/************************************************************************* * This function resets the buckets **************************************************************************/ void PQueueReset(PQueueType *queue) { int i, j; queue->nnodes = 0; if (queue->type == 1) { queue->maxgain = -queue->ngainspan; j = queue->ngainspan+queue->pgainspan+1; queue->buckets -= queue->ngainspan; for (i=0; i<j; i++) queue->buckets[i] = NULL; queue->buckets += queue->ngainspan; } else { idxset(queue->maxnodes, -1, queue->locator); } }
/************************************************************************* * This function takes a graph and produces a bisection by using a region * growing algorithm. The resulting partition is returned in * graph->where **************************************************************************/ void MocGrowBisection(CtrlType *ctrl, GraphType *graph, float *tpwgts, float ubfactor) { int i, j, k, nvtxs, ncon, from, bestcut, mincut, nbfs; idxtype *bestwhere, *where; nvtxs = graph->nvtxs; MocAllocate2WayPartitionMemory(ctrl, graph); where = graph->where; bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere"); nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS); bestcut = idxsum(graph->nedges, graph->adjwgt); for (; nbfs>0; nbfs--) { idxset(nvtxs, 1, where); where[RandomInRange(nvtxs)] = 0; MocCompute2WayPartitionParams(ctrl, graph); MocInit2WayBalance(ctrl, graph, tpwgts); MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 4); MocBalance2Way(ctrl, graph, tpwgts, 1.02); MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 4); if (bestcut >= graph->mincut) { bestcut = graph->mincut; idxcopy(nvtxs, where, bestwhere); if (bestcut == 0) break; } } graph->mincut = bestcut; idxcopy(nvtxs, bestwhere, where); /*GKfree(&bestwhere, LTERM);*/ GKfree1((void**)&bestwhere); }
/************************************************************************* * This function computes the boundary definition for balancing **************************************************************************/ void ComputeKWayBalanceBoundary(CtrlType *ctrl, GraphType *graph, int nparts) { int i, nvtxs, nbnd; idxtype *bndind, *bndptr; nvtxs = graph->nvtxs; bndind = graph->bndind; bndptr = idxset(nvtxs, -1, graph->bndptr); /*------------------------------------------------------------ / Compute the new boundary /------------------------------------------------------------*/ nbnd = 0; for (i=0; i<nvtxs; i++) { if (graph->rinfo[i].ed > 0) BNDInsert(nbnd, bndind, bndptr, i); } graph->nbnd = nbnd; }
/************************************************************************* * This function takes a graph and produces a bisection by using a region * growing algorithm. The resulting partition is returned in * graph->where **************************************************************************/ void MocGrowBisection2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec) { int /*i, j, k,*/ nvtxs, /*ncon, from,*/ bestcut, /*mincut,*/ nbfs; idxtype *bestwhere, *where; nvtxs = graph->nvtxs; MocAllocate2WayPartitionMemory(ctrl, graph); where = graph->where; bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere"); nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS); bestcut = idxsum(graph->nedges, graph->adjwgt); for (; nbfs>0; nbfs--) { idxset(nvtxs, 1, where); where[RandomInRange(nvtxs)] = 0; MocCompute2WayPartitionParams(ctrl, graph); MocBalance2Way2(ctrl, graph, tpwgts, ubvec); MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4); MocBalance2Way2(ctrl, graph, tpwgts, ubvec); MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4); if (bestcut > graph->mincut) { bestcut = graph->mincut; idxcopy(nvtxs, where, bestwhere); if (bestcut == 0) break; } } graph->mincut = bestcut; idxcopy(nvtxs, bestwhere, where); GKfree((void**)&bestwhere, LTERM); }
/************************************************************************* * This function computes the subdomain graph **************************************************************************/ void ComputeSubDomainGraph(GraphType *graph, int nparts, idxtype *pmat, idxtype *ndoms) { int i, j, k, me, nvtxs, ndegrees; idxtype *xadj, *adjncy, *adjwgt, *where; RInfoType *rinfo; EDegreeType *edegrees; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; rinfo = graph->rinfo; idxset(nparts*nparts, 0, pmat); for (i=0; i<nvtxs; i++) { if (rinfo[i].ed > 0) { me = where[i]; ndegrees = rinfo[i].ndegrees; edegrees = rinfo[i].edegrees; k = me*nparts; for (j=0; j<ndegrees; j++) pmat[k+edegrees[j].pid] += edegrees[j].ed; } } for (i=0; i<nparts; i++) { ndoms[i] = 0; for (j=0; j<nparts; j++) { if (pmat[i*nparts+j] > 0) ndoms[i]++; } } }
/************************************************************************* * This function takes a graph and produces a bisection by using a region * growing algorithm. The resulting partition is returned in * graph->where **************************************************************************/ void MocGrowBisectionNew2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec) { idxtype i, j, k, nvtxs, ncon, from, bestcut, mincut, nbfs, inbfs; idxtype *bestwhere, *where; nvtxs = graph->nvtxs; MocAllocate2WayPartitionMemory(ctrl, graph); where = graph->where; bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere"); nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS); for (inbfs=0; inbfs<nbfs; inbfs++) { idxset(nvtxs, 1, where); where[RandomInRange(nvtxs)] = 0; MocCompute2WayPartitionParams(ctrl, graph); MocInit2WayBalance2(ctrl, graph, tpwgts, ubvec); MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4); if (inbfs == 0 || bestcut > graph->mincut) { bestcut = graph->mincut; idxcopy(nvtxs, where, bestwhere); if (bestcut == 0) break; } } graph->mincut = bestcut; idxcopy(nvtxs, bestwhere, where); gk_free((void **)&bestwhere, LTERM); }
/************************************************************************* * This function finds all the connected components induced by the * partitioning vector in wgraph->where and tries to push them around to * remove some of them **************************************************************************/ void EliminateComponents(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor) { int i, ii, j, jj, k, me, nvtxs, tvwgt, first, last, nleft, ncmps, cwgt, other, target, deltawgt; idxtype *xadj, *adjncy, *vwgt, *adjwgt, *where, *pwgts, *maxpwgt; idxtype *cpvec, *touched, *perm, *todo, *cind, *cptr, *npcmps; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; vwgt = graph->vwgt; adjwgt = graph->adjwgt; where = graph->where; pwgts = graph->pwgts; touched = idxset(nvtxs, 0, idxwspacemalloc(ctrl, nvtxs)); cptr = idxwspacemalloc(ctrl, nvtxs); cind = idxwspacemalloc(ctrl, nvtxs); perm = idxwspacemalloc(ctrl, nvtxs); todo = idxwspacemalloc(ctrl, nvtxs); maxpwgt = idxwspacemalloc(ctrl, nparts); cpvec = idxwspacemalloc(ctrl, nparts); npcmps = idxset(nparts, 0, idxwspacemalloc(ctrl, nparts)); for (i=0; i<nvtxs; i++) perm[i] = todo[i] = i; /* Find the connected componends induced by the partition */ ncmps = -1; first = last = 0; nleft = nvtxs; while (nleft > 0) { if (first == last) { /* Find another starting vertex */ cptr[++ncmps] = first; ASSERT(touched[todo[0]] == 0); i = todo[0]; cind[last++] = i; touched[i] = 1; me = where[i]; npcmps[me]++; } i = cind[first++]; k = perm[i]; j = todo[k] = todo[--nleft]; perm[j] = k; for (j=xadj[i]; j<xadj[i+1]; j++) { k = adjncy[j]; if (where[k] == me && !touched[k]) { cind[last++] = k; touched[k] = 1; } } } cptr[++ncmps] = first; /* printf("I found %d components, for this %d-way partition\n", ncmps, nparts); */ if (ncmps > nparts) { /* There are more components than processors */ /* First determine the max allowed load imbalance */ tvwgt = idxsum(nparts, pwgts); for (i=0; i<nparts; i++) maxpwgt[i] = ubfactor*tpwgts[i]*tvwgt; deltawgt = 5; for (i=0; i<ncmps; i++) { me = where[cind[cptr[i]]]; /* Get the domain of this component */ if (npcmps[me] == 1) continue; /* Skip it because it is contigous */ /*printf("Trying to move %d from %d\n", i, me); */ /* Determine the weight of the block to be moved and abort if too high */ for (cwgt=0, j=cptr[i]; j<cptr[i+1]; j++) cwgt += vwgt[cind[j]]; if (cwgt > .30*pwgts[me]) continue; /* Skip the component if it is over 30% of the weight */ /* Determine the connectivity */ idxset(nparts, 0, cpvec); for (j=cptr[i]; j<cptr[i+1]; j++) { ii = cind[j]; for (jj=xadj[ii]; jj<xadj[ii+1]; jj++) cpvec[where[adjncy[jj]]] += adjwgt[jj]; } cpvec[me] = 0; target = -1; for (j=0; j<nparts; j++) { if (cpvec[j] > 0 && (cwgt < deltawgt || pwgts[j] + cwgt < maxpwgt[j])) { if (target == -1 || cpvec[target] < cpvec[j]) target = j; } } /* printf("\tMoving it to %d [%d]\n", target, cpvec[target]);*/ if (target != -1) { /* Assign all the vertices of 'me' to 'target' and update data structures */ INC_DEC(pwgts[target], pwgts[me], cwgt); npcmps[me]--; MoveGroup(ctrl, graph, nparts, target, i, cptr, cind); } } } idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function creates the coarser graph **************************************************************************/ void CreateCoarseGraphNoMask(CtrlType *ctrl, GraphType *graph, int cnvtxs, idxtype *match, idxtype *perm) { int i, j, k, m, istart, iend, nvtxs, nedges, ncon, cnedges, v, u, dovsize; idxtype *xadj, *vwgt, *vsize, *adjncy, *adjwgt, *adjwgtsum, *auxadj; idxtype *cmap, *htable; idxtype *cxadj, *cvwgt, *cvsize, *cadjncy, *cadjwgt, *cadjwgtsum; float *nvwgt, *cnvwgt; GraphType *cgraph; dovsize = (ctrl->optype == OP_KVMETIS ? 1 : 0); IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr)); nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; vwgt = graph->vwgt; vsize = graph->vsize; nvwgt = graph->nvwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; adjwgtsum = graph->adjwgtsum; cmap = graph->cmap; /* Initialize the coarser graph */ cgraph = SetUpCoarseGraph(graph, cnvtxs, dovsize); cxadj = cgraph->xadj; cvwgt = cgraph->vwgt; cvsize = cgraph->vsize; cnvwgt = cgraph->nvwgt; cadjwgtsum = cgraph->adjwgtsum; cadjncy = cgraph->adjncy; cadjwgt = cgraph->adjwgt; htable = idxset(cnvtxs, -1, idxwspacemalloc(ctrl, cnvtxs)); iend = xadj[nvtxs]; auxadj = ctrl->wspace.auxcore; memcpy(auxadj, adjncy, iend*sizeof(idxtype)); for (i=0; i<iend; i++) auxadj[i] = cmap[auxadj[i]]; cxadj[0] = cnvtxs = cnedges = 0; for (i=0; i<nvtxs; i++) { v = perm[i]; if (cmap[v] != cnvtxs) continue; u = match[v]; if (ncon == 1) cvwgt[cnvtxs] = vwgt[v]; else scopy(ncon, nvwgt+v*ncon, cnvwgt+cnvtxs*ncon); if (dovsize) cvsize[cnvtxs] = vsize[v]; cadjwgtsum[cnvtxs] = adjwgtsum[v]; nedges = 0; istart = xadj[v]; iend = xadj[v+1]; for (j=istart; j<iend; j++) { k = auxadj[j]; if ((m = htable[k]) == -1) { cadjncy[nedges] = k; cadjwgt[nedges] = adjwgt[j]; htable[k] = nedges++; } else { cadjwgt[m] += adjwgt[j]; } } if (v != u) { if (ncon == 1) cvwgt[cnvtxs] += vwgt[u]; else saxpy(ncon, 1.0, nvwgt+u*ncon, 1, cnvwgt+cnvtxs*ncon, 1); if (dovsize) cvsize[cnvtxs] += vsize[u]; cadjwgtsum[cnvtxs] += adjwgtsum[u]; istart = xadj[u]; iend = xadj[u+1]; for (j=istart; j<iend; j++) { k = auxadj[j]; if ((m = htable[k]) == -1) { cadjncy[nedges] = k; cadjwgt[nedges] = adjwgt[j]; htable[k] = nedges++; } else { cadjwgt[m] += adjwgt[j]; } } /* Remove the contracted adjacency weight */ if ((j = htable[cnvtxs]) != -1) { ASSERT(cadjncy[j] == cnvtxs); cadjwgtsum[cnvtxs] -= cadjwgt[j]; cadjncy[j] = cadjncy[--nedges]; cadjwgt[j] = cadjwgt[nedges]; htable[cnvtxs] = -1; } } ASSERTP(cadjwgtsum[cnvtxs] == idxsum(nedges, cadjwgt), ("%d %d\n", cadjwgtsum[cnvtxs], idxsum(nedges, cadjwgt))); for (j=0; j<nedges; j++) htable[cadjncy[j]] = -1; /* Zero out the htable */ cnedges += nedges; cxadj[++cnvtxs] = cnedges; cadjncy += nedges; cadjwgt += nedges; } cgraph->nedges = cnedges; ReAdjustMemory(graph, cgraph, dovsize); IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr)); idxwspacefree(ctrl, cnvtxs); }
