sanchayanmaity/gem5
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

gpu-compute: Refactoring Wavefront::dynWaveId

Browse source
This commit is contained in:
Alexandru Dutu 2016-09-16 12:31:46 -04:00
parent 498d0e63e5
commit e9b14d5111
4 changed files with 16 additions and 15 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/arch/hsail/gen.py
View file

@ -755,7 +755,7 @@ gen_special('GridSize', 'w->gridSz[src0]')
gen_special('GridGroups', gen_special('GridGroups',
'divCeil(w->gridSz[src0],w->workGroupSz[src0])') 'divCeil(w->gridSz[src0],w->workGroupSz[src0])')
gen_special('LaneId', 'lane') gen_special('LaneId', 'lane')
gen_special('WaveId', 'w->dynWaveId') gen_special('WaveId', 'w->wfId')
gen_special('Clock', 'w->computeUnit->shader->tick_cnt', 'U64') gen_special('Clock', 'w->computeUnit->shader->tick_cnt', 'U64')
# gen_special('CU'', ') # gen_special('CU'', ')

24
src/gpu-compute/compute_unit.cc
View file

@ -221,7 +221,7 @@ ComputeUnit::updateEvents() {
void void
ComputeUnit::StartWF(Wavefront *w, int trueWgSize[], int trueWgSizeTotal, ComputeUnit::StartWF(Wavefront *w, int trueWgSize[], int trueWgSizeTotal,
int cnt, LdsChunk *ldsChunk, NDRange *ndr) int waveId, LdsChunk *ldsChunk, NDRange *ndr)
{ {
static int _n_wave = 0; static int _n_wave = 0;
@ -232,20 +232,20 @@ ComputeUnit::StartWF(Wavefront *w, int trueWgSize[], int trueWgSizeTotal,
init_mask.reset(); init_mask.reset();
for (int k = 0; k < wfSize(); ++k) { for (int k = 0; k < wfSize(); ++k) {
if (k + cnt * wfSize() < trueWgSizeTotal) if (k + waveId * wfSize() < trueWgSizeTotal)
init_mask[k] = 1; init_mask[k] = 1;
} }
w->kernId = ndr->dispatchId; w->kernId = ndr->dispatchId;
w->dynWaveId = cnt; w->wfId = waveId;
w->initMask = init_mask.to_ullong(); w->initMask = init_mask.to_ullong();
for (int k = 0; k < wfSize(); ++k) { for (int k = 0; k < wfSize(); ++k) {
w->workItemId[0][k] = (k+cnt*wfSize()) % trueWgSize[0]; w->workItemId[0][k] = (k + waveId * wfSize()) % trueWgSize[0];
w->workItemId[1][k] = w->workItemId[1][k] =
((k + cnt * wfSize()) / trueWgSize[0]) % trueWgSize[1]; ((k + waveId * wfSize()) / trueWgSize[0]) % trueWgSize[1];
w->workItemId[2][k] = w->workItemId[2][k] =
(k + cnt * wfSize()) / (trueWgSize[0] * trueWgSize[1]); (k + waveId * wfSize()) / (trueWgSize[0] * trueWgSize[1]);
w->workItemFlatId[k] = w->workItemId[2][k] * trueWgSize[0] * w->workItemFlatId[k] = w->workItemId[2][k] * trueWgSize[0] *
trueWgSize[1] + w->workItemId[1][k] * trueWgSize[0] + trueWgSize[1] + w->workItemId[1][k] * trueWgSize[0] +
@ -294,8 +294,8 @@ ComputeUnit::StartWF(Wavefront *w, int trueWgSize[], int trueWgSizeTotal,
// is this the last wavefront in the workgroup // is this the last wavefront in the workgroup
// if set the spillWidth to be the remaining work-items // if set the spillWidth to be the remaining work-items
// so that the vector access is correct // so that the vector access is correct
if ((cnt + 1) * wfSize() >= trueWgSizeTotal) { if ((waveId + 1) * wfSize() >= trueWgSizeTotal) {
w->spillWidth = trueWgSizeTotal - (cnt * wfSize()); w->spillWidth = trueWgSizeTotal - (waveId * wfSize());
} else { } else {
w->spillWidth = wfSize(); w->spillWidth = wfSize();
} }
@ -341,7 +341,7 @@ ComputeUnit::StartWorkgroup(NDRange *ndr)
// calculate the number of 32-bit vector registers required by wavefront // calculate the number of 32-bit vector registers required by wavefront
int vregDemand = ndr->q.sRegCount + (2 * ndr->q.dRegCount); int vregDemand = ndr->q.sRegCount + (2 * ndr->q.dRegCount);
int cnt = 0; int wave_id = 0;
// Assign WFs by spreading them across SIMDs, 1 WF per SIMD at a time // Assign WFs by spreading them across SIMDs, 1 WF per SIMD at a time
for (int m = 0; m < shader->n_wf * numSIMDs; ++m) { for (int m = 0; m < shader->n_wf * numSIMDs; ++m) {
@ -352,7 +352,7 @@ ComputeUnit::StartWorkgroup(NDRange *ndr)
if (w->status == Wavefront::S_STOPPED) { if (w->status == Wavefront::S_STOPPED) {
// if we have scheduled all work items then stop // if we have scheduled all work items then stop
// scheduling wavefronts // scheduling wavefronts
if (cnt * wfSize() >= trueWgSizeTotal) if (wave_id * wfSize() >= trueWgSizeTotal)
break; break;
// reserve vector registers for the scheduled wavefront // reserve vector registers for the scheduled wavefront
@ -365,8 +365,8 @@ ComputeUnit::StartWorkgroup(NDRange *ndr)
w->reservedVectorRegs = normSize; w->reservedVectorRegs = normSize;
vectorRegsReserved[m % numSIMDs] += w->reservedVectorRegs; vectorRegsReserved[m % numSIMDs] += w->reservedVectorRegs;
StartWF(w, trueWgSize, trueWgSizeTotal, cnt, ldsChunk, ndr); StartWF(w, trueWgSize, trueWgSizeTotal, wave_id, ldsChunk, ndr);
++cnt; ++wave_id;
} }
} }
++barrier_id; ++barrier_id;

