#include <gpu_dyn_inst.hh>

Inheritance diagram for gem5::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)

const std::vector< OperandInfo > &	srcVecRegOperands () const

const std::vector< OperandInfo > &	dstVecRegOperands () const

const std::vector< OperandInfo > &	srcScalarRegOperands () const

const std::vector< OperandInfo > &	dstScalarRegOperands () const

int	numSrcRegOperands ()

int	numDstRegOperands ()

int	numSrcVecRegOperands () const

int	numDstVecRegOperands () const

int	maxSrcVecRegOperandSize ()

int	numSrcVecDWords ()

int	numDstVecDWords ()

int	numSrcScalarRegOperands () const

int	numDstScalarRegOperands () const

int	maxSrcScalarRegOperandSize ()

int	numSrcScalarDWords ()

int	numDstScalarDWords ()

int	maxOperandSize ()

int	getNumOperands () const

bool	hasSourceSgpr () const

bool	hasDestinationSgpr () const

bool	hasSourceVgpr () const

bool	hasDestinationVgpr () const

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

bool	isOpcode (const std::string &opcodeStr, const std::string &extStr) 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	isSleep () 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 gem5::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

int	maxSrcVecRegOpSize

int	maxSrcScalarRegOpSize

Tick	accessTime = -1

std::vector< Tick >	roundTripTime

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

Additional Inherited Members
Protected Attributes inherited from gem5::GPUExecContext
ComputeUnit *	cu

Wavefront *	wf

TheGpuISA::GPUISA *	gpuISA

Detailed Description

Definition at line 118 of file gpu_dyn_inst.hh.

Member Typedef Documentation

◆ StatusVector

typedef std::unordered_map<Addr, std::vector<int> > gem5::GPUDynInst::StatusVector

Definition at line 451 of file gpu_dyn_inst.hh.

Constructor & Destructor Documentation

◆ GPUDynInst()

gem5::GPUDynInst::GPUDynInst	(	ComputeUnit *	_cu,
		Wavefront *	_wf,
		GPUStaticInst *	static_inst,
		uint64_t	instSeqNum
	)

◆ ~GPUDynInst()

gem5::GPUDynInst::~GPUDynInst ( )

Definition at line 98 of file gpu_dyn_inst.cc.

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

Member Function Documentation

◆ allLanesZero()

bool gem5::GPUDynInst::allLanesZero ( ) const

inline

Definition at line 409 of file gpu_dyn_inst.hh.

References addr, cu_id, DPRINTF, gem5::ArmISA::NumVecElemPerVecReg, simdId, statusVector, and wfSlotId.

◆ completeAcc()

void gem5::GPUDynInst::completeAcc ( GPUDynInstPtr gpuDynInst )

Definition at line 304 of file gpu_dyn_inst.cc.

References _staticInst, gem5::GPUStaticInst::completeAcc(), gem5::GPUExecContext::cu, gem5::ComputeUnit::cu_id, DPRINTF, exec_mask, simdId, and wfSlotId.

◆ decrementStatusVector()

void gem5::GPUDynInst::decrementStatusVector ( int lane )

inline

Definition at line 390 of file gpu_dyn_inst.hh.

References statusVector.

◆ disassemble()

const std::string & gem5::GPUDynInst::disassemble ( ) const

Definition at line 276 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::disassemble().

Referenced by GPUDynInst().

◆ doApertureCheck()

void gem5::GPUDynInst::doApertureCheck ( const VectorMask & mask )

Definition at line 737 of file gpu_dyn_inst.cc.

References addr, gem5::bits(), gem5::GPUExecContext::computeUnit(), gem5::GPUStaticInst::executed_as, executedAs(), fatal, gem5::Shader::isLdsApe(), gem5::mask(), sc_core::SC_NONE, gem5::ComputeUnit::shader, staticInstruction(), and gem5::ComputeUnit::wfSize().

Referenced by resolveFlatSegment().

◆ dstScalarRegOperands()

const std::vector< OperandInfo > & gem5::GPUDynInst::dstScalarRegOperands ( ) const

Definition at line 132 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::dstScalarRegOperands().