/************************************************************************* * This function compresses a graph by merging identical vertices * The compression should lead to at least 10% reduction. **************************************************************************/ void CompressGraph(CtrlType *ctrl, GraphType *graph, int nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *cptr, idxtype *cind) { int i, ii, iii, j, jj, k, l, cnvtxs, cnedges; idxtype *cxadj, *cadjncy, *cvwgt, *mark, *map; KeyValueType *keys; mark = idxsmalloc(nvtxs, -1, "CompressGraph: mark"); map = idxsmalloc(nvtxs, -1, "CompressGraph: map"); keys = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "CompressGraph: keys"); /* Compute a key for each adjacency list */ for (i=0; i<nvtxs; i++) { k = 0; for (j=xadj[i]; j<xadj[i+1]; j++) k += adjncy[j]; keys[i].key = k+i; /* Add the diagonal entry as well */ keys[i].val = i; } ikeysort(nvtxs, keys); l = cptr[0] = 0; for (cnvtxs=i=0; i<nvtxs; i++) { ii = keys[i].val; if (map[ii] == -1) { mark[ii] = i; /* Add the diagonal entry */ for (j=xadj[ii]; j<xadj[ii+1]; j++) mark[adjncy[j]] = i; cind[l++] = ii; map[ii] = cnvtxs; for (j=i+1; j<nvtxs; j++) { iii = keys[j].val; if (keys[i].key != keys[j].key || xadj[ii+1]-xadj[ii] != xadj[iii+1]-xadj[iii]) break; /* Break if keys or degrees are different */ if (map[iii] == -1) { /* Do a comparison if iii has not been mapped */ for (jj=xadj[iii]; jj<xadj[iii+1]; jj++) { if (mark[adjncy[jj]] != i) break; } if (jj == xadj[iii+1]) { /* Identical adjacency structure */ map[iii] = cnvtxs; cind[l++] = iii; } } } cptr[++cnvtxs] = l; } } /* printf("Original: %6d, Compressed: %6d\n", nvtxs, cnvtxs); */ InitGraph(graph); if (cnvtxs >= COMPRESSION_FRACTION*nvtxs) { graph->nvtxs = nvtxs; graph->nedges = xadj[nvtxs]; graph->ncon = 1; graph->xadj = xadj; graph->adjncy = adjncy; graph->gdata = idxmalloc(3*nvtxs+graph->nedges, "CompressGraph: gdata"); graph->vwgt = graph->gdata; graph->adjwgtsum = graph->gdata+nvtxs; graph->cmap = graph->gdata+2*nvtxs; graph->adjwgt = graph->gdata+3*nvtxs; idxset(nvtxs, 1, graph->vwgt); idxset(graph->nedges, 1, graph->adjwgt); for (i=0; i<nvtxs; i++) graph->adjwgtsum[i] = xadj[i+1]-xadj[i]; graph->label = idxmalloc(nvtxs, "CompressGraph: label"); for (i=0; i<nvtxs; i++) graph->label[i] = i; } else { /* Ok, form the compressed graph */ cnedges = 0; for (i=0; i<cnvtxs; i++) { ii = cind[cptr[i]]; cnedges += xadj[ii+1]-xadj[ii]; } /* Allocate memory for the compressed graph*/ graph->gdata = idxmalloc(4*cnvtxs+1 + 2*cnedges, "CompressGraph: gdata"); cxadj = graph->xadj = graph->gdata; cvwgt = graph->vwgt = graph->gdata + cnvtxs+1; graph->adjwgtsum = graph->gdata + 2*cnvtxs+1; graph->cmap = graph->gdata + 3*cnvtxs+1; cadjncy = graph->adjncy = graph->gdata + 4*cnvtxs+1; graph->adjwgt = graph->gdata + 4*cnvtxs+1 + cnedges; /* Now go and compress the graph */ idxset(nvtxs, -1, mark); l = cxadj[0] = 0; for (i=0; i<cnvtxs; i++) { cvwgt[i] = cptr[i+1]-cptr[i]; mark[i] = i; /* Remove any dioganal entries in the compressed graph */ for (j=cptr[i]; j<cptr[i+1]; j++) { ii = cind[j]; for (jj=xadj[ii]; jj<xadj[ii+1]; jj++) { k = map[adjncy[jj]]; if (mark[k] != i) cadjncy[l++] = k; mark[k] = i; } } cxadj[i+1] = l; } graph->nvtxs = cnvtxs; graph->nedges = l; graph->ncon = 1; idxset(graph->nedges, 1, graph->adjwgt); for (i=0; i<cnvtxs; i++) graph->adjwgtsum[i] = cxadj[i+1]-cxadj[i]; graph->label = idxmalloc(cnvtxs, "CompressGraph: label"); for (i=0; i<cnvtxs; i++) graph->label[i] = i; } GKfree(&keys, &map, &mark, LTERM); }
/************************************************************************* * This function performs k-way refinement **************************************************************************/ void MCGreedy_KWayEdgeBalanceHorizontal(CtrlType *ctrl, GraphType *graph, int nparts, float *ubvec, int npasses) { int i, ii, /*iii,*/ j, /*jj,*/ k, /*l,*/ pass, nvtxs, ncon, nbnd, myndegrees, oldgain, gain, nmoves; int from, me, to, oldcut; idxtype *xadj, *adjncy, *adjwgt; idxtype *where, *perm, *bndptr, *bndind, *moved; EDegreeType *myedegrees; RInfoType *myrinfo; PQueueType queue; float *npwgts, *nvwgt, *minwgt, *maxwgt, tvec[MAXNCON]; nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; bndind = graph->bndind; bndptr = graph->bndptr; where = graph->where; npwgts = graph->npwgts; /* Setup the weight intervals of the various subdomains */ minwgt = fwspacemalloc(ctrl, ncon*nparts); maxwgt = fwspacemalloc(ctrl, ncon*nparts); for (i=0; i<nparts; i++) { for (j=0; j<ncon; j++) { maxwgt[i*ncon+j] = ubvec[j]/nparts; minwgt[i*ncon+j] = 1.0/(ubvec[j]*nparts); } } perm = idxwspacemalloc(ctrl, nvtxs); moved = idxwspacemalloc(ctrl, nvtxs); PQueueInit(ctrl, &queue, nvtxs, graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]); if (ctrl->dbglvl&DBG_REFINE) { printf("Partitions: [%5.4f %5.4f], Nv-Nb[%6d %6d]. Cut: %6d, LB: ", npwgts[samin(ncon*nparts, npwgts)], npwgts[samax(ncon*nparts, npwgts)], graph->nvtxs, graph->nbnd, graph->mincut); ComputeHKWayLoadImbalance(ncon, nparts, npwgts, tvec); for (i=0; i<ncon; i++) printf("%.3f ", tvec[i]); printf("[B]\n"); } for (pass=0; pass<npasses; pass++) { ASSERT(ComputeCut(graph, where) == graph->mincut); /* Check to see if things are out of balance, given the tolerance */ if (MocIsHBalanced(ncon, nparts, npwgts, ubvec)) break; PQueueReset(&queue); idxset(nvtxs, -1, moved); oldcut = graph->mincut; nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; PQueueInsert(&queue, i, graph->rinfo[i].ed - graph->rinfo[i].id); moved[i] = 2; } nmoves = 0; for (;;) { if ((i = PQueueGetMax(&queue)) == -1) break; moved[i] = 1; myrinfo = graph->rinfo+i; from = where[i]; nvwgt = graph->nvwgt+i*ncon; if (AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, -1.0, nvwgt, minwgt+from*ncon)) continue; /* This cannot be moved! */ myedegrees = myrinfo->edegrees; myndegrees = myrinfo->ndegrees; for (k=0; k<myndegrees; k++) { to = myedegrees[k].pid; if (IsHBalanceBetterFT(ncon, nparts, npwgts+from*ncon, npwgts+to*ncon, nvwgt, ubvec)) break; } if (k == myndegrees) continue; /* break out if you did not find a candidate */ for (j=k+1; j<myndegrees; j++) { to = myedegrees[j].pid; if (IsHBalanceBetterTT(ncon, nparts, npwgts+myedegrees[k].pid*ncon, npwgts+to*ncon, nvwgt, ubvec)) k = j; } to = myedegrees[k].pid; j = 0; if (!AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, maxwgt+from*ncon)) j++; if (myedegrees[k].ed-myrinfo->id >= 0) j++; if (!AreAllHVwgtsAbove(ncon, 1.0, npwgts+to*ncon, 0.0, nvwgt, minwgt+to*ncon) && AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon)) j++; if (j == 0) continue; /* DELETE if (myedegrees[k].ed-myrinfo->id < 0 && AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, maxwgt+from*ncon) && AreAllHVwgtsAbove(ncon, 1.0, npwgts+to*ncon, 0.0, nvwgt, minwgt+to*ncon) && AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon)) continue; */ /*===================================================================== * If we got here, we can now move the vertex from 'from' to 'to' *======================================================================*/ graph->mincut -= myedegrees[k].ed-myrinfo->id; IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut)); /* Update where, weight, and ID/ED information of the vertex you moved */ saxpy(ncon, 1.0, nvwgt, 1, npwgts+to*ncon, 1); saxpy(ncon, -1.0, nvwgt, 1, npwgts+from*ncon, 1); where[i] = to; myrinfo->ed += myrinfo->id-myedegrees[k].ed; SWAP(myrinfo->id, myedegrees[k].ed, j); if (myedegrees[k].ed == 0) myedegrees[k] = myedegrees[--myrinfo->ndegrees]; else myedegrees[k].pid = from; if (myrinfo->ed == 0) BNDDelete(nbnd, bndind, bndptr, i); /* Update the degrees of adjacent vertices */ for (j=xadj[i]; j<xadj[i+1]; j++) { ii = adjncy[j]; me = where[ii]; myrinfo = graph->rinfo+ii; if (myrinfo->edegrees == NULL) { myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree; ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii]; } myedegrees = myrinfo->edegrees; ASSERT(CheckRInfo(myrinfo)); oldgain = (myrinfo->ed-myrinfo->id); if (me == from) { INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]); if (myrinfo->ed > 0 && bndptr[ii] == -1) BNDInsert(nbnd, bndind, bndptr, ii); } else if (me == to) { INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]); if (myrinfo->ed == 0 && bndptr[ii] != -1) BNDDelete(nbnd, bndind, bndptr, ii); } /* Remove contribution from the .ed of 'from' */ if (me != from) { for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == from) { if (myedegrees[k].ed == adjwgt[j]) myedegrees[k] = myedegrees[--myrinfo->ndegrees]; else myedegrees[k].ed -= adjwgt[j]; break; } } } /* Add contribution to the .ed of 'to' */ if (me != to) { for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == to) { myedegrees[k].ed += adjwgt[j]; break; } } if (k == myrinfo->ndegrees) { myedegrees[myrinfo->ndegrees].pid = to; myedegrees[myrinfo->ndegrees++].ed = adjwgt[j]; } } /* Update the queue */ if (me == to || me == from) { gain = myrinfo->ed-myrinfo->id; if (moved[ii] == 2) { if (myrinfo->ed > 0) PQueueUpdate(&queue, ii, oldgain, gain); else { PQueueDelete(&queue, ii, oldgain); moved[ii] = -1; } } else if (moved[ii] == -1 && myrinfo->ed > 0) { PQueueInsert(&queue, ii, gain); moved[ii] = 2; } } ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]); ASSERT(CheckRInfo(myrinfo)); } nmoves++; } graph->nbnd = nbnd; if (ctrl->dbglvl&DBG_REFINE) { printf("\t [%5.4f %5.4f], Nb: %6d, Nmoves: %5d, Cut: %6d, LB: ", npwgts[samin(ncon*nparts, npwgts)], npwgts[samax(ncon*nparts, npwgts)], nbnd, nmoves, graph->mincut); ComputeHKWayLoadImbalance(ncon, nparts, npwgts, tvec); for (i=0; i<ncon; i++) printf("%.3f ", tvec[i]); printf("\n"); } if (nmoves == 0) break; } PQueueFree(ctrl, &queue); fwspacefree(ctrl, ncon*nparts); fwspacefree(ctrl, ncon*nparts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function projects a partition, and at the same time computes the * parameters for refinement. **************************************************************************/ void MocProject2WayPartition(CtrlType *ctrl, GraphType *graph) { int i, j, k, nvtxs, nbnd, me; idxtype *xadj, *adjncy, *adjwgt, *adjwgtsum; idxtype *cmap, *where, *id, *ed, *bndptr, *bndind; idxtype *cwhere, *cid, *ced, *cbndptr; GraphType *cgraph; cgraph = graph->coarser; cwhere = cgraph->where; cid = cgraph->id; ced = cgraph->ed; cbndptr = cgraph->bndptr; nvtxs = graph->nvtxs; cmap = graph->cmap; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; adjwgtsum = graph->adjwgtsum; MocAllocate2WayPartitionMemory(ctrl, graph); where = graph->where; id = idxset(nvtxs, 0, graph->id); ed = idxset(nvtxs, 0, graph->ed); bndptr = idxset(nvtxs, -1, graph->bndptr); bndind = graph->bndind; /* Go through and project partition and compute id/ed for the nodes */ for (i=0; i<nvtxs; i++) { k = cmap[i]; where[i] = cwhere[k]; cmap[i] = cbndptr[k]; } for (nbnd=0, i=0; i<nvtxs; i++) { me = where[i]; id[i] = adjwgtsum[i]; if (xadj[i] == xadj[i+1]) { bndptr[i] = nbnd; bndind[nbnd++] = i; } else { if (cmap[i] != -1) { /* If it is an interface node. Note that cmap[i] = cbndptr[cmap[i]] */ for (j=xadj[i]; j<xadj[i+1]; j++) { if (me != where[adjncy[j]]) ed[i] += adjwgt[j]; } id[i] -= ed[i]; if (ed[i] > 0 || xadj[i] == xadj[i+1]) { bndptr[i] = nbnd; bndind[nbnd++] = i; } } } } graph->mincut = cgraph->mincut; graph->nbnd = nbnd; scopy(2*graph->ncon, cgraph->npwgts, graph->npwgts); FreeGraph(graph->coarser); graph->coarser = NULL; }
