#include <gpu_dyn_inst.hh>
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 |
| Addr | pc () |
| void | pc (Addr _pc) |
| 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 | |
| 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 | isFlatGlobal () const |
| bool | isFlatScratch () 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 | needsToken () 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 | isAtomicPkAddBF16 () 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 | isI8 () const |
| bool | isF16 () const |
| bool | isF32 () const |
| bool | isF64 () const |
| bool | isFMA () const |
| bool | isMAC () const |
| bool | isMAD () const |
| bool | isMFMA () 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 |
| bool | isSystemReq () |
| void | setSystemReq () |
| 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 |
| bool | systemReq = false |
| 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 143 of file gpu_dyn_inst.hh.
Member Typedef Documentation
◆ StatusVector
| typedef std::unordered_map<Addr, std::vector<int> > gem5::GPUDynInst::StatusVector |
Definition at line 488 of file gpu_dyn_inst.hh.
Constructor & Destructor Documentation
◆ GPUDynInst()
| gem5::GPUDynInst::GPUDynInst | ( | ComputeUnit * | _cu, |
| Wavefront * | _wf, | ||
| GPUStaticInst * | static_inst, | ||
| uint64_t | instSeqNum ) |
Definition at line 44 of file gpu_dyn_inst.cc.
References _seqNum, _staticInst, a_data, addr, gem5::GPUExecContext::computeUnit(), gem5::ComputeUnit::cu_id, cu_id, d_data, disassemble(), DPRINTF, getNumOperands(), gem5::GPUExecContext::GPUExecContext(), gem5::ArmISA::i, isSaveRestore, kern_id, gem5::Wavefront::kernId, maxSrcScalarRegOpSize, maxSrcVecRegOpSize, numScalarReqs, scalar_data, scalarAddr, simdId, gem5::Wavefront::simdId, statusVector, time, tlbHitLevel, gem5::GPUExecContext::wavefront(), wfDynId, gem5::Wavefront::wfDynId, gem5::ComputeUnit::wfSize(), wfSlotId, gem5::Wavefront::wfSlotId, wg_id, gem5::Wavefront::wgId, and x_data.
◆ ~GPUDynInst()
| gem5::GPUDynInst::~GPUDynInst | ( | ) |
Definition at line 96 of file gpu_dyn_inst.cc.
References _staticInst, a_data, d_data, scalar_data, and x_data.
Member Function Documentation
◆ allLanesZero()
|
inline |
Definition at line 446 of file gpu_dyn_inst.hh.
References addr, cu_id, DPRINTF, simdId, statusVector, and wfSlotId.
◆ completeAcc()
| void gem5::GPUDynInst::completeAcc | ( | GPUDynInstPtr | gpuDynInst | ) |
Definition at line 314 of file gpu_dyn_inst.cc.
References _staticInst, gem5::GPUExecContext::cu, DPRINTF, exec_mask, simdId, and wfSlotId.
◆ decrementStatusVector()
|
inline |
Definition at line 427 of file gpu_dyn_inst.hh.
References statusVector.
◆ disassemble()
| const std::string & gem5::GPUDynInst::disassemble | ( | ) | const |
◆ doApertureCheck()
| void gem5::GPUDynInst::doApertureCheck | ( | const VectorMask & | mask | ) |
Definition at line 783 of file gpu_dyn_inst.cc.
References addr, gem5::bits(), gem5::GPUExecContext::computeUnit(), gem5::GPUStaticInst::executed_as, executedAs(), fatal, gem5::ArmISA::mask, staticInstruction(), and gem5::ComputeUnit::wfSize().
Referenced by resolveFlatSegment().
◆ dstScalarRegOperands()
| const std::vector< OperandInfo > & gem5::GPUDynInst::dstScalarRegOperands | ( | ) | const |
Definition at line 130 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by hasDestinationSgpr(), and writesFlatScratch().
◆ dstVecRegOperands()
| const std::vector< OperandInfo > & gem5::GPUDynInst::dstVecRegOperands | ( | ) | const |
Definition at line 118 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by hasDestinationVgpr().
◆ execute()
| void gem5::GPUDynInst::execute | ( | GPUDynInstPtr | gpuDynInst | ) |
Definition at line 106 of file gpu_dyn_inst.cc.