Referenced by hasDestinationSgpr(), and writesFlatScratch().

◆ dstVecRegOperands()

const std::vector< OperandInfo > & gem5::GPUDynInst::dstVecRegOperands ( ) const

Definition at line 120 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::dstVecRegOperands().

Referenced by hasDestinationVgpr().

◆ execute()

void gem5::GPUDynInst::execute ( GPUDynInstPtr gpuDynInst )

Definition at line 108 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::execute().

◆ executedAs()

enums::StorageClassType gem5::GPUDynInst::executedAs ( )

Definition at line 288 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::executed_as.

Referenced by doApertureCheck(), and resolveFlatSegment().

◆ getAccessTime()

Tick gem5::GPUDynInst::getAccessTime ( ) const

inline

Definition at line 464 of file gpu_dyn_inst.hh.

References accessTime.

◆ getLaneStatus()

int gem5::GPUDynInst::getLaneStatus ( int lane ) const

inline

Definition at line 401 of file gpu_dyn_inst.hh.

References statusVector.

◆ getLineAddressTime()

const std::map<Addr, std::vector<Tick> >& gem5::GPUDynInst::getLineAddressTime ( ) const

inline

Definition at line 472 of file gpu_dyn_inst.hh.

References lineAddressTime.

◆ getNumOperands()

int gem5::GPUDynInst::getNumOperands ( ) const

Definition at line 232 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::getNumOperands().

Referenced by GPUDynInst().

◆ getRoundTripTime()

std::vector<Tick> gem5::GPUDynInst::getRoundTripTime ( ) const

inline

Definition at line 469 of file gpu_dyn_inst.hh.

References roundTripTime.

◆ hasDestinationSgpr()

bool gem5::GPUDynInst::hasDestinationSgpr ( ) const

Definition at line 256 of file gpu_dyn_inst.cc.

References dstScalarRegOperands().

◆ hasDestinationVgpr()

bool gem5::GPUDynInst::hasDestinationVgpr ( ) const

Definition at line 244 of file gpu_dyn_inst.cc.

References dstVecRegOperands().

◆ hasSourceSgpr()

bool gem5::GPUDynInst::hasSourceSgpr ( ) const

Definition at line 250 of file gpu_dyn_inst.cc.

References srcScalarRegOperands().

◆ hasSourceVgpr()

bool gem5::GPUDynInst::hasSourceVgpr ( ) const

Definition at line 238 of file gpu_dyn_inst.cc.

References srcVecRegOperands().

◆ ignoreExec()

bool gem5::GPUDynInst::ignoreExec ( ) const

Definition at line 518 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::ignoreExec().

◆ initiateAcc()

void gem5::GPUDynInst::initiateAcc ( GPUDynInstPtr gpuDynInst )

Definition at line 295 of file gpu_dyn_inst.cc.

References _staticInst, gem5::GPUExecContext::cu, gem5::ComputeUnit::cu_id, DPRINTF, exec_mask, gem5::GPUStaticInst::initiateAcc(), simdId, and wfSlotId.

◆ isALU()

bool gem5::GPUDynInst::isALU ( ) const

accessor methods for the attributes of the underlying GPU static instruction

Definition at line 318 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isALU().

◆ isArgLoad()

bool gem5::GPUDynInst::isArgLoad ( ) const

Definition at line 629 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isArgLoad().

◆ isArgSeg()

bool gem5::GPUDynInst::isArgSeg ( ) const

Definition at line 647 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isArgSeg().

◆ isAtomic()

bool gem5::GPUDynInst::isAtomic ( ) const

Definition at line 432 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomic().

Referenced by resolveFlatSegment().

◆ isAtomicAdd()

bool gem5::GPUDynInst::isAtomicAdd ( ) const

Definition at line 593 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicAdd().

Referenced by makeAtomicOpFunctor().

◆ isAtomicAnd()

bool gem5::GPUDynInst::isAtomicAnd ( ) const

Definition at line 564 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicAnd().

Referenced by makeAtomicOpFunctor().

◆ isAtomicCAS()

bool gem5::GPUDynInst::isAtomicCAS ( ) const

Definition at line 582 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicCAS().

