick currentTime, int hopId)

const std::map< Addr, std::vector< Tick > > &	getLineAddressTime () const

Public Member Functions inherited from GPUExecContext
	GPUExecContext (ComputeUnit _cu, Wavefront _wf)

Wavefront *	wavefront ()

ComputeUnit *	computeUnit ()

template<typename T >
T	readConstVal (int opIdx) const

RegVal	readMiscReg (int opIdx) const

void	writeMiscReg (int opIdx, RegVal operandVal)

Public Attributes
Addr	scalarAddr

std::vector< Addr >	addr

Addr	pAddr

uint8_t *	d_data

uint8_t *	scalar_data

uint8_t *	a_data

uint8_t *	x_data

VectorMask	exec_mask

int	simdId

int	wfDynId

int	kern_id

int	cu_id

int	wg_id

int	wfSlotId

int	execUnitId

Tick	time

WaitClass	latency

StatusVector	memStatusVector

std::vector< int >	statusVector

std::vector< int >	tlbHitLevel

int	numScalarReqs

bool	isSaveRestore

Private Attributes
GPUStaticInst *	_staticInst

const InstSeqNum	_seqNum

Tick	accessTime = -1

std::vector< Tick >	roundTripTime

std::map< Addr, std::vector< Tick > >	lineAddressTime

Additional Inherited Members
Protected Attributes inherited from GPUExecContext
ComputeUnit *	cu

Wavefront *	wf

TheGpuISA::GPUISA *	gpuISA

Detailed Description

Definition at line 76 of file gpu_dyn_inst.hh.

Member Typedef Documentation

◆ StatusVector

typedef std::unordered_map<Addr, std::vector<int> > GPUDynInst::StatusVector

Definition at line 403 of file gpu_dyn_inst.hh.

Constructor & Destructor Documentation

◆ GPUDynInst()

GPUDynInst::GPUDynInst	(	ComputeUnit *	_cu,
		Wavefront *	_wf,
		GPUStaticInst *	static_inst,
		uint64_t	instSeqNum
	)

Definition at line 42 of file gpu_dyn_inst.cc.

References a_data, GPUExecContext::computeUnit(), cu_id, ComputeUnit::cu_id, d_data, ArmISA::i, kern_id, Wavefront::kernId, ArmISA::NumVecElemPerVecReg, scalar_data, Wavefront::simdId, simdId, statusVector, time, tlbHitLevel, wfDynId, Wavefront::wfDynId, ComputeUnit::wfSize(), Wavefront::wfSlotId, wfSlotId, wg_id, Wavefront::wgId, and x_data.

◆ ~GPUDynInst()

GPUDynInst::~GPUDynInst ( )

Definition at line 85 of file gpu_dyn_inst.cc.

References _staticInst, a_data, d_data, scalar_data, and x_data.

Member Function Documentation

◆ allLanesZero()

bool GPUDynInst::allLanesZero ( ) const

inline

Definition at line 361 of file gpu_dyn_inst.hh.

References addr, cu_id, DPRINTF, ArmISA::NumVecElemPerVecReg, simdId, statusVector, and wfSlotId.

◆ completeAcc()

void GPUDynInst::completeAcc ( GPUDynInstPtr gpuDynInst )

Definition at line 316 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::completeAcc(), GPUExecContext::cu, ComputeUnit::cu_id, DPRINTF, exec_mask, simdId, and wfSlotId.

◆ decrementStatusVector()

void GPUDynInst::decrementStatusVector ( int lane )

inline

Definition at line 342 of file gpu_dyn_inst.hh.

References statusVector.

◆ disassemble()

const std::string & GPUDynInst::disassemble ( ) const

Definition at line 254 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::disassemble().

◆ doApertureCheck()

void GPUDynInst::doApertureCheck ( const VectorMask & mask )

Definition at line 737 of file gpu_dyn_inst.cc.

References addr, bits(), GPUExecContext::computeUnit(), GPUStaticInst::executed_as, executedAs(), fatal, Shader::isLdsApe(), ArmISA::mask, sc_core::SC_NONE, ComputeUnit::shader, staticInstruction(), and ComputeUnit::wfSize().

Referenced by resolveFlatSegment().

◆ execute()

void GPUDynInst::execute ( GPUDynInstPtr gpuDynInst )

Definition at line 95 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::execute().

