/************************************************************************* * 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 performs a node-based FM refinement. This is the * one-way version **************************************************************************/ void FM_2WayNodeRefine_OneSided(CtrlType *ctrl, GraphType *graph, float ubfactor, 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, *swaps, *perm; PQueueType parts; NRInfoType *rinfo; idxtype higain, oldgain, mincut, initcut, mincutorder; idxtype pass, to, other, limit; idxtype badmaxpwgt, mindiff, newdiff; 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; PQueueInit(ctrl, &parts, nvtxs, ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt)); perm = idxwspacemalloc(ctrl, nvtxs); swaps = idxwspacemalloc(ctrl, nvtxs); mptr = idxwspacemalloc(ctrl, nvtxs+1); mind = idxwspacemalloc(ctrl, nvtxs); IFSET(ctrl->dbglvl, DBG_REFINE, mprintf("Partitions-N1: [%6D %6D] Nv-Nb[%6D %6D]. ISep: %6D\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); badmaxpwgt = (int)(ubfactor*(pwgts[0]+pwgts[1]+pwgts[2])/2); to = (pwgts[0] < pwgts[1] ? 1 : 0); for (pass=0; pass<npasses; pass++) { other = to; to = (to+1)%2; PQueueReset(&parts); 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, i, vwgt[i]-rinfo[i].edegrees[other]); } 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]); for (nswaps=0; nswaps<nvtxs; nswaps++) { if ((higain = PQueueGetMax(&parts)) == -1) break; ASSERT(bndptr[higain] != -1); if (pwgts[to]+vwgt[higain] > badmaxpwgt) break; /* No point going any further. Balance will be bad */ 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; 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] */ 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); 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]; /* Since the moves are one-sided this vertex has not been moved yet */ PQueueUpdateUp(&parts, kk, oldgain, oldgain+vwgt[k]); } } /* Insert the new vertex into the priority queue. Safe due to one-sided moves */ PQueueInsert(&parts, k, vwgt[k]-edegrees[other]); } } mptr[nswaps+1] = nmind; IFSET(ctrl->dbglvl, DBG_MOVEINFO, mprintf("Moved %6D to %3D, Gain: %5D [%5D] \t[%5D %5D %5D] [%3D %2D]\n", higain, to, (vwgt[higain]-rinfo[higain].edegrees[other]), vwgt[higain], pwgts[0], pwgts[1], pwgts[2], nswaps, limit)); } /**************************************************************** * Roll back computation *****************************************************************/ for (nswaps--; nswaps>mincutorder; nswaps--) { higain = swaps[nswaps]; ASSERT(CheckNodePartitionParams(graph)); ASSERT(where[higain] == to); 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 (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut)) break; } PQueueFree(ctrl, &parts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs+1); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * 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); }
void FM_2WayNodeRefine1Sided(ctrl_t *ctrl, graph_t *graph, idx_t niter) { idx_t i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind, iend; idx_t *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr; idx_t *mptr, *mind, *swaps; rpq_t *queue; nrinfo_t *rinfo; idx_t higain, mincut, initcut, mincutorder; idx_t pass, to, other, limit; idx_t badmaxpwgt, mindiff, newdiff; real_t mult; WCOREPUSH; 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; queue = rpqCreate(nvtxs); swaps = iwspacemalloc(ctrl, nvtxs); mptr = iwspacemalloc(ctrl, nvtxs+1); mind = iwspacemalloc(ctrl, 2*nvtxs); mult = 0.5*ctrl->ubfactors[0]; badmaxpwgt = (idx_t)(mult*(pwgts[0]+pwgts[1]+pwgts[2])); IFSET(ctrl->dbglvl, METIS_DBG_REFINE, printf("Partitions-N1: [%6"PRIDX" %6"PRIDX"] Nv-Nb[%6"PRIDX" %6"PRIDX"]. ISep: %6"PRIDX"\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); to = (pwgts[0] < pwgts[1] ? 1 : 0); for (pass=0; pass<2*niter; pass++) { /* the 2*niter is for the two sides */ other = to; to = (to+1)%2; rpqReset(queue); mincutorder = -1; initcut = mincut = graph->mincut; nbnd = graph->nbnd; /* use the swaps array in place of the traditional perm array to save memory */ my_irandArrayPermute_r(nbnd, swaps, nbnd, 1, &ctrl->curseed); for (ii=0; ii<nbnd; ii++) { i = bndind[swaps[ii]]; ASSERT(where[i] == 2); rpqInsert(queue, i, vwgt[i]-rinfo[i].edegrees[other]); } ASSERT(CheckNodeBnd(graph, nbnd)); ASSERT(CheckNodePartitionParams(graph)); limit = (ctrl->compress ? gk_min(5*nbnd, 500) : gk_min(3*nbnd, 300)); /****************************************************** * Get into the FM loop *******************************************************/ IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux3Tmr)); mptr[0] = nmind = 0; mindiff = iabs(pwgts[0]-pwgts[1]); for (nswaps=0; nswaps<nvtxs; nswaps++) { if ((higain = rpqGetTop(queue)) == -1) break; ASSERT(bndptr[higain] != -1); /* The following check is to ensure we break out if there is a posibility of over-running the mind array. */ if (nmind + xadj[higain+1]-xadj[higain] >= 2*nvtxs-1) break; if (pwgts[to]+vwgt[higain] > badmaxpwgt) break; /* No point going any further. Balance will be bad */ pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]); newdiff = iabs(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 > 3*limit || (nswaps - mincutorder > limit && pwgts[2] > 1.10*mincut)) { 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; swaps[nswaps] = higain; /********************************************************** * Update the degrees of the affected nodes ***********************************************************/ IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux1Tmr)); 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, ("%"PRIDX" %"PRIDX" %"PRIDX"\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], iend=xadj[k+1]; jj<iend; jj++) { kk = adjncy[jj]; if (where[kk] != 2) edegrees[where[kk]] += vwgt[kk]; else { rinfo[kk].edegrees[other] -= vwgt[k]; /* Since the moves are one-sided this vertex has not been moved yet */ rpqUpdate(queue, kk, vwgt[kk]-rinfo[kk].edegrees[other]); } } /* Insert the new vertex into the priority queue. Safe due to one-sided moves */ rpqInsert(queue, k, vwgt[k]-edegrees[other]); } } mptr[nswaps+1] = nmind; IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux1Tmr)); IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO, printf("Moved %6"PRIDX" to %3"PRIDX", Gain: %5"PRIDX" [%5"PRIDX"] \t[%5"PRIDX" %5"PRIDX" %5"PRIDX"] [%3"PRIDX" %2"PRIDX"]\n", higain, to, (vwgt[higain]-rinfo[higain].edegrees[other]), vwgt[higain], pwgts[0], pwgts[1], pwgts[2], nswaps, limit)); } IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux3Tmr)); /**************************************************************** * Roll back computation *****************************************************************/ IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux2Tmr)); for (nswaps--; nswaps>mincutorder; nswaps--) { higain = swaps[nswaps]; ASSERT(CheckNodePartitionParams(graph)); ASSERT(where[higain] == to); 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], iend=xadj[k+1]; jj<iend; jj++) { kk = adjncy[jj]; if (where[kk] == 2) rinfo[kk].edegrees[other] += vwgt[k]; } } } IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux2Tmr)); ASSERT(mincut == pwgts[2]); IFSET(ctrl->dbglvl, METIS_DBG_REFINE, printf("\tMinimum sep: %6"PRIDX" at %5"PRIDX", PWGTS: [%6"PRIDX" %6"PRIDX"], NBND: %6"PRIDX"\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd)); graph->mincut = mincut; graph->nbnd = nbnd; if (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut)) break; } rpqDestroy(queue); WCOREPOP; }