Referenced by makeAtomicOpFunctor().

◆ isAtomicDec()

bool gem5::GPUDynInst::isAtomicDec ( ) const

Definition at line 611 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicDec().

Referenced by makeAtomicOpFunctor().

◆ isAtomicExch()

bool gem5::GPUDynInst::isAtomicExch ( ) const

Definition at line 587 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicExch().

Referenced by makeAtomicOpFunctor().

◆ isAtomicInc()

bool gem5::GPUDynInst::isAtomicInc ( ) const

Definition at line 605 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicInc().

Referenced by makeAtomicOpFunctor().

◆ isAtomicMax()

bool gem5::GPUDynInst::isAtomicMax ( ) const

Definition at line 617 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicMax().

Referenced by makeAtomicOpFunctor().

◆ isAtomicMin()

bool gem5::GPUDynInst::isAtomicMin ( ) const

Definition at line 623 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicMin().

Referenced by makeAtomicOpFunctor().

◆ isAtomicNoRet()

bool gem5::GPUDynInst::isAtomicNoRet ( ) const

Definition at line 438 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicNoRet().

Referenced by setRequestFlags().

◆ isAtomicOr()

bool gem5::GPUDynInst::isAtomicOr ( ) const

Definition at line 570 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicOr().

Referenced by makeAtomicOpFunctor().

◆ isAtomicRet()

bool gem5::GPUDynInst::isAtomicRet ( ) const

Definition at line 444 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicRet().

Referenced by setRequestFlags().

◆ isAtomicSub()

bool gem5::GPUDynInst::isAtomicSub ( ) const

Definition at line 599 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicSub().

Referenced by makeAtomicOpFunctor().

◆ isAtomicXor()

bool gem5::GPUDynInst::isAtomicXor ( ) const

Definition at line 576 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isAtomicXor().

Referenced by makeAtomicOpFunctor().

◆ isBarrier()

bool gem5::GPUDynInst::isBarrier ( ) const

Definition at line 396 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isBarrier().

◆ isBranch()

bool gem5::GPUDynInst::isBranch ( ) const

Definition at line 324 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isBranch().

◆ isCondBranch()

bool gem5::GPUDynInst::isCondBranch ( ) const

Definition at line 330 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isCondBranch().

◆ isDPPInst()

bool gem5::GPUDynInst::isDPPInst ( ) const

Definition at line 360 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isDPPInst().

◆ isEndOfKernel()

bool gem5::GPUDynInst::isEndOfKernel ( ) const

Definition at line 342 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isEndOfKernel().

Referenced by setRequestFlags().

◆ isF16()

bool gem5::GPUDynInst::isF16 ( ) const

Definition at line 701 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isF16().

◆ isF32()

bool gem5::GPUDynInst::isF32 ( ) const

Definition at line 707 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isF32().

◆ isF64()

bool gem5::GPUDynInst::isF64 ( ) const

Definition at line 713 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isF64().

◆ isFlat()

bool gem5::GPUDynInst::isFlat ( ) const

Definition at line 414 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isFlat().

◆ isFMA()

bool gem5::GPUDynInst::isFMA ( ) const

Definition at line 719 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isFMA().

◆ isGloballyCoherent()

bool gem5::GPUDynInst::isGloballyCoherent ( ) const

Definition at line 689 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isGloballyCoherent().

Referenced by setRequestFlags().

◆ isGlobalMem()

bool gem5::GPUDynInst::isGlobalMem ( ) const

Definition at line 635 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isGlobalMem().

◆ isGlobalSeg()

bool gem5::GPUDynInst::isGlobalSeg ( ) const

Definition at line 653 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isGlobalSeg().

◆ isGroupSeg()

bool gem5::GPUDynInst::isGroupSeg ( ) const

Definition at line 659 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isGroupSeg().

◆ isKernArgSeg()

bool gem5::GPUDynInst::isKernArgSeg ( ) const

Definition at line 665 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isKernArgSeg().

◆ isKernelLaunch()

bool gem5::GPUDynInst::isKernelLaunch ( ) const

Definition at line 348 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isKernelLaunch().

Referenced by setRequestFlags().