◆ executedAs()

Enums::StorageClassType GPUDynInst::executedAs ( )

Definition at line 266 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::executed_as.

Referenced by doApertureCheck(), and resolveFlatSegment().

◆ getAccessTime()

Tick GPUDynInst::getAccessTime ( ) const

inline

Definition at line 416 of file gpu_dyn_inst.hh.

References accessTime.

◆ getLaneStatus()

int GPUDynInst::getLaneStatus ( int lane ) const

inline

Definition at line 353 of file gpu_dyn_inst.hh.

References statusVector.

◆ getLineAddressTime()

const std::map<Addr, std::vector<Tick> >& GPUDynInst::getLineAddressTime ( ) const

inline

Definition at line 424 of file gpu_dyn_inst.hh.

References lineAddressTime.

◆ getNumOperands()

int GPUDynInst::getNumOperands ( )

Definition at line 143 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::getNumOperands().

Referenced by hasSgprRawDependence(), hasVgprRawDependence(), and srcIsVgpr().

◆ getOperandSize()

int GPUDynInst::getOperandSize ( int operandIdx )

Definition at line 167 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::getOperandSize().

◆ getRegisterIndex()

int GPUDynInst::getRegisterIndex	(	int	operandIdx,
		GPUDynInstPtr	gpuDynInst
	)

Definition at line 161 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::getRegisterIndex().

◆ getRoundTripTime()

std::vector<Tick> GPUDynInst::getRoundTripTime ( ) const

inline

Definition at line 421 of file gpu_dyn_inst.hh.

References roundTripTime.

◆ hasDestinationSgpr()

bool GPUDynInst::hasDestinationSgpr ( ) const

Definition at line 207 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isDstOperand(), and GPUStaticInst::isScalarRegister().

◆ hasDestinationVgpr()

bool GPUDynInst::hasDestinationVgpr ( ) const

Definition at line 229 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isDstOperand(), and GPUStaticInst::isVectorRegister().

◆ hasSgprRawDependence()

bool GPUDynInst::hasSgprRawDependence ( GPUDynInstPtr s )

Definition at line 289 of file gpu_dyn_inst.cc.

References getNumOperands(), ArmISA::i, isScalarRegister(), isSrcOperand(), ArmISA::j, and ArmISA::s.

◆ hasSourceSgpr()

bool GPUDynInst::hasSourceSgpr ( ) const

Definition at line 185 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isScalarRegister(), and GPUStaticInst::isSrcOperand().

◆ hasSourceVgpr()

bool GPUDynInst::hasSourceVgpr ( ) const

Definition at line 196 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isSrcOperand(), and GPUStaticInst::isVectorRegister().

◆ hasVgprRawDependence()

bool GPUDynInst::hasVgprRawDependence ( GPUDynInstPtr s )

Definition at line 272 of file gpu_dyn_inst.cc.

References getNumOperands(), ArmISA::i, isSrcOperand(), isVectorRegister(), ArmISA::j, and ArmISA::s.

◆ ignoreExec()

bool GPUDynInst::ignoreExec ( ) const

Definition at line 516 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::ignoreExec().

◆ initiateAcc()

void GPUDynInst::initiateAcc ( GPUDynInstPtr gpuDynInst )

Definition at line 307 of file gpu_dyn_inst.cc.

References _staticInst, GPUExecContext::cu, ComputeUnit::cu_id, DPRINTF, exec_mask, GPUStaticInst::initiateAcc(), simdId, and wfSlotId.

◆ isALU()

bool GPUDynInst::isALU ( ) const

accessor methods for the attributes of the underlying GPU static instruction

Definition at line 330 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isALU().

◆ isArgLoad()

bool GPUDynInst::isArgLoad ( ) const

Definition at line 629 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isArgLoad().

◆ isArgSeg()

bool GPUDynInst::isArgSeg ( ) const

Definition at line 647 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isArgSeg().

◆ isAtomic()

bool GPUDynInst::isAtomic ( ) const

Definition at line 438 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomic().

Referenced by resolveFlatSegment().

◆ isAtomicAdd()

bool GPUDynInst::isAtomicAdd ( ) const

Definition at line 593 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicAdd().

Referenced by makeAtomicOpFunctor().