void FM_2WayCutRefine(ctrl_t *ctrl, graph_t *graph, real_t *ntpwgts, idx_t niter) { idx_t i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, limit, tmp; idx_t *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts; idx_t *moved, *swaps, *perm; rpq_t *queues[2]; idx_t higain, mincut, mindiff, origdiff, initcut, newcut, mincutorder, avgvwgt; idx_t tpwgts[2]; WCOREPUSH; nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; id = graph->id; ed = graph->ed; pwgts = graph->pwgts; bndptr = graph->bndptr; bndind = graph->bndind; moved = iwspacemalloc(ctrl, nvtxs); swaps = iwspacemalloc(ctrl, nvtxs); perm = iwspacemalloc(ctrl, nvtxs); tpwgts[0] = graph->tvwgt[0]*ntpwgts[0]; tpwgts[1] = graph->tvwgt[0]-tpwgts[0]; limit = gk_min(gk_max(0.01*nvtxs, 15), 100); avgvwgt = gk_min((pwgts[0]+pwgts[1])/20, 2*(pwgts[0]+pwgts[1])/nvtxs); queues[0] = rpqCreate(nvtxs); queues[1] = rpqCreate(nvtxs); IFSET(ctrl->dbglvl, METIS_DBG_REFINE, Print2WayRefineStats(ctrl, graph, ntpwgts, 0, -2)); origdiff = iabs(tpwgts[0]-pwgts[0]); iset(nvtxs, -1, moved); for (pass=0; pass<niter; pass++) { /* Do a number of passes */ rpqReset(queues[0]); rpqReset(queues[1]); mincutorder = -1; newcut = mincut = initcut = graph->mincut; mindiff = iabs(tpwgts[0]-pwgts[0]); ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); /* Insert boundary nodes in the priority queues */ nbnd = graph->nbnd; irandArrayPermute(nbnd, perm, nbnd, 1); for (ii=0; ii<nbnd; ii++) { i = perm[ii]; ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0); ASSERT(bndptr[bndind[i]] != -1); rpqInsert(queues[where[bndind[i]]], bndind[i], ed[bndind[i]]-id[bndind[i]]); } for (nswaps=0; nswaps<nvtxs; nswaps++) { from = (tpwgts[0]-pwgts[0] < tpwgts[1]-pwgts[1] ? 0 : 1); to = (from+1)%2; if ((higain = rpqGetTop(queues[from])) == -1) break; ASSERT(bndptr[higain] != -1); newcut -= (ed[higain]-id[higain]); INC_DEC(pwgts[to], pwgts[from], vwgt[higain]); if ((newcut < mincut && iabs(tpwgts[0]-pwgts[0]) <= origdiff+avgvwgt) || (newcut == mincut && iabs(tpwgts[0]-pwgts[0]) < mindiff)) { mincut = newcut; mindiff = iabs(tpwgts[0]-pwgts[0]); mincutorder = nswaps; } else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */ newcut += (ed[higain]-id[higain]); INC_DEC(pwgts[from], pwgts[to], vwgt[higain]); break; } where[higain] = to; moved[higain] = nswaps; swaps[nswaps] = higain; IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO, printf("Moved %6"PRIDX" from %"PRIDX". [%3"PRIDX" %3"PRIDX"] %5"PRIDX" [%4"PRIDX" %4"PRIDX"]\n", higain, from, ed[higain]-id[higain], vwgt[higain], newcut, pwgts[0], pwgts[1])); /************************************************************** * Update the id[i]/ed[i] values of the affected nodes ***************************************************************/ SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && xadj[higain] < xadj[higain+1]) BNDDelete(nbnd, bndind, bndptr, higain); 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); /* Update its boundary information and queue position */ if (bndptr[k] != -1) { /* If k was a boundary vertex */ if (ed[k] == 0) { /* Not a boundary vertex any more */ BNDDelete(nbnd, bndind, bndptr, k); if (moved[k] == -1) /* Remove it if in the queues */ rpqDelete(queues[where[k]], k); } else { /* If it has not been moved, update its position in the queue */ if (moved[k] == -1) rpqUpdate(queues[where[k]], k, ed[k]-id[k]); } } else { if (ed[k] > 0) { /* It will now become a boundary vertex */ BNDInsert(nbnd, bndind, bndptr, k); if (moved[k] == -1) rpqInsert(queues[where[k]], k, ed[k]-id[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); INC_DEC(pwgts[to], pwgts[(to+1)%2], vwgt[higain]); 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); } } graph->mincut = mincut; graph->nbnd = nbnd; IFSET(ctrl->dbglvl, METIS_DBG_REFINE, Print2WayRefineStats(ctrl, graph, ntpwgts, 0, mincutorder)); if (mincutorder <= 0 || mincut == initcut) break; } rpqDestroy(queues[0]); rpqDestroy(queues[1]); WCOREPOP; }
/************************************************************************* * 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 performs an edge-based FM refinement **************************************************************************/ void FM_2WayEdgeRefine(CtrlType *ctrl, GraphType *graph, int *tpwgts, int npasses) { int i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, limit, tmp; idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts; idxtype *moved, *swaps, *perm; PQueueType parts[2]; int higain, oldgain, mincut, mindiff, origdiff, initcut, newcut, mincutorder, avgvwgt; nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; id = graph->id; ed = graph->ed; pwgts = graph->pwgts; bndptr = graph->bndptr; bndind = graph->bndind; moved = idxwspacemalloc(ctrl, nvtxs); swaps = idxwspacemalloc(ctrl, nvtxs); perm = idxwspacemalloc(ctrl, nvtxs); limit = (int) amin(amax(0.01*nvtxs, 15), 100); avgvwgt = amin((pwgts[0]+pwgts[1])/20, 2*(pwgts[0]+pwgts[1])/nvtxs); tmp = graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]; PQueueInit(ctrl, &parts[0], nvtxs, tmp); PQueueInit(ctrl, &parts[1], nvtxs, tmp); IFSET(ctrl->dbglvl, DBG_REFINE, printf("Partitions: [%6d %6d] T[%6d %6d], Nv-Nb[%6d %6d]. ICut: %6d\n", pwgts[0], pwgts[1], tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); origdiff = abs(tpwgts[0]-pwgts[0]); idxset(nvtxs, -1, moved); for (pass=0; pass<npasses; pass++) { /* Do a number of passes */ PQueueReset(&parts[0]); PQueueReset(&parts[1]); mincutorder = -1; newcut = mincut = initcut = graph->mincut; mindiff = abs(tpwgts[0]-pwgts[0]); ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); /* Insert boundary nodes in the priority queues */ nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = perm[ii]; ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0); ASSERT(bndptr[bndind[i]] != -1); PQueueInsert(&parts[where[bndind[i]]], bndind[i], ed[bndind[i]]-id[bndind[i]]); } for (nswaps=0; nswaps<nvtxs; nswaps++) { from = (tpwgts[0]-pwgts[0] < tpwgts[1]-pwgts[1] ? 