◆ isLoad()

bool gem5::GPUDynInst::isLoad ( ) const

Definition at line 420 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isLoad().

Referenced by resolveFlatSegment().

◆ isLocalMem()

bool gem5::GPUDynInst::isLocalMem ( ) const

Definition at line 641 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isLocalMem().

◆ isMAC()

bool gem5::GPUDynInst::isMAC ( ) const

Definition at line 725 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isMAC().

◆ isMAD()

bool gem5::GPUDynInst::isMAD ( ) const

Definition at line 731 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isMAD().

◆ isMemRef()

bool gem5::GPUDynInst::isMemRef ( ) const

Definition at line 408 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isMemRef().

◆ isMemSync()

bool gem5::GPUDynInst::isMemSync ( ) const

Definition at line 402 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isMemSync().

Referenced by resolveFlatSegment(), and setRequestFlags().

◆ isNop()

bool gem5::GPUDynInst::isNop ( ) const

Definition at line 336 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isNop().

◆ isOpcode() [1/2]

bool gem5::GPUDynInst::isOpcode ( const std::string & opcodeStr ) const

Definition at line 270 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::opcode().

◆ isOpcode() [2/2]

bool gem5::GPUDynInst::isOpcode	(	const std::string &	opcodeStr,
		const std::string &	extStr
	)		const

Definition at line 262 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::opcode().

◆ isPrivateSeg()

bool gem5::GPUDynInst::isPrivateSeg ( ) const

Definition at line 671 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isPrivateSeg().

◆ isReadOnlySeg()

bool gem5::GPUDynInst::isReadOnlySeg ( ) const

Definition at line 677 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isReadOnlySeg().

◆ isReturn()

bool gem5::GPUDynInst::isReturn ( ) const

Definition at line 366 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isReturn().

◆ isScalar()

bool gem5::GPUDynInst::isScalar ( ) const

Definition at line 456 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isScalar().

◆ isSDWAInst()

bool gem5::GPUDynInst::isSDWAInst ( ) const

Definition at line 354 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isSDWAInst().

◆ isSleep()

bool gem5::GPUDynInst::isSleep ( ) const

Definition at line 390 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isSleep().

◆ isSpecialOp()

bool gem5::GPUDynInst::isSpecialOp ( ) const

Definition at line 378 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isSpecialOp().

◆ isSpillSeg()

bool gem5::GPUDynInst::isSpillSeg ( ) const

Definition at line 683 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isSpillSeg().

◆ isStore()

bool gem5::GPUDynInst::isStore ( ) const

Definition at line 426 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isStore().

Referenced by resolveFlatSegment().

◆ isSystemCoherent()

bool gem5::GPUDynInst::isSystemCoherent ( ) const

Definition at line 695 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isSystemCoherent().

Referenced by setRequestFlags().

◆ isUnconditionalJump()

bool gem5::GPUDynInst::isUnconditionalJump ( ) const

Definition at line 372 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isUnconditionalJump().

◆ isVector()

bool gem5::GPUDynInst::isVector ( ) const

Definition at line 450 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isScalar().

◆ isWaitcnt()

bool gem5::GPUDynInst::isWaitcnt ( ) const

Definition at line 384 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::isWaitcnt().

◆ makeAtomicOpFunctor()

template<typename c0 >

AtomicOpFunctorPtr gem5::GPUDynInst::makeAtomicOpFunctor	(	c0 *	reg0,
		c0 *	reg1
	)

inline

Definition at line 303 of file gpu_dyn_inst.hh.

References gem5::GPUExecContext::cu, fatal, isAtomicAdd(), isAtomicAnd(), isAtomicCAS(), isAtomicDec(), isAtomicExch(), isAtomicInc(), isAtomicMax(), isAtomicMin(), isAtomicOr(), isAtomicSub(), and isAtomicXor().

◆ maxOperandSize()

int gem5::GPUDynInst::maxOperandSize ( )

Definition at line 226 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::maxOperandSize().

◆ maxSrcScalarRegOperandSize()

int gem5::GPUDynInst::maxSrcScalarRegOperandSize ( )

Definition at line 200 of file gpu_dyn_inst.cc.