/************************************************************************* * This function performs k-way refinement **************************************************************************/ void Greedy_KWayEdgeRefine(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor, int npasses) { int i, ii, iii, j, jj, k, l, pass, nvtxs, nbnd, tvwgt, myndegrees, oldgain, gain; int from, me, to, oldcut, vwgt; idxtype *xadj, *adjncy, *adjwgt; idxtype *where, *pwgts, *perm, *bndptr, *bndind, *minwgt, *maxwgt, *moved, *itpwgts; EDegreeType *myedegrees; RInfoType *myrinfo; PQueueType queue; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; bndind = graph->bndind; bndptr = graph->bndptr; where = graph->where; pwgts = graph->pwgts; /* Setup the weight intervals of the various subdomains */ minwgt = idxwspacemalloc(ctrl, nparts); maxwgt = idxwspacemalloc(ctrl, nparts); itpwgts = idxwspacemalloc(ctrl, nparts); tvwgt = idxsum(nparts, pwgts); ASSERT(tvwgt == idxsum(nvtxs, graph->vwgt)); for (i=0; i<nparts; i++) { itpwgts[i] = tpwgts[i]*tvwgt; maxwgt[i] = tpwgts[i]*tvwgt*ubfactor; minwgt[i] = tpwgts[i]*tvwgt*(1.0/ubfactor); } perm = idxwspacemalloc(ctrl, nvtxs); moved = idxwspacemalloc(ctrl, nvtxs); PQueueInit(ctrl, &queue, nvtxs, graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]); IFSET(ctrl->dbglvl, DBG_REFINE, printf("Partitions: [%6d %6d]-[%6d %6d], Balance: %5.3f, Nv-Nb[%6d %6d]. Cut: %6d\n", pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], minwgt[0], maxwgt[0], 1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nvtxs, graph->nbnd, graph->mincut)); for (pass=0; pass<npasses; pass++) { ASSERT(ComputeCut(graph, where) == graph->mincut); PQueueReset(&queue); idxset(nvtxs, -1, moved); oldcut = graph->mincut; nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; PQueueInsert(&queue, i, graph->rinfo[i].ed - graph->rinfo[i].id); moved[i] = 2; } for (iii=0;; iii++) { if ((i = PQueueGetMax(&queue)) == -1) break; moved[i] = 1; myrinfo = graph->rinfo+i; from = where[i]; vwgt = graph->vwgt[i]; if (pwgts[from]-vwgt < minwgt[from]) continue; /* This cannot be moved! */ myedegrees = myrinfo->edegrees; myndegrees = myrinfo->ndegrees; j = myrinfo->id; for (k=0; k<myndegrees; k++) { to = myedegrees[k].pid; gain = myedegrees[k].ed-j; /* j = myrinfo->id. Allow good nodes to move */ if (pwgts[to]+vwgt <= maxwgt[to]+gain && gain >= 0) break; } if (k == myndegrees) continue; /* break out if you did not find a candidate */ for (j=k+1; j<myndegrees; j++) { to = myedegrees[j].pid; if ((myedegrees[j].ed > myedegrees[k].ed && pwgts[to]+vwgt <= maxwgt[to]) || (myedegrees[j].ed == myedegrees[k].ed && itpwgts[myedegrees[k].pid]*pwgts[to] < itpwgts[to]*pwgts[myedegrees[k].pid])) k = j; } to = myedegrees[k].pid; j = 0; if (myedegrees[k].ed-myrinfo->id > 0) j = 1; else if (myedegrees[k].ed-myrinfo->id == 0) { if ((iii&7) == 0 || pwgts[from] >= maxwgt[from] || itpwgts[from]*(pwgts[to]+vwgt) < itpwgts[to]*pwgts[from]) j = 1; } if (j == 0) continue; /*===================================================================== * If we got here, we can now move the vertex from 'from' to 'to' *======================================================================*/ graph->mincut -= myedegrees[k].ed-myrinfo->id; IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut)); /* Update where, weight, and ID/ED information of the vertex you moved */ where[i] = to; INC_DEC(pwgts[to], pwgts[from], vwgt); myrinfo->ed += myrinfo->id-myedegrees[k].ed; SWAP(myrinfo->id, myedegrees[k].ed, j); if (myedegrees[k].ed == 0) myedegrees[k] = myedegrees[--myrinfo->ndegrees]; else myedegrees[k].pid = from; if (myrinfo->ed < myrinfo->id) BNDDelete(nbnd, bndind, bndptr, i); /* Update the degrees of adjacent vertices */ for (j=xadj[i]; j<xadj[i+1]; j++) { ii = adjncy[j]; me = where[ii]; myrinfo = graph->rinfo+ii; if (myrinfo->edegrees == NULL) { myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree; ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii]; } myedegrees = myrinfo->edegrees; ASSERT(CheckRInfo(myrinfo)); oldgain = (myrinfo->ed-myrinfo->id); if (me == from) { INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]); if (myrinfo->ed-myrinfo->id >= 0 && bndptr[ii] == -1) BNDInsert(nbnd, bndind, bndptr, ii); } else if (me == to) { INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]); if (myrinfo->ed-myrinfo->id < 0 && bndptr[ii] != -1) BNDDelete(nbnd, bndind, bndptr, ii); } /* Remove contribution from the .ed of 'from' */ if (me != from) { for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == from) { if (myedegrees[k].ed == adjwgt[j]) myedegrees[k] = myedegrees[--myrinfo->ndegrees]; else myedegrees[k].ed -= adjwgt[j]; break; } } } /* Add contribution to the .ed of 'to' */ if (me != to) { for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == to) { myedegrees[k].ed += adjwgt[j]; break; } } if (k == myrinfo->ndegrees) { myedegrees[myrinfo->ndegrees].pid = to; myedegrees[myrinfo->ndegrees++].ed = adjwgt[j]; } } /* Update the queue */ if (me == to || me == from) { gain = myrinfo->ed-myrinfo->id; if (moved[ii] == 2) { if (gain >= 0) PQueueUpdate(&queue, ii, oldgain, gain); else { PQueueDelete(&queue, ii, oldgain); moved[ii] = -1; } } else if (moved[ii] == -1 && gain >= 0) { PQueueInsert(&queue, ii, gain); moved[ii] = 2; } } ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]); ASSERT(CheckRInfo(myrinfo)); } } graph->nbnd = nbnd; IFSET(ctrl->dbglvl, DBG_REFINE, printf("\t[%6d %6d], Balance: %5.3f, Nb: %6d. Cut: %6d\n", pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], 1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nbnd, graph->mincut)); if (graph->mincut == oldcut) break; } PQueueFree(ctrl, &queue); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function is the entry point for the node ND code for ParMETIS **************************************************************************/ void METIS_NodeNDP(int nvtxs, idxtype *xadj, idxtype *adjncy, int npes, int *options, idxtype *perm, idxtype *iperm, idxtype *sizes) { int i, ii, j, l, wflag, nflag; GraphType graph; CtrlType ctrl; idxtype *cptr, *cind; if (options[0] == 0) { /* Use the default parameters */ ctrl.CType = ONMETIS_CTYPE; ctrl.IType = ONMETIS_ITYPE; ctrl.RType = ONMETIS_RTYPE; ctrl.dbglvl = ONMETIS_DBGLVL; ctrl.oflags = ONMETIS_OFLAGS; ctrl.pfactor = ONMETIS_PFACTOR; ctrl.nseps = ONMETIS_NSEPS; } else { ctrl.CType = options[OPTION_CTYPE]; ctrl.IType = options[OPTION_ITYPE]; ctrl.RType = options[OPTION_RTYPE]; ctrl.dbglvl = options[OPTION_DBGLVL]; ctrl.oflags = options[OPTION_OFLAGS]; ctrl.pfactor = options[OPTION_PFACTOR]; ctrl.nseps = options[OPTION_NSEPS]; } if (ctrl.nseps < 1) ctrl.nseps = 1; ctrl.optype = OP_ONMETIS; ctrl.CoarsenTo = 100; IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl)); IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr)); InitRandom(-1); if (ctrl.oflags&OFLAG_COMPRESS) { /*============================================================ * Compress the graph ==============================================================*/ cptr = idxmalloc(nvtxs+1, "ONMETIS: cptr"); cind = idxmalloc(nvtxs, "ONMETIS: cind"); CompressGraph(&ctrl, &graph, nvtxs, xadj, adjncy, cptr, cind); if (graph.nvtxs >= COMPRESSION_FRACTION*(nvtxs)) { ctrl.oflags--; /* We actually performed no compression */ GKfree((void**)&cptr, &cind, LTERM); } else if (2*graph.nvtxs < nvtxs && ctrl.nseps == 1) ctrl.nseps = 2; } else { SetUpGraph(&graph, OP_ONMETIS, nvtxs, 1, xadj, adjncy, NULL, NULL, 0); } /*============================================================= * Do the nested dissection ordering --=============================================================*/ ctrl.maxvwgt = 1.5*(idxsum(graph.nvtxs, graph.vwgt)/ctrl.CoarsenTo); AllocateWorkSpace(&ctrl, &graph, 2); idxset(2*npes-1, 0, sizes); MlevelNestedDissectionP(&ctrl, &graph, iperm, graph.nvtxs, npes, 0, sizes); FreeWorkSpace(&ctrl, &graph); if (ctrl.oflags&OFLAG_COMPRESS) { /* Uncompress the ordering */ if (graph.nvtxs < COMPRESSION_FRACTION*(nvtxs)) { /* construct perm from iperm */ for (i=0; i<graph.nvtxs; i++) perm[iperm[i]] = i; for (l=ii=0; ii<graph.nvtxs; ii++) { i = perm[ii]; for (j=cptr[i]; j<cptr[i+1]; j++) iperm[cind[j]] = l++; } } GKfree((void**)&cptr, &cind, LTERM); } for (i=0; i<nvtxs; i++) perm[iperm[i]] = i; IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr)); IFSET(ctrl.dbglvl, DBG_TIME, PrintTimers(&ctrl)); }
/************************************************************************* * This function projects a partition, and at the same time computes the * parameters for refinement. **************************************************************************/ void ProjectKWayPartition(CtrlType *ctrl, GraphType *graph, int nparts) { int i, j, k, nvtxs, nbnd, me, other, istart, iend, ndegrees; idxtype *xadj, *adjncy, *adjwgt, *adjwgtsum; idxtype *cmap, *where, *bndptr, *bndind; idxtype *cwhere; GraphType *cgraph; RInfoType *crinfo, *rinfo, *myrinfo; EDegreeType *myedegrees; idxtype *htable; cgraph = graph->coarser; cwhere = cgraph->where; crinfo = cgraph->rinfo; nvtxs = graph->nvtxs; cmap = graph->cmap; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; adjwgtsum = graph->adjwgtsum; AllocateKWayPartitionMemory(ctrl, graph, nparts); where = graph->where; rinfo = graph->rinfo; bndind = graph->bndind; bndptr = idxset(nvtxs, -1, graph->bndptr); /* Go through and project partition and compute id/ed for the nodes */ for (i=0; i<nvtxs; i++) { k = cmap[i]; where[i] = cwhere[k]; cmap[i] = crinfo[k].ed; /* For optimization */ } htable = idxset(nparts, -1, idxwspacemalloc(ctrl, nparts)); ctrl->wspace.cdegree = 0; for (nbnd=0, i=0; i<nvtxs; i++) { me = where[i]; myrinfo = rinfo+i; myrinfo->id = myrinfo->ed = myrinfo->ndegrees = 0; myrinfo->edegrees = NULL; myrinfo->id = adjwgtsum[i]; if (cmap[i] > 0) { /* If it is an interface node. Note cmap[i] = crinfo[cmap[i]].ed */ istart = xadj[i]; iend = xadj[i+1]; myedegrees = myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree; ctrl->wspace.cdegree += iend-istart; ndegrees = 0; for (j=istart; j<iend; j++) { other = where[adjncy[j]]; if (me != other) { myrinfo->ed += adjwgt[j]; if ((k = htable[other]) == -1) { htable[other] = ndegrees; myedegrees[ndegrees].pid = other; myedegrees[ndegrees++].ed = adjwgt[j]; } else { myedegrees[k].ed += adjwgt[j]; } } } myrinfo->id -= myrinfo->ed; /* Remove space for edegrees if it was interior */ if (myrinfo->ed == 0) { myrinfo->edegrees = NULL; ctrl->wspace.cdegree -= iend-istart; } else { if (myrinfo->ed-myrinfo->id >= 0) BNDInsert(nbnd, bndind, bndptr, i); myrinfo->ndegrees = ndegrees; for (j=0; j<ndegrees; j++) htable[myedegrees[j].pid] = -1; } } } idxcopy(nparts, cgraph->pwgts, graph->pwgts); graph->mincut = cgraph->mincut; graph->nbnd = nbnd; FreeGraph(graph->coarser); graph->coarser = NULL; idxwspacefree(ctrl, nparts); ASSERT(CheckBnd2(graph)); }
