src/gpu-compute/compute_unit.cc

181 ComputeUnit::fillKernelState(Wavefront *w, NDRange *ndr)
183     w->resizeRegFiles(ndr->q.cRegCount, ndr->q.sRegCount, ndr->q.dRegCount);
185     w->workGroupSz[0] = ndr->q.wgSize[0];
186     w->workGroupSz[1] = ndr->q.wgSize[1];
187     w->workGroupSz[2] = ndr->q.wgSize[2];
189     w->gridSz[0] = ndr->q.gdSize[0];
190     w->gridSz[1] = ndr->q.gdSize[1];
191     w->gridSz[2] = ndr->q.gdSize[2];
192     w->kernelArgs = ndr->q.args;
193     w->privSizePerItem = ndr->q.privMemPerItem;
194     w->spillSizePerItem = ndr->q.spillMemPerItem;
195     w->roBase = ndr->q.roMemStart;
196     w->roSize = ndr->q.roMemTotal;
197     w->computeActualWgSz(ndr);
229                             NDRange *ndr)
241     w->kernId = ndr->dispatchId;
265     w->privBase = ndr->q.privMemStart;
266     ndr->q.privMemStart += ndr->q.privMemPerItem * wfSize();
268     w->spillBase = ndr->q.spillMemStart;
269     ndr->q.spillMemStart += ndr->q.spillMemPerItem * wfSize();
274     w->wgId = ndr->globalWgId;
275     w->dispatchId = ndr->dispatchId;
276     w->workGroupId[0] = w->wgId % ndr->numWg[0];
277     w->workGroupId[1] = (w->wgId / ndr->numWg[0]) % ndr->numWg[1];
278     w->workGroupId[2] = w->wgId / (ndr->numWg[0] * ndr->numWg[1]);
308     w->start(++_n_wave, ndr->q.code_ptr);
312 ComputeUnit::StartWorkgroup(NDRange *ndr)
317     LdsChunk *ldsChunk = lds.reserveSpace(ndr->dispatchId, ndr->globalWgId,
318                                           ndr->q.ldsSize);
332     int vregDemand = ndr->q.sRegCount + (2 * ndr->q.dRegCount);
342             fillKernelState(w, ndr);
358             startWavefront(w, wave_id, ldsChunk, ndr);
366 ComputeUnit::ReadyWorkgroup(NDRange *ndr)
373         trueWgSize[d] = std::min(ndr->q.wgSize[d], ndr->q.gdSize[d] -
374                                  ndr->wgId[d] * ndr->q.wgSize[d]);
384     int vregDemandPerWI = ndr->q.sRegCount + (2 * ndr->q.dRegCount);
437     if (!lds.canReserve(ndr->q.ldsSize)) {
444     return freeWfSlots >= numWfs && vregAvail && lds.canReserve(ndr->q.ldsSize);