References _staticInst.
◆ executedAs()
| enums::StorageClassType gem5::GPUDynInst::executedAs | ( | ) |
Definition at line 298 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by doApertureCheck(), and resolveFlatSegment().
◆ getAccessTime()
|
inline |
Definition at line 501 of file gpu_dyn_inst.hh.
References accessTime.
◆ getLaneStatus()
|
inline |
Definition at line 438 of file gpu_dyn_inst.hh.
References statusVector.
◆ getLineAddressTime()
|
inline |
Definition at line 509 of file gpu_dyn_inst.hh.
References lineAddressTime.
◆ getNumOperands()
| int gem5::GPUDynInst::getNumOperands | ( | ) | const |
◆ getRoundTripTime()
|
inline |
Definition at line 506 of file gpu_dyn_inst.hh.
References roundTripTime.
◆ hasDestinationSgpr()
| bool gem5::GPUDynInst::hasDestinationSgpr | ( | ) | const |
Definition at line 254 of file gpu_dyn_inst.cc.
References dstScalarRegOperands().
◆ hasDestinationVgpr()
| bool gem5::GPUDynInst::hasDestinationVgpr | ( | ) | const |
Definition at line 242 of file gpu_dyn_inst.cc.
References dstVecRegOperands().
◆ hasSourceSgpr()
| bool gem5::GPUDynInst::hasSourceSgpr | ( | ) | const |
Definition at line 248 of file gpu_dyn_inst.cc.
References srcScalarRegOperands().
◆ hasSourceVgpr()
| bool gem5::GPUDynInst::hasSourceVgpr | ( | ) | const |
Definition at line 236 of file gpu_dyn_inst.cc.
References srcVecRegOperands().
◆ ignoreExec()
| bool gem5::GPUDynInst::ignoreExec | ( | ) | const |
Definition at line 540 of file gpu_dyn_inst.cc.
References _staticInst.
◆ initiateAcc()
| void gem5::GPUDynInst::initiateAcc | ( | GPUDynInstPtr | gpuDynInst | ) |
Definition at line 305 of file gpu_dyn_inst.cc.
References _staticInst, gem5::GPUExecContext::cu, DPRINTF, exec_mask, simdId, and wfSlotId.
◆ isALU()
| bool gem5::GPUDynInst::isALU | ( | ) | const |
accessor methods for the attributes of the underlying GPU static instruction
Definition at line 328 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isArgLoad()
| bool gem5::GPUDynInst::isArgLoad | ( | ) | const |
Definition at line 663 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isArgSeg()
| bool gem5::GPUDynInst::isArgSeg | ( | ) | const |
Definition at line 681 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isAtomic()
| bool gem5::GPUDynInst::isAtomic | ( | ) | const |
Definition at line 454 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by resolveFlatSegment().
◆ isAtomicAdd()
| bool gem5::GPUDynInst::isAtomicAdd | ( | ) | const |
Definition at line 621 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicAnd()
| bool gem5::GPUDynInst::isAtomicAnd | ( | ) | const |
Definition at line 592 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicCAS()
| bool gem5::GPUDynInst::isAtomicCAS | ( | ) | const |
Definition at line 610 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicDec()
| bool gem5::GPUDynInst::isAtomicDec | ( | ) | const |
Definition at line 639 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicExch()
| bool gem5::GPUDynInst::isAtomicExch | ( | ) | const |
Definition at line 615 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicInc()
| bool gem5::GPUDynInst::isAtomicInc | ( | ) | const |
Definition at line 633 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicMax()
| bool gem5::GPUDynInst::isAtomicMax | ( | ) | const |
Definition at line 645 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicMin()
| bool gem5::GPUDynInst::isAtomicMin | ( | ) | const |
Definition at line 651 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicNoRet()
| bool gem5::GPUDynInst::isAtomicNoRet | ( | ) | const |
Definition at line 460 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by setRequestFlags().
◆ isAtomicOr()
| bool gem5::GPUDynInst::isAtomicOr | ( | ) | const |
Definition at line 598 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicPkAddBF16()
| bool gem5::GPUDynInst::isAtomicPkAddBF16 | ( | ) | const |
Definition at line 657 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicRet()
| bool gem5::GPUDynInst::isAtomicRet | ( | ) | const |
Definition at line 466 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by setRequestFlags().