/************************************************************************* * This function computes the initial id/ed **************************************************************************/ void ComputeKWayPartitionParams(CtrlType *ctrl, GraphType *graph, int nparts) { int i, j, k, nvtxs, nbnd, mincut, me, other; idxtype *xadj, *vwgt, *adjncy, *adjwgt, *pwgts, *where, *bndind, *bndptr; RInfoType *rinfo, *myrinfo; EDegreeType *myedegrees; nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; pwgts = idxset(nparts, 0, graph->pwgts); bndind = graph->bndind; bndptr = idxset(nvtxs, -1, graph->bndptr); rinfo = graph->rinfo; /*------------------------------------------------------------ / Compute now the id/ed degrees /------------------------------------------------------------*/ ctrl->wspace.cdegree = 0; nbnd = mincut = 0; for (i=0; i<nvtxs; i++) { me = where[i]; pwgts[me] += vwgt[i]; myrinfo = rinfo+i; myrinfo->id = myrinfo->ed = myrinfo->ndegrees = 0; myrinfo->edegrees = NULL; for (j=xadj[i]; j<xadj[i+1]; j++) { if (me != where[adjncy[j]]) myrinfo->ed += adjwgt[j]; } myrinfo->id = graph->adjwgtsum[i] - myrinfo->ed; if (myrinfo->ed > 0) mincut += myrinfo->ed; if (myrinfo->ed-myrinfo->id >= 0) BNDInsert(nbnd, bndind, bndptr, i); /* Time to compute the particular external degrees */ if (myrinfo->ed > 0) { myedegrees = myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree; ctrl->wspace.cdegree += xadj[i+1]-xadj[i]; for (j=xadj[i]; j<xadj[i+1]; j++) { other = where[adjncy[j]]; if (me != other) { for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == other) { myedegrees[k].ed += adjwgt[j]; break; } } if (k == myrinfo->ndegrees) { myedegrees[myrinfo->ndegrees].pid = other; myedegrees[myrinfo->ndegrees++].ed = adjwgt[j]; } } } ASSERT(myrinfo->ndegrees <= xadj[i+1]-xadj[i]); } } graph->mincut = mincut/2; graph->nbnd = nbnd; }
void METIS_PartMeshDual_WV(int *ne, int *nn, idxtype *elmnts, int *etype, int *numflag, int *nparts, int *edgecut, idxtype *epart, idxtype *npart, idxtype *vwgts) { int i, j, k, me; idxtype *xadj, *adjncy, *pwgts, *nptr, *nind; int options[10], pnumflag=0, wgtflag=2; int nnbrs, nbrind[200], nbrwgt[200], maxpwgt; int esize, esizes[] = {-1, 3, 4, 8, 4}; esize = esizes[*etype]; if (*numflag == 1) ChangeMesh2CNumbering((*ne)*esize, elmnts); xadj = idxmalloc(*ne+1, "METIS_MESHPARTNODAL: xadj"); adjncy = idxmalloc(esize*(*ne), "METIS_MESHPARTNODAL: adjncy"); METIS_MeshToDual(ne, nn, elmnts, etype, &pnumflag, xadj, adjncy); options[0] = 0; METIS_PartGraphKway(ne, xadj, adjncy, vwgts, NULL, &wgtflag, &pnumflag, nparts, options, edgecut, epart); /* Construct the node-element list */ nptr = idxsmalloc(*nn+1, 0, "METIS_MESHPARTDUAL: nptr"); for (j=esize*(*ne), i=0; i<j; i++) nptr[elmnts[i]]++; MAKECSR(i, *nn, nptr); nind = idxmalloc(nptr[*nn], "METIS_MESHPARTDUAL: nind"); for (k=i=0; i<(*ne); i++) { for (j=0; j<esize; j++, k++) nind[nptr[elmnts[k]]++] = i; } for (i=(*nn); i>0; i--) nptr[i] = nptr[i-1]; nptr[0] = 0; /* OK, now compute a nodal partition based on the element partition npart */ idxset(*nn, -1, npart); pwgts = idxsmalloc(*nparts, 0, "METIS_MESHPARTDUAL: pwgts"); for (i=0; i<*nn; i++) { me = epart[nind[nptr[i]]]; for (j=nptr[i]+1; j<nptr[i+1]; j++) { if (epart[nind[j]] != me) break; } if (j == nptr[i+1]) { npart[i] = me; pwgts[me]++; } } maxpwgt = 1.03*(*nn)/(*nparts); for (i=0; i<*nn; i++) { if (npart[i] == -1) { /* Assign the boundary element */ nnbrs = 0; for (j=nptr[i]; j<nptr[i+1]; j++) { me = epart[nind[j]]; for (k=0; k<nnbrs; k++) { if (nbrind[k] == me) { nbrwgt[k]++; break; } } if (k == nnbrs) { nbrind[nnbrs] = me; nbrwgt[nnbrs++] = 1; } } /* Try to assign it first to the domain with most things in common */ j = iamax(nnbrs, nbrwgt); if (pwgts[nbrind[j]] < maxpwgt) { npart[i] = nbrind[j]; } else { /* If that fails, assign it to a light domain */ npart[i] = nbrind[0]; for (j=0; j<nnbrs; j++) { if (pwgts[nbrind[j]] < maxpwgt) { npart[i] = nbrind[j]; break; } } } pwgts[npart[i]]++; } } if (*numflag == 1) ChangeMesh2FNumbering2((*ne)*esize, elmnts, *ne, *nn, epart, npart); GKfree(&xadj, &adjncy, &pwgts, &nptr, &nind, LTERM); }
/************************************************************************* * This function partitions a finite element mesh by partitioning its nodal * graph using KMETIS and then assigning elements in a load balanced fashion. **************************************************************************/ void METIS_PartMeshNodal(int *ne, int *nn, idxtype *elmnts, int *etype, int *numflag, int *nparts, int *edgecut, idxtype *epart, idxtype *npart) { int i, j, k, me; idxtype *xadj, *adjncy, *pwgts; int options[10], pnumflag=0, wgtflag=0; int nnbrs, nbrind[200], nbrwgt[200], maxpwgt; int esize, esizes[] = {-1, 3, 4, 8, 4}; esize = esizes[*etype]; if (*numflag == 1) ChangeMesh2CNumbering((*ne)*esize, elmnts); xadj = idxmalloc(*nn+1, "METIS_MESHPARTNODAL: xadj"); adjncy = idxmalloc(20*(*nn), "METIS_MESHPARTNODAL: adjncy"); METIS_MeshToNodal(ne, nn, elmnts, etype, &pnumflag, xadj, adjncy); adjncy = realloc(adjncy, xadj[*nn]*sizeof(idxtype)); options[0] = 0; METIS_PartGraphKway(nn, xadj, adjncy, NULL, NULL, &wgtflag, &pnumflag, nparts, options, edgecut, npart); /* OK, now compute an element partition based on the nodal partition npart */ idxset(*ne, -1, epart); pwgts = idxsmalloc(*nparts, 0, "METIS_MESHPARTNODAL: pwgts"); for (i=0; i<*ne; i++) { me = npart[elmnts[i*esize]]; for (j=1; j<esize; j++) { if (npart[elmnts[i*esize+j]] != me) break; } if (j == esize) { epart[i] = me; pwgts[me]++; } } maxpwgt = 1.03*(*ne)/(*nparts); for (i=0; i<*ne; i++) { if (epart[i] == -1) { /* Assign the boundary element */ nnbrs = 0; for (j=0; j<esize; j++) { me = npart[elmnts[i*esize+j]]; for (k=0; k<nnbrs; k++) { if (nbrind[k] == me) { nbrwgt[k]++; break; } } if (k == nnbrs) { nbrind[nnbrs] = me; nbrwgt[nnbrs++] = 1; } } /* Try to assign it first to the domain with most things in common */ j = iamax(nnbrs, nbrwgt); if (pwgts[nbrind[j]] < maxpwgt) { epart[i] = nbrind[j]; } else { /* If that fails, assign it to a light domain */ for (j=0; j<nnbrs; j++) { if (pwgts[nbrind[j]] < maxpwgt) { epart[i] = nbrind[j]; break; } } if (j == nnbrs) epart[i] = nbrind[iamax(nnbrs, nbrwgt)]; } pwgts[epart[i]]++; } } if (*numflag == 1) ChangeMesh2FNumbering2((*ne)*esize, elmnts, *ne, *nn, epart, npart); GKfree(&xadj, &adjncy, &pwgts, LTERM); }
/************************************************************************* * This function takes a graph and produces a bisection by using a region * growing algorithm. The resulting partition is returned in * graph->where **************************************************************************/ void GrowBisectionNode(CtrlType *ctrl, GraphType *graph, float ubfactor) { int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], tpwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs; idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *bndind; idxtype *queue, *touched, *gain, *bestwhere; nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere"); queue = idxmalloc(nvtxs, "BisectGraph: queue"); touched = idxmalloc(nvtxs, "BisectGraph: touched"); tpwgts[0] = idxsum(nvtxs, vwgt); tpwgts[1] = tpwgts[0]/2; tpwgts[0] -= tpwgts[1]; maxpwgt[0] = ubfactor*tpwgts[0]; maxpwgt[1] = ubfactor*tpwgts[1]; minpwgt[0] = (1.0/ubfactor)*tpwgts[0]; minpwgt[1] = (1.0/ubfactor)*tpwgts[1]; /* Allocate memory for graph->rdata. Allocate sufficient memory for both edge and node */ graph->rdata = idxmalloc(5*nvtxs+3, "GrowBisectionNode: graph->rdata"); graph->pwgts = graph->rdata; graph->where = graph->rdata + 3; graph->bndptr = graph->rdata + nvtxs + 3; graph->bndind = graph->rdata + 2*nvtxs + 3; graph->nrinfo = (NRInfoType *)(graph->rdata + 3*nvtxs + 3); graph->id = graph->rdata + 3*nvtxs + 3; graph->ed = graph->rdata + 4*nvtxs + 3; where = graph->where; bndind = graph->bndind; nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS); bestcut = tpwgts[0]+tpwgts[1]; for (nbfs++; nbfs>0; nbfs--) { idxset(nvtxs, 0, touched); pwgts[1] = tpwgts[0]+tpwgts[1]; pwgts[0] = 0; idxset(nvtxs, 1, where); queue[0] = RandomInRange(nvtxs); touched[queue[0]] = 1; first = 0; last = 1; nleft = nvtxs-1; drain = 0; /* Start the BFS from queue to get a partition */ if (nbfs >= 1) { for (;;) { if (first == last) { /* Empty. Disconnected graph! */ if (nleft == 0 || drain) break; k = RandomInRange(nleft); for (i=0; i<nvtxs; i++) { if (touched[i] == 0) { if (k == 0) break; else k--; } } queue[0] = i; touched[i] = 1; first = 0; last = 1;; nleft--; } i = queue[first++]; if (pwgts[1]-vwgt[i] < minpwgt[1]) { drain = 1; continue; } where[i] = 0; INC_DEC(pwgts[0], pwgts[1], vwgt[i]); if (pwgts[1] <= maxpwgt[1]) break; drain = 0; for (j=xadj[i]; j<xadj[i+1]; j++) { k = adjncy[j]; if (touched[k] == 0) { queue[last++] = k; touched[k] = 1; nleft--; } } } } /************************************************************* * Do some partition refinement **************************************************************/ Compute2WayPartitionParams(ctrl, graph); Balance2Way(ctrl, graph, tpwgts, ubfactor); FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4); /* Construct and refine the vertex separator */ for (i=0; i<graph->nbnd; i++) where[bndind[i]] = 2; Compute2WayNodePartitionParams(ctrl, graph); FM_2WayNodeRefine(ctrl, graph, ubfactor, 6); /* printf("ISep: [%d %d %d] %d\n", graph->pwgts[0], graph->pwgts[1], graph->pwgts[2], bestcut); */ if (bestcut > graph->mincut) { bestcut = graph->mincut; idxcopy(nvtxs, where, bestwhere); } } graph->mincut = bestcut; idxcopy(nvtxs, bestwhere, where); Compute2WayNodePartitionParams(ctrl, graph); GKfree(&bestwhere, &queue, &touched, LTERM); }