References maxSrcScalarRegOpSize, and srcScalarRegOperands().

◆ maxSrcVecRegOperandSize()

int gem5::GPUDynInst::maxSrcVecRegOperandSize ( )

Definition at line 162 of file gpu_dyn_inst.cc.

References maxSrcVecRegOpSize, and srcVecRegOperands().

◆ numDstRegOperands()

int gem5::GPUDynInst::numDstRegOperands ( )

Definition at line 144 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numDstRegOperands().

◆ numDstScalarDWords()

int gem5::GPUDynInst::numDstScalarDWords ( )

Definition at line 220 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numDstScalarDWords().

◆ numDstScalarRegOperands()

int gem5::GPUDynInst::numDstScalarRegOperands ( ) const

Definition at line 194 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numDstScalarOperands().

◆ numDstVecDWords()

int gem5::GPUDynInst::numDstVecDWords ( )

Definition at line 182 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numDstVecDWords().

◆ numDstVecRegOperands()

int gem5::GPUDynInst::numDstVecRegOperands ( ) const

Definition at line 156 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numDstVecOperands().

◆ numSrcRegOperands()

int gem5::GPUDynInst::numSrcRegOperands ( )

Definition at line 138 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numSrcRegOperands().

◆ numSrcScalarDWords()

int gem5::GPUDynInst::numSrcScalarDWords ( )

Definition at line 214 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numSrcScalarDWords().

◆ numSrcScalarRegOperands()

int gem5::GPUDynInst::numSrcScalarRegOperands ( ) const

Definition at line 188 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numSrcScalarOperands().

◆ numSrcVecDWords()

int gem5::GPUDynInst::numSrcVecDWords ( )

Definition at line 176 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numSrcVecDWords().

◆ numSrcVecRegOperands()

int gem5::GPUDynInst::numSrcVecRegOperands ( ) const

Definition at line 150 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::numSrcVecOperands().

◆ printStatusVector()

std::string gem5::GPUDynInst::printStatusVector ( ) const

inline

Definition at line 435 of file gpu_dyn_inst.hh.

References gem5::ArmISA::NumVecElemPerVecReg, statusVector, and sc_dt::to_string().

◆ profileLineAddressTime()

void gem5::GPUDynInst::profileLineAddressTime	(	Addr	addr,
		Tick	currentTime,
		int	hopId
	)

Definition at line 980 of file gpu_dyn_inst.cc.

References addr, and lineAddressTime.

◆ profileRoundTripTime()

void gem5::GPUDynInst::profileRoundTripTime	(	Tick	currentTime,
		int	hopId
	)

Definition at line 970 of file gpu_dyn_inst.cc.

References roundTripTime.

◆ readsExec()

bool gem5::GPUDynInst::readsExec ( ) const

Definition at line 506 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::readsEXEC().

◆ readsExecMask()

bool gem5::GPUDynInst::readsExecMask ( ) const

Definition at line 534 of file gpu_dyn_inst.cc.

References _staticInst, gem5::GPUStaticInst::readsEXEC(), and gem5::GPUStaticInst::srcOperands().

◆ readsFlatScratch()

bool gem5::GPUDynInst::readsFlatScratch ( ) const

Definition at line 554 of file gpu_dyn_inst.cc.

References srcScalarRegOperands().

◆ readsMode()

bool gem5::GPUDynInst::readsMode ( ) const

Definition at line 494 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::readsMode().

◆ readsSCC()

bool gem5::GPUDynInst::readsSCC ( ) const

Definition at line 462 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::readsSCC().

◆ readsVCC()

bool gem5::GPUDynInst::readsVCC ( ) const

Definition at line 474 of file gpu_dyn_inst.cc.

References _staticInst, gem5::GPUStaticInst::readsVCC(), and gem5::GPUStaticInst::srcOperands().

◆ resetEntireStatusVector()

void gem5::GPUDynInst::resetEntireStatusVector ( )

inline

Definition at line 362 of file gpu_dyn_inst.hh.

References gem5::ArmISA::NumVecElemPerVecReg, resetStatusVector(), and statusVector.