◆ isAtomicAnd()

bool GPUDynInst::isAtomicAnd ( ) const

Definition at line 564 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicAnd().

Referenced by makeAtomicOpFunctor().

◆ isAtomicCAS()

bool GPUDynInst::isAtomicCAS ( ) const

Definition at line 582 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicCAS().

Referenced by makeAtomicOpFunctor().

◆ isAtomicDec()

bool GPUDynInst::isAtomicDec ( ) const

Definition at line 611 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicDec().

Referenced by makeAtomicOpFunctor().

◆ isAtomicExch()

bool GPUDynInst::isAtomicExch ( ) const

Definition at line 587 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicExch().

Referenced by makeAtomicOpFunctor().

◆ isAtomicInc()

bool GPUDynInst::isAtomicInc ( ) const

Definition at line 605 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicInc().

Referenced by makeAtomicOpFunctor().

◆ isAtomicMax()

bool GPUDynInst::isAtomicMax ( ) const

Definition at line 617 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicMax().

Referenced by makeAtomicOpFunctor().

◆ isAtomicMin()

bool GPUDynInst::isAtomicMin ( ) const

Definition at line 623 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicMin().

Referenced by makeAtomicOpFunctor().

◆ isAtomicNoRet()

bool GPUDynInst::isAtomicNoRet ( ) const

Definition at line 444 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicNoRet().

Referenced by setRequestFlags().

◆ isAtomicOr()

bool GPUDynInst::isAtomicOr ( ) const

Definition at line 570 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicOr().

Referenced by makeAtomicOpFunctor().

◆ isAtomicRet()

bool GPUDynInst::isAtomicRet ( ) const

Definition at line 450 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicRet().

Referenced by setRequestFlags().

◆ isAtomicSub()

bool GPUDynInst::isAtomicSub ( ) const

Definition at line 599 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicSub().

Referenced by makeAtomicOpFunctor().

◆ isAtomicXor()

bool GPUDynInst::isAtomicXor ( ) const

Definition at line 576 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isAtomicXor().

Referenced by makeAtomicOpFunctor().

◆ isBarrier()

bool GPUDynInst::isBarrier ( ) const

Definition at line 402 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isBarrier().

◆ isBranch()

bool GPUDynInst::isBranch ( ) const

Definition at line 336 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isBranch().

◆ isCondBranch()

bool GPUDynInst::isCondBranch ( ) const

Definition at line 342 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isCondBranch().

◆ isDPPInst()

bool GPUDynInst::isDPPInst ( ) const

Definition at line 372 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isDPPInst().

◆ isDstOperand()

bool GPUDynInst::isDstOperand ( int operandIdx )

Definition at line 173 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isDstOperand().

◆ isEndOfKernel()

bool GPUDynInst::isEndOfKernel ( ) const

Definition at line 354 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isEndOfKernel().

Referenced by setRequestFlags().

◆ isF16()

bool GPUDynInst::isF16 ( ) const

Definition at line 701 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isF16().

◆ isF32()

bool GPUDynInst::isF32 ( ) const

Definition at line 707 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isF32().

◆ isF64()

bool GPUDynInst::isF64 ( ) const

Definition at line 713 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isF64().

◆ isFlat()

bool GPUDynInst::isFlat ( ) const

Definition at line 420 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isFlat().

◆ isFMA()

bool GPUDynInst::isFMA ( ) const

Definition at line 719 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isFMA().

◆ isGloballyCoherent()

bool GPUDynInst::isGloballyCoherent ( ) const

Definition at line 689 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isGloballyCoherent().

Referenced by setRequestFlags().

◆ isGlobalMem()

bool GPUDynInst::isGlobalMem ( ) const

Definition at line 635 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isGlobalMem().

◆ isGlobalSeg()

bool GPUDynInst::isGlobalSeg ( ) const

Definition at line 653 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isGlobalSeg().

◆ isGroupSeg()

bool GPUDynInst::isGroupSeg ( ) const

Definition at line 659 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isGroupSeg().

◆ isKernArgSeg()

bool GPUDynInst::isKernArgSeg ( ) const

Definition at line 665 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isKernArgSeg().

◆ isKernelLaunch()

bool GPUDynInst::isKernelLaunch ( ) const

Definition at line 360 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isKernelLaunch().

Referenced by setRequestFlags().