/************************************************************************* * This function performs an edge-based FM refinement **************************************************************************/ void MocGeneral2WayBalance2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec) { int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, limit, tmp, cnum; idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind; idxtype *moved, *swaps, *perm, *qnum; float *nvwgt, *npwgts, origbal[MAXNCON], minbal[MAXNCON], newbal[MAXNCON]; PQueueType parts[MAXNCON][2]; int higain, oldgain, mincut, newcut, mincutorder; float *maxwgt, *minwgt, tvec[MAXNCON]; nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; nvwgt = graph->nvwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; id = graph->id; ed = graph->ed; npwgts = graph->npwgts; bndptr = graph->bndptr; bndind = graph->bndind; moved = idxwspacemalloc(ctrl, nvtxs); swaps = idxwspacemalloc(ctrl, nvtxs); perm = idxwspacemalloc(ctrl, nvtxs); qnum = idxwspacemalloc(ctrl, nvtxs); limit = amin(amax(0.01*nvtxs, 15), 100); /* Setup the weight intervals of the two subdomains */ minwgt = fwspacemalloc(ctrl, 2*ncon); maxwgt = fwspacemalloc(ctrl, 2*ncon); for (i=0; i<2; i++) { for (j=0; j<ncon; j++) { maxwgt[i*ncon+j] = tpwgts[i]*ubvec[j]; minwgt[i*ncon+j] = tpwgts[i]*(1.0/ubvec[j]); } } /* Initialize the queues */ for (i=0; i<ncon; i++) { PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1); PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1); } for (i=0; i<nvtxs; i++) qnum[i] = samax(ncon, nvwgt+i*ncon); Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, origbal); for (i=0; i<ncon; i++) minbal[i] = origbal[i]; newcut = mincut = graph->mincut; mincutorder = -1; if (ctrl->dbglvl&DBG_REFINE) { printf("Parts: ["); for (l=0; l<ncon; l++) printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]); printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: ", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut); for (i=0; i<ncon; i++) printf("%.3f ", origbal[i]); printf("[B]\n"); } idxset(nvtxs, -1, moved); ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); /* Insert all nodes in the priority queues */ nbnd = graph->nbnd; RandomPermute(nvtxs, perm, 1); for (ii=0; ii<nvtxs; ii++) { i = perm[ii]; PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]); } for (nswaps=0; nswaps<nvtxs; nswaps++) { if (AreAllBelow(ncon, minbal, ubvec)) break; SelectQueue3(ncon, npwgts, tpwgts, &from, &cnum, parts, maxwgt); to = (from+1)%2; if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1) break; saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); newcut -= (ed[higain]-id[higain]); Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, newbal); if (IsBetter2wayBalance(ncon, newbal, minbal, ubvec) || (IsBetter2wayBalance(ncon, newbal, origbal, ubvec) && newcut < mincut)) { mincut = newcut; for (i=0; i<ncon; i++) minbal[i] = newbal[i]; mincutorder = nswaps; } else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */ newcut += (ed[higain]-id[higain]); saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); break; } where[higain] = to; moved[higain] = nswaps; swaps[nswaps] = higain; if (ctrl->dbglvl&DBG_MOVEINFO) { printf("Moved %6d from %d(%d). Gain: %5d, Cut: %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut); for (i=0; i<ncon; i++) printf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]); Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec); printf(", LB: "); for (i=0; i<ncon; i++) printf("%.3f ", tvec[i]); if (mincutorder == nswaps) printf(" *\n"); else printf("\n"); } /************************************************************** * Update the id[i]/ed[i] values of the affected nodes ***************************************************************/ SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1]) BNDDelete(nbnd, bndind, bndptr, higain); if (ed[higain] > 0 && bndptr[higain] == -1) BNDInsert(nbnd, bndind, bndptr, higain); for (j=xadj[higain]; j<xadj[higain+1]; j++) { k = adjncy[j]; oldgain = ed[k]-id[k]; kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]); INC_DEC(id[k], ed[k], kwgt); /* Update the queue position */ if (moved[k] == -1) PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]); /* Update its boundary information */ if (ed[k] == 0 && bndptr[k] != -1) BNDDelete(nbnd, bndind, bndptr, k); else if (ed[k] > 0 && bndptr[k] == -1) BNDInsert(nbnd, bndind, bndptr, k); } } /**************************************************************** * Roll back computations *****************************************************************/ for (i=0; i<nswaps; i++) moved[swaps[i]] = -1; /* reset moved array */ for (nswaps--; nswaps>mincutorder; nswaps--) { higain = swaps[nswaps]; to = where[higain] = (where[higain]+1)%2; SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1]) BNDDelete(nbnd, bndind, bndptr, higain); else if (ed[higain] > 0 && bndptr[higain] == -1) BNDInsert(nbnd, bndind, bndptr, higain); saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1); for (j=xadj[higain]; j<xadj[higain+1]; j++) { k = adjncy[j]; kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]); INC_DEC(id[k], ed[k], kwgt); if (bndptr[k] != -1 && ed[k] == 0) BNDDelete(nbnd, bndind, bndptr, k); if (bndptr[k] == -1 && ed[k] > 0) BNDInsert(nbnd, bndind, bndptr, k); } } if (ctrl->dbglvl&DBG_REFINE) { printf("\tMincut: %6d at %5d, NBND: %6d, NPwgts: [", mincut, mincutorder, nbnd); for (i=0; i<ncon; i++) printf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]); printf("], LB: "); Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec); for (i=0; i<ncon; i++) printf("%.3f ", tvec[i]); printf("\n"); } graph->mincut = mincut; graph->nbnd = nbnd; for (i=0; i<ncon; i++) { PQueueFree(ctrl, &parts[i][0]); PQueueFree(ctrl, &parts[i][1]); } idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); fwspacefree(ctrl, 2*ncon); fwspacefree(ctrl, 2*ncon); }
/************************************************************************* * This function takes a graph and produces a bisection by using a region * growing algorithm. The resulting partition is returned in * graph->where **************************************************************************/ void RandomBisection(CtrlType *ctrl, GraphType *graph, int *tpwgts, float ubfactor) { int i, ii, j, k, nvtxs, pwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs; idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where; idxtype *perm, *bestwhere; nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; Allocate2WayPartitionMemory(ctrl, graph); where = graph->where; bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere"); perm = idxmalloc(nvtxs, "BisectGraph: queue"); ASSERTP(tpwgts[0]+tpwgts[1] == idxsum(nvtxs, vwgt), ("%d %d\n", tpwgts[0]+tpwgts[1], idxsum(nvtxs, vwgt))); maxpwgt[0] = ubfactor*tpwgts[0]; maxpwgt[1] = ubfactor*tpwgts[1]; minpwgt[0] = (1.0/ubfactor)*tpwgts[0]; minpwgt[1] = (1.0/ubfactor)*tpwgts[1]; nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS); bestcut = idxsum(nvtxs, graph->adjwgtsum)+1; /* The +1 is for the 0 edges case */ for (; nbfs>0; nbfs--) { RandomPermute(nvtxs, perm, 1); idxset(nvtxs, 1, where); pwgts[1] = tpwgts[0]+tpwgts[1]; pwgts[0] = 0; if (nbfs != 1) { for (ii=0; ii<nvtxs; ii++) { i = perm[ii]; if (pwgts[0]+vwgt[i] < maxpwgt[0]) { where[i] = 0; pwgts[0] += vwgt[i]; pwgts[1] -= vwgt[i]; if (pwgts[0] > minpwgt[0]) break; } } } /************************************************************* * Do some partition refinement **************************************************************/ Compute2WayPartitionParams(ctrl, graph); /* printf("IPART: %3d [%5d %5d] [%5d %5d] %5d\n", graph->nvtxs, pwgts[0], pwgts[1], graph->pwgts[0], graph->pwgts[1], graph->mincut); */ Balance2Way(ctrl, graph, tpwgts, ubfactor); /* printf("BPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut); */ FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4); /* printf("RPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut); */ if (bestcut > graph->mincut) { bestcut = graph->mincut; idxcopy(nvtxs, where, bestwhere); if (bestcut == 0) break; } } graph->mincut = bestcut; idxcopy(nvtxs, bestwhere, where); GKfree(&bestwhere, &perm, LTERM); }
/************************************************************************* * This function prunes all the vertices in a graph with degree greater * than factor*average **************************************************************************/ void PruneGraph(CtrlType *ctrl, GraphType *graph, int nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *iperm, float factor) { int i, j, k, l, nlarge, pnvtxs, pnedges; idxtype *pxadj, *padjncy; idxtype *perm; perm = idxmalloc(nvtxs, "PruneGraph: perm"); factor = factor*xadj[nvtxs]/nvtxs; pnvtxs = pnedges = nlarge = 0; for (i=0; i<nvtxs; i++) { if (xadj[i+1]-xadj[i] < factor) { perm[i] = pnvtxs; iperm[pnvtxs++] = i; pnedges += xadj[i+1]-xadj[i]; } else { perm[i] = nvtxs - ++nlarge; iperm[nvtxs-nlarge] = i; } } /* printf("Pruned %d vertices\n", nlarge); */ InitGraph(graph); if (nlarge == 0) { /* No prunning */ graph->nvtxs = nvtxs; graph->nedges = xadj[nvtxs]; graph->ncon = 1; graph->xadj = xadj; graph->adjncy = adjncy; graph->gdata = idxmalloc(3*nvtxs+graph->nedges, "CompressGraph: gdata"); graph->vwgt = graph->gdata; graph->adjwgtsum = graph->gdata+nvtxs; graph->cmap = graph->gdata+2*nvtxs; graph->adjwgt = graph->gdata+3*nvtxs; idxset(nvtxs, 1, graph->vwgt); idxset(graph->nedges, 1, graph->adjwgt); for (i=0; i<nvtxs; i++) graph->adjwgtsum[i] = xadj[i+1]-xadj[i]; graph->label = idxmalloc(nvtxs, "CompressGraph: label"); for (i=0; i<nvtxs; i++) graph->label[i] = i; } else { /* Prune the graph */ /* Allocate memory for the compressed graph*/ graph->gdata = idxmalloc(4*pnvtxs+1 + 2*pnedges, "PruneGraph: gdata"); pxadj = graph->xadj = graph->gdata; graph->vwgt = graph->gdata + pnvtxs+1; graph->adjwgtsum = graph->gdata + 2*pnvtxs+1; graph->cmap = graph->gdata + 3*pnvtxs+1; padjncy = graph->adjncy = graph->gdata + 4*pnvtxs+1; graph->adjwgt = graph->gdata + 4*pnvtxs+1 + pnedges; pxadj[0] = pnedges = l = 0; for (i=0; i<nvtxs; i++) { if (xadj[i+1]-xadj[i] < factor) { for (j=xadj[i]; j<xadj[i+1]; j++) { k = perm[adjncy[j]]; if (k < pnvtxs) padjncy[pnedges++] = k; } pxadj[++l] = pnedges; } } graph->nvtxs = pnvtxs; graph->nedges = pnedges; graph->ncon = 1; idxset(pnvtxs, 1, graph->vwgt); idxset(pnedges, 1, graph->adjwgt); for (i=0; i<pnvtxs; i++) graph->adjwgtsum[i] = pxadj[i+1]-pxadj[i]; graph->label = idxmalloc(pnvtxs, "CompressGraph: label"); for (i=0; i<pnvtxs; i++) graph->label[i] = i; } free(perm); }
/************************************************************************* * This function takes a graph and produces a bisection by using a region * growing algorithm. The resulting partition is returned in * graph->where **************************************************************************/ void GrowBisection(CtrlType *ctrl, GraphType *graph, int *tpwgts, float ubfactor) { int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs; idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where; idxtype *queue, *touched, *gain, *bestwhere; nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; Allocate2WayPartitionMemory(ctrl, graph); where = graph->where; bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere"); queue = idxmalloc(nvtxs, "BisectGraph: queue"); touched = idxmalloc(nvtxs, "BisectGraph: touched"); ASSERTP(tpwgts[0]+tpwgts[1] == idxsum(nvtxs, vwgt), ("%d %d\n", tpwgts[0]+tpwgts[1], idxsum(nvtxs, vwgt))); maxpwgt[0] = ubfactor*tpwgts[0]; maxpwgt[1] = ubfactor*tpwgts[1]; minpwgt[0] = (1.0/ubfactor)*tpwgts[0]; minpwgt[1] = (1.0/ubfactor)*tpwgts[1]; nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS); bestcut = idxsum(nvtxs, graph->adjwgtsum)+1; /* The +1 is for the 0 edges case */ for (; nbfs>0; nbfs--) { idxset(nvtxs, 0, touched); pwgts[1] = tpwgts[0]+tpwgts[1]; pwgts[0] = 0; idxset(nvtxs, 1, where); queue[0] = RandomInRange(nvtxs); touched[queue[0]] = 1; first = 0; last = 1; nleft = nvtxs-1; drain = 0; /* Start the BFS from queue to get a partition */ for (;;) { if (first == last) { /* Empty. Disconnected graph! */ if (nleft == 0 || drain) break; k = RandomInRange(nleft); for (i=0; i<nvtxs; i++) { if (touched[i] == 0) { if (k == 0) break; else k--; } } queue[0] = i; touched[i] = 1; first = 0; last = 1;; nleft--; } i = queue[first++]; if (pwgts[0] > 0 && pwgts[1]-vwgt[i] < minpwgt[1]) { drain = 1; continue; } where[i] = 0; INC_DEC(pwgts[0], pwgts[1], vwgt[i]); if (pwgts[1] <= maxpwgt[1]) break; drain = 0; for (j=xadj[i]; j<xadj[i+1]; j++) { k = adjncy[j]; if (touched[k] == 0) { queue[last++] = k; touched[k] = 1; nleft--; } } } /* Check to see if we hit any bad limiting cases */ if (pwgts[1] == 0) { i = RandomInRange(nvtxs); where[i] = 1; INC_DEC(pwgts[1], pwgts[0], vwgt[i]); } /************************************************************* * Do some partition refinement **************************************************************/ Compute2WayPartitionParams(ctrl, graph); /*printf("IPART: %3d [%5d %5d] [%5d %5d] %5d\n", graph->nvtxs, pwgts[0], pwgts[1], graph->pwgts[0], graph->pwgts[1], graph->mincut); */ Balance2Way(ctrl, graph, tpwgts, ubfactor); /*printf("BPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/ FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4); /*printf("RPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/ if (bestcut > graph->mincut) { bestcut = graph->mincut; idxcopy(nvtxs, where, bestwhere); if (bestcut == 0) break; } } graph->mincut = bestcut; idxcopy(nvtxs, bestwhere, where); GKfree(&bestwhere, &queue, &touched, LTERM); }