2
src/gpu-compute/wavefront.cc
View file

@ -935,7 +935,7 @@ Wavefront::pc(uint32_t new_pc)
uint32_t uint32_t
Wavefront::getStaticContextSize() const Wavefront::getStaticContextSize() const
{ {
return barCnt.size() * sizeof(int) + sizeof(dynWaveId) + sizeof(maxBarCnt) + return barCnt.size() * sizeof(int) + sizeof(wfId) + sizeof(maxBarCnt) +
sizeof(oldBarrierCnt) + sizeof(barrierCnt) + sizeof(wgId) + sizeof(oldBarrierCnt) + sizeof(barrierCnt) + sizeof(wgId) +
sizeof(computeUnit->cu_id) + sizeof(barrierId) + sizeof(initMask) + sizeof(computeUnit->cu_id) + sizeof(barrierId) + sizeof(initMask) +
sizeof(privBase) + sizeof(spillBase) + sizeof(ldsChunk) + sizeof(privBase) + sizeof(spillBase) + sizeof(ldsChunk) +

3
src/gpu-compute/wavefront.hh
View file

@ -194,7 +194,8 @@ class Wavefront : public SimObject
uint32_t gridSz[3]; uint32_t gridSz[3];
uint32_t wgId; uint32_t wgId;
uint32_t wgSz; uint32_t wgSz;
uint32_t dynWaveId; // wavefront id within a workgroup
uint32_t wfId;
uint32_t maxDynWaveId; uint32_t maxDynWaveId;
uint32_t dispatchId; uint32_t dispatchId;
// outstanding global+local memory requests // outstanding global+local memory requests