◆ isAtomicSub()
| bool gem5::GPUDynInst::isAtomicSub | ( | ) | const |
Definition at line 627 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isAtomicXor()
| bool gem5::GPUDynInst::isAtomicXor | ( | ) | const |
Definition at line 604 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by makeAtomicOpFunctor().
◆ isBarrier()
| bool gem5::GPUDynInst::isBarrier | ( | ) | const |
Definition at line 406 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isBranch()
| bool gem5::GPUDynInst::isBranch | ( | ) | const |
Definition at line 334 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isCondBranch()
| bool gem5::GPUDynInst::isCondBranch | ( | ) | const |
Definition at line 340 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isDPPInst()
| bool gem5::GPUDynInst::isDPPInst | ( | ) | const |
Definition at line 370 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isEndOfKernel()
| bool gem5::GPUDynInst::isEndOfKernel | ( | ) | const |
Definition at line 352 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by setRequestFlags().
◆ isF16()
| bool gem5::GPUDynInst::isF16 | ( | ) | const |
Definition at line 741 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isF32()
| bool gem5::GPUDynInst::isF32 | ( | ) | const |
Definition at line 747 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isF64()
| bool gem5::GPUDynInst::isF64 | ( | ) | const |
Definition at line 753 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isFlat()
| bool gem5::GPUDynInst::isFlat | ( | ) | const |
Definition at line 424 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by needsToken(), and resolveFlatSegment().
◆ isFlatGlobal()
| bool gem5::GPUDynInst::isFlatGlobal | ( | ) | const |
◆ isFlatScratch()
| bool gem5::GPUDynInst::isFlatScratch | ( | ) | const |
◆ isFMA()
| bool gem5::GPUDynInst::isFMA | ( | ) | const |
Definition at line 759 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isGloballyCoherent()
| bool gem5::GPUDynInst::isGloballyCoherent | ( | ) | const |
Definition at line 723 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by setRequestFlags().
◆ isGlobalMem()
| bool gem5::GPUDynInst::isGlobalMem | ( | ) | const |
◆ isGlobalSeg()
| bool gem5::GPUDynInst::isGlobalSeg | ( | ) | const |
Definition at line 687 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isGroupSeg()
| bool gem5::GPUDynInst::isGroupSeg | ( | ) | const |
Definition at line 693 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isI8()
| bool gem5::GPUDynInst::isI8 | ( | ) | const |
Definition at line 735 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isKernArgSeg()
| bool gem5::GPUDynInst::isKernArgSeg | ( | ) | const |
Definition at line 699 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isKernelLaunch()
| bool gem5::GPUDynInst::isKernelLaunch | ( | ) | const |
Definition at line 358 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by setRequestFlags().
◆ isLoad()
| bool gem5::GPUDynInst::isLoad | ( | ) | const |
Definition at line 442 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by resolveFlatSegment().
◆ isLocalMem()
| bool gem5::GPUDynInst::isLocalMem | ( | ) | const |
Definition at line 675 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isMAC()
| bool gem5::GPUDynInst::isMAC | ( | ) | const |
Definition at line 765 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isMAD()
| bool gem5::GPUDynInst::isMAD | ( | ) | const |
Definition at line 771 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isMemRef()
| bool gem5::GPUDynInst::isMemRef | ( | ) | const |
Definition at line 418 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isMemSync()
| bool gem5::GPUDynInst::isMemSync | ( | ) | const |
Definition at line 412 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by resolveFlatSegment(), and setRequestFlags().