◆ isLoad()

bool GPUDynInst::isLoad ( ) const

Definition at line 426 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isLoad().

Referenced by resolveFlatSegment().

◆ isLocalMem()

bool GPUDynInst::isLocalMem ( ) const

Definition at line 641 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isLocalMem().

◆ isMAC()

bool GPUDynInst::isMAC ( ) const

Definition at line 725 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isMAC().

◆ isMAD()

bool GPUDynInst::isMAD ( ) const

Definition at line 731 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isMAD().

◆ isMemRef()

bool GPUDynInst::isMemRef ( ) const

Definition at line 414 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isMemRef().

◆ isMemSync()

bool GPUDynInst::isMemSync ( ) const

Definition at line 408 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isMemSync().

Referenced by resolveFlatSegment(), and setRequestFlags().

◆ isNop()

bool GPUDynInst::isNop ( ) const

Definition at line 348 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isNop().

◆ isOpcode() [1/2]

bool GPUDynInst::isOpcode ( const std::string & opcodeStr ) const

Definition at line 248 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::opcode().

◆ isOpcode() [2/2]

bool GPUDynInst::isOpcode	(	const std::string &	opcodeStr,
		const std::string &	extStr
	)		const

Definition at line 240 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::opcode().

◆ isPrivateSeg()

bool GPUDynInst::isPrivateSeg ( ) const

Definition at line 671 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isPrivateSeg().

◆ isReadOnlySeg()

bool GPUDynInst::isReadOnlySeg ( ) const

Definition at line 677 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isReadOnlySeg().

◆ isReturn()

bool GPUDynInst::isReturn ( ) const

Definition at line 378 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isReturn().

◆ isScalar()

bool GPUDynInst::isScalar ( ) const

Definition at line 462 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isScalar().

◆ isScalarRegister()

bool GPUDynInst::isScalarRegister ( int operandIdx )

Definition at line 155 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isScalarRegister().

Referenced by hasSgprRawDependence().

◆ isSDWAInst()

bool GPUDynInst::isSDWAInst ( ) const

Definition at line 366 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isSDWAInst().

◆ isSpecialOp()

bool GPUDynInst::isSpecialOp ( ) const

Definition at line 390 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isSpecialOp().

◆ isSpillSeg()

bool GPUDynInst::isSpillSeg ( ) const

Definition at line 683 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isSpillSeg().

◆ isSrcOperand()

bool GPUDynInst::isSrcOperand ( int operandIdx )

Definition at line 179 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isSrcOperand().

Referenced by hasSgprRawDependence(), and hasVgprRawDependence().

◆ isStore()

bool GPUDynInst::isStore ( ) const

Definition at line 432 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isStore().

Referenced by resolveFlatSegment().

◆ isSystemCoherent()

bool GPUDynInst::isSystemCoherent ( ) const

Definition at line 695 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isSystemCoherent().

Referenced by setRequestFlags().

◆ isUnconditionalJump()

bool GPUDynInst::isUnconditionalJump ( ) const

Definition at line 384 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isUnconditionalJump().

◆ isVector()

bool GPUDynInst::isVector ( ) const

Definition at line 456 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isScalar().

◆ isVectorRegister()

bool GPUDynInst::isVectorRegister ( int operandIdx )

Definition at line 149 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isVectorRegister().

Referenced by hasVgprRawDependence().

◆ isWaitcnt()

bool GPUDynInst::isWaitcnt ( ) const

Definition at line 396 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::isWaitcnt().

◆ makeAtomicOpFunctor()

template<typename c0 >

AtomicOpFunctorPtr GPUDynInst::makeAtomicOpFunctor	(	c0 *	reg0,
		c0 *	reg1
	)

inline

Definition at line 255 of file gpu_dyn_inst.hh.

References GPUExecContext::cu, fatal, isAtomicAdd(), isAtomicAnd(), isAtomicCAS(), isAtomicDec(), isAtomicExch(), isAtomicInc(), isAtomicMax(), isAtomicMin(), isAtomicOr(), isAtomicSub(), and isAtomicXor().

◆ numDstRegOperands()

int GPUDynInst::numDstRegOperands ( )

Definition at line 107 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::numDstRegOperands().