/************************************************************************* * This function finds a matching using the HEM heuristic **************************************************************************/ void Match_SHEM(CtrlType *ctrl, GraphType *graph) { int i, ii, j, k, nvtxs, cnvtxs, maxidx, maxwgt, avgdegree; idxtype *xadj, *vwgt, *adjncy, *adjwgt; idxtype *match, *cmap, *degrees, *perm, *tperm; IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr)); nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; cmap = graph->cmap; match = idxset(nvtxs, UNMATCHED, idxwspacemalloc(ctrl, nvtxs)); perm = idxwspacemalloc(ctrl, nvtxs); tperm = idxwspacemalloc(ctrl, nvtxs); degrees = idxwspacemalloc(ctrl, nvtxs); RandomPermute(nvtxs, tperm, 1); avgdegree = 0.7*(xadj[nvtxs]/nvtxs); for (i=0; i<nvtxs; i++) degrees[i] = (xadj[i+1]-xadj[i] > avgdegree ? avgdegree : xadj[i+1]-xadj[i]); BucketSortKeysInc(nvtxs, avgdegree, degrees, tperm, perm); cnvtxs = 0; /* Take care any islands. Islands are matched with non-islands due to coarsening */ for (ii=0; ii<nvtxs; ii++) { i = perm[ii]; if (match[i] == UNMATCHED) { /* Unmatched */ if (xadj[i] < xadj[i+1]) break; maxidx = i; for (j=nvtxs-1; j>ii; j--) { k = perm[j]; if (match[k] == UNMATCHED && xadj[k] < xadj[k+1]) { maxidx = k; break; } } cmap[i] = cmap[maxidx] = cnvtxs++; match[i] = maxidx; match[maxidx] = i; } } /* Continue with normal matching */ for (; ii<nvtxs; ii++) { i = perm[ii]; if (match[i] == UNMATCHED) { /* Unmatched */ maxidx = i; maxwgt = 0; /* Find a heavy-edge matching, subject to maxvwgt constraints */ for (j=xadj[i]; j<xadj[i+1]; j++) { if (match[adjncy[j]] == UNMATCHED && maxwgt < adjwgt[j] && vwgt[i]+vwgt[adjncy[j]] <= ctrl->maxvwgt) { maxwgt = adjwgt[j]; maxidx = adjncy[j]; } } cmap[i] = cmap[maxidx] = cnvtxs++; match[i] = maxidx; match[maxidx] = i; } } IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr)); idxwspacefree(ctrl, nvtxs); /* degrees */ idxwspacefree(ctrl, nvtxs); /* tperm */ CreateCoarseGraph(ctrl, graph, cnvtxs, match, perm); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function sets up the graph from the user input **************************************************************************/ void SetUpGraph(GraphType *graph, int OpType, int nvtxs, int ncon, idxtype *xadj, idxtype *adjncy, idxtype *vwgt, idxtype *adjwgt, int wgtflag) { int i, j, k, sum, gsize; float *nvwgt; idxtype tvwgt[MAXNCON]; if (OpType == OP_KMETIS && ncon == 1 && (wgtflag&2) == 0 && (wgtflag&1) == 0) { SetUpGraphKway(graph, nvtxs, xadj, adjncy); return; } InitGraph(graph); graph->nvtxs = nvtxs; graph->nedges = xadj[nvtxs]; graph->ncon = ncon; graph->xadj = xadj; graph->adjncy = adjncy; if (ncon == 1) { /* We are in the non mC mode */ gsize = 0; if ((wgtflag&2) == 0) gsize += nvtxs; if ((wgtflag&1) == 0) gsize += graph->nedges; gsize += 2*nvtxs; graph->gdata = idxmalloc(gsize, "SetUpGraph: gdata"); /* Create the vertex/edge weight vectors if they are not supplied */ gsize = 0; if ((wgtflag&2) == 0) { vwgt = graph->vwgt = idxset(nvtxs, 1, graph->gdata); gsize += nvtxs; } else graph->vwgt = vwgt; if ((wgtflag&1) == 0) { adjwgt = graph->adjwgt = idxset(graph->nedges, 1, graph->gdata+gsize); gsize += graph->nedges; } else graph->adjwgt = adjwgt; /* Compute the initial values of the adjwgtsum */ graph->adjwgtsum = graph->gdata + gsize; gsize += nvtxs; for (i=0; i<nvtxs; i++) { sum = 0; for (j=xadj[i]; j<xadj[i+1]; j++) sum += adjwgt[j]; graph->adjwgtsum[i] = sum; } graph->cmap = graph->gdata + gsize; gsize += nvtxs; } else { /* Set up the graph in MOC mode */ gsize = 0; if ((wgtflag&1) == 0) gsize += graph->nedges; gsize += 2*nvtxs; graph->gdata = idxmalloc(gsize, "SetUpGraph: gdata"); gsize = 0; for (i=0; i<ncon; i++) tvwgt[i] = idxsum_strd(nvtxs, vwgt+i, ncon); nvwgt = graph->nvwgt = fmalloc(ncon*nvtxs, "SetUpGraph: nvwgt"); for (i=0; i<nvtxs; i++) { for (j=0; j<ncon; j++) nvwgt[i*ncon+j] = (1.0*vwgt[i*ncon+j])/(1.0*tvwgt[j]); } /* Create the edge weight vectors if they are not supplied */ if ((wgtflag&1) == 0) { adjwgt = graph->adjwgt = idxset(graph->nedges, 1, graph->gdata+gsize); gsize += graph->nedges; } else graph->adjwgt = adjwgt; /* Compute the initial values of the adjwgtsum */ graph->adjwgtsum = graph->gdata + gsize; gsize += nvtxs; for (i=0; i<nvtxs; i++) { sum = 0; for (j=xadj[i]; j<xadj[i+1]; j++) sum += adjwgt[j]; graph->adjwgtsum[i] = sum; } graph->cmap = graph->gdata + gsize; gsize += nvtxs; } if (OpType != OP_KMETIS && OpType != OP_KVMETIS) { graph->label = idxmalloc(nvtxs, "SetUpGraph: label"); for (i=0; i<nvtxs; i++) graph->label[i] = i; } }
/************************************************************************* * This function creates the coarser graph **************************************************************************/ void CreateCoarseGraph_NVW(CtrlType *ctrl, GraphType *graph, int cnvtxs, idxtype *match, idxtype *perm) { int i, j, jj, k, kk, l, m, istart, iend, nvtxs, nedges, ncon, cnedges, v, u, mask; idxtype *xadj, *adjncy, *adjwgtsum, *auxadj; idxtype *cmap, *htable; idxtype *cxadj, *cvwgt, *cadjncy, *cadjwgt, *cadjwgtsum; float *nvwgt, *cnvwgt; GraphType *cgraph; IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr)); nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; nvwgt = graph->nvwgt; adjncy = graph->adjncy; adjwgtsum = graph->adjwgtsum; cmap = graph->cmap; /* Initialize the coarser graph */ cgraph = SetUpCoarseGraph(graph, cnvtxs, 0); cxadj = cgraph->xadj; cvwgt = cgraph->vwgt; cnvwgt = cgraph->nvwgt; cadjwgtsum = cgraph->adjwgtsum; cadjncy = cgraph->adjncy; cadjwgt = cgraph->adjwgt; iend = xadj[nvtxs]; auxadj = ctrl->wspace.auxcore; memcpy(auxadj, adjncy, iend*sizeof(idxtype)); for (i=0; i<iend; i++) auxadj[i] = cmap[auxadj[i]]; mask = HTLENGTH; htable = idxset(mask+1, -1, idxwspacemalloc(ctrl, mask+1)); cxadj[0] = cnvtxs = cnedges = 0; for (i=0; i<nvtxs; i++) { v = perm[i]; if (cmap[v] != cnvtxs) continue; u = match[v]; cvwgt[cnvtxs] = 1; cadjwgtsum[cnvtxs] = adjwgtsum[v]; nedges = 0; istart = xadj[v]; iend = xadj[v+1]; for (j=istart; j<iend; j++) { k = auxadj[j]; kk = k&mask; if ((m = htable[kk]) == -1) { cadjncy[nedges] = k; cadjwgt[nedges] = 1; htable[kk] = nedges++; } else if (cadjncy[m] == k) { cadjwgt[m]++; } else { for (jj=0; jj<nedges; jj++) { if (cadjncy[jj] == k) { cadjwgt[jj]++; break; } } if (jj == nedges) { cadjncy[nedges] = k; cadjwgt[nedges++] = 1; } } } if (v != u) { cvwgt[cnvtxs]++; cadjwgtsum[cnvtxs] += adjwgtsum[u]; istart = xadj[u]; iend = xadj[u+1]; for (j=istart; j<iend; j++) { k = auxadj[j]; kk = k&mask; if ((m = htable[kk]) == -1) { cadjncy[nedges] = k; cadjwgt[nedges] = 1; htable[kk] = nedges++; } else if (cadjncy[m] == k) { cadjwgt[m]++; } else { for (jj=0; jj<nedges; jj++) { if (cadjncy[jj] == k) { cadjwgt[jj]++; break; } } if (jj == nedges) { cadjncy[nedges] = k; cadjwgt[nedges++] = 1; } } } /* Remove the contracted adjacency weight */ jj = htable[cnvtxs&mask]; if (jj >= 0 && cadjncy[jj] != cnvtxs) { for (jj=0; jj<nedges; jj++) { if (cadjncy[jj] == cnvtxs) break; } } if (jj >= 0 && cadjncy[jj] == cnvtxs) { /* This 2nd check is needed for non-adjacent matchings */ cadjwgtsum[cnvtxs] -= cadjwgt[jj]; cadjncy[jj] = cadjncy[--nedges]; cadjwgt[jj] = cadjwgt[nedges]; } } ASSERTP(cadjwgtsum[cnvtxs] == idxsum(nedges, cadjwgt), ("%d %d %d %d %d\n", cnvtxs, cadjwgtsum[cnvtxs], idxsum(nedges, cadjwgt), adjwgtsum[u], adjwgtsum[v])); for (j=0; j<nedges; j++) htable[cadjncy[j]&mask] = -1; /* Zero out the htable */ htable[cnvtxs&mask] = -1; cnedges += nedges; cxadj[++cnvtxs] = cnedges; cadjncy += nedges; cadjwgt += nedges; } cgraph->nedges = cnedges; ReAdjustMemory(graph, cgraph, 0); IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr)); idxwspacefree(ctrl, mask+1); }
/************************************************************************* * This function sets up the graph from the user input **************************************************************************/ void VolSetUpGraph(GraphType *graph, int OpType, int nvtxs, int ncon, idxtype *xadj, idxtype *adjncy, idxtype *vwgt, idxtype *vsize, int wgtflag) { int i, j, k, sum, gsize; idxtype *adjwgt; float *nvwgt; idxtype tvwgt[MAXNCON]; InitGraph(graph); graph->nvtxs = nvtxs; graph->nedges = xadj[nvtxs]; graph->ncon = ncon; graph->xadj = xadj; graph->adjncy = adjncy; if (ncon == 1) { /* We are in the non mC mode */ gsize = graph->nedges; /* This is for the edge weights */ if ((wgtflag&2) == 0) gsize += nvtxs; /* vwgts */ if ((wgtflag&1) == 0) gsize += nvtxs; /* vsize */ gsize += 2*nvtxs; graph->gdata = idxmalloc(gsize, "SetUpGraph: gdata"); /* Create the vertex/edge weight vectors if they are not supplied */ gsize = 0; if ((wgtflag&2) == 0) { vwgt = graph->vwgt = idxset(nvtxs, 1, graph->gdata); gsize += nvtxs; } else graph->vwgt = vwgt; if ((wgtflag&1) == 0) { vsize = graph->vsize = idxset(nvtxs, 1, graph->gdata); gsize += nvtxs; } else graph->vsize = vsize; /* Allocate memory for edge weights and initialize them to the sum of the vsize */ adjwgt = graph->adjwgt = graph->gdata+gsize; gsize += graph->nedges; for (i=0; i<nvtxs; i++) { for (j=xadj[i]; j<xadj[i+1]; j++) adjwgt[j] = 1+vsize[i]+vsize[adjncy[j]]; } /* Compute the initial values of the adjwgtsum */ graph->adjwgtsum = graph->gdata + gsize; gsize += nvtxs; for (i=0; i<nvtxs; i++) { sum = 0; for (j=xadj[i]; j<xadj[i+1]; j++) sum += adjwgt[j]; graph->adjwgtsum[i] = sum; } graph->cmap = graph->gdata + gsize; gsize += nvtxs; } else { /* Set up the graph in MOC mode */ gsize = graph->nedges; if ((wgtflag&1) == 0) gsize += nvtxs; gsize += 2*nvtxs; graph->gdata = idxmalloc(gsize, "SetUpGraph: gdata"); gsize = 0; /* Create the normalized vertex weights along each constrain */ if ((wgtflag&2) == 0) vwgt = idxsmalloc(nvtxs, 1, "SetUpGraph: vwgt"); for (i=0; i<ncon; i++) tvwgt[i] = idxsum_strd(nvtxs, vwgt+i, ncon); nvwgt = graph->nvwgt = fmalloc(ncon*nvtxs, "SetUpGraph: nvwgt"); for (i=0; i<nvtxs; i++) { for (j=0; j<ncon; j++) nvwgt[i*ncon+j] = (1.0*vwgt[i*ncon+j])/(1.0*tvwgt[j]); } if ((wgtflag&2) == 0) free(vwgt); /* Create the vsize vector if it is not supplied */ if ((wgtflag&1) == 0) { vsize = graph->vsize = idxset(nvtxs, 1, graph->gdata); gsize += nvtxs; } else graph->vsize = vsize; /* Allocate memory for edge weights and initialize them to the sum of the vsize */ adjwgt = graph->adjwgt = graph->gdata+gsize; gsize += graph->nedges; for (i=0; i<nvtxs; i++) { for (j=xadj[i]; j<xadj[i+1]; j++) adjwgt[j] = 1+vsize[i]+vsize[adjncy[j]]; } /* Compute the initial values of the adjwgtsum */ graph->adjwgtsum = graph->gdata + gsize; gsize += nvtxs; for (i=0; i<nvtxs; i++) { sum = 0; for (j=xadj[i]; j<xadj[i+1]; j++) sum += adjwgt[j]; graph->adjwgtsum[i] = sum; } graph->cmap = graph->gdata + gsize; gsize += nvtxs; } if (OpType != OP_KVMETIS) { graph->label = idxmalloc(nvtxs, "SetUpGraph: label"); for (i=0; i<nvtxs; i++) graph->label[i] = i; } }