0 : 1); to = (from+1)%2; if ((higain = PQueueGetMax(&parts[from])) == -1) break; ASSERT(bndptr[higain] != -1); newcut -= (ed[higain]-id[higain]); INC_DEC(pwgts[to], pwgts[from], vwgt[higain]); if ((newcut < mincut && abs(tpwgts[0]-pwgts[0]) <= origdiff+avgvwgt) || (newcut == mincut && abs(tpwgts[0]-pwgts[0]) < mindiff)) { mincut = newcut; mindiff = abs(tpwgts[0]-pwgts[0]); mincutorder = nswaps; } else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */ newcut += (ed[higain]-id[higain]); INC_DEC(pwgts[from], pwgts[to], vwgt[higain]); break; } where[higain] = to; moved[higain] = nswaps; swaps[nswaps] = higain; IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("Moved %6d from %d. [%3d %3d] %5d [%4d %4d]\n", higain, from, ed[higain]-id[higain], vwgt[higain], newcut, pwgts[0], pwgts[1])); /************************************************************** * Update the id[i]/ed[i] values of the affected nodes ***************************************************************/ SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && xadj[higain] < xadj[higain+1]) BNDDelete(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 its boundary information and queue position */ if (bndptr[k] != -1) { /* If k was a boundary vertex */ if (ed[k] == 0) { /* Not a boundary vertex any more */ BNDDelete(nbnd, bndind, bndptr, k); if (moved[k] == -1) /* Remove it if in the queues */ PQueueDelete(&parts[where[k]], k, oldgain); } else { /* If it has not been moved, update its position in the queue */ if (moved[k] == -1) PQueueUpdate(&parts[where[k]], k, oldgain, ed[k]-id[k]); } } else { if (ed[k] > 0) { /* It will now become a boundary vertex */ BNDInsert(nbnd, bndind, bndptr, k); if (moved[k] == -1) PQueueInsert(&parts[where[k]], k, ed[k]-id[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); INC_DEC(pwgts[to], pwgts[(to+1)%2], vwgt[higain]); 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); } } IFSET(ctrl->dbglvl, DBG_REFINE, printf("\tMinimum cut: %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); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * 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); }
void FM_2WayNodeRefine_TwoSidedP(CtrlType *ctrl, GraphType *graph, idxtype *hmarker, float ubfactor, int npasses) { int 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; int higain, oldgain, mincut, initcut, mincutorder; int pass, to, other, limit; int badmaxpwgt, mindiff, newdiff; int 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, printf("Partitions: [%6d %6d] Nv-Nb[%6d %6d]. ISep: %6d\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); badmaxpwgt = (int)(ubfactor*amax(pwgts[0], pwgts[1])); 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); if (hmarker[i] == -1) { PQueueInsert(&parts[0], i, vwgt[i]-rinfo[i].edegrees[1]); PQueueInsert(&parts[1], i, vwgt[i]-rinfo[i].edegrees[0]); moved[i] = -5; } else if (hmarker[i] != 2) { PQueueInsert(&parts[hmarker[i]], i, vwgt[i]-rinfo[i].edegrees[(hmarker[i]+1)%2]); moved[i] = -(10+hmarker[i]); } } ASSERT(CheckNodeBnd(graph, nbnd)); ASSERT(CheckNodePartitionParams(graph)); limit = nbnd; /****************************************************** * Get into the FM loop *******************************************************/ mptr[0] = nmind = 0; mindiff = abs(pwgts[0]-pwgts[1]); to = (pwgts[0] < pwgts[1] ? 0 : 1); for (nswaps=0; nswaps<nvtxs; nswaps++) { 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)); if (pwgts[to]+vwgt[u[to]] > badmaxpwgt) to = (to+1)%2; } else if (u[0] == -1 && u[1] == -1) { break; } else if (u[0] != -1 && pwgts[0]+vwgt[u[0]] <= badmaxpwgt) { to = 0; } else if (u[1] != -1 && pwgts[1]+vwgt[u[1]] <= badmaxpwgt) { to = 1; } else break; other = (to+1)%2; higain = PQueueGetMax(&parts[to]); /* Delete its matching entry in the other queue */ if (moved[higain] == -5) PQueueDelete(&parts[other], higain, vwgt[higain]-rinfo[higain].edegrees[to]); ASSERT(bndptr[higain] != -1); pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]); newdiff = 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] == -5 || moved[k] == -(10+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] == -5 || moved[kk] == -(10+to)) PQueueUpdate(&parts[to], kk, oldgain, oldgain+vwgt[k]); } } /* Insert the new vertex into the priority queue (if it has not been moved). */ if (moved[k] == -1 && (hmarker[k] == -1 || hmarker[k] == to)) { PQueueInsert(&parts[to], k, vwgt[k]-edegrees[other]); moved[k] = -(10+to); } #ifdef FULLMOVES /* this does not work as well as the above partial one */ if (moved[k] == -1) { if (hmarker[k] == -1) { PQueueInsert(&parts[0], k, vwgt[k]-edegrees[1]); PQueueInsert(&parts[1], k, vwgt[k]-edegrees[0]); moved[k] = -5; } else if (hmarker[k] != 2) { PQueueInsert(&parts[hmarker[k]], k, vwgt[k]-edegrees[(hmarker[k]+1)%2]); moved[k] = -(10+hmarker[k]); } } #endif } } mptr[nswaps+1] = nmind; IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("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, printf("\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 performs an edge-based FM refinement **************************************************************************/ void MocFM_2WayEdgeRefine(CtrlType *ctrl, GraphType *graph, float *tpwgts, int npasses) { int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, me, limit, tmp, cnum; idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind; idxtype *moved, *swaps, *perm, *qnum; float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal; PQueueType parts[MAXNCON][2]; int higain, oldgain, mincut, initcut, newcut, mincutorder; float rtpwgts[2]; 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, 25), 150); /* 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); origbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); rtpwgts[0] = origbal*tpwgts[0]; rtpwgts[1] = origbal*tpwgts[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: %.3f\n", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut, origbal); } */ idxset(nvtxs, -1, moved); for (pass=0; pass<npasses; pass++) { /* Do a number of passes */ for (i=0; i<ncon; i++) { PQueueReset(&parts[i][0]); PQueueReset(&parts[i][1]); } mincutorder = -1; newcut = mincut = initcut = graph->mincut; for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[0]-npwgts[i]); minbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); /* Insert boundary nodes in the priority queues */ nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; ASSERT(ed[i] > 0 || id[i] == 0); ASSERT(bndptr[i] != -1); PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]); } for (nswaps=0; nswaps<nvtxs; nswaps++) { SelectQueue(ncon, npwgts, rtpwgts, &from, &cnum, parts); to = (from+1)%2; if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1) break; ASSERT(bndptr[higain] != -1); 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]); newbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); if ((newcut < mincut && newbal-origbal <= .00001) || (newcut == mincut && (newbal < minbal || (newbal == minbal && BetterBalance(ncon, npwgts, tpwgts, mindiff))))) { mincut = newcut; minbal = newbal; mincutorder = nswaps; for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[0]-npwgts[i]); } 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 (l=0; l<ncon; l++) printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]); printf(", %.3f LB: %.3f\n", minbal, newbal); } */ /************************************************************** * Update the id[i]/ed[i] values of the affected nodes ***************************************************************/ SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && xadj[higain] < xadj[higain+1]) BNDDelete(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 its boundary information and queue position */ if (bndptr[k] != -1) { /* If k was a boundary vertex */ if (ed[k] == 0) { /* Not a boundary vertex any more */ BNDDelete(nbnd, bndind, bndptr, k); if (moved[k] == -1) /* Remove it if in the queues */ PQueueDelete(&parts[qnum[k]][where[k]], k, oldgain); } else { /* If it has not been moved, update its position in the queue */ if (moved[k] == -1) PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]); } } else { if (ed[k] > 0) { /* It will now become a boundary vertex */ BNDInsert(nbnd, bndind, bndptr, k); if (moved[k] == -1) PQueueInsert(&parts[qnum[k]][where[k]], k, ed[k]-id[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 (l=0; l<ncon; l++) printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]); printf("], LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts)); } */ graph->mincut = mincut; graph->nbnd = nbnd; if (mincutorder == -1 || mincut == initcut) break; } 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); }
void FM_2WayNodeRefine_OneSidedP(CtrlType *ctrl, GraphType *graph, idxtype *hmarker, float ubfactor, int npasses) { int i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind, nbad, qsize; idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr; idxtype *mptr, *mind, *swaps, *perm, *inqueue; PQueueType parts; NRInfoType *rinfo; int higain, oldgain, mincut, initcut, mincutorder; int pass, from, to, limit; int badmaxpwgt, mindiff, newdiff; ASSERT(graph->mincut == graph->pwgts[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; PQueueInit(ctrl, &parts, nvtxs, ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt)); perm = idxwspacemalloc(ctrl, nvtxs); swaps = idxwspacemalloc(ctrl, nvtxs); mptr = idxwspacemalloc(ctrl, nvtxs+1); mind = idxwspacemalloc(ctrl, nvtxs); inqueue = idxwspacemalloc(ctrl, nvtxs); idxset(nvtxs, -1, inqueue); badmaxpwgt = (int)(ubfactor*amax(pwgts[0], pwgts[1])); IFSET(ctrl->dbglvl, DBG_REFINE, printf("Partitions-N1: [%6d %6d] Nv-Nb[%6d %6d] MaxPwgt[%6d]. ISep: %6d\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, badmaxpwgt, graph->mincut)); to = (pwgts[0] < pwgts[1] ? 1 : 0); for (pass=0; pass<npasses; pass++) { from = to; to = (from+1)%2; PQueueReset(&parts); 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); if (hmarker[i] == -1 || hmarker[i] == to) { PQueueInsert(&parts, i, vwgt[i]-rinfo[i].edegrees[from]); inqueue[i] = pass; } } qsize = parts.nnodes; ASSERT(CheckNodeBnd(graph, nbnd)); ASSERT(CheckNodePartitionParams(graph)); limit = nbnd; /****************************************************** * Get into the FM loop *******************************************************/ mptr[0] = nmind = nbad = 0; mindiff = abs(pwgts[0]-pwgts[1]); for (nswaps=0; nswaps<nvtxs; nswaps++) { if ((higain = PQueueGetMax(&parts)) == -1) break; inqueue[higain] = -1; ASSERT(bndptr[higain] != -1); if (pwgts[to]+vwgt[higain] > badmaxpwgt) { /* Skip this vertex */ if (nbad++ > limit) break; else { nswaps--; continue; } } pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[from]); newdiff = abs(pwgts[to]+vwgt[higain] - (pwgts[from]-rinfo[higain].edegrees[from])); if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) { mincut = pwgts[2]; mincutorder = nswaps; mindiff = newdiff; nbad = 0; } else { if (nbad++ > limit) { pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[from]); break; /* No further improvement, break out */ } } BNDDelete(nbnd, bndind, bndptr, higain); pwgts[to] += vwgt[higain]; where[higain] = to; 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] */ rinfo[k].edegrees[to] += vwgt[higain]; } else if (where[k] == from) { /* 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[from] -= 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[from]; rinfo[kk].edegrees[from] -= vwgt[k]; /* Update the gain of this node if it was skipped */ if (inqueue[kk] == pass) PQueueUpdateUp(&parts, kk, oldgain, oldgain+vwgt[k]); } } /* Insert the new vertex into the priority queue. Safe due to one-sided moves */ if (hmarker[k] == -1 || hmarker[k] == to) { PQueueInsert(&parts, k, vwgt[k]-edegrees[from]); inqueue[k] = pass; } } } mptr[nswaps+1] = nmind; IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("Moved %6d to %3d, Gain: %5d [%5d] \t[%5d %5d %5d] [%3d %2d]\n", higain, to, (vwgt[higain]-rinfo[higain].edegrees[from]), vwgt[higain], pwgts[0], pwgts[1], pwgts[2], nswaps, limit)); } /**************************************************************** * Roll back computation *****************************************************************/ for (nswaps--; nswaps>mincutorder; nswaps--) { higain = swaps[nswaps]; ASSERT(CheckNodePartitionParams(graph)); ASSERT(where[higain] == to); 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] = from; INC_DEC(pwgts[from], 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[from] += vwgt[k]; } } } ASSERT(mincut == pwgts[2]); IFSET(ctrl->dbglvl, DBG_REFINE, printf("\tMinimum sep: %6d at %5d, PWGTS: [%6d %6d], NBND: %6d, QSIZE: %6d\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd, qsize)); graph->mincut = mincut; graph->nbnd = nbnd; if (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut)) break; } PQueueFree(ctrl, &parts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs+1); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function balances two partitions by moving boundary nodes * from the domain that is overweight to the one that is underweight. **************************************************************************/ void Bnd2WayBalance(CtrlType *ctrl, GraphType *graph, int *tpwgts) { int i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, tmp; idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts; idxtype *moved, *perm; PQueueType parts; int higain, oldgain, mincut, mindiff; nvtxs = graph->nvtxs; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; id = graph->id; ed = graph->ed; pwgts = graph->pwgts; bndptr = graph->bndptr; bndind = graph->bndind; moved = idxwspacemalloc(ctrl, nvtxs); perm = idxwspacemalloc(ctrl, nvtxs); /* Determine from which domain you will be moving data */ mindiff = abs(tpwgts[0]-pwgts[0]); from = (pwgts[0] < tpwgts[0] ? 