From bd65ec07441cbb19fc2d3570da32a0205dad52c2 Mon Sep 17 00:00:00 2001 From: Alexandru Dutu Date: Fri, 16 Sep 2016 12:32:36 -0400 Subject: [PATCH] gpu-compute: Adding context serialization methods to Wavefront This patch adds methods to serialize the context of a particular wavefront to the simulated system memory. Context serialization is used when a wavefront is preempeted (i.e. context switch). --- src/gpu-compute/wavefront.cc | 125 +++++++++++++++++++++++++++++++++++ src/gpu-compute/wavefront.hh | 12 ++++ 2 files changed, 137 insertions(+) diff --git a/src/gpu-compute/wavefront.cc b/src/gpu-compute/wavefront.cc index c2356b47f..42739a7b0 100644 --- a/src/gpu-compute/wavefront.cc +++ b/src/gpu-compute/wavefront.cc @@ -941,3 +941,128 @@ Wavefront::getStaticContextSize() const sizeof(privBase) + sizeof(spillBase) + sizeof(ldsChunk) + computeUnit->wfSize() * sizeof(ReconvergenceStackEntry); } + +void +Wavefront::getContext(const void *out) +{ + uint8_t *iter = (uint8_t *)out; + for (int i = 0; i < barCnt.size(); i++) { + *(int *)iter = barCnt[i]; iter += sizeof(barCnt[i]); + } + *(int *)iter = wfId; iter += sizeof(wfId); + *(int *)iter = maxBarCnt; iter += sizeof(maxBarCnt); + *(int *)iter = oldBarrierCnt; iter += sizeof(oldBarrierCnt); + *(int *)iter = barrierCnt; iter += sizeof(barrierCnt); + *(int *)iter = computeUnit->cu_id; iter += sizeof(computeUnit->cu_id); + *(uint32_t *)iter = wgId; iter += sizeof(wgId); + *(uint32_t *)iter = barrierId; iter += sizeof(barrierId); + *(uint64_t *)iter = initMask.to_ullong(); iter += sizeof(initMask.to_ullong()); + *(Addr *)iter = privBase; iter += sizeof(privBase); + *(Addr *)iter = spillBase; iter += sizeof(spillBase); + + int stackSize = reconvergenceStack.size(); + ReconvergenceStackEntry empty = {std::numeric_limits::max(), + std::numeric_limits::max(), + std::numeric_limits::max()}; + for (int i = 0; i < workItemId[0].size(); i++) { + if (i < stackSize) { + *(ReconvergenceStackEntry *)iter = *reconvergenceStack.back(); + iter += sizeof(ReconvergenceStackEntry); + reconvergenceStack.pop_back(); + } else { + *(ReconvergenceStackEntry *)iter = empty; + iter += sizeof(ReconvergenceStackEntry); + } + } + + int wf_size = computeUnit->wfSize(); + for (int i = 0; i < maxSpVgprs; i++) { + uint32_t vgprIdx = remap(i, sizeof(uint32_t), 1); + for (int lane = 0; lane < wf_size; lane++) { + uint32_t regVal = computeUnit->vrf[simdId]-> + read(vgprIdx,lane); + *(uint32_t *)iter = regVal; iter += sizeof(regVal); + } + } + + for (int i = 0; i < maxDpVgprs; i++) { + uint32_t vgprIdx = remap(i, sizeof(uint64_t), 1); + for (int lane = 0; lane < wf_size; lane++) { + uint64_t regVal = computeUnit->vrf[simdId]-> + read(vgprIdx,lane); + *(uint64_t *)iter = regVal; iter += sizeof(regVal); + } + } + + for (int i = 0; i < condRegState->numRegs(); i++) { + for (int lane = 0; lane < wf_size; lane++) { + uint64_t regVal = condRegState->read(i, lane); + *(uint64_t *)iter = regVal; iter += sizeof(regVal); + } + } + + /* saving LDS content */ + if (ldsChunk) + for (int i = 0; i < ldsChunk->size(); i++) { + char val = ldsChunk->read(i); + *(char *) iter = val; iter += sizeof(val); + } +} + +void +Wavefront::setContext(const void *in) +{ + uint8_t *iter = (uint8_t *)in; + for (int i = 0; i < barCnt.size(); i++) { + barCnt[i] = *(int *)iter; iter += sizeof(barCnt[i]); + } + wfId = *(int *)iter; iter += sizeof(wfId); + maxBarCnt = *(int *)iter; iter += sizeof(maxBarCnt); + oldBarrierCnt = *(int *)iter; iter += sizeof(oldBarrierCnt); + barrierCnt = *(int *)iter; iter += sizeof(barrierCnt); + computeUnit->cu_id = *(int *)iter; iter += sizeof(computeUnit->cu_id); + wgId = *(uint32_t *)iter; iter += sizeof(wgId); + barrierId = *(uint32_t *)iter; iter += sizeof(barrierId); + initMask = VectorMask(*(uint64_t *)iter); iter += sizeof(initMask); + privBase = *(Addr *)iter; iter += sizeof(privBase); + spillBase = *(Addr *)iter; iter += sizeof(spillBase); + + for (int i = 0; i < workItemId[0].size(); i++) { + ReconvergenceStackEntry newEntry = *(ReconvergenceStackEntry *)iter; + iter += sizeof(ReconvergenceStackEntry); + if (newEntry.pc != std::numeric_limits::max()) { + pushToReconvergenceStack(newEntry.pc, newEntry.rpc, + newEntry.execMask); + } + } + int wf_size = computeUnit->wfSize(); + + for (int i = 0; i < maxSpVgprs; i++) { + uint32_t vgprIdx = remap(i, sizeof(uint32_t), 1); + for (int lane = 0; lane < wf_size; lane++) { + uint32_t regVal = *(uint32_t *)iter; iter += sizeof(regVal); + computeUnit->vrf[simdId]->write(vgprIdx, regVal, lane); + } + } + + for (int i = 0; i < maxDpVgprs; i++) { + uint32_t vgprIdx = remap(i, sizeof(uint64_t), 1); + for (int lane = 0; lane < wf_size; lane++) { + uint64_t regVal = *(uint64_t *)iter; iter += sizeof(regVal); + computeUnit->vrf[simdId]->write(vgprIdx, regVal, lane); + } + } + + for (int i = 0; i < condRegState->numRegs(); i++) { + for (int lane = 0; lane < wf_size; lane++) { + uint64_t regVal = *(uint64_t *)iter; iter += sizeof(regVal); + condRegState->write(i, lane, regVal); + } + } + /** Restoring LDS contents */ + if (ldsChunk) + for (int i = 0; i < ldsChunk->size(); i++) { + char val = *(char *) iter; iter += sizeof(val); + ldsChunk->write(i, val); + } +} diff --git a/src/gpu-compute/wavefront.hh b/src/gpu-compute/wavefront.hh index 8caeda91e..ef8c80989 100644 --- a/src/gpu-compute/wavefront.hh +++ b/src/gpu-compute/wavefront.hh @@ -354,6 +354,18 @@ class Wavefront : public SimObject */ uint32_t getStaticContextSize() const; + /** + * Returns the hardware context as a stream of bytes + * This method is designed for HSAIL execution + */ + void getContext(const void *out); + + /** + * Sets the hardware context fromt a stream of bytes + * This method is designed for HSAIL execution + */ + void setContext(const void *in); + private: /** * Stack containing Control Flow Graph nodes (i.e., kernel instructions)