/************************************************************************* * This function performs k-way refinement **************************************************************************/ void Greedy_KWayEdgeBalanceMConn(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor, int npasses) { int i, ii, iii, j, jj, k, l, pass, nvtxs, nbnd, tvwgt, myndegrees, oldgain, gain, nmoves; int from, me, to, oldcut, vwgt, maxndoms, nadd; idxtype *xadj, *adjncy, *adjwgt; idxtype *where, *pwgts, *perm, *bndptr, *bndind, *minwgt, *maxwgt, *moved, *itpwgts; idxtype *phtable, *pmat, *pmatptr, *ndoms; EDegreeType *myedegrees; RInfoType *myrinfo; PQueueType queue; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; bndind = graph->bndind; bndptr = graph->bndptr; where = graph->where; pwgts = graph->pwgts; pmat = ctrl->wspace.pmat; phtable = idxwspacemalloc(ctrl, nparts); ndoms = idxwspacemalloc(ctrl, nparts); ComputeSubDomainGraph(graph, nparts, pmat, ndoms); /* Setup the weight intervals of the various subdomains */ minwgt = idxwspacemalloc(ctrl, nparts); maxwgt = idxwspacemalloc(ctrl, nparts); itpwgts = idxwspacemalloc(ctrl, nparts); tvwgt = idxsum(nparts, pwgts); ASSERT(tvwgt == idxsum(nvtxs, graph->vwgt)); for (i=0; i<nparts; i++) { itpwgts[i] = tpwgts[i]*tvwgt; maxwgt[i] = tpwgts[i]*tvwgt*ubfactor; minwgt[i] = tpwgts[i]*tvwgt*(1.0/ubfactor); } perm = idxwspacemalloc(ctrl, nvtxs); moved = idxwspacemalloc(ctrl, nvtxs); PQueueInit(ctrl, &queue, nvtxs, graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]); IFSET(ctrl->dbglvl, DBG_REFINE, printf("Partitions: [%6d %6d]-[%6d %6d], Balance: %5.3f, Nv-Nb[%6d %6d]. Cut: %6d [B]\n", pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], minwgt[0], maxwgt[0], 1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nvtxs, graph->nbnd, graph->mincut)); for (pass=0; pass<npasses; pass++) { ASSERT(ComputeCut(graph, where) == graph->mincut); /* Check to see if things are out of balance, given the tolerance */ for (i=0; i<nparts; i++) { if (pwgts[i] > maxwgt[i]) break; } if (i == nparts) /* Things are balanced. Return right away */ break; PQueueReset(&queue); idxset(nvtxs, -1, moved); oldcut = graph->mincut; nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; PQueueInsert(&queue, i, graph->rinfo[i].ed - graph->rinfo[i].id); moved[i] = 2; } maxndoms = ndoms[idxamax(nparts, ndoms)]; for (nmoves=0;;) { if ((i = PQueueGetMax(&queue)) == -1) break; moved[i] = 1; myrinfo = graph->rinfo+i; from = where[i]; vwgt = graph->vwgt[i]; if (pwgts[from]-vwgt < minwgt[from]) continue; /* This cannot be moved! */ myedegrees = myrinfo->edegrees; myndegrees = myrinfo->ndegrees; /* Determine the valid domains */ for (j=0; j<myndegrees; j++) { to = myedegrees[j].pid; phtable[to] = 1; pmatptr = pmat + to*nparts; for (nadd=0, k=0; k<myndegrees; k++) { if (k == j) continue; l = myedegrees[k].pid; if (pmatptr[l] == 0) { if (ndoms[l] > maxndoms-1) { phtable[to] = 0; nadd = maxndoms; break; } nadd++; } } if (ndoms[to]+nadd > maxndoms) phtable[to] = 0; } for (k=0; k<myndegrees; k++) { to = myedegrees[k].pid; if (!phtable[to]) continue; if (pwgts[to]+vwgt <= maxwgt[to] || itpwgts[from]*(pwgts[to]+vwgt) <= itpwgts[to]*pwgts[from]) break; } if (k == myndegrees) continue; /* break out if you did not find a candidate */ for (j=k+1; j<myndegrees; j++) { to = myedegrees[j].pid; if (!phtable[to]) continue; if (itpwgts[myedegrees[k].pid]*pwgts[to] < itpwgts[to]*pwgts[myedegrees[k].pid]) k = j; } to = myedegrees[k].pid; if (pwgts[from] < maxwgt[from] && pwgts[to] > minwgt[to] && myedegrees[k].ed-myrinfo->id < 0) continue; /*===================================================================== * If we got here, we can now move the vertex from 'from' to 'to' *======================================================================*/ graph->mincut -= myedegrees[k].ed-myrinfo->id; IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut)); /* Update pmat to reflect the move of 'i' */ pmat[from*nparts+to] += (myrinfo->id-myedegrees[k].ed); pmat[to*nparts+from] += (myrinfo->id-myedegrees[k].ed); if (pmat[from*nparts+to] == 0) { ndoms[from]--; if (ndoms[from]+1 == maxndoms) maxndoms = ndoms[idxamax(nparts, ndoms)]; } if (pmat[to*nparts+from] == 0) { ndoms[to]--; if (ndoms[to]+1 == maxndoms) maxndoms = ndoms[idxamax(nparts, ndoms)]; } /* Update where, weight, and ID/ED information of the vertex you moved */ where[i] = to; INC_DEC(pwgts[to], pwgts[from], vwgt); myrinfo->ed += myrinfo->id-myedegrees[k].ed; SWAP(myrinfo->id, myedegrees[k].ed, j); if (myedegrees[k].ed == 0) myedegrees[k] = myedegrees[--myrinfo->ndegrees]; else myedegrees[k].pid = from; if (myrinfo->ed == 0) BNDDelete(nbnd, bndind, bndptr, i); /* Update the degrees of adjacent vertices */ for (j=xadj[i]; j<xadj[i+1]; j++) { ii = adjncy[j]; me = where[ii]; myrinfo = graph->rinfo+ii; if (myrinfo->edegrees == NULL) { myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree; ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii]; } myedegrees = myrinfo->edegrees; ASSERT(CheckRInfo(myrinfo)); oldgain = (myrinfo->ed-myrinfo->id); if (me == from) { INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]); if (myrinfo->ed > 0 && bndptr[ii] == -1) BNDInsert(nbnd, bndind, bndptr, ii); } else if (me == to) { INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]); if (myrinfo->ed == 0 && bndptr[ii] != -1) BNDDelete(nbnd, bndind, bndptr, ii); } /* Remove contribution from the .ed of 'from' */ if (me != from) { for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == from) { if (myedegrees[k].ed == adjwgt[j]) myedegrees[k] = myedegrees[--myrinfo->ndegrees]; else myedegrees[k].ed -= adjwgt[j]; break; } } } /* Add contribution to the .ed of 'to' */ if (me != to) { for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == to) { myedegrees[k].ed += adjwgt[j]; break; } } if (k == myrinfo->ndegrees) { myedegrees[myrinfo->ndegrees].pid = to; myedegrees[myrinfo->ndegrees++].ed = adjwgt[j]; } } /* Update pmat to reflect the move of 'i' for domains other than 'from' and 'to' */ if (me != from && me != to) { pmat[me*nparts+from] -= adjwgt[j]; pmat[from*nparts+me] -= adjwgt[j]; if (pmat[me*nparts+from] == 0) { ndoms[me]--; if (ndoms[me]+1 == maxndoms) maxndoms = ndoms[idxamax(nparts, ndoms)]; } if (pmat[from*nparts+me] == 0) { ndoms[from]--; if (ndoms[from]+1 == maxndoms) maxndoms = ndoms[idxamax(nparts, ndoms)]; } if (pmat[me*nparts+to] == 0) { ndoms[me]++; if (ndoms[me] > maxndoms) { printf("You just increased the maxndoms: %d %d\n", ndoms[me], maxndoms); maxndoms = ndoms[me]; } } if (pmat[to*nparts+me] == 0) { ndoms[to]++; if (ndoms[to] > maxndoms) { printf("You just increased the maxndoms: %d %d\n", ndoms[to], maxndoms); maxndoms = ndoms[to]; } } pmat[me*nparts+to] += adjwgt[j]; pmat[to*nparts+me] += adjwgt[j]; } /* Update the queue */ if (me == to || me == from) { gain = myrinfo->ed-myrinfo->id; if (moved[ii] == 2) { if (myrinfo->ed > 0) PQueueUpdate(&queue, ii, oldgain, gain); else { PQueueDelete(&queue, ii, oldgain); moved[ii] = -1; } } else if (moved[ii] == -1 && myrinfo->ed > 0) { PQueueInsert(&queue, ii, gain); moved[ii] = 2; } } ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]); ASSERT(CheckRInfo(myrinfo)); } nmoves++; } graph->nbnd = nbnd; IFSET(ctrl->dbglvl, DBG_REFINE, printf("\t[%6d %6d], Balance: %5.3f, Nb: %6d. Nmoves: %5d, Cut: %6d, %d\n", pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], 1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nbnd, nmoves, graph->mincut,idxsum(nparts, ndoms))); } PQueueFree(ctrl, &queue); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function performs a node-based FM refinement **************************************************************************/ void FM_2WayNodeRefineEqWgt(CtrlType *ctrl, GraphType *graph, idxtype npasses) { idxtype i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind; idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr; idxtype *mptr, *mind, *moved, *swaps, *perm; PQueueType parts[2]; NRInfoType *rinfo; idxtype higain, oldgain, mincut, initcut, mincutorder; idxtype pass, to, other, limit; idxtype mindiff, newdiff; idxtype u[2], g[2]; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; vwgt = graph->vwgt; bndind = graph->bndind; bndptr = graph->bndptr; where = graph->where; pwgts = graph->pwgts; rinfo = graph->nrinfo; i = ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt); PQueueInit(ctrl, &parts[0], nvtxs, i); PQueueInit(ctrl, &parts[1], nvtxs, i); moved = idxwspacemalloc(ctrl, nvtxs); swaps = idxwspacemalloc(ctrl, nvtxs); mptr = idxwspacemalloc(ctrl, nvtxs+1); mind = idxwspacemalloc(ctrl, nvtxs); perm = idxwspacemalloc(ctrl, nvtxs); IFSET(ctrl->dbglvl, DBG_REFINE, mprintf("Partitions: [%6D %6D] Nv-Nb[%6D %6D]. ISep: %6D\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); for (pass=0; pass<npasses; pass++) { idxset(nvtxs, -1, moved); PQueueReset(&parts[0]); PQueueReset(&parts[1]); mincutorder = -1; initcut = mincut = graph->mincut; nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; ASSERT(where[i] == 2); PQueueInsert(&parts[0], i, vwgt[i]-rinfo[i].edegrees[1]); PQueueInsert(&parts[1], i, vwgt[i]-rinfo[i].edegrees[0]); } ASSERT(CheckNodeBnd(graph, nbnd)); ASSERT(CheckNodePartitionParams(graph)); limit = (ctrl->oflags&OFLAG_COMPRESS ? amin(5*nbnd, 400) : amin(2*nbnd, 300)); /****************************************************** * Get into the FM loop *******************************************************/ mptr[0] = nmind = 0; mindiff = idxtype_abs(pwgts[0]-pwgts[1]); to = (pwgts[0] < pwgts[1] ? 0 : 1); for (nswaps=0; nswaps<nvtxs; nswaps++) { to = (pwgts[0] < pwgts[1] ? 0 : 1); if (pwgts[0] == pwgts[1]) { u[0] = PQueueSeeMax(&parts[0]); u[1] = PQueueSeeMax(&parts[1]); if (u[0] != -1 && u[1] != -1) { g[0] = vwgt[u[0]]-rinfo[u[0]].edegrees[1]; g[1] = vwgt[u[1]]-rinfo[u[1]].edegrees[0]; to = (g[0] > g[1] ? 0 : (g[0] < g[1] ? 