◆ isMFMA()
| bool gem5::GPUDynInst::isMFMA | ( | ) | const |
Definition at line 777 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isNop()
| bool gem5::GPUDynInst::isNop | ( | ) | const |
Definition at line 346 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isOpcode() [1/2]
| bool gem5::GPUDynInst::isOpcode | ( | const std::string & | opcodeStr | ) | const |
Definition at line 268 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isOpcode() [2/2]
| bool gem5::GPUDynInst::isOpcode | ( | const std::string & | opcodeStr, |
| const std::string & | extStr ) const |
Definition at line 260 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isPrivateSeg()
| bool gem5::GPUDynInst::isPrivateSeg | ( | ) | const |
Definition at line 705 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isReadOnlySeg()
| bool gem5::GPUDynInst::isReadOnlySeg | ( | ) | const |
Definition at line 711 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isReturn()
| bool gem5::GPUDynInst::isReturn | ( | ) | const |
Definition at line 376 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isScalar()
| bool gem5::GPUDynInst::isScalar | ( | ) | const |
Definition at line 478 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isSDWAInst()
| bool gem5::GPUDynInst::isSDWAInst | ( | ) | const |
Definition at line 364 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isSleep()
| bool gem5::GPUDynInst::isSleep | ( | ) | const |
Definition at line 400 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isSpecialOp()
| bool gem5::GPUDynInst::isSpecialOp | ( | ) | const |
Definition at line 388 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isSpillSeg()
| bool gem5::GPUDynInst::isSpillSeg | ( | ) | const |
Definition at line 717 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isStore()
| bool gem5::GPUDynInst::isStore | ( | ) | const |
Definition at line 448 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by resolveFlatSegment().
◆ isSystemCoherent()
| bool gem5::GPUDynInst::isSystemCoherent | ( | ) | const |
Definition at line 729 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by setRequestFlags().
◆ isSystemReq()
|
inline |
Definition at line 515 of file gpu_dyn_inst.hh.
References systemReq.
◆ isUnconditionalJump()
| bool gem5::GPUDynInst::isUnconditionalJump | ( | ) | const |
Definition at line 382 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isVector()
| bool gem5::GPUDynInst::isVector | ( | ) | const |
Definition at line 472 of file gpu_dyn_inst.cc.
References _staticInst.
◆ isWaitcnt()
| bool gem5::GPUDynInst::isWaitcnt | ( | ) | const |
Definition at line 394 of file gpu_dyn_inst.cc.
References _staticInst.
◆ makeAtomicOpFunctor()
|
inline |
Definition at line 337 of file gpu_dyn_inst.hh.
References gem5::GPUExecContext::cu, fatal, isAtomicAdd(), isAtomicAnd(), isAtomicCAS(), isAtomicDec(), isAtomicExch(), isAtomicInc(), isAtomicMax(), isAtomicMin(), isAtomicOr(), isAtomicPkAddBF16(), isAtomicSub(), and isAtomicXor().
◆ maxOperandSize()
| int gem5::GPUDynInst::maxOperandSize | ( | ) |
Definition at line 224 of file gpu_dyn_inst.cc.
References _staticInst.
◆ maxSrcScalarRegOperandSize()
| int gem5::GPUDynInst::maxSrcScalarRegOperandSize | ( | ) |
Definition at line 198 of file gpu_dyn_inst.cc.
References maxSrcScalarRegOpSize, and srcScalarRegOperands().
◆ maxSrcVecRegOperandSize()
| int gem5::GPUDynInst::maxSrcVecRegOperandSize | ( | ) |
Definition at line 160 of file gpu_dyn_inst.cc.
References maxSrcVecRegOpSize, and srcVecRegOperands().
◆ needsToken()
| bool gem5::GPUDynInst::needsToken | ( | ) | const |
Definition at line 586 of file gpu_dyn_inst.cc.
References isFlat(), isFlatGlobal(), isFlatScratch(), and isGlobalMem().
◆ numDstRegOperands()
| int gem5::GPUDynInst::numDstRegOperands | ( | ) |
Definition at line 142 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numDstScalarDWords()
| int gem5::GPUDynInst::numDstScalarDWords | ( | ) |
Definition at line 218 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numDstScalarRegOperands()
| int gem5::GPUDynInst::numDstScalarRegOperands | ( | ) | const |
Definition at line 192 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numDstVecDWords()
| int gem5::GPUDynInst::numDstVecDWords | ( | ) |
Definition at line 180 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numDstVecRegOperands()
| int gem5::GPUDynInst::numDstVecRegOperands | ( | ) | const |
Definition at line 154 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numSrcRegOperands()
| int gem5::GPUDynInst::numSrcRegOperands | ( | ) |
Definition at line 136 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numSrcScalarDWords()
| int gem5::GPUDynInst::numSrcScalarDWords | ( | ) |
Definition at line 212 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numSrcScalarRegOperands()
| int gem5::GPUDynInst::numSrcScalarRegOperands | ( | ) | const |
Definition at line 186 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numSrcVecDWords()
| int gem5::GPUDynInst::numSrcVecDWords | ( | ) |
Definition at line 174 of file gpu_dyn_inst.cc.
References _staticInst.
◆ numSrcVecRegOperands()
| int gem5::GPUDynInst::numSrcVecRegOperands | ( | ) | const |
Definition at line 148 of file gpu_dyn_inst.cc.
References _staticInst.
◆ pc() [1/2]
| Addr gem5::GPUDynInst::pc | ( | ) |
Definition at line 286 of file gpu_dyn_inst.cc.
References gem5::Wavefront::pc(), and gem5::GPUExecContext::wavefront().
◆ pc() [2/2]
| void gem5::GPUDynInst::pc | ( | Addr | _pc | ) |
Definition at line 292 of file gpu_dyn_inst.cc.
References gem5::Wavefront::pc(), and gem5::GPUExecContext::wavefront().
◆ printStatusVector()
|
inline |
Definition at line 472 of file gpu_dyn_inst.hh.
References statusVector.
◆ profileLineAddressTime()
Definition at line 1054 of file gpu_dyn_inst.cc.
References addr, and lineAddressTime.
◆ profileRoundTripTime()
| void gem5::GPUDynInst::profileRoundTripTime | ( | Tick | currentTime, |
| int | hopId ) |
Definition at line 1044 of file gpu_dyn_inst.cc.
References roundTripTime.
◆ readsExec()
| bool gem5::GPUDynInst::readsExec | ( | ) | const |
Definition at line 528 of file gpu_dyn_inst.cc.
References _staticInst.
◆ readsExecMask()
| bool gem5::GPUDynInst::readsExecMask | ( | ) | const |
Definition at line 556 of file gpu_dyn_inst.cc.
References _staticInst.
◆ readsFlatScratch()
| bool gem5::GPUDynInst::readsFlatScratch | ( | ) | const |
Definition at line 576 of file gpu_dyn_inst.cc.
References srcScalarRegOperands().
◆ readsMode()
| bool gem5::GPUDynInst::readsMode | ( | ) | const |
Definition at line 516 of file gpu_dyn_inst.cc.
References _staticInst.
◆ readsSCC()
| bool gem5::GPUDynInst::readsSCC | ( | ) | const |
Definition at line 484 of file gpu_dyn_inst.cc.
References _staticInst.
◆ readsVCC()
| bool gem5::GPUDynInst::readsVCC | ( | ) | const |
Definition at line 496 of file gpu_dyn_inst.cc.
References _staticInst.
◆ resetEntireStatusVector()
|
inline |
Definition at line 398 of file gpu_dyn_inst.hh.
References resetStatusVector(), and statusVector.
◆ resetStatusVector()
|
inline |
Definition at line 408 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 855 of file gpu_dyn_inst.cc.
References addr, gem5::Wavefront::archFlatScratchAddr, gem5::ApertureRegister::base, gem5::GPUExecContext::computeUnit(), gem5::Wavefront::computeUnit, gem5::GPUExecContext::cu, gem5::Wavefront::decLGKMInstsIssued(), gem5::Wavefront::decVMemInstsIssued(), doApertureCheck(), gem5::Wavefront::execUnitId, executedAs(), gem5::Wavefront::flatGmUnitId, gem5::Wavefront::flatLmUnitId, isAtomic(), isFlat(), isLoad(), isMemSync(), isStore(), gem5::Shader::ldsApe(), gem5::Wavefront::lgkmIssued, gem5::ArmISA::mask, gem5::Wavefront::maxSgprs, gem5::ArmISA::offset, panic, gem5::Wavefront::rdGmReqsInPipe, gem5::Wavefront::rdLmReqsInPipe, seqNum(), gem5::ComputeUnit::shader, simdId, gem5::Wavefront::vmemIssued, gem5::GPUExecContext::wavefront(), gem5::ComputeUnit::wfSize(), gem5::Wavefront::wrGmReqsInPipe, and gem5::Wavefront::wrLmReqsInPipe.
◆ seqNum()
| InstSeqNum gem5::GPUDynInst::seqNum | ( | ) | const |
Definition at line 280 of file gpu_dyn_inst.cc.
References _seqNum.
Referenced by resolveFlatSegment().
◆ setAccessTime()
|
inline |
Definition at line 503 of file gpu_dyn_inst.hh.
References accessTime.
◆ setRequestFlags()
|
inline |
Definition at line 369 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()
|
inline |
Definition at line 415 of file gpu_dyn_inst.hh.
References statusVector.
Referenced by resetStatusVector().
◆ setSystemReq()
|
inline |
Definition at line 516 of file gpu_dyn_inst.hh.
References systemReq.
◆ srcLiteral()
| TheGpuISA::ScalarRegU32 gem5::GPUDynInst::srcLiteral | ( | ) | const |
Definition at line 997 of file gpu_dyn_inst.cc.
References _staticInst.
◆ srcScalarRegOperands()
| const std::vector< OperandInfo > & gem5::GPUDynInst::srcScalarRegOperands | ( | ) | const |
Definition at line 124 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by hasSourceSgpr(), maxSrcScalarRegOperandSize(), and readsFlatScratch().
◆ srcVecRegOperands()
| const std::vector< OperandInfo > & gem5::GPUDynInst::srcVecRegOperands | ( | ) | const |
Definition at line 112 of file gpu_dyn_inst.cc.
References _staticInst.
Referenced by hasSourceVgpr(), and maxSrcVecRegOperandSize().
◆ staticInstruction()
|
inline |
Definition at line 241 of file gpu_dyn_inst.hh.
References _staticInst.
Referenced by doApertureCheck().
◆ updateStats()
| void gem5::GPUDynInst::updateStats | ( | ) |
Definition at line 1003 of file gpu_dyn_inst.cc.
References _staticInst, and gem5::GPUExecContext::cu.
◆ writesExec()
| bool gem5::GPUDynInst::writesExec | ( | ) | const |
Definition at line 534 of file gpu_dyn_inst.cc.
References _staticInst.
◆ writesExecMask()
| bool gem5::GPUDynInst::writesExecMask | ( | ) | const |
Definition at line 546 of file gpu_dyn_inst.cc.
References _staticInst.
◆ writesFlatScratch()
| bool gem5::GPUDynInst::writesFlatScratch | ( | ) | const |
Definition at line 566 of file gpu_dyn_inst.cc.
References dstScalarRegOperands().
◆ writesMode()
| bool gem5::GPUDynInst::writesMode | ( | ) | const |
Definition at line 522 of file gpu_dyn_inst.cc.
References _staticInst.
◆ writesSCC()
| bool gem5::GPUDynInst::writesSCC | ( | ) | const |
Definition at line 490 of file gpu_dyn_inst.cc.
References _staticInst.
◆ writesVCC()
| bool gem5::GPUDynInst::writesVCC | ( | ) | const |
Definition at line 506 of file gpu_dyn_inst.cc.
References _staticInst.
Member Data Documentation
◆ _seqNum
|
private |
Definition at line 520 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst(), and seqNum().
◆ _staticInst
|
private |
Definition at line 519 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(), isAtomicPkAddBF16(), isAtomicRet(), isAtomicSub(), isAtomicXor(), isBarrier(), isBranch(), isCondBranch(), isDPPInst(), isEndOfKernel(), isF16(), isF32(), isF64(), isFlat(), isFlatGlobal(), isFlatScratch(), isFMA(), isGloballyCoherent(), isGlobalMem(), isGlobalSeg(), isGroupSeg(), isI8(), isKernArgSeg(), isKernelLaunch(), isLoad(), isLocalMem(), isMAC(), isMAD(), isMemRef(), isMemSync(), isMFMA(), isNop(), isOpcode(), 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 206 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst(), and ~GPUDynInst().
◆ accessTime
|
private |
Definition at line 526 of file gpu_dyn_inst.hh.
Referenced by getAccessTime(), and setAccessTime().
◆ addr
| std::vector<Addr> gem5::GPUDynInst::addr |
Definition at line 198 of file gpu_dyn_inst.hh.
Referenced by allLanesZero(), doApertureCheck(), GPUDynInst(), profileLineAddressTime(), and resolveFlatSegment().
◆ cu_id
| int gem5::GPUDynInst::cu_id |
Definition at line 219 of file gpu_dyn_inst.hh.
Referenced by allLanesZero(), and GPUDynInst().
◆ d_data
| uint8_t* gem5::GPUDynInst::d_data |
Definition at line 202 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst(), and ~GPUDynInst().
◆ exec_mask
| VectorMask gem5::GPUDynInst::exec_mask |
Definition at line 210 of file gpu_dyn_inst.hh.
Referenced by completeAcc(), and initiateAcc().
◆ execUnitId
| int gem5::GPUDynInst::execUnitId |
Definition at line 225 of file gpu_dyn_inst.hh.
◆ isSaveRestore
| bool gem5::GPUDynInst::isSaveRestore |
Definition at line 513 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ kern_id
| int gem5::GPUDynInst::kern_id |
Definition at line 217 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ latency
| WaitClass gem5::GPUDynInst::latency |
Definition at line 229 of file gpu_dyn_inst.hh.
◆ lineAddressTime
|
private |
Definition at line 534 of file gpu_dyn_inst.hh.
Referenced by getLineAddressTime(), and profileLineAddressTime().
◆ maxSrcScalarRegOpSize
|
private |
Definition at line 522 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst(), and maxSrcScalarRegOperandSize().
◆ maxSrcVecRegOpSize
|
private |
Definition at line 521 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst(), and maxSrcVecRegOperandSize().
◆ memStatusVector
| StatusVector gem5::GPUDynInst::memStatusVector |
Definition at line 489 of file gpu_dyn_inst.hh.
◆ numScalarReqs
| int gem5::GPUDynInst::numScalarReqs |
Definition at line 499 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ pAddr
| Addr gem5::GPUDynInst::pAddr |
Definition at line 199 of file gpu_dyn_inst.hh.
◆ roundTripTime
|
private |
Definition at line 530 of file gpu_dyn_inst.hh.
Referenced by getRoundTripTime(), and profileRoundTripTime().
◆ scalar_data
| uint8_t* gem5::GPUDynInst::scalar_data |
Definition at line 204 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst(), and ~GPUDynInst().
◆ scalarAddr
| Addr gem5::GPUDynInst::scalarAddr |
Definition at line 196 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ simdId
| int gem5::GPUDynInst::simdId |
Definition at line 213 of file gpu_dyn_inst.hh.
Referenced by allLanesZero(), completeAcc(), GPUDynInst(), initiateAcc(), and resolveFlatSegment().
◆ statusVector
| std::vector<int> gem5::GPUDynInst::statusVector |
Definition at line 493 of file gpu_dyn_inst.hh.
Referenced by allLanesZero(), decrementStatusVector(), getLaneStatus(), GPUDynInst(), printStatusVector(), resetEntireStatusVector(), and setStatusVector().
◆ systemReq
|
private |
Definition at line 523 of file gpu_dyn_inst.hh.
Referenced by isSystemReq(), and setSystemReq().
◆ time
| Tick gem5::GPUDynInst::time |
Definition at line 227 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ tlbHitLevel
| std::vector<int> gem5::GPUDynInst::tlbHitLevel |
Definition at line 495 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ wfDynId
| int gem5::GPUDynInst::wfDynId |
Definition at line 215 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ wfSlotId
| int gem5::GPUDynInst::wfSlotId |
Definition at line 223 of file gpu_dyn_inst.hh.
Referenced by allLanesZero(), completeAcc(), GPUDynInst(), and initiateAcc().
◆ wg_id
| int gem5::GPUDynInst::wg_id |
Definition at line 221 of file gpu_dyn_inst.hh.
Referenced by GPUDynInst().
◆ x_data
| uint8_t* gem5::GPUDynInst::x_data |
Definition at line 208 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
Generated on Mon Oct 27 2025 04:13:11 for gem5 by doxygen 1.14.0