1 : 0); to = (from+1)%2; IFSET(ctrl->dbglvl, DBG_REFINE, printf("Partitions: [%6d %6d] T[%6d %6d], Nv-Nb[%6d %6d]. ICut: %6d [B]\n", pwgts[0], pwgts[1], tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); tmp = graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]; PQueueInit(ctrl, &parts, nvtxs, tmp); idxset(nvtxs, -1, moved); ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); /* Insert the boundary nodes of the proper partition whose size is OK in the priority queue */ nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (ii=0; ii<nbnd; ii++) { i = perm[ii]; ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0); ASSERT(bndptr[bndind[i]] != -1); if (where[bndind[i]] == from && vwgt[bndind[i]] <= mindiff) PQueueInsert(&parts, bndind[i], ed[bndind[i]]-id[bndind[i]]); } mincut = graph->mincut; for (nswaps=0; nswaps<nvtxs; nswaps++) { if ((higain = PQueueGetMax(&parts)) == -1) break; ASSERT(bndptr[higain] != -1); if (pwgts[to]+vwgt[higain] > tpwgts[to]) break; mincut -= (ed[higain]-id[higain]); INC_DEC(pwgts[to], pwgts[from], vwgt[higain]); where[higain] = to; moved[higain] = nswaps; IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("Moved %6d from %d. [%3d %3d] %5d [%4d %4d]\n", higain, from, ed[higain]-id[higain], vwgt[higain], mincut, pwgts[0], pwgts[1])); /************************************************************** * Update the id[i]/ed[i] values of the affected nodes ***************************************************************/ SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && xadj[higain] < xadj[higain+1]) BNDDelete(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 its boundary information and queue position */ if (bndptr[k] != -1) { /* If k was a boundary vertex */ if (ed[k] == 0) { /* Not a boundary vertex any more */ BNDDelete(nbnd, bndind, bndptr, k); if (moved[k] == -1 && where[k] == from && vwgt[k] <= mindiff) /* Remove it if in the queues */ PQueueDelete(&parts, k, oldgain); } else { /* If it has not been moved, update its position in the queue */ if (moved[k] == -1 && where[k] == from && vwgt[k] <= mindiff) PQueueUpdate(&parts, k, oldgain, ed[k]-id[k]); } } else { if (ed[k] > 0) { /* It will now become a boundary vertex */ BNDInsert(nbnd, bndind, bndptr, k); if (moved[k] == -1 && where[k] == from && vwgt[k] <= mindiff) PQueueInsert(&parts, k, ed[k]-id[k]); } } } } IFSET(ctrl->dbglvl, DBG_REFINE, printf("\tMinimum cut: %6d, PWGTS: [%6d %6d], NBND: %6d\n", mincut, pwgts[0], pwgts[1], nbnd)); graph->mincut = mincut; graph->nbnd = nbnd; PQueueFree(ctrl, &parts); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
void FM_2WayNodeBalance(ctrl_t *ctrl, graph_t *graph) { idx_t i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, gain; idx_t badmaxpwgt, higain, oldgain, to, other; idx_t *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr; idx_t *perm, *moved; rpq_t *queue; nrinfo_t *rinfo; real_t mult; 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; mult = 0.5*ctrl->ubfactors[0]; badmaxpwgt = (idx_t)(mult*(pwgts[0]+pwgts[1])); if (gk_max(pwgts[0], pwgts[1]) < badmaxpwgt) return; if (iabs(pwgts[0]-pwgts[1]) < 3*graph->tvwgt[0]/nvtxs) return; WCOREPUSH; to = (pwgts[0] < pwgts[1] ? 0 : 1); other = (to+1)%2; queue = rpqCreate(nvtxs); perm = iwspacemalloc(ctrl, nvtxs); moved = iset(nvtxs, -1, iwspacemalloc(ctrl, nvtxs)); IFSET(ctrl->dbglvl, METIS_DBG_REFINE, printf("Partitions: [%6"PRIDX" %6"PRIDX"] Nv-Nb[%6"PRIDX" %6"PRIDX"]. ISep: %6"PRIDX" [B]\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); nbnd = graph->nbnd; my_irandArrayPermute_r(nbnd, perm, nbnd, 1, &ctrl->curseed); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; ASSERT(where[i] == 2); rpqInsert(queue, 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 = rpqGetTop(queue)) == -1) break; moved[higain] = 1; gain = vwgt[higain]-rinfo[higain].edegrees[other]; badmaxpwgt = (idx_t)(mult*(pwgts[0]+pwgts[1])); /* break if other is now underwight */ if (pwgts[to] > pwgts[other]) break; /* break if balance is achieved and no +ve or zero gain */ if (gain < 0 && pwgts[other] < badmaxpwgt) break; /* skip this vertex if it will violate balance on the other side */ if (pwgts[to]+vwgt[higain] > badmaxpwgt) continue; ASSERT(bndptr[higain] != -1); pwgts[2] -= gain; BNDDelete(nbnd, bndind, bndptr, higain); pwgts[to] += vwgt[higain]; where[higain] = to; IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO, printf("Moved %6"PRIDX" to %3"PRIDX", Gain: %3"PRIDX", \t[%5"PRIDX" %5"PRIDX" %5"PRIDX"]\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, ("%"PRIDX" %"PRIDX" %"PRIDX"\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) rpqUpdate(queue, kk, oldgain+vwgt[k]); } } /* Insert the new vertex into the priority queue */ rpqInsert(queue, k, vwgt[k]-edegrees[other]); } } } IFSET(ctrl->dbglvl, METIS_DBG_REFINE, printf("\tBalanced sep: %6"PRIDX" at %4"PRIDX", PWGTS: [%6"PRIDX" %6"PRIDX"], NBND: %6"PRIDX"\n", pwgts[2], nswaps, pwgts[0], pwgts[1], nbnd)); graph->mincut = pwgts[2]; graph->nbnd = nbnd; rpqDestroy(queue); WCOREPOP; }
/************************************************************************* * 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); }
/************************************************************************* * This function balances two partitions by moving the highest gain * (including negative gain) vertices to the other domain. * It is used only when tha unbalance is due to non contigous * subdomains. That is, the are no boundary vertices. * It moves vertices from the domain that is overweight to the one that * is underweight. **************************************************************************/ void Mc_Serial_Init2WayBalance(GraphType *graph, float *tpwgts) { int i, ii, j, k; int kwgt, nvtxs, nbnd, ncon, nswaps, from, to, cnum, tmp; idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind; idxtype *qnum; float *nvwgt, *npwgts; FPQueueType parts[MAXNCON][2]; int higain, oldgain, mincut; KeyValueType *cand; nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; adjncy = graph->adjncy; nvwgt = graph->nvwgt; adjwgt = graph->adjwgt; where = graph->where; id = graph->sendind; ed = graph->recvind; npwgts = graph->gnpwgts; bndptr = graph->sendptr; bndind = graph->recvptr; qnum = idxmalloc(nvtxs, "qnum"); cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand"); /* This is called for initial partitioning so we know from where to pick nodes */ from = 1; to = (from+1)%2; for (i=0; i<ncon; i++) { FPQueueInit(&parts[i][0], nvtxs); FPQueueInit(&parts[i][1], nvtxs); } /* Compute the queues in which each vertex will be assigned to */ for (i=0; i<nvtxs; i++) qnum[i] = samax(ncon, nvwgt+i*ncon); for (i=0; i<nvtxs; i++) { cand[i].key = id[i]-ed[i]; cand[i].val = i; } ikeysort(nvtxs, cand); /* Insert the nodes of the proper partition in the appropriate priority queue */ for (ii=0; ii<nvtxs; ii++) { i = cand[ii].val; if (where[i] == from) { if (ed[i] > 0) FPQueueInsert(&parts[qnum[i]][0], i, (float)(ed[i]-id[i])); else FPQueueInsert(&parts[qnum[i]][1], i, (float)(ed[i]-id[i])); } } mincut = graph->mincut; nbnd = graph->gnvtxs; for (nswaps=0; nswaps<nvtxs; nswaps++) { if (Serial_AreAnyVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, tpwgts+from*ncon)) break; if ((cnum = Serial_SelectQueueOneWay(ncon, npwgts, tpwgts, from, parts)) == -1) break; if ((higain = FPQueueGetMax(&parts[cnum][0])) == -1) higain = FPQueueGetMax(&parts[cnum][1]); mincut -= (ed[higain]-id[higain]); saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); where[higain] = to; /************************************************************** * 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 (where[k] == from) { if (ed[k] > 0 && bndptr[k] == -1) { /* It moves in boundary */ FPQueueDelete(&parts[qnum[k]][1], k); FPQueueInsert(&parts[qnum[k]][0], k, (float)(ed[k]-id[k])); } else { /* It must be in the boundary already */ FPQueueUpdate(&parts[qnum[k]][0], k, (float)(oldgain), (float)(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); } } graph->mincut = mincut; graph->gnvtxs = nbnd; for (i=0; i<ncon; i++) { FPQueueFree(&parts[i][0]); FPQueueFree(&parts[i][1]); } GKfree((void **)&cand, (void **)&qnum, LTERM); }