◆ numDstVecDWORDs()

int GPUDynInst::numDstVecDWORDs ( )

Definition at line 131 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::numDstVecDWORDs().

◆ numDstVecOperands()

int GPUDynInst::numDstVecOperands ( )

Definition at line 119 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::numDstVecOperands().

◆ numOpdDWORDs()

int GPUDynInst::numOpdDWORDs ( int operandIdx )

Definition at line 137 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::numOpdDWORDs().

◆ numSrcRegOperands()

int GPUDynInst::numSrcRegOperands ( )

Definition at line 101 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::numSrcRegOperands().

◆ numSrcVecDWORDs()

int GPUDynInst::numSrcVecDWORDs ( )

Definition at line 125 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::numSrcVecDWORDs().

◆ numSrcVecOperands()

int GPUDynInst::numSrcVecOperands ( )

Definition at line 113 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::numSrcVecOperands().

◆ printStatusVector()

std::string GPUDynInst::printStatusVector ( ) const

inline

Definition at line 387 of file gpu_dyn_inst.hh.

References ArmISA::NumVecElemPerVecReg, statusVector, and sc_dt::to_string().

◆ profileLineAddressTime()

void GPUDynInst::profileLineAddressTime	(	Addr	addr,
		Tick	currentTime,
		int	hopId
	)

Definition at line 981 of file gpu_dyn_inst.cc.

References addr, and lineAddressTime.

◆ profileRoundTripTime()

void GPUDynInst::profileRoundTripTime	(	Tick	currentTime,
		int	hopId
	)

Definition at line 971 of file gpu_dyn_inst.cc.

References roundTripTime.

◆ readsEXEC()

bool GPUDynInst::readsEXEC ( ) const

Definition at line 504 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::readsEXEC().

◆ readsExecMask()

bool GPUDynInst::readsExecMask ( ) const

Definition at line 532 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isExecMaskRegister(), and GPUStaticInst::isSrcOperand().

◆ readsFlatScratch()

bool GPUDynInst::readsFlatScratch ( ) const

Definition at line 553 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isFlatScratchRegister(), GPUStaticInst::isScalarRegister(), and GPUStaticInst::isSrcOperand().

◆ readsMode()

bool GPUDynInst::readsMode ( ) const

Definition at line 492 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::readsMode().

◆ readsSCC()

bool GPUDynInst::readsSCC ( ) const

Definition at line 468 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::readsSCC().

◆ readsVCC()

bool GPUDynInst::readsVCC ( ) const

Definition at line 480 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::readsVCC().

◆ resetEntireStatusVector()

void GPUDynInst::resetEntireStatusVector ( )

inline

Definition at line 314 of file gpu_dyn_inst.hh.

References ArmISA::NumVecElemPerVecReg, resetStatusVector(), and statusVector.

◆ resetStatusVector()

void GPUDynInst::resetStatusVector ( int lane )

inline

Definition at line 324 of file gpu_dyn_inst.hh.

References setStatusVector().

Referenced by resetEntireStatusVector().

◆ resolveFlatSegment()

void GPUDynInst::resolveFlatSegment ( const VectorMask & mask )

Flat instructions may resolve to the private segment (scratch), which is backed by main memory and provides per-lane scratch memory. Flat addressing uses apertures - registers that specify the address range in the VA space where LDS/private memory is mapped. The value of which is set by the kernel mode driver. These apertures use addresses that are not used by x86 CPUs. When the address of a Flat operation falls into one of the apertures, the Flat operation is redirected to either LDS or to the private memory segment.

For private memory the SW runtime will allocate some space in the VA space for each AQL queue. The base address of which is stored in scalar registers per the AMD GPU ABI. The amd_queue_t scratch_backing_memory_location provides the base address in memory for the queue's private segment. Various other fields loaded into register state during kernel launch specify per-WF and per-work-item offsets so that individual lanes may access their private segment allocation.

For more details about flat addressing see: http://rocm-documentation.readthedocs.io/en/latest/ ROCm_Compiler_SDK/ROCm-Native-ISA.html#flat-scratch

https://github.com/ROCm-Developer-Tools/ ROCm-ComputeABI-Doc/blob/master/AMDGPU-ABI.md #flat-addressing

Definition at line 809 of file gpu_dyn_inst.cc.