◆ resetStatusVector()

void gem5::GPUDynInst::resetStatusVector ( int lane )

inline

Definition at line 372 of file gpu_dyn_inst.hh.

References setStatusVector().

Referenced by resetEntireStatusVector().

◆ resolveFlatSegment()

void gem5::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, gem5::Wavefront::computeUnit, gem5::Wavefront::decLGKMInstsIssued(), gem5::Wavefront::decVMemInstsIssued(), doApertureCheck(), gem5::Wavefront::execUnitId, executedAs(), gem5::Wavefront::flatGmUnitId, gem5::Wavefront::flatLmUnitId, gem5::Shader::getHiddenPrivateBase(), gem5::Shader::getScratchBase(), isAtomic(), isLoad(), isMemSync(), isStore(), gem5::RegisterManager::mapSgpr(), gem5::mask(), gem5::Wavefront::maxSgprs, gem5::ArmISA::offset, panic, gem5::Wavefront::rdGmReqsInPipe, gem5::Wavefront::rdLmReqsInPipe, gem5::ComputeUnit::registerManager, gem5::ComputeUnit::shader, simdId, gem5::ComputeUnit::srf, gem5::GPUExecContext::wavefront(), gem5::ComputeUnit::wfSize(), gem5::Wavefront::wrGmReqsInPipe, and gem5::Wavefront::wrLmReqsInPipe.

◆ seqNum()

InstSeqNum gem5::GPUDynInst::seqNum ( ) const

Definition at line 282 of file gpu_dyn_inst.cc.

References _seqNum.

◆ setAccessTime()

void gem5::GPUDynInst::setAccessTime ( Tick currentTime )

inline

Definition at line 466 of file gpu_dyn_inst.hh.

References accessTime.

◆ setRequestFlags()

void gem5::GPUDynInst::setRequestFlags ( RequestPtr req ) const

inline

Definition at line 333 of file gpu_dyn_inst.hh.

References gem5::Request::ATOMIC_NO_RETURN_OP, gem5::Request::ATOMIC_RETURN_OP, gem5::Request::GLC_BIT, gem5::Request::INV_L1, isAtomicNoRet(), isAtomicRet(), isEndOfKernel(), isGloballyCoherent(), isKernelLaunch(), isMemSync(), isSystemCoherent(), and gem5::Request::SLC_BIT.

◆ setStatusVector()

void gem5::GPUDynInst::setStatusVector	(	int	lane,
		int	newVal
	)

inline

Definition at line 379 of file gpu_dyn_inst.hh.

References statusVector.

Referenced by resetStatusVector().

◆ srcLiteral()

TheGpuISA::ScalarRegU32 gem5::GPUDynInst::srcLiteral ( ) const

Definition at line 923 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::srcLiteral().

◆ srcScalarRegOperands()

const std::vector< OperandInfo > & gem5::GPUDynInst::srcScalarRegOperands ( ) const

Definition at line 126 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::srcScalarRegOperands().

Referenced by hasSourceSgpr(), maxSrcScalarRegOperandSize(), and readsFlatScratch().

◆ srcVecRegOperands()

const std::vector< OperandInfo > & gem5::GPUDynInst::srcVecRegOperands ( ) const

Definition at line 114 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::srcVecRegOperands().

Referenced by hasSourceVgpr(), and maxSrcVecRegOperandSize().

◆ staticInstruction()

GPUStaticInst* gem5::GPUDynInst::staticInstruction ( )

inline

Definition at line 213 of file gpu_dyn_inst.hh.

References _staticInst.

Referenced by doApertureCheck().

◆ updateStats()

void gem5::GPUDynInst::updateStats ( )

Definition at line 929 of file gpu_dyn_inst.cc.

References _staticInst, gem5::GPUExecContext::cu, gem5::ComputeUnit::ComputeUnitStats::dynamicFlatMemInstrCnt, gem5::ComputeUnit::ComputeUnitStats::dynamicGMemInstrCnt, gem5::ComputeUnit::ComputeUnitStats::dynamicLMemInstrCnt, gem5::GPUStaticInst::isFlat(), gem5::GPUStaticInst::isLocalMem(), gem5::ComputeUnit::pageAccesses, gem5::ComputeUnit::ComputeUnitStats::pageDivergenceDist, gem5::ComputeUnit::pagesTouched, gem5::statistics::DistBase< Derived, Stor >::sample(), and gem5::ComputeUnit::stats.