/************************************************************************* * This function balances two partitions by moving the highest gain * (including negative gain) vertices to the other domain. * It is used only when tha unbalance is due to non contigous * subdomains. That is, the are no boundary vertices. * It moves vertices from the domain that is overweight to the one that * is underweight. **************************************************************************/ void MocInit2WayBalance(CtrlType *ctrl, GraphType *graph, float *tpwgts) { int i, ii, j, k, l, kwgt, nvtxs, nbnd, ncon, nswaps, from, to, pass, me, cnum, tmp; idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind; idxtype *perm, *qnum; float *nvwgt, *npwgts; PQueueType parts[MAXNCON][2]; int higain, oldgain, mincut; nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; adjncy = graph->adjncy; nvwgt = graph->nvwgt; adjwgt = graph->adjwgt; where = graph->where; id = graph->id; ed = graph->ed; npwgts = graph->npwgts; bndptr = graph->bndptr; bndind = graph->bndind; perm = idxwspacemalloc(ctrl, nvtxs); qnum = idxwspacemalloc(ctrl, nvtxs); /* This is called for initial partitioning so we know from where to pick nodes */ from = 1; to = (from+1)%2; 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: %.3f [B]\n", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut, Compute2WayHLoadImbalance(ncon, npwgts, tpwgts)); } for (i=0; i<ncon; i++) { PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1); PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1); } ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); ASSERT(CheckGraph(graph)); /* Compute the queues in which each vertex will be assigned to */ for (i=0; i<nvtxs; i++) qnum[i] = samax(ncon, nvwgt+i*ncon); /* Insert the nodes of the proper partition in the appropriate priority queue */ RandomPermute(nvtxs, perm, 1); for (ii=0; ii<nvtxs; ii++) { i = perm[ii]; if (where[i] == from) { if (ed[i] > 0) PQueueInsert(&parts[qnum[i]][0], i, ed[i]-id[i]); else PQueueInsert(&parts[qnum[i]][1], i, ed[i]-id[i]); } } mincut = graph->mincut; nbnd = graph->nbnd; for (nswaps=0; nswaps<nvtxs; nswaps++) { if (AreAnyVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, tpwgts[from])) break; if ((cnum = SelectQueueOneWay(ncon, npwgts, tpwgts, from, parts)) == -1) break; if ((higain = PQueueGetMax(&parts[cnum][0])) == -1) higain = PQueueGetMax(&parts[cnum][1]); mincut -= (ed[higain]-id[higain]); saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); where[higain] = to; if (ctrl->dbglvl&DBG_MOVEINFO) { printf("Moved %6d from %d(%d). [%5d] %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], mincut); for (l=0; l<ncon; l++) printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]); printf(", LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts)); if (ed[higain] == 0 && id[higain] > 0) printf("\t Pulled from the interior!\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 (where[k] == from) { if (ed[k] > 0 && bndptr[k] == -1) { /* It moves in boundary */ PQueueDelete(&parts[qnum[k]][1], k, oldgain); PQueueInsert(&parts[qnum[k]][0], k, ed[k]-id[k]); } else { /* It must be in the boundary already */ if (bndptr[k] == -1) printf("What you thought was wrong!\n"); PQueueUpdate(&parts[qnum[k]][0], 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); } ASSERTP(ComputeCut(graph, where) == mincut, ("%d != %d\n", ComputeCut(graph, where), mincut)); } if (ctrl->dbglvl&DBG_REFINE) { printf("\tMincut: %6d, NBND: %6d, NPwgts: ", mincut, nbnd); for (l=0; l<ncon; l++) printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]); printf(", LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts)); } graph->mincut = mincut; graph->nbnd = nbnd; for (i=0; i<ncon; i++) { PQueueFree(ctrl, &parts[i][0]); PQueueFree(ctrl, &parts[i][1]); } ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); idxwspacefree(ctrl, nvtxs); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function performs an edge-based FM refinement **************************************************************************/ void Mc_Serial_FM_2WayRefine(GraphType *graph, float *tpwgts, int npasses) { int i, ii, j, k; int kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, limit, tmp, cnum; idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind; idxtype *moved, *swaps, *qnum; float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal; FPQueueType parts[MAXNCON][2]; int higain, oldgain, mincut, initcut, newcut, mincutorder; float rtpwgts[MAXNCON*2]; KeyValueType *cand; int mype; MPI_Comm_rank(MPI_COMM_WORLD, &mype); nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; nvwgt = graph->nvwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; id = graph->sendind; ed = graph->recvind; npwgts = graph->gnpwgts; bndptr = graph->sendptr; bndind = graph->recvptr; moved = idxmalloc(nvtxs, "moved"); swaps = idxmalloc(nvtxs, "swaps"); qnum = idxmalloc(nvtxs, "qnum"); cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand"); limit = amin(amax(0.01*nvtxs, 25), 150); /* Initialize the queues */ for (i=0; i<ncon; i++) { FPQueueInit(&parts[i][0], nvtxs); FPQueueInit(&parts[i][1], nvtxs); } for (i=0; i<nvtxs; i++) qnum[i] = samax(ncon, nvwgt+i*ncon); origbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); for (i=0; i<ncon; i++) { rtpwgts[i] = origbal*tpwgts[i]; rtpwgts[ncon+i] = origbal*tpwgts[ncon+i]; } idxset(nvtxs, -1, moved); for (pass=0; pass<npasses; pass++) { /* Do a number of passes */ for (i=0; i<ncon; i++) { FPQueueReset(&parts[i][0]); FPQueueReset(&parts[i][1]); } mincutorder = -1; newcut = mincut = initcut = graph->mincut; for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[i]-npwgts[i]); minbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); /* Insert boundary nodes in the priority queues */ nbnd = graph->gnvtxs; for (i=0; i<nbnd; i++) { cand[i].key = id[i]-ed[i]; cand[i].val = i; } ikeysort(nbnd, cand); for (ii=0; ii<nbnd; ii++) { i = bndind[cand[ii].val]; FPQueueInsert(&parts[qnum[i]][where[i]], i, (float)(ed[i]-id[i])); } for (nswaps=0; nswaps<nvtxs; nswaps++) { Serial_SelectQueue(ncon, npwgts, rtpwgts, &from, &cnum, parts); to = (from+1)%2; if (from == -1 || (higain = FPQueueGetMax(&parts[cnum][from])) == -1) break; saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); newcut -= (ed[higain]-id[higain]); newbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); if ((newcut < mincut && newbal-origbal <= .00001) || (newcut == mincut && (newbal < minbal || (newbal == minbal && Serial_BetterBalance(ncon, npwgts, tpwgts, mindiff))))) { mincut = newcut; minbal = newbal; mincutorder = nswaps; for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[i]-npwgts[i]); } else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */ newcut += (ed[higain]-id[higain]); saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); break; } where[higain] = to; moved[higain] = nswaps; swaps[nswaps] = higain; /************************************************************** * Update the id[i]/ed[i] values of the affected nodes ***************************************************************/ SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && xadj[higain] < xadj[higain+1]) BNDDelete(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 its boundary information and queue position */ if (bndptr[k] != -1) { /* If k was a boundary vertex */ if (ed[k] == 0) { /* Not a boundary vertex any more */ BNDDelete(nbnd, bndind, bndptr, k); if (moved[k] == -1) /* Remove it if in the queues */ FPQueueDelete(&parts[qnum[k]][where[k]], k); } else { /* If it has not been moved, update its position in the queue */ if (moved[k] == -1) FPQueueUpdate(&parts[qnum[k]][where[k]], k, (float)oldgain, (float)(ed[k]-id[k])); } } else { if (ed[k] > 0) { /* It will now become a boundary vertex */ BNDInsert(nbnd, bndind, bndptr, k); if (moved[k] == -1) FPQueueInsert(&parts[qnum[k]][where[k]], k, (float)(ed[k]-id[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); saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy2(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); } } graph->mincut = mincut; graph->gnvtxs = nbnd; if (mincutorder == -1 || mincut == initcut) break; } for (i=0; i<ncon; i++) { FPQueueFree(&parts[i][0]); FPQueueFree(&parts[i][1]); } GKfree((void **)&cand, (void **)&qnum, (void **)&moved, (void **)&swaps, LTERM); return; }
void FM_Mc2WayCutRefine(ctrl_t *ctrl, graph_t *graph, real_t *ntpwgts, idx_t niter) { idx_t i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, me, limit, tmp, cnum; idx_t *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *id, *ed, *bndptr, *bndind; idx_t *moved, *swaps, *perm, *qnum; idx_t higain, mincut, initcut, newcut, mincutorder; real_t *invtvwgt, *ubfactors, *minbalv, *newbalv; real_t origbal, minbal, newbal, rgain, ffactor; rpq_t **queues; WCOREPUSH; nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; vwgt = graph->vwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; invtvwgt = graph->invtvwgt; where = graph->where; id = graph->id; ed = graph->ed; pwgts = graph->pwgts; bndptr = graph->bndptr; bndind = graph->bndind; moved = iwspacemalloc(ctrl, nvtxs); swaps = iwspacemalloc(ctrl, nvtxs); perm = iwspacemalloc(ctrl, nvtxs); qnum = iwspacemalloc(ctrl, nvtxs); ubfactors = rwspacemalloc(ctrl, ncon); newbalv = rwspacemalloc(ctrl, ncon); minbalv = rwspacemalloc(ctrl, ncon); limit = gk_min(gk_max(0.01*nvtxs, 25), 150); /* Determine a fudge factor to allow the refinement routines to get out of tight balancing constraints. */ ffactor = .5/gk_max(20, nvtxs); /* Initialize the queues */ queues = (rpq_t **)wspacemalloc(ctrl, 2*ncon*sizeof(rpq_t *)); for (i=0; i<2*ncon; i++) queues[i] = rpqCreate(nvtxs); for (i=0; i<nvtxs; i++) qnum[i] = iargmax_nrm(ncon, vwgt+i*ncon, invtvwgt); /* Determine the unbalance tolerance for each constraint. The tolerance is equal to the maximum of the original load imbalance and the user-supplied allowed tolerance. The rationale behind this approach is to allow the refinement routine to improve the cut, without having to worry about fixing load imbalance problems. The load imbalance is addressed by the balancing routines. */ origbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ctrl->ubfactors, ubfactors); for (i=0; i<ncon; i++) ubfactors[i] = (ubfactors[i] > 0 ? ctrl->ubfactors[i]+ubfactors[i] : ctrl->ubfactors[i]); IFSET(ctrl->dbglvl, METIS_DBG_REFINE, Print2WayRefineStats(ctrl, graph, ntpwgts, origbal, -2)); iset(nvtxs, -1, moved); for (pass=0; pass<niter; pass++) { /* Do a number of passes */ for (i=0; i<2*ncon; i++) rpqReset(queues[i]); mincutorder = -1; newcut = mincut = initcut = graph->mincut; minbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ubfactors, minbalv); ASSERT(ComputeCut(graph, where) == graph->mincut); ASSERT(CheckBnd(graph)); /* Insert boundary nodes in the priority queues */ nbnd = graph->nbnd; irandArrayPermute(nbnd, perm, nbnd/5, 1); for (ii=0; ii<nbnd; ii++) { i = bndind[perm[ii]]; ASSERT(ed[i] > 0 || id[i] == 0); ASSERT(bndptr[i] != -1); //rgain = 1.0*(ed[i]-id[i])/sqrt(vwgt[i*ncon+qnum[i]]+1); //rgain = (ed[i]-id[i] > 0 ? 1.0*(ed[i]-id[i])/sqrt(vwgt[i*ncon+qnum[i]]+1) : ed[i]-id[i]); rgain = ed[i]-id[i]; rpqInsert(queues[2*qnum[i]+where[i]], i, rgain); } for (nswaps=0; nswaps<nvtxs; nswaps++) { SelectQueue(graph, ctrl->pijbm, ubfactors, queues, &from, &cnum); to = (from+1)%2; if (from == -1 || (higain = rpqGetTop(queues[2*cnum+from])) == -1) break; ASSERT(bndptr[higain] != -1); newcut -= (ed[higain]-id[higain]); iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1); iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+from*ncon, 1); newbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ubfactors, newbalv); if ((newcut < mincut && newbal <= ffactor) || (newcut == mincut && (newbal < minbal || (newbal == minbal && BetterBalance2Way(ncon, minbalv, newbalv))))) { mincut = newcut; minbal = newbal; mincutorder = nswaps; rcopy(ncon, newbalv, minbalv); } else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */ newcut += (ed[higain]-id[higain]); iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+from*ncon, 1); iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1); break; } where[higain] = to; moved[higain] = nswaps; swaps[nswaps] = higain; if (ctrl->dbglvl&METIS_DBG_MOVEINFO) { printf("Moved%6"PRIDX" from %"PRIDX"(%"PRIDX") Gain:%5"PRIDX", " "Cut:%5"PRIDX", NPwgts:", higain, from, cnum, ed[higain]-id[higain], newcut); for (l=0; l<ncon; l++) printf("(%.3"PRREAL" %.3"PRREAL")", pwgts[l]*invtvwgt[l], pwgts[ncon+l]*invtvwgt[l]); printf(" %+.3"PRREAL" LB: %.3"PRREAL"(%+.3"PRREAL")\n", minbal, ComputeLoadImbalance(graph, 2, ctrl->pijbm), newbal); } /************************************************************** * Update the id[i]/ed[i] values of the affected nodes ***************************************************************/ SWAP(id[higain], ed[higain], tmp); if (ed[higain] == 0 && xadj[higain] < xadj[higain+1]) BNDDelete(nbnd, bndind, bndptr, higain); 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); /* Update its boundary information and queue position */ if (bndptr[k] != -1) { /* If k was a boundary vertex */ if (ed[k] == 0) { /* Not a boundary vertex any more */ BNDDelete(nbnd, bndind, bndptr, k); if (moved[k] == -1) /* Remove it if in the queues */ rpqDelete(queues[2*qnum[k]+where[k]], k); } else { /* If it has not been moved, update its position in the queue */ if (moved[k] == -1) { //rgain = 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1); //rgain = (ed[k]-id[k] > 0 ? // 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1) : ed[k]-id[k]); rgain = ed[k]-id[k]; rpqUpdate(queues[2*qnum[k]+where[k]], k, rgain); } } } else { if (ed[k] > 0) { /* It will now become a boundary vertex */ BNDInsert(nbnd, bndind, bndptr, k); if (moved[k] == -1) { //rgain = 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1); //rgain = (ed[k]-id[k] > 0 ? // 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1) : ed[k]-id[k]); rgain = ed[k]-id[k]; rpqInsert(queues[2*qnum[k]+where[k]], k, rgain); } } } } } /**************************************************************** * 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); iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1); iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+((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); } } graph->mincut = mincut; graph->nbnd = nbnd; IFSET(ctrl->dbglvl, METIS_DBG_REFINE, Print2WayRefineStats(ctrl, graph, ntpwgts, minbal, mincutorder)); if (mincutorder <= 0 || mincut == initcut) break; } for (i=0; i<2*ncon; i++) rpqDestroy(queues[i]); WCOREPOP; }