References addr, Wavefront::computeUnit, Wavefront::decLGKMInstsIssued(), Wavefront::decVMemInstsIssued(), doApertureCheck(), Wavefront::execUnitId, executedAs(), Wavefront::flatGmUnitId, Wavefront::flatLmUnitId, Shader::getHiddenPrivateBase(), Shader::getScratchBase(), isAtomic(), isLoad(), isMemSync(), isStore(), RegisterManager::mapSgpr(), ArmISA::mask, Wavefront::maxSgprs, ArmISA::offset, panic, Wavefront::rdGmReqsInPipe, Wavefront::rdLmReqsInPipe, ComputeUnit::registerManager, ComputeUnit::shader, simdId, ComputeUnit::srf, GPUExecContext::wavefront(), ComputeUnit::wfSize(), Wavefront::wrGmReqsInPipe, and Wavefront::wrLmReqsInPipe.

◆ seqNum()

InstSeqNum GPUDynInst::seqNum ( ) const

Definition at line 260 of file gpu_dyn_inst.cc.

References _seqNum.

◆ setAccessTime()

void GPUDynInst::setAccessTime ( Tick currentTime )

inline

Definition at line 418 of file gpu_dyn_inst.hh.

References accessTime.

◆ setRequestFlags()

void GPUDynInst::setRequestFlags ( RequestPtr req ) const

inline

Definition at line 285 of file gpu_dyn_inst.hh.

References Request::ACQUIRE, Request::ATOMIC_NO_RETURN_OP, Request::ATOMIC_RETURN_OP, Request::GLC_BIT, isAtomicNoRet(), isAtomicRet(), isEndOfKernel(), isGloballyCoherent(), isKernelLaunch(), isMemSync(), isSystemCoherent(), and Request::SLC_BIT.

◆ setStatusVector()

void GPUDynInst::setStatusVector	(	int	lane,
		int	newVal
	)

inline

Definition at line 331 of file gpu_dyn_inst.hh.

References statusVector.

Referenced by resetStatusVector().

◆ srcIsVgpr()

bool GPUDynInst::srcIsVgpr ( int index ) const

Definition at line 218 of file gpu_dyn_inst.cc.

References _staticInst, getNumOperands(), MipsISA::index, GPUStaticInst::isSrcOperand(), and GPUStaticInst::isVectorRegister().

◆ srcLiteral()

TheGpuISA::ScalarRegU32 GPUDynInst::srcLiteral ( ) const

Definition at line 924 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::srcLiteral().

◆ staticInstruction()

GPUStaticInst* GPUDynInst::staticInstruction ( )

inline

Definition at line 166 of file gpu_dyn_inst.hh.

References _staticInst.

Referenced by doApertureCheck().

◆ updateStats()

void GPUDynInst::updateStats ( )

Definition at line 930 of file gpu_dyn_inst.cc.

References _staticInst, GPUExecContext::cu, ComputeUnit::dynamicFlatMemInstrCnt, ComputeUnit::dynamicGMemInstrCnt, ComputeUnit::dynamicLMemInstrCnt, GPUStaticInst::isFlat(), GPUStaticInst::isLocalMem(), ComputeUnit::pageAccesses, ComputeUnit::pageDivergenceDist, ComputeUnit::pagesTouched, and Stats::DistBase< Derived, Stor >::sample().

◆ writesEXEC()

bool GPUDynInst::writesEXEC ( ) const

Definition at line 510 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::writesEXEC().

◆ writesExecMask()

bool GPUDynInst::writesExecMask ( ) const

Definition at line 522 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isDstOperand(), and GPUStaticInst::isExecMaskRegister().

◆ writesFlatScratch()

bool GPUDynInst::writesFlatScratch ( ) const

Definition at line 542 of file gpu_dyn_inst.cc.

References _staticInst, GPUStaticInst::getNumOperands(), ArmISA::i, GPUStaticInst::isDstOperand(), GPUStaticInst::isFlatScratchRegister(), and GPUStaticInst::isScalarRegister().

◆ writesMode()

bool GPUDynInst::writesMode ( ) const

Definition at line 498 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::writesMode().

◆ writesSCC()

bool GPUDynInst::writesSCC ( ) const

Definition at line 474 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::writesSCC().

◆ writesVCC()

bool GPUDynInst::writesVCC ( ) const

Definition at line 486 of file gpu_dyn_inst.cc.