◆ writesExec()

bool gem5::GPUDynInst::writesExec ( ) const

Definition at line 512 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::writesEXEC().

◆ writesExecMask()

bool gem5::GPUDynInst::writesExecMask ( ) const

Definition at line 524 of file gpu_dyn_inst.cc.

References _staticInst, gem5::GPUStaticInst::dstOperands(), and gem5::GPUStaticInst::writesEXEC().

◆ writesFlatScratch()

bool gem5::GPUDynInst::writesFlatScratch ( ) const

Definition at line 544 of file gpu_dyn_inst.cc.

References dstScalarRegOperands().

◆ writesMode()

bool gem5::GPUDynInst::writesMode ( ) const

Definition at line 500 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::writesMode().

◆ writesSCC()

bool gem5::GPUDynInst::writesSCC ( ) const

Definition at line 468 of file gpu_dyn_inst.cc.

References _staticInst, and gem5::GPUStaticInst::writesSCC().

◆ writesVCC()

bool gem5::GPUDynInst::writesVCC ( ) const

Definition at line 484 of file gpu_dyn_inst.cc.

References _staticInst, gem5::GPUStaticInst::dstOperands(), and gem5::GPUStaticInst::writesVCC().

Member Data Documentation

◆ _seqNum

const InstSeqNum gem5::GPUDynInst::_seqNum

private

Definition at line 479 of file gpu_dyn_inst.hh.

Referenced by seqNum().

◆ _staticInst

GPUStaticInst* gem5::GPUDynInst::_staticInst

private

Definition at line 478 of file gpu_dyn_inst.hh.

Referenced by completeAcc(), disassemble(), dstScalarRegOperands(), dstVecRegOperands(), execute(), executedAs(), getNumOperands(), GPUDynInst(), ignoreExec(), initiateAcc(), isALU(), isArgLoad(), isArgSeg(), isAtomic(), isAtomicAdd(), isAtomicAnd(), isAtomicCAS(), isAtomicDec(), isAtomicExch(), isAtomicInc(), isAtomicMax(), isAtomicMin(), isAtomicNoRet(), isAtomicOr(), isAtomicRet(), isAtomicSub(), isAtomicXor(), isBarrier(), isBranch(), isCondBranch(), isDPPInst(), isEndOfKernel(), isF16(), isF32(), isF64(), isFlat(), isFMA(), isGloballyCoherent(), isGlobalMem(), isGlobalSeg(), isGroupSeg(), isKernArgSeg(), isKernelLaunch(), isLoad(), isLocalMem(), isMAC(), isMAD(), isMemRef(), isMemSync(), isNop(), isOpcode(), isPrivateSeg(), isReadOnlySeg(), isReturn(), isScalar(), isSDWAInst(), isSleep(), isSpecialOp(), isSpillSeg(), isStore(), isSystemCoherent(), isUnconditionalJump(), isVector(), isWaitcnt(), maxOperandSize(), numDstRegOperands(), numDstScalarDWords(), numDstScalarRegOperands(), numDstVecDWords(), numDstVecRegOperands(), numSrcRegOperands(), numSrcScalarDWords(), numSrcScalarRegOperands(), numSrcVecDWords(), numSrcVecRegOperands(), readsExec(), readsExecMask(), readsMode(), readsSCC(), readsVCC(), srcLiteral(), srcScalarRegOperands(), srcVecRegOperands(), staticInstruction(), updateStats(), writesExec(), writesExecMask(), writesMode(), writesSCC(), writesVCC(), and ~GPUDynInst().

◆ a_data

uint8_t* gem5::GPUDynInst::a_data

Definition at line 178 of file gpu_dyn_inst.hh.

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

◆ accessTime

Tick gem5::GPUDynInst::accessTime = -1

private

Definition at line 484 of file gpu_dyn_inst.hh.

Referenced by getAccessTime(), and setAccessTime().

