diff options
Diffstat (limited to 'benchmarks/CUDA/DG/3rdParty/ParMetis-3.1/ParMETISLib/sync')
| -rw-r--r-- | benchmarks/CUDA/DG/3rdParty/ParMetis-3.1/ParMETISLib/sync | 186 |
1 files changed, 186 insertions, 0 deletions
diff --git a/benchmarks/CUDA/DG/3rdParty/ParMetis-3.1/ParMETISLib/sync b/benchmarks/CUDA/DG/3rdParty/ParMetis-3.1/ParMETISLib/sync new file mode 100644 index 0000000..11a0cfb --- /dev/null +++ b/benchmarks/CUDA/DG/3rdParty/ParMetis-3.1/ParMETISLib/sync @@ -0,0 +1,186 @@ +adrivers.c: ubavg = savg(graph->ncon, ctrl->ubvec); +adrivers.c: ctrl->redist_factor = ctrl->redist_base * ((gtewgt/gtvsize)/ ctrl->edge_size_ratio); +adrivers.c: IFSET(ctrl->dbglvl, DBG_PROGRESS, rprintf(ctrl, "[%6d %8d %5d %5d][%d]\n", +adrivers.c: graph->gnvtxs, GlobalSESum(ctrl, graph->nedges), GlobalSEMin(ctrl, graph->nvtxs), GlobalSEMax(ctrl, graph->nvtxs), ctrl->CoarsenTo)); +adrivers.c: if (graph->gnvtxs < 1.3*ctrl->CoarsenTo || +adrivers.c: if (lbavg > ubavg + 0.035 && ctrl->partType != REFINE_PARTITION) +adrivers.c: if (ctrl->dbglvl&DBG_PROGRESS) { +adrivers.c: switch (ctrl->ps_relation) { +adrivers.c: if (ctrl->dbglvl&DBG_PROGRESS) { +akwayfm.c: int npes = ctrl->npes, mype = ctrl->mype, nparts = ctrl->nparts; +akwayfm.c: IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->KWayTmr)); +akwayfm.c: ubvec = ctrl->ubvec; +akwayfm.c: tpwgts = ctrl->tpwgts; +akwayfm.c: ipc_factor = ctrl->ipc_factor; +akwayfm.c: redist_factor = ctrl->redist_factor; +akwayfm.c: MPI_Bcast((void *)pperm, nparts, IDX_DATATYPE, 0, ctrl->comm); +akwayfm.c: switch (ctrl->ps_relation) { +akwayfm.c: switch (ctrl->ps_relation) { +akwayfm.c: MPI_Allreduce((void *)lnpwgts, (void *)pgnpwgts, nparts*ncon, MPI_FLOAT, MPI_SUM, ctrl->comm); +akwayfm.c: IFSET(ctrl->dbglvl, DBG_RMOVEINFO, rprintf(ctrl, "\t[%d %d], [%.4f], [%d %d %d]\n", +akwayfm.c: MPI_Irecv((void *)(rupdate+sendptr[i]), sendptr[i+1]-sendptr[i], IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->rreq+i); +akwayfm.c: MPI_Isend((void *)(supdate+j), k-j, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i); +akwayfm.c: MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses); +akwayfm.c: MPI_Get_count(ctrl->statuses+i, IDX_DATATYPE, nupds_pe+i); +akwayfm.c: MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses); +akwayfm.c: MPI_Allreduce((void *)lnpwgts, (void *)gnpwgts, nparts*ncon, MPI_FLOAT, MPI_SUM, ctrl->comm); +akwayfm.c: IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->KWayTmr)); +balancemylink.c: ipc_factor = ctrl->ipc_factor; +balancemylink.c: redist_factor = ctrl->redist_factor; +coarsen.c: int npes=ctrl->npes, mype=ctrl->mype; +coarsen.c: MPI_Allgather((void *)(cvtxdist+npes), 1, IDX_DATATYPE, (void *)cvtxdist, 1, IDX_DATATYPE, ctrl->comm); +coarsen.c: MPI_Irecv((void *)(rsizes+i), 1, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->rreq+i); +coarsen.c: MPI_Isend((void *)(ssizes+i), 1, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i); +coarsen.c: MPI_Wait(ctrl->rreq+i, &ctrl->status); +coarsen.c: MPI_Wait(ctrl->sreq+i, &ctrl->status); +coarsen.c: MPI_Irecv((void *)(rgraph+l), (4+ncon)*(rlens[i+1]-rlens[i])+2*rsizes[i], IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->rreq+i); +coarsen.c: sgraph[ll++] = (ctrl->partType == STATIC_PARTITION) ? -1 : vsize[ii]; +coarsen.c: sgraph[ll++] = (ctrl->partType == STATIC_PARTITION) ? -1 : home[ii]; +coarsen.c: MPI_Isend((void *)(sgraph+l), ll-l, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i); +coarsen.c: MPI_Wait(ctrl->rreq+i, &ctrl->status); +coarsen.c: MPI_Wait(ctrl->sreq+i, &ctrl->status); +coarsen.c: if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) { +coarsen.c: if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) { +coarsen.c: if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) { +coarsen.c: if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) { +coarsen.c: cgraph->nvwgt[j*ncon+h] = (float)(cvwgt[j*ncon+h])/(float)(ctrl->tvwgts[h]); +comm.c: firstvtx = graph->vtxdist[ctrl->mype]; +comm.c: peind[i], 1, ctrl->comm, ctrl->rreq+i); +comm.c: peind[i], 1, ctrl->comm, ctrl->sreq+i); +comm.c: MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses); +comm.c: MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses); +comm.c: firstvtx = graph->vtxdist[ctrl->mype]; +comm.c: peind[i], 1, ctrl->comm, ctrl->rreq+i); +comm.c: idxcopy(ctrl->npes, sendptr, psendptr); +comm.c: peind[i], 1, ctrl->comm, ctrl->sreq+i); +comm.c: MPI_Isend((void *)(sendpairs), 0, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i); +comm.c: MPI_Wait(ctrl->rreq+i, &(ctrl->status)); +comm.c: MPI_Get_count(&ctrl->status, IDX_DATATYPE, &n); +comm.c: MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses); +comm.c: MPI_Allreduce((void *)&value, (void *)&max, 1, MPI_INT, MPI_MAX, ctrl->comm); +comm.c: MPI_Allreduce((void *)&value, (void *)&max, 1, MPI_DOUBLE, MPI_MAX, ctrl->comm); +comm.c: MPI_Allreduce((void *)&value, (void *)&min, 1, MPI_INT, MPI_MIN, ctrl->comm); +comm.c: MPI_Allreduce((void *)&value, (void *)&sum, 1, MPI_INT, MPI_SUM, ctrl->comm); +comm.c: MPI_Allreduce((void *)&value, (void *)&max, 1, MPI_FLOAT, MPI_MAX, ctrl->comm); +comm.c: MPI_Allreduce((void *)&value, (void *)&min, 1, MPI_FLOAT, MPI_MIN, ctrl->comm); +comm.c: MPI_Allreduce((void *)&value, (void *)&sum, 1, MPI_FLOAT, MPI_SUM, ctrl->comm); +debug.c: for (penum=0; penum<ctrl->npes; penum++) { +debug.c: if (ctrl->mype == penum) { +debug.c: if (ctrl->mype == 0) +debug.c: printf("\t%3d. ", ctrl->mype); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: for (penum=0; penum<ctrl->npes; penum++) { +debug.c: if (ctrl->mype == penum) { +debug.c: if (ctrl->mype == 0) +debug.c: printf("\t%3d. ", ctrl->mype); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: for (penum=0; penum<ctrl->npes; penum++) { +debug.c: if (ctrl->mype == penum) { +debug.c: if (ctrl->mype == 0) +debug.c: printf("\t%3d. ", ctrl->mype); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: firstvtx = graph->vtxdist[ctrl->mype]; +debug.c: for (penum=0; penum<ctrl->npes; penum++) { +debug.c: if (ctrl->mype == penum) { +debug.c: MPI_Barrier(ctrl->comm); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: firstvtx = graph->vtxdist[ctrl->mype]; +debug.c: for (penum=0; penum<ctrl->npes; penum++) { +debug.c: if (ctrl->mype == penum) { +debug.c: MPI_Barrier(ctrl->comm); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: for (penum=0; penum<ctrl->npes; penum++) { +debug.c: if (ctrl->mype == penum) { +debug.c: printf("PE: %d, nnbrs: %d\n", ctrl->mype, graph->nnbrs); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: MPI_Barrier(ctrl->comm); +debug.c: for (penum=0; penum<ctrl->npes; penum++) { +debug.c: if (ctrl->mype == penum) { +debug.c: printf("PE: %d, nnbrs: %d", ctrl->mype, nnbrs); +debug.c: MPI_Barrier(ctrl->comm); +diffutil.c: nparts = ctrl->nparts; +diffutil.c: myhome = (ctrl->ps_relation == COUPLED) ? ctrl->mype : graph->home[i]; +diffutil.c: /* PrintVector(ctrl, ctrl->npes, 0, lend, "Lend: "); */ +diffutil.c: MPI_Allreduce((void *)lstart, (void *)gstart, nparts, IDX_DATATYPE, MPI_SUM, ctrl->comm); +diffutil.c: MPI_Allreduce((void *)lleft, (void *)gleft, nparts, IDX_DATATYPE, MPI_SUM, ctrl->comm); +diffutil.c: MPI_Allreduce((void *)lend, (void *)gend, nparts, IDX_DATATYPE, MPI_SUM, ctrl->comm); +grsetup.c: graph->gnvtxs = vtxdist[ctrl->npes]; +grsetup.c: graph->nvtxs = vtxdist[ctrl->mype+1]-vtxdist[ctrl->mype]; +grsetup.c: ctrl->tvwgts[j] = GlobalSESum(ctrl, ltvwgts[j]); +grsetup.c: if (ctrl->tvwgts[i] == 0) { +grsetup.c: graph->nvwgt[i*ncon+j] = (float)(graph->vwgt[i*ncon+j]) / (float)(ctrl->tvwgts[j]); +grsetup.c: srand(ctrl->seed); +grsetup.c: MPI_Comm_dup(comm, &(ctrl->gcomm)); +grsetup.c: MPI_Comm_rank(ctrl->gcomm, &ctrl->mype); +grsetup.c: MPI_Comm_size(ctrl->gcomm, &ctrl->npes); +grsetup.c: ctrl->dbglvl = dbglvl; +grsetup.c: ctrl->nparts = nparts; /* Set the # of partitions is de-coupled from the # of domains */ +grsetup.c: ctrl->comm = ctrl->gcomm; +grsetup.c: ctrl->xyztype = XYZ_SPFILL; +grsetup.c: srand(ctrl->mype); +grsetup.c: lpvtxs = idxsmalloc(ctrl->nparts, 0, "ComputeMoveStatistics: lpvtxs"); +grsetup.c: gpvtxs = idxsmalloc(ctrl->nparts, 0, "ComputeMoveStatistics: gpvtxs"); +grsetup.c: if (where[i] != ctrl->mype) +grsetup.c: /* PrintVector(ctrl, ctrl->npes, 0, lpvtxs, "Lpvtxs: "); */ +grsetup.c: MPI_Allreduce((void *)lpvtxs, (void *)gpvtxs, ctrl->nparts, IDX_DATATYPE, MPI_SUM, ctrl->comm); +grsetup.c: *maxin = GlobalSEMax(ctrl, gpvtxs[ctrl->mype]-(nvtxs-j)); +initbalance.c: IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr)); +initbalance.c: mytpwgts = fsmalloc(ctrl->nparts, 0.0, "mytpwgts"); +initbalance.c: for (i=0; i<ctrl->nparts; i++) +initbalance.c: mytpwgts[i] += ctrl->tpwgts[i*ncon+j]; +initbalance.c: for (i=0; i<ctrl->nparts; i++) +initbalance.c: if (ctrl->ps_relation == DISCOUPLED) { +initbalance.c: rcounts = imalloc(ctrl->npes, "rcounts"); +initbalance.c: rdispls = imalloc(ctrl->npes+1, "rdispls"); +initbalance.c: for (i=0; i<ctrl->npes; i++) { +initbalance.c: MAKECSR(i, ctrl->npes, rdispls); +initbalance.c: (void *)part, rcounts, rdispls, IDX_DATATYPE, ctrl->comm); +initbalance.c: for (i=0; i<ctrl->npes; i++) +initbalance.c: if (part[i] >= ctrl->nparts) +initbalance.c: part[i] = home[i] = part[i] % ctrl->nparts; +initbalance.c: part[i] = home[i] = (-1*part[i]) % ctrl->nparts; +initbalance.c: IFSET(ctrl->dbglvl, DBG_REFINEINFO, Moc_ComputeSerialBalance(ctrl, agraph, agraph->where, lbvec)); +initbalance.c: IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "input cut: %d, balance: ", ComputeSerialEdgeCut(agraph))); +initbalance.c: IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "%.3f ", lbvec[i])); +initbalance.c: IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "\n")); +initbalance.c: sr = (ctrl->mype % 2 == 0) ? 1 : 0; +initbalance.c: gd = (ctrl->mype % 2 == 1) ? 1 : 0; +initbalance.c: if (graph->ncon > MAX_NCON_FOR_DIFFUSION || ctrl->npes == 1) { +initbalance.c: MPI_Comm_split(ctrl->gcomm, sr, 0, &ipcomm); +initbalance.c: myctrl.sync = ctrl->sync; +initbalance.c: myctrl.seed = ctrl->seed; +initbalance.c: myctrl.nparts = ctrl->nparts; +initbalance.c: myctrl.ipc_factor = ctrl->ipc_factor; +initbalance.c: myctrl.redist_factor = ctrl->redist_base; +initbalance.c: myctrl.tpwgts = ctrl->tpwgts; +initbalance.c: icopy(ncon, ctrl->tvwgts, myctrl.tvwgts); +initbalance.c: icopy(ncon, ctrl->ubvec, myctrl.ubvec); +initbalance.c: moptions[7] = ctrl->sync + (mype % ngroups) + 1; +initbalance.c: lnparts = ctrl->nparts; +initbalance.c: lpecost.rank = ctrl->mype; +initbalance.c: if (ctrl->mype == gpecost.rank && ctrl->mype != sr_pe) { +initbalance.c: MPI_Send((void *)part, nvtxs, IDX_DATATYPE, sr_pe, 1, ctrl->comm); +initbalance.c: if (ctrl->mype != gpecost.rank && ctrl->mype == sr_pe) { +initbalance.c: MPI_Recv((void *)part, nvtxs, IDX_DATATYPE, gpecost.rank, 1, ctrl->comm, &status); +initbalance.c: if (ctrl->mype == sr_pe) { +initbalance.c: SerialRemap(&cgraph, ctrl->nparts, home, lwhere, part, ctrl->tpwgts); +initbalance.c: lpecost.rank = ctrl->mype; +initbalance.c: if (ctrl->mype == gpecost.rank && ctrl->mype != gd_pe) +initbalance.c: MPI_Send((void *)part, nvtxs, IDX_DATATYPE, gd_pe, 1, ctrl->comm); +initbalance.c: if (ctrl->mype != gpecost.rank && ctrl->mype == gd_pe) +initbalance.c: MPI_Recv((void *)part, nvtxs, IDX_DATATYPE, gpecost.rank, 1, ctrl->comm, &status); +initbalance.c: if (ctrl->mype == gd_pe) { +initbalance.c: SerialRemap(&cgraph, ctrl->nparts, home, lwhere, part, ctrl->tpwgts); +initbalance.c: if (ctrl->mype == sr_pe || ctrl->mype == gd_pe) { +initbalance.c: my_cost = ctrl->ipc_factor * my_cut + REDIST_WGT * ctrl->redist_base * my_totalv; +initbalance.c: IFSET(ctrl->dbglvl, DBG_REFINEINFO, printf("%s initial cut: %.1f, totalv: %.1f, balance: %.3f\n", +initbalance.c: (ctrl->mype == sr_pe ? "scratch-remap" : "diffusion"), my_cut, my_totalv, my_balance)); +initbalance.c: if (ctrl->mype == gd_pe) { +initbalance.c: MPI_Send((void *)buffer, 2, MPI_FLOAT, sr_pe, 1, ctrl->comm); +initbalance.c: MPI_Recv((void *)buffer, 2, MPI_FLOAT, gd_pe, 1, ctrl->comm, &status); +initbalance.c: if (ctrl->mype == sr_pe) { +initbalance.c: MPI_Bcast((void *)&who_wins, 1, MPI_INT, sr_pe, ctrl->comm); +initbalance.c: MPI_Bcast((void *)part, nvtxs, IDX_DATATYPE, who_wins, ctrl->comm); +initbalance.c: idxcopy(graph->nvtxs, part+vtxdist[ctrl->mype], graph->where); +initbalance.c: IFSET(ctrl->dbglvl, DBG_TIME, stoptim
\ No newline at end of file |