References _staticInst, and GPUStaticInst::writesVCC().

Member Data Documentation

◆ _seqNum

const InstSeqNum GPUDynInst::_seqNum

private

Definition at line 431 of file gpu_dyn_inst.hh.

Referenced by seqNum().

◆ _staticInst

GPUStaticInst* GPUDynInst::_staticInst

private

Definition at line 430 of file gpu_dyn_inst.hh.

Referenced by completeAcc(), disassemble(), execute(), executedAs(), getNumOperands(), getOperandSize(), getRegisterIndex(), hasDestinationSgpr(), hasDestinationVgpr(), hasSourceSgpr(), hasSourceVgpr(), ignoreExec(), initiateAcc(), isALU(), isArgLoad(), isArgSeg(), isAtomic(), isAtomicAdd(), isAtomicAnd(), isAtomicCAS(), isAtomicDec(), isAtomicExch(), isAtomicInc(), isAtomicMax(), isAtomicMin(), isAtomicNoRet(), isAtomicOr(), isAtomicRet(), isAtomicSub(), isAtomicXor(), isBarrier(), isBranch(), isCondBranch(), isDPPInst(), isDstOperand(), isEndOfKernel(), isF16(), isF32(), isF64(), isFlat(), isFMA(), isGloballyCoherent(), isGlobalMem(), isGlobalSeg(), isGroupSeg(), isKernArgSeg(), isKernelLaunch(), isLoad(), isLocalMem(), isMAC(), isMAD(), isMemRef(), isMemSync(), isNop(), isOpcode(), isPrivateSeg(), isReadOnlySeg(), isReturn(), isScalar(), isScalarRegister(), isSDWAInst(), isSpecialOp(), isSpillSeg(), isSrcOperand(), isStore(), isSystemCoherent(), isUnconditionalJump(), isVector(), isVectorRegister(), isWaitcnt(), numDstRegOperands(), numDstVecDWORDs(), numDstVecOperands(), numOpdDWORDs(), numSrcRegOperands(), numSrcVecDWORDs(), numSrcVecOperands(), readsEXEC(), readsExecMask(), readsFlatScratch(), readsMode(), readsSCC(), readsVCC(), srcIsVgpr(), srcLiteral(), staticInstruction(), updateStats(), writesEXEC(), writesExecMask(), writesFlatScratch(), writesMode(), writesSCC(), writesVCC(), and ~GPUDynInst().

◆ a_data

uint8_t* GPUDynInst::a_data

Definition at line 131 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst(), and ~GPUDynInst().

◆ accessTime

Tick GPUDynInst::accessTime = -1

private

Definition at line 434 of file gpu_dyn_inst.hh.

Referenced by getAccessTime(), and setAccessTime().

◆ addr

std::vector<Addr> GPUDynInst::addr

Definition at line 123 of file gpu_dyn_inst.hh.

Referenced by allLanesZero(), doApertureCheck(), profileLineAddressTime(), and resolveFlatSegment().

◆ cu_id

int GPUDynInst::cu_id

Definition at line 144 of file gpu_dyn_inst.hh.

Referenced by allLanesZero(), and GPUDynInst().

◆ d_data

uint8_t* GPUDynInst::d_data

Definition at line 127 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst(), and ~GPUDynInst().

◆ exec_mask

VectorMask GPUDynInst::exec_mask

Definition at line 135 of file gpu_dyn_inst.hh.

Referenced by completeAcc(), and initiateAcc().

◆ execUnitId

int GPUDynInst::execUnitId

Definition at line 150 of file gpu_dyn_inst.hh.

◆ isSaveRestore

bool GPUDynInst::isSaveRestore

Definition at line 428 of file gpu_dyn_inst.hh.

◆ kern_id

int GPUDynInst::kern_id

Definition at line 142 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ latency

WaitClass GPUDynInst::latency

Definition at line 154 of file gpu_dyn_inst.hh.

◆ lineAddressTime

std::map<Addr, std::vector<Tick> > GPUDynInst::lineAddressTime

private

Definition at line 442 of file gpu_dyn_inst.hh.

Referenced by getLineAddressTime(), and profileLineAddressTime().

◆ memStatusVector

StatusVector GPUDynInst::memStatusVector

Definition at line 404 of file gpu_dyn_inst.hh.

◆ numScalarReqs

int GPUDynInst::numScalarReqs

Definition at line 414 of file gpu_dyn_inst.hh.

◆ pAddr

Addr GPUDynInst::pAddr

Definition at line 124 of file gpu_dyn_inst.hh.

◆ roundTripTime

std::vector<Tick> GPUDynInst::roundTripTime

private

Definition at line 438 of file gpu_dyn_inst.hh.

Referenced by getRoundTripTime(), and profileRoundTripTime().

◆ scalar_data

uint8_t* GPUDynInst::scalar_data

Definition at line 129 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst(), and ~GPUDynInst().

◆ scalarAddr

Addr GPUDynInst::scalarAddr

Definition at line 121 of file gpu_dyn_inst.hh.

◆ simdId

int GPUDynInst::simdId

Definition at line 138 of file gpu_dyn_inst.hh.

Referenced by allLanesZero(), completeAcc(), GPUDynInst(), initiateAcc(), and resolveFlatSegment().

◆ statusVector

std::vector<int> GPUDynInst::statusVector

Definition at line 408 of file gpu_dyn_inst.hh.

Referenced by allLanesZero(), decrementStatusVector(), getLaneStatus(), GPUDynInst(), printStatusVector(), resetEntireStatusVector(), and setStatusVector().

◆ time

Tick GPUDynInst::time

Definition at line 152 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ tlbHitLevel

std::vector<int> GPUDynInst::tlbHitLevel

Definition at line 410 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ wfDynId

int GPUDynInst::wfDynId

Definition at line 140 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ wfSlotId

int GPUDynInst::wfSlotId

Definition at line 148 of file gpu_dyn_inst.hh.

Referenced by allLanesZero(), completeAcc(), GPUDynInst(), and initiateAcc().

◆ wg_id

int GPUDynInst::wg_id

Definition at line 146 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ x_data

uint8_t* GPUDynInst::x_data

Definition at line 133 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst(), and ~GPUDynInst().

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

gpu-compute/gpu_dyn_inst.hh
gpu-compute/gpu_dyn_inst.cc

Public Types

Public Member Functions

Public Attributes

Private Attributes

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ StatusVector

Constructor & Destructor Documentation

◆ GPUDynInst()

◆ ~GPUDynInst()

Member Function Documentation

◆ allLanesZero()

◆ completeAcc()

◆ decrementStatusVector()

◆ disassemble()

◆ doApertureCheck()

◆ execute()

◆ executedAs()

◆ getAccessTime()

◆ getLaneStatus()

◆ getLineAddressTime()

◆ getNumOperands()

◆ getOperandSize()

◆ getRegisterIndex()

◆ getRoundTripTime()

◆ hasDestinationSgpr()

◆ hasDestinationVgpr()

◆ hasSgprRawDependence()

◆ hasSourceSgpr()

◆ hasSourceVgpr()

◆ hasVgprRawDependence()

◆ ignoreExec()

◆ initiateAcc()

◆ isALU()

◆ isArgLoad()

◆ isArgSeg()

◆ isAtomic()

◆ isAtomicAdd()

◆ isAtomicAnd()

◆ isAtomicCAS()

◆ isAtomicDec()

◆ isAtomicExch()

◆ isAtomicInc()

◆ isAtomicMax()

◆ isAtomicMin()

◆ isAtomicNoRet()

◆ isAtomicOr()

◆ isAtomicRet()

◆ isAtomicSub()

◆ isAtomicXor()

◆ isBarrier()

◆ isBranch()

◆ isCondBranch()

◆ isDPPInst()

◆ isDstOperand()

◆ isEndOfKernel()

◆ isF16()

◆ isF32()

◆ isF64()

◆ isFlat()

◆ isFMA()

◆ isGloballyCoherent()

◆ isGlobalMem()

◆ isGlobalSeg()

◆ isGroupSeg()

◆ isKernArgSeg()

◆ isKernelLaunch()

◆ isLoad()

◆ isLocalMem()

◆ isMAC()

◆ isMAD()

◆ isMemRef()

◆ isMemSync()

◆ isNop()

◆ isOpcode() [1/2]

◆ isOpcode() [2/2]

◆ isPrivateSeg()

◆ isReadOnlySeg()

◆ isReturn()