◆ addr

std::vector<Addr> gem5::GPUDynInst::addr

Definition at line 170 of file gpu_dyn_inst.hh.

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

◆ cu_id

int gem5::GPUDynInst::cu_id

Definition at line 191 of file gpu_dyn_inst.hh.

Referenced by allLanesZero(), and GPUDynInst().

◆ d_data

uint8_t* gem5::GPUDynInst::d_data

Definition at line 174 of file gpu_dyn_inst.hh.

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

◆ exec_mask

VectorMask gem5::GPUDynInst::exec_mask

Definition at line 182 of file gpu_dyn_inst.hh.

Referenced by completeAcc(), and initiateAcc().

◆ execUnitId

int gem5::GPUDynInst::execUnitId

Definition at line 197 of file gpu_dyn_inst.hh.

◆ isSaveRestore

bool gem5::GPUDynInst::isSaveRestore

Definition at line 476 of file gpu_dyn_inst.hh.

◆ kern_id

int gem5::GPUDynInst::kern_id

Definition at line 189 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ latency

WaitClass gem5::GPUDynInst::latency

Definition at line 201 of file gpu_dyn_inst.hh.

◆ lineAddressTime

std::map<Addr, std::vector<Tick> > gem5::GPUDynInst::lineAddressTime

private

Definition at line 492 of file gpu_dyn_inst.hh.

Referenced by getLineAddressTime(), and profileLineAddressTime().

◆ maxSrcScalarRegOpSize

int gem5::GPUDynInst::maxSrcScalarRegOpSize

private

Definition at line 481 of file gpu_dyn_inst.hh.

Referenced by maxSrcScalarRegOperandSize().

◆ maxSrcVecRegOpSize

int gem5::GPUDynInst::maxSrcVecRegOpSize

private

Definition at line 480 of file gpu_dyn_inst.hh.

Referenced by maxSrcVecRegOperandSize().

◆ memStatusVector

StatusVector gem5::GPUDynInst::memStatusVector

Definition at line 452 of file gpu_dyn_inst.hh.

◆ numScalarReqs

int gem5::GPUDynInst::numScalarReqs

Definition at line 462 of file gpu_dyn_inst.hh.

◆ pAddr

Addr gem5::GPUDynInst::pAddr

Definition at line 171 of file gpu_dyn_inst.hh.

◆ roundTripTime

std::vector<Tick> gem5::GPUDynInst::roundTripTime

private

Definition at line 488 of file gpu_dyn_inst.hh.

Referenced by getRoundTripTime(), and profileRoundTripTime().

◆ scalar_data

uint8_t* gem5::GPUDynInst::scalar_data

Definition at line 176 of file gpu_dyn_inst.hh.

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

◆ scalarAddr

Addr gem5::GPUDynInst::scalarAddr

Definition at line 168 of file gpu_dyn_inst.hh.

◆ simdId

int gem5::GPUDynInst::simdId

Definition at line 185 of file gpu_dyn_inst.hh.

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

◆ statusVector

std::vector<int> gem5::GPUDynInst::statusVector

Definition at line 456 of file gpu_dyn_inst.hh.

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

◆ time

Tick gem5::GPUDynInst::time

Definition at line 199 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ tlbHitLevel

std::vector<int> gem5::GPUDynInst::tlbHitLevel

Definition at line 458 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ wfDynId

int gem5::GPUDynInst::wfDynId

Definition at line 187 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ wfSlotId

int gem5::GPUDynInst::wfSlotId

Definition at line 195 of file gpu_dyn_inst.hh.

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

◆ wg_id

int gem5::GPUDynInst::wg_id

Definition at line 193 of file gpu_dyn_inst.hh.

Referenced by GPUDynInst().

◆ x_data

uint8_t* gem5::GPUDynInst::x_data

Definition at line 180 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()

◆ dstScalarRegOperands()

◆ dstVecRegOperands()

◆ execute()

◆ executedAs()

◆ getAccessTime()

◆ getLaneStatus()

◆ getLineAddressTime()

◆ getNumOperands()

◆ getRoundTripTime()

◆ 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()

◆ 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()

◆ isScalar()

◆ isSDWAInst()

◆ isSleep()