src/gpu-compute/wavefront.cc

165 Wavefront::isGmInstruction(GPUDynInstPtr ii)
167     if (ii->isGlobalMem() || ii->isFlat())
174 Wavefront::isLmInstruction(GPUDynInstPtr ii)
176     if (ii->isLocalMem()) {
187     GPUDynInstPtr ii = instructionBuffer.front();
189     if (status != S_STOPPED && (ii->isNop() ||
190         ii->isReturn() || ii->isBranch() ||
191         ii->isALU() || (ii->isKernArgSeg() && ii->isLoad()))) {
202     GPUDynInstPtr ii = instructionBuffer.front();
204     if (status != S_STOPPED && ii->isBarrier()) {
215     GPUDynInstPtr ii = instructionBuffer.front();
217     if (status != S_STOPPED && ii->isGlobalMem()) {
228     GPUDynInstPtr ii = instructionBuffer.front();
230     if (status != S_STOPPED && ii->isLocalMem()) {
241     GPUDynInstPtr ii = instructionBuffer.front();
243     if (status != S_STOPPED && ii->isPrivateSeg()) {
254     GPUDynInstPtr ii = instructionBuffer.front();
256     if (status != S_STOPPED && ii->isFlat()) {
269         GPUDynInstPtr ii = it;
271         if (ii->isReturn() || ii->isBranch()) {
325     GPUDynInstPtr ii = instructionBuffer.front();
353     if (!(ii->isBarrier() || ii->isNop() || ii->isReturn() || ii->isBranch() ||
354         ii->isALU() || ii->isLoad() || ii->isStore() || ii->isAtomic() ||
355         ii->isMemFence() || ii->isFlat())) {
356         panic("next instruction: %s is of unknown type\n", ii->disassemble());
360             computeUnit->cu_id, simdId, wfSlotId, ii->disassemble());
362     if (type == I_ALU && ii->isBarrier()) {
375     } else if (type == I_ALU && ii->isNop()) {
383     } else if (type == I_ALU && ii->isReturn()) {
396     } else if (type == I_ALU && (ii->isBranch() ||
397                ii->isALU() ||
398                (ii->isKernArgSeg() && ii->isLoad()) ||
399                ii->isArgSeg())) {
405         if (!computeUnit->vrf[simdId]->vrfOperandAccessReady(this, ii,
410         if (!computeUnit->vrf[simdId]->operandsReady(this, ii)) {
414     } else if (type == I_GLOBAL && ii->isGlobalMem()) {
416         if (ii->isLoad() || ii->isAtomic() || ii->isMemFence()) {
423         if (ii->isStore() || ii->isAtomic() || ii->isMemFence()) {
445         if (!computeUnit->vrf[simdId]->vrfOperandAccessReady(this, ii,
449         if (!computeUnit->vrf[simdId]->operandsReady(this, ii)) {
453     } else if (type == I_SHARED && ii->isLocalMem()) {
455         if (ii->isLoad() || ii->isAtomic() || ii->isMemFence()) {
461         if (ii->isStore() || ii->isAtomic() || ii->isMemFence()) {
482         if (!computeUnit->vrf[simdId]->vrfOperandAccessReady(this, ii,
486         if (!computeUnit->vrf[simdId]->operandsReady(this, ii)) {
490     } else if (type == I_FLAT && ii->isFlat()) {
521         if (!computeUnit->vrf[simdId]->vrfOperandAccessReady(this, ii,
526         if (!computeUnit->vrf[simdId]->operandsReady(this, ii)) {
537             simdId, wfSlotId, ii->disassemble());
545     GPUDynInstPtr ii = instructionBuffer.front();
546     assert(ii);
547     computeUnit->vrf[simdId]->updateResources(this, ii);
549     if (ii->isALU() || ii->isSpecialOp() ||
550         ii->isBranch() ||
554         (ii->isKernArgSeg() && ii->isLoad()) || ii->isArgSeg() ||
555         ii->isReturn()) {
561     } else if (ii->isBarrier()) {
564     } else if (ii->isLoad() && ii->isFlat()) {
565         assert(Enums::SC_NONE != ii->executedAs());
568         if ( Enums::SC_SHARED == ii->executedAs() ) {
579     } else if (ii->isStore() && ii->isFlat()) {
580         assert(Enums::SC_NONE != ii->executedAs());
583         if (Enums::SC_SHARED == ii->executedAs()) {
594     } else if (ii->isLoad() && ii->isGlobalMem()) {
601     } else if (ii->isStore() && ii->isGlobalMem()) {
608     } else if ((ii->isAtomic() || ii->isMemFence()) && ii->isGlobalMem()) {
616     } else if (ii->isLoad() && ii->isLocalMem()) {
623     } else if (ii->isStore() && ii->isLocalMem()) {
630     } else if ((ii->isAtomic() || ii->isMemFence()) && ii->isLocalMem()) {
653     GPUDynInstPtr ii = instructionBuffer.front();
658             ii->disassemble(), old_pc);
662     ii->execute(ii);
663     computeUnit->updateInstStats(ii);
665     computeUnit->vrf[simdId]->exec(ii, this);
666     srcRegOpDist.sample(ii->numSrcRegOperands());
667     dstRegOpDist.sample(ii->numDstRegOperands());
673         uint32_t new_pc = _gpuISA.advancePC(old_pc, ii);
690         if (isGmInstruction(ii)) {
692         } else if (isLmInstruction(ii)) {
699     if (ii->isALU() || ii->isSpecialOp() ||
700         ii->isBranch() ||
704         (ii->isKernArgSeg() && ii->isLoad()) ||
705         ii->isArgSeg() ||
706         ii->isReturn()) {
713     } else if (ii->isBarrier()) {
716     } else if (ii->isLoad() && ii->isFlat()) {
717         assert(Enums::SC_NONE != ii->executedAs());
719         if (Enums::SC_SHARED == ii->executedAs()) {
730     } else if (ii->isStore() && ii->isFlat()) {
731         assert(Enums::SC_NONE != ii->executedAs());
732         if (Enums::SC_SHARED == ii->executedAs()) {
743     } else if (ii->isLoad() && ii->isGlobalMem()) {
748     } else if (ii->isStore() && ii->isGlobalMem()) {
753     } else if ((ii->isAtomic() || ii->isMemFence()) && ii->isGlobalMem()) {
758     } else if (ii->isLoad() && ii->isLocalMem()) {
763     } else if (ii->isStore() && ii->isLocalMem()) {
768     } else if ((ii->isAtomic() || ii->isMemFence()) && ii->isLocalMem()) {