1 : pass%2)); } } other = (to+1)%2; if ((higain = PQueueGetMax(&parts[to])) == -1) break; if (moved[higain] == -1) /* Delete if it was in the separator originally */ PQueueDelete(&parts[other], higain, vwgt[higain]-rinfo[higain].edegrees[to]); ASSERT(bndptr[higain] != -1); pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]); newdiff = idxtype_abs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other])); if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) { mincut = pwgts[2]; mincutorder = nswaps; mindiff = newdiff; } else { if (nswaps - mincutorder > limit) { pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]); break; /* No further improvement, break out */ } } BNDDelete(nbnd, bndind, bndptr, higain); pwgts[to] += vwgt[higain]; where[higain] = to; moved[higain] = nswaps; swaps[nswaps] = higain; /********************************************************** * Update the degrees of the affected nodes ***********************************************************/ for (j=xadj[higain]; j<xadj[higain+1]; j++) { k = adjncy[j]; if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */ oldgain = vwgt[k]-rinfo[k].edegrees[to]; rinfo[k].edegrees[to] += vwgt[higain]; if (moved[k] == -1 || moved[k] == -(2+other)) PQueueUpdate(&parts[other], k, oldgain, oldgain-vwgt[higain]); } else if (where[k] == other) { /* This vertex is pulled into the separator */ ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k])); BNDInsert(nbnd, bndind, bndptr, k); mind[nmind++] = k; /* Keep track for rollback */ where[k] = 2; pwgts[other] -= vwgt[k]; edegrees = rinfo[k].edegrees; edegrees[0] = edegrees[1] = 0; for (jj=xadj[k]; jj<xadj[k+1]; jj++) { kk = adjncy[jj]; if (where[kk] != 2) edegrees[where[kk]] += vwgt[kk]; else { oldgain = vwgt[kk]-rinfo[kk].edegrees[other]; rinfo[kk].edegrees[other] -= vwgt[k]; if (moved[kk] == -1 || moved[kk] == -(2+to)) PQueueUpdate(&parts[to], kk, oldgain, oldgain+vwgt[k]); } } /* Insert the new vertex into the priority queue. Only one side! */ if (moved[k] == -1) { PQueueInsert(&parts[to], k, vwgt[k]-edegrees[other]); moved[k] = -(2+to); } } } mptr[nswaps+1] = nmind; IFSET(ctrl->dbglvl, DBG_MOVEINFO, mprintf("Moved %6D to %3D, Gain: %5D [%5D] [%4D %4D] \t[%5D %5D %5D]\n", higain, to, g[to], g[other], vwgt[u[to]], vwgt[u[other]], pwgts[0], pwgts[1], pwgts[2])); } /**************************************************************** * Roll back computation *****************************************************************/ for (nswaps--; nswaps>mincutorder; nswaps--) { higain = swaps[nswaps]; ASSERT(CheckNodePartitionParams(graph)); to = where[higain]; other = (to+1)%2; INC_DEC(pwgts[2], pwgts[to], vwgt[higain]); where[higain] = 2; BNDInsert(nbnd, bndind, bndptr, higain); edegrees = rinfo[higain].edegrees; edegrees[0] = edegrees[1] = 0; for (j=xadj[higain]; j<xadj[higain+1]; j++) { k = adjncy[j]; if (where[k] == 2) rinfo[k].edegrees[to] -= vwgt[higain]; else edegrees[where[k]] += vwgt[k]; } /* Push nodes out of the separator */ for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) { k = mind[j]; ASSERT(where[k] == 2); where[k] = other; INC_DEC(pwgts[other], pwgts[2], vwgt[k]); BNDDelete(nbnd, bndind, bndptr, k); for (jj=xadj[k]; jj<xadj[k+1]; jj++) { kk = adjncy[jj]; if (where[kk] == 2) rinfo[kk].edegrees[other] += vwgt[k]; } } } ASSERT(mincut == pwgts[2]); IFSET(ctrl->dbglvl, DBG_REFINE, mprintf("\tMinimum sep: %6D at %5D, PWGTS: [%6D %6D], NBND: %6D\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd)); graph->mincut = mincut; graph->nbnd = nbnd; if (mincutorder == -1 || mincut >= initcut) break; } PQueueFree(ctrl, &parts[0]); PQueueFree(ctrl, &parts[1]); idxwspacefree(ctrl, nvtxs+1); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function computes the subdomain graph **************************************************************************/ void EliminateSubDomainEdges(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts) { int i, ii, j, k, me, other, nvtxs, total, max, avg, totalout, nind, ncand, ncand2, target, target2, nadd; int min, move, cpwgt, tvwgt; idxtype *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *maxpwgt, *pmat, *ndoms, *mypmat, *otherpmat, *ind; KeyValueType *cand, *cand2; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; vwgt = graph->vwgt; adjwgt = graph->adjwgt; where = graph->where; pwgts = graph->pwgts; /* We assume that this is properly initialized */ maxpwgt = idxwspacemalloc(ctrl, nparts); ndoms = idxwspacemalloc(ctrl, nparts); otherpmat = idxwspacemalloc(ctrl, nparts); ind = idxwspacemalloc(ctrl, nvtxs); pmat = ctrl->wspace.pmat; cand = (KeyValueType *)GKmalloc(nparts*sizeof(KeyValueType), "EliminateSubDomainEdges: cand"); cand2 = (KeyValueType *)GKmalloc(nparts*sizeof(KeyValueType), "EliminateSubDomainEdges: cand"); /* Compute the pmat matrix and ndoms */ ComputeSubDomainGraph(graph, nparts, pmat, ndoms); /* Compute the maximum allowed weight for each domain */ tvwgt = idxsum(nparts, pwgts); for (i=0; i<nparts; i++) maxpwgt[i] = 1.25*tpwgts[i]*tvwgt; /* Get into the loop eliminating subdomain connections */ for (;;) { total = idxsum(nparts, ndoms); avg = total/nparts; max = ndoms[idxamax(nparts, ndoms)]; /* printf("Adjacent Subdomain Stats: Total: %3d, Max: %3d, Avg: %3d [%5d]\n", total, max, avg, idxsum(nparts*nparts, pmat)); */ if (max < 1.4*avg) break; me = idxamax(nparts, ndoms); mypmat = pmat + me*nparts; totalout = idxsum(nparts, mypmat); /*printf("Me: %d, TotalOut: %d,\n", me, totalout);*/ /* Sort the connections according to their cut */ for (ncand2=0, i=0; i<nparts; i++) { if (mypmat[i] > 0) { cand2[ncand2].key = mypmat[i]; cand2[ncand2++].val = i; } } ikeysort(ncand2, cand2); move = 0; for (min=0; min<ncand2; min++) { if (cand2[min].key > totalout/(2*ndoms[me])) break; other = cand2[min].val; /*printf("\tMinOut: %d to %d\n", mypmat[other], other);*/ idxset(nparts, 0, otherpmat); /* Go and find the vertices in 'other' that are connected in 'me' */ for (nind=0, i=0; i<nvtxs; i++) { if (where[i] == other) { for (j=xadj[i]; j<xadj[i+1]; j++) { if (where[adjncy[j]] == me) { ind[nind++] = i; break; } } } } /* Go and construct the otherpmat to see where these nind vertices are connected to */ for (cpwgt=0, ii=0; ii<nind; ii++) { i = ind[ii]; cpwgt += vwgt[i]; for (j=xadj[i]; j<xadj[i+1]; j++) otherpmat[where[adjncy[j]]] += adjwgt[j]; } otherpmat[other] = 0; for (ncand=0, i=0; i<nparts; i++) { if (otherpmat[i] > 0) { cand[ncand].key = -otherpmat[i]; cand[ncand++].val = i; } } ikeysort(ncand, cand); /* * Go through and the select the first domain that is common with 'me', and * does not increase the ndoms[target] higher than my ndoms, subject to the * maxpwgt constraint. Traversal is done from the mostly connected to the least. */ target = target2 = -1; for (i=0; i<ncand; i++) { k = cand[i].val; if (mypmat[k] > 0) { if (pwgts[k] + cpwgt > maxpwgt[k]) /* Check if balance will go off */ continue; for (j=0; j<nparts; j++) { if (otherpmat[j] > 0 && ndoms[j] >= ndoms[me]-1 && pmat[nparts*j+k] == 0) break; } if (j == nparts) { /* No bad second level effects */ for (nadd=0, j=0; j<nparts; j++) { if (otherpmat[j] > 0 && pmat[nparts*k+j] == 0) nadd++; } /*printf("\t\tto=%d, nadd=%d, %d\n", k, nadd, ndoms[k]);*/ if (target2 == -1 && ndoms[k]+nadd < ndoms[me]) { target2 = k; } if (nadd == 0) { target = k; break; } } } } if (target == -1 && target2 != -1) target = target2; if (target == -1) { /* printf("\t\tCould not make the move\n");*/ continue; } /*printf("\t\tMoving to %d\n", target);*/ /* Update the partition weights */ INC_DEC(pwgts[target], pwgts[other], cpwgt); MoveGroupMConn(ctrl, graph, ndoms, pmat, nparts, target, nind, ind); move = 1; break; } if (move == 0) break; } idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nvtxs); GKfree(&cand, &cand2, LTERM); }
/************************************************************************* * This function performs a node-based FM refinement **************************************************************************/ void FM_2WayNodeBalance(CtrlType *ctrl, GraphType *graph, float ubfactor) { idxtype i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps; idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr; idxtype *perm, *moved; PQueueType parts; NRInfoType *rinfo; idxtype higain, oldgain; idxtype pass, to, other; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; vwgt = graph->vwgt; bndind = graph->bndind; bndptr = graph->bndptr; where = graph->where; pwgts = graph->pwgts; rinfo = graph->nrinfo; if (idxtype_abs(pwgts[0]-pwgts[1]) < (int)((ubfactor-1.0)*(pwgts[0]+pwgts[1]))) return; if (idxtype_abs(pwgts[0]-pwgts[1]) < 3*idxsum(nvtxs, vwgt, 1)/nvtxs) return; to = (pwgts[0] < pwgts[1] ? 0 : 1); other = (to+1)%2; PQueueInit(ctrl, &parts, nvtxs, ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt)); perm = idxwspacemalloc(ctrl, nvtxs); moved = idxset(nvtxs, -1, idxwspacemalloc(ctrl, nvtxs)); IFSET(ctrl->dbglvl, DBG_REFINE, mprintf("Partitions: [%6D %6D] Nv-Nb[%6D %6D]. ISep: %6D [B]\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; ASSERT(where[i] == 2); PQueueInsert(&parts, i, vwgt[i]-rinfo[i].edegrees[other]); } ASSERT(CheckNodeBnd(graph, nbnd)); ASSERT(CheckNodePartitionParams(graph)); /****************************************************** * Get into the FM loop *******************************************************/ for (nswaps=0; nswaps<nvtxs; nswaps++) { if ((higain = PQueueGetMax(&parts)) == -1) break; moved[higain] = 1; if (pwgts[other] - rinfo[higain].edegrees[other] < (pwgts[0]+pwgts[1])/2) continue; #ifdef XXX if (pwgts[other] - rinfo[higain].edegrees[other] < pwgts[to]+vwgt[higain]) break; #endif ASSERT(bndptr[higain] != -1); pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]); BNDDelete(nbnd, bndind, bndptr, higain); pwgts[to] += vwgt[higain]; where[higain] = to; IFSET(ctrl->dbglvl, DBG_MOVEINFO, mprintf("Moved %6D to %3D, Gain: %3D, \t[%5D %5D %5D]\n", higain, to, vwgt[higain]-rinfo[higain].edegrees[other], pwgts[0], pwgts[1], pwgts[2])); /********************************************************** * Update the degrees of the affected nodes ***********************************************************/ for (j=xadj[higain]; j<xadj[higain+1]; j++) { k = adjncy[j]; if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */ rinfo[k].edegrees[to] += vwgt[higain]; } else if (where[k] == other) { /* This vertex is pulled into the separator */ ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k])); BNDInsert(nbnd, bndind, bndptr, k); where[k] = 2; pwgts[other] -= vwgt[k]; edegrees = rinfo[k].edegrees; edegrees[0] = edegrees[1] = 0; for (jj=xadj[k]; jj<xadj[k+1]; jj++) { kk = adjncy[jj]; if (where[kk] != 2) edegrees[where[kk]] += vwgt[kk]; else { ASSERT(bndptr[kk] != -1); oldgain = vwgt[kk]-rinfo[kk].edegrees[other]; rinfo[kk].edegrees[other] -= vwgt[k]; if (moved[kk] == -1) PQueueUpdateUp(&parts, kk, oldgain, oldgain+vwgt[k]); } } /* Insert the new vertex into the priority queue */ PQueueInsert(&parts, k, vwgt[k]-edegrees[other]); } } if (pwgts[to] > pwgts[other]) break; } IFSET(ctrl->dbglvl, DBG_REFINE, mprintf("\tBalanced sep: %6D at %4D, PWGTS: [%6D %6D], NBND: %6D\n", pwgts[2], nswaps, pwgts[0], pwgts[1], nbnd)); graph->mincut = pwgts[2]; graph->nbnd = nbnd; PQueueFree(ctrl, &parts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }