From 00a6346c91f6e09eb9c0f4cf61a7d72932a1282f Mon Sep 17 00:00:00 2001 From: Tony Gutierrez Date: Wed, 26 Oct 2016 22:47:19 -0400 Subject: [PATCH] hsail, gpu-compute: remove doGm/SmReturn add completeAcc we are removing doGmReturn from the GM pipe, and adding completeAcc() implementations for the HSAIL mem ops. the behavior in doGmReturn is dependent on HSAIL and HSAIL mem ops, however the completion phase of memory ops in machine ISA can be very different, even amongst individual machine ISA mem ops. so we remove this functionality from the pipeline and allow it to be implemented by the individual instructions. --- src/arch/hsail/insts/mem.hh | 140 ++++++++++++++++++++ src/gpu-compute/global_memory_pipeline.cc | 153 +++++----------------- src/gpu-compute/global_memory_pipeline.hh | 8 +- src/gpu-compute/gpu_dyn_inst.cc | 6 + src/gpu-compute/gpu_dyn_inst.hh | 4 + src/gpu-compute/local_memory_pipeline.cc | 126 ++++-------------- src/gpu-compute/local_memory_pipeline.hh | 8 +- 7 files changed, 225 insertions(+), 220 deletions(-) diff --git a/src/arch/hsail/insts/mem.hh b/src/arch/hsail/insts/mem.hh index e223c7cf5..2e7dfcd1c 100644 --- a/src/arch/hsail/insts/mem.hh +++ b/src/arch/hsail/insts/mem.hh @@ -36,9 +36,12 @@ #ifndef __ARCH_HSAIL_INSTS_MEM_HH__ #define __ARCH_HSAIL_INSTS_MEM_HH__ +#include + #include "arch/hsail/insts/decl.hh" #include "arch/hsail/insts/gpu_static_inst.hh" #include "arch/hsail/operand.hh" +#include "gpu-compute/compute_unit.hh" namespace HsailISA { @@ -491,6 +494,86 @@ namespace HsailISA gpuDynInst->updateStats(); } + void + completeAcc(GPUDynInstPtr gpuDynInst) override + { + typedef typename MemDataType::CType c1; + + constexpr bool is_vt_32 = DestDataType::vgprType == VT_32; + + /** + * this code essentially replaces the long if-else chain + * that was in used GlobalMemPipeline::exec() to infer the + * size (single/double) and type (floating point/integer) of + * the destination register. this is needed for load + * instructions because the loaded value and the + * destination type can be of different sizes, and we also + * need to know if the value we're writing back is floating + * point and signed/unsigned, so we can properly cast the + * writeback value + */ + typedef typename std::conditional::value, + float, typename std::conditional::value, + int32_t, uint32_t>::type>::type, + typename std::conditional::value, + double, typename std::conditional::value, + int64_t, uint64_t>::type>::type>::type c0; + + + Wavefront *w = gpuDynInst->wavefront(); + + std::vector regVec; + // iterate over number of destination register operands since + // this is a load + for (int k = 0; k < num_dest_operands; ++k) { + assert((sizeof(c1) * num_dest_operands) + <= MAX_WIDTH_FOR_MEM_INST); + + int dst = this->dest.regIndex() + k; + if (num_dest_operands > MAX_REGS_FOR_NON_VEC_MEM_INST) + dst = dest_vect[k].regIndex(); + // virtual->physical VGPR mapping + int physVgpr = w->remap(dst, sizeof(c0), 1); + // save the physical VGPR index + regVec.push_back(physVgpr); + + c1 *p1 = + &((c1*)gpuDynInst->d_data)[k * w->computeUnit->wfSize()]; + + for (int i = 0; i < w->computeUnit->wfSize(); ++i) { + if (gpuDynInst->exec_mask[i]) { + DPRINTF(GPUReg, "CU%d, WF[%d][%d], lane %d: " + "$%s%d <- %d global ld done (src = wavefront " + "ld inst)\n", w->computeUnit->cu_id, w->simdId, + w->wfSlotId, i, sizeof(c0) == 4 ? "s" : "d", + dst, *p1); + // write the value into the physical VGPR. This is a + // purely functional operation. No timing is modeled. + w->computeUnit->vrf[w->simdId]->write(physVgpr, + *p1, i); + } + ++p1; + } + } + + // Schedule the write operation of the load data on the VRF. + // This simply models the timing aspect of the VRF write operation. + // It does not modify the physical VGPR. + int loadVrfBankConflictCycles = gpuDynInst->computeUnit()-> + vrf[w->simdId]->exec(gpuDynInst->seqNum(), w, regVec, + sizeof(c0), gpuDynInst->time); + + if (this->isGlobalMem()) { + gpuDynInst->computeUnit()->globalMemoryPipe + .incLoadVRFBankConflictCycles(loadVrfBankConflictCycles); + } else { + assert(this->isLocalMem()); + gpuDynInst->computeUnit()->localMemoryPipe + .incLoadVRFBankConflictCycles(loadVrfBankConflictCycles); + } + } + private: void execLdAcq(GPUDynInstPtr gpuDynInst) override @@ -941,6 +1024,11 @@ namespace HsailISA execSt(gpuDynInst); } + // stores don't write anything back, so there is nothing + // to do here. we only override this method to avoid the + // fatal in the base class implementation + void completeAcc(GPUDynInstPtr gpuDynInst) override { } + private: // execSt may be called through a continuation // if the store had release semantics. see comment for @@ -1409,6 +1497,58 @@ namespace HsailISA } + void + completeAcc(GPUDynInstPtr gpuDynInst) override + { + // if this is not an atomic return op, then we + // have nothing more to do. + if (this->isAtomicRet()) { + // the size of the src operands and the + // memory being operated on must match + // for HSAIL atomics - this assumption may + // not apply to all ISAs + typedef typename MemDataType::CType CType; + + Wavefront *w = gpuDynInst->wavefront(); + int dst = this->dest.regIndex(); + std::vector regVec; + // virtual->physical VGPR mapping + int physVgpr = w->remap(dst, sizeof(CType), 1); + regVec.push_back(physVgpr); + CType *p1 = &((CType*)gpuDynInst->d_data)[0]; + + for (int i = 0; i < w->computeUnit->wfSize(); ++i) { + if (gpuDynInst->exec_mask[i]) { + DPRINTF(GPUReg, "CU%d, WF[%d][%d], lane %d: " + "$%s%d <- %d global ld done (src = wavefront " + "ld inst)\n", w->computeUnit->cu_id, w->simdId, + w->wfSlotId, i, sizeof(CType) == 4 ? "s" : "d", + dst, *p1); + // write the value into the physical VGPR. This is a + // purely functional operation. No timing is modeled. + w->computeUnit->vrf[w->simdId]->write(physVgpr, *p1, i); + } + ++p1; + } + + // Schedule the write operation of the load data on the VRF. + // This simply models the timing aspect of the VRF write operation. + // It does not modify the physical VGPR. + int loadVrfBankConflictCycles = gpuDynInst->computeUnit()-> + vrf[w->simdId]->exec(gpuDynInst->seqNum(), w, regVec, + sizeof(CType), gpuDynInst->time); + + if (this->isGlobalMem()) { + gpuDynInst->computeUnit()->globalMemoryPipe + .incLoadVRFBankConflictCycles(loadVrfBankConflictCycles); + } else { + assert(this->isLocalMem()); + gpuDynInst->computeUnit()->localMemoryPipe + .incLoadVRFBankConflictCycles(loadVrfBankConflictCycles); + } + } + } + void execute(GPUDynInstPtr gpuDynInst) override; private: diff --git a/src/gpu-compute/global_memory_pipeline.cc b/src/gpu-compute/global_memory_pipeline.cc index ab3e8c47e..f48af5a6f 100644 --- a/src/gpu-compute/global_memory_pipeline.cc +++ b/src/gpu-compute/global_memory_pipeline.cc @@ -65,13 +65,15 @@ GlobalMemPipeline::exec() !gmReturnedStores.empty() ? gmReturnedStores.front() : nullptr; bool accessVrf = true; + Wavefront *w = nullptr; + // check the VRF to see if the operands of a load (or load component // of an atomic) are accessible if ((m) && (m->isLoad() || m->isAtomicRet())) { - Wavefront *w = computeUnit->wfList[m->simdId][m->wfSlotId]; + w = m->wavefront(); accessVrf = - w->computeUnit->vrf[m->simdId]-> + w->computeUnit->vrf[w->simdId]-> vrfOperandAccessReady(m->seqNum(), w, m, VrfAccessType::WRITE); } @@ -82,44 +84,38 @@ GlobalMemPipeline::exec() (computeUnit->shader->coissue_return || computeUnit->wfWait.at(m->pipeId).rdy())) { - if (m->v_type == VT_32 && m->m_type == Enums::M_U8) - doGmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_U16) - doGmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_U32) - doGmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_S8) - doGmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_S16) - doGmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_S32) - doGmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_F16) - doGmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_F32) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U8) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U16) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U32) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U64) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S8) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S16) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S32) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S64) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_F16) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_F32) - doGmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_F64) - doGmReturn(m); + w = m->wavefront(); + + m->completeAcc(m); + + if (m->isLoad() || m->isAtomic()) { + gmReturnedLoads.pop(); + assert(inflightLoads > 0); + --inflightLoads; + } else { + assert(m->isStore()); + gmReturnedStores.pop(); + assert(inflightStores > 0); + --inflightStores; + } + + // Decrement outstanding register count + computeUnit->shader->ScheduleAdd(&w->outstandingReqs, m->time, -1); + + if (m->isStore() || m->isAtomic()) { + computeUnit->shader->ScheduleAdd(&w->outstandingReqsWrGm, + m->time, -1); + } + + if (m->isLoad() || m->isAtomic()) { + computeUnit->shader->ScheduleAdd(&w->outstandingReqsRdGm, + m->time, -1); + } + + // Mark write bus busy for appropriate amount of time + computeUnit->glbMemToVrfBus.set(m->time); + if (!computeUnit->shader->coissue_return) + w->computeUnit->wfWait.at(m->pipeId).set(m->time); } // If pipeline has executed a global memory instruction @@ -149,83 +145,6 @@ GlobalMemPipeline::exec() } } -template -void -GlobalMemPipeline::doGmReturn(GPUDynInstPtr m) -{ - Wavefront *w = computeUnit->wfList[m->simdId][m->wfSlotId]; - - // Return data to registers - if (m->isLoad() || m->isAtomic()) { - gmReturnedLoads.pop(); - assert(inflightLoads > 0); - --inflightLoads; - - if (m->isLoad() || m->isAtomicRet()) { - std::vector regVec; - // iterate over number of destination register operands since - // this is a load or atomic operation - for (int k = 0; k < m->n_reg; ++k) { - assert((sizeof(c1) * m->n_reg) <= MAX_WIDTH_FOR_MEM_INST); - int dst = m->dst_reg + k; - - if (m->n_reg > MAX_REGS_FOR_NON_VEC_MEM_INST) - dst = m->dst_reg_vec[k]; - // virtual->physical VGPR mapping - int physVgpr = w->remap(dst, sizeof(c0), 1); - // save the physical VGPR index - regVec.push_back(physVgpr); - c1 *p1 = &((c1 *)m->d_data)[k * w->computeUnit->wfSize()]; - - for (int i = 0; i < w->computeUnit->wfSize(); ++i) { - if (m->exec_mask[i]) { - DPRINTF(GPUReg, "CU%d, WF[%d][%d], lane %d: " - "$%s%d <- %d global ld done (src = wavefront " - "ld inst)\n", w->computeUnit->cu_id, w->simdId, - w->wfSlotId, i, sizeof(c0) == 4 ? "s" : "d", - dst, *p1); - // write the value into the physical VGPR. This is a - // purely functional operation. No timing is modeled. - w->computeUnit->vrf[w->simdId]->write(physVgpr, - *p1, i); - } - ++p1; - } - } - - // Schedule the write operation of the load data on the VRF. - // This simply models the timing aspect of the VRF write operation. - // It does not modify the physical VGPR. - loadVrfBankConflictCycles += - w->computeUnit->vrf[w->simdId]->exec(m->seqNum(), - w, regVec, sizeof(c0), - m->time); - } - } else { - gmReturnedStores.pop(); - assert(inflightStores > 0); - --inflightStores; - } - - // Decrement outstanding register count - computeUnit->shader->ScheduleAdd(&w->outstandingReqs, m->time, -1); - - if (m->isStore() || m->isAtomic()) { - computeUnit->shader->ScheduleAdd(&w->outstandingReqsWrGm, m->time, - -1); - } - - if (m->isLoad() || m->isAtomic()) { - computeUnit->shader->ScheduleAdd(&w->outstandingReqsRdGm, m->time, - -1); - } - - // Mark write bus busy for appropriate amount of time - computeUnit->glbMemToVrfBus.set(m->time); - if (!computeUnit->shader->coissue_return) - w->computeUnit->wfWait.at(m->pipeId).set(m->time); -} - void GlobalMemPipeline::regStats() { diff --git a/src/gpu-compute/global_memory_pipeline.hh b/src/gpu-compute/global_memory_pipeline.hh index ed49f6f6b..368a15079 100644 --- a/src/gpu-compute/global_memory_pipeline.hh +++ b/src/gpu-compute/global_memory_pipeline.hh @@ -62,8 +62,6 @@ class GlobalMemPipeline void init(ComputeUnit *cu); void exec(); - template void doGmReturn(GPUDynInstPtr m); - std::queue &getGMReqFIFO() { return gmIssuedRequests; } std::queue &getGMStRespFIFO() { return gmReturnedStores; } std::queue &getGMLdRespFIFO() { return gmReturnedLoads; } @@ -89,6 +87,12 @@ class GlobalMemPipeline const std::string &name() const { return _name; } void regStats(); + void + incLoadVRFBankConflictCycles(int num_cycles) + { + loadVrfBankConflictCycles += num_cycles; + } + private: ComputeUnit *computeUnit; std::string _name; diff --git a/src/gpu-compute/gpu_dyn_inst.cc b/src/gpu-compute/gpu_dyn_inst.cc index ec6340360..7092a7a40 100644 --- a/src/gpu-compute/gpu_dyn_inst.cc +++ b/src/gpu-compute/gpu_dyn_inst.cc @@ -155,6 +155,12 @@ GPUDynInst::initiateAcc(GPUDynInstPtr gpuDynInst) time = 0; } +void +GPUDynInst::completeAcc(GPUDynInstPtr gpuDynInst) +{ + _staticInst->completeAcc(gpuDynInst); +} + /** * accessor methods for the attributes of * the underlying GPU static instruction diff --git a/src/gpu-compute/gpu_dyn_inst.hh b/src/gpu-compute/gpu_dyn_inst.hh index c07d85d78..527b87b4c 100644 --- a/src/gpu-compute/gpu_dyn_inst.hh +++ b/src/gpu-compute/gpu_dyn_inst.hh @@ -258,6 +258,10 @@ class GPUDynInst : public GPUExecContext // Initiate the specified memory operation, by creating a // memory request and sending it off to the memory system. void initiateAcc(GPUDynInstPtr gpuDynInst); + // Complete the specified memory operation, by writing + // value back to the RF in the case of a load or atomic + // return or, in the case of a store, we do nothing + void completeAcc(GPUDynInstPtr gpuDynInst); void updateStats(); diff --git a/src/gpu-compute/local_memory_pipeline.cc b/src/gpu-compute/local_memory_pipeline.cc index 80dad6fcd..9e7dc6fb3 100644 --- a/src/gpu-compute/local_memory_pipeline.cc +++ b/src/gpu-compute/local_memory_pipeline.cc @@ -62,11 +62,13 @@ LocalMemPipeline::exec() lmReturnedRequests.front() : nullptr; bool accessVrf = true; + Wavefront *w = nullptr; + if ((m) && (m->isLoad() || m->isAtomicRet())) { - Wavefront *w = computeUnit->wfList[m->simdId][m->wfSlotId]; + w = m->wavefront(); accessVrf = - w->computeUnit->vrf[m->simdId]-> + w->computeUnit->vrf[w->simdId]-> vrfOperandAccessReady(m->seqNum(), w, m, VrfAccessType::WRITE); } @@ -74,44 +76,29 @@ LocalMemPipeline::exec() if (!lmReturnedRequests.empty() && m->latency.rdy() && accessVrf && computeUnit->locMemToVrfBus.rdy() && (computeUnit->shader->coissue_return || computeUnit->wfWait.at(m->pipeId).rdy())) { - if (m->v_type == VT_32 && m->m_type == Enums::M_U8) - doSmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_U16) - doSmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_U32) - doSmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_S8) - doSmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_S16) - doSmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_S32) - doSmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_F16) - doSmReturn(m); - else if (m->v_type == VT_32 && m->m_type == Enums::M_F32) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U8) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U16) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U32) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_U64) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S8) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S16) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S32) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_S64) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_F16) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_F32) - doSmReturn(m); - else if (m->v_type == VT_64 && m->m_type == Enums::M_F64) - doSmReturn(m); + + lmReturnedRequests.pop(); + w = m->wavefront(); + + m->completeAcc(m); + + // Decrement outstanding request count + computeUnit->shader->ScheduleAdd(&w->outstandingReqs, m->time, -1); + + if (m->isStore() || m->isAtomic()) { + computeUnit->shader->ScheduleAdd(&w->outstandingReqsWrLm, + m->time, -1); + } + + if (m->isLoad() || m->isAtomic()) { + computeUnit->shader->ScheduleAdd(&w->outstandingReqsRdLm, + m->time, -1); + } + + // Mark write bus busy for appropriate amount of time + computeUnit->locMemToVrfBus.set(m->time); + if (computeUnit->shader->coissue_return == 0) + w->computeUnit->wfWait.at(m->pipeId).set(m->time); } // If pipeline has executed a local memory instruction @@ -129,65 +116,6 @@ LocalMemPipeline::exec() } } -template -void -LocalMemPipeline::doSmReturn(GPUDynInstPtr m) -{ - lmReturnedRequests.pop(); - Wavefront *w = computeUnit->wfList[m->simdId][m->wfSlotId]; - - // Return data to registers - if (m->isLoad() || m->isAtomicRet()) { - std::vector regVec; - for (int k = 0; k < m->n_reg; ++k) { - int dst = m->dst_reg+k; - - if (m->n_reg > MAX_REGS_FOR_NON_VEC_MEM_INST) - dst = m->dst_reg_vec[k]; - // virtual->physical VGPR mapping - int physVgpr = w->remap(dst,sizeof(c0),1); - // save the physical VGPR index - regVec.push_back(physVgpr); - c1 *p1 = &((c1 *)m->d_data)[k * w->computeUnit->wfSize()]; - - for (int i = 0; i < w->computeUnit->wfSize(); ++i) { - if (m->exec_mask[i]) { - // write the value into the physical VGPR. This is a purely - // functional operation. No timing is modeled. - w->computeUnit->vrf[w->simdId]->write(physVgpr, - *p1, i); - } - ++p1; - } - } - - // Schedule the write operation of the load data on the VRF. This simply - // models the timing aspect of the VRF write operation. It does not - // modify the physical VGPR. - loadVrfBankConflictCycles += - w->computeUnit->vrf[w->simdId]->exec(m->seqNum(), w, - regVec, sizeof(c0), m->time); - } - - // Decrement outstanding request count - computeUnit->shader->ScheduleAdd(&w->outstandingReqs, m->time, -1); - - if (m->isStore() || m->isAtomic()) { - computeUnit->shader->ScheduleAdd(&w->outstandingReqsWrLm, - m->time, -1); - } - - if (m->isLoad() || m->isAtomic()) { - computeUnit->shader->ScheduleAdd(&w->outstandingReqsRdLm, - m->time, -1); - } - - // Mark write bus busy for appropriate amount of time - computeUnit->locMemToVrfBus.set(m->time); - if (computeUnit->shader->coissue_return == 0) - w->computeUnit->wfWait.at(m->pipeId).set(m->time); -} - void LocalMemPipeline::regStats() { diff --git a/src/gpu-compute/local_memory_pipeline.hh b/src/gpu-compute/local_memory_pipeline.hh index a63d867d0..e0a21fd82 100644 --- a/src/gpu-compute/local_memory_pipeline.hh +++ b/src/gpu-compute/local_memory_pipeline.hh @@ -61,8 +61,6 @@ class LocalMemPipeline void init(ComputeUnit *cu); void exec(); - template void doSmReturn(GPUDynInstPtr m); - std::queue &getLMReqFIFO() { return lmIssuedRequests; } std::queue &getLMRespFIFO() { return lmReturnedRequests; } @@ -81,6 +79,12 @@ class LocalMemPipeline const std::string& name() const { return _name; } void regStats(); + void + incLoadVRFBankConflictCycles(int num_cycles) + { + loadVrfBankConflictCycles += num_cycles; + } + private: ComputeUnit *computeUnit; std::string _name;