/************************************************************************* * This function performs an edge-based FM refinement **************************************************************************/ void Mc_Serial_Balance2Way(GraphType *graph, float *tpwgts, float lbfactor) { int i, ii, j, k, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, limit, tmp, cnum; idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind; idxtype *moved, *swaps, *qnum; float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal; FPQueueType parts[MAXNCON][2]; int higain, oldgain, mincut, newcut, mincutorder; int qsizes[MAXNCON][2]; KeyValueType *cand; nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; nvwgt = graph->nvwgt; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; id = graph->sendind; ed = graph->recvind; npwgts = graph->gnpwgts; bndptr = graph->sendptr; bndind = graph->recvptr; moved = idxmalloc(nvtxs, "moved"); swaps = idxmalloc(nvtxs, "swaps"); qnum = idxmalloc(nvtxs, "qnum"); cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand"); limit = amin(amax(0.01*nvtxs, 15), 100); /* Initialize the queues */ for (i=0; i<ncon; i++) { FPQueueInit(&parts[i][0], nvtxs); FPQueueInit(&parts[i][1], nvtxs); qsizes[i][0] = qsizes[i][1] = 0; } for (i=0; i<nvtxs; i++) { qnum[i] = samax(ncon, nvwgt+i*ncon); qsizes[qnum[i]][where[i]]++; } for (from=0; from<2; from++) { for (j=0; j<ncon; j++) { if (qsizes[j][from] == 0) { for (i=0; i<nvtxs; i++) { if (where[i] != from) continue; k = samax2(ncon, nvwgt+i*ncon); if (k == j && qsizes[qnum[i]][from] > qsizes[j][from] && nvwgt[i*ncon+qnum[i]] < 1.3*nvwgt[i*ncon+j]) { qsizes[qnum[i]][from]--; qsizes[j][from]++; qnum[i] = j; } } } } } for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[i]-npwgts[i]); minbal = origbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); newcut = mincut = graph->mincut; mincutorder = -1; idxset(nvtxs, -1, moved); /* Insert all nodes in the priority queues */ nbnd = graph->gnvtxs; for (i=0; i<nvtxs; i++) { cand[i].key = id[i]-ed[i]; cand[i].val = i; } ikeysort(nvtxs, cand); for (ii=0; ii<nvtxs; ii++) { i = cand[ii].val; FPQueueInsert(&parts[qnum[i]][where[i]], i, (float)(ed[i]-id[i])); } for (nswaps=0; nswaps<nvtxs; nswaps++) { if (minbal < lbfactor) break; Serial_SelectQueue(ncon, npwgts, tpwgts, &from, &cnum, parts); to = (from+1)%2; if (from == -1 || (higain = FPQueueGetMax(&parts[cnum][from])) == -1) break; saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); newcut -= (ed[higain]-id[higain]); newbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); if (newbal < minbal || (newbal == minbal && (newcut < mincut || (newcut == mincut && Serial_BetterBalance(ncon, npwgts, tpwgts, mindiff))))) { mincut = newcut; minbal = newbal; mincutorder = nswaps; for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[i]-npwgts[i]); } else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */ newcut += (ed[higain]-id[higain]); saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1); saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); break; } where[higain] = to; moved[higain] = nswaps; swaps[nswaps] = higain; /************************************************************** * 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) FPQueueUpdate(&parts[qnum[k]][where[k]], k, (float)(oldgain), (float)(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 (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); saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1); saxpy2(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); } } graph->mincut = mincut; graph->gnvtxs = nbnd; for (i=0; i<ncon; i++) { FPQueueFree(&parts[i][0]); FPQueueFree(&parts[i][1]); } GKfree((void **)&cand, (void **)&qnum, (void **)&moved, (void **)&swaps, LTERM); return; }
/************************************************************************* * This function performs k-way refinement **************************************************************************/ void MCRandom_KWayEdgeRefineHorizontal(CtrlType *ctrl, GraphType *graph, int nparts, float *orgubvec, int npasses) { int i, ii, iii, j, /*jj,*/ k, /*l,*/ pass, nvtxs, ncon, nmoves, nbnd, myndegrees, same; int from, me, to, oldcut, gain; idxtype *xadj, *adjncy, *adjwgt; idxtype *where, *perm, *bndptr, *bndind; EDegreeType *myedegrees; RInfoType *myrinfo; float *npwgts, *nvwgt, *minwgt, *maxwgt, maxlb, minlb, ubvec[MAXNCON], tvec[MAXNCON]; nvtxs = graph->nvtxs; ncon = graph->ncon; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; bndptr = graph->bndptr; bndind = graph->bndind; where = graph->where; npwgts = graph->npwgts; /* Setup the weight intervals of the various subdomains */ minwgt = fwspacemalloc(ctrl, nparts*ncon); maxwgt = fwspacemalloc(ctrl, nparts*ncon); /* See if the orgubvec consists of identical constraints */ maxlb = minlb = orgubvec[0]; for (i=1; i<ncon; i++) { minlb = (orgubvec[i] < minlb ? orgubvec[i] : minlb); maxlb = (orgubvec[i] > maxlb ? orgubvec[i] : maxlb); } same = (fabs(maxlb-minlb) < .01 ? 1 : 0); /* Let's not get very optimistic. Let Balancing do the work */ ComputeHKWayLoadImbalance(ncon, nparts, npwgts, ubvec); for (i=0; i<ncon; i++) ubvec[i] = amax(ubvec[i], orgubvec[i]); if (!same) { 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); } } } else { maxlb = ubvec[0]; for (i=1; i<ncon; i++) maxlb = (ubvec[i] > maxlb ? ubvec[i] : maxlb); for (i=0; i<nparts; i++) { for (j=0; j<ncon; j++) { maxwgt[i*ncon+j] = maxlb/nparts; minwgt[i*ncon+j] = 1.0/(maxlb*nparts); } } } perm = idxwspacemalloc(ctrl, nvtxs); 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("\n"); } for (pass=0; pass<npasses; pass++) { ASSERT(ComputeCut(graph, where) == graph->mincut); oldcut = graph->mincut; nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (nmoves=iii=0; iii<graph->nbnd; iii++) { ii = perm[iii]; if (ii >= nbnd) continue; i = bndind[ii]; myrinfo = graph->rinfo+i; if (myrinfo->ed >= myrinfo->id) { /* Total ED is too high */ from = where[i]; nvwgt = graph->nvwgt+i*ncon; if (myrinfo->id > 0 && 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; gain = myedegrees[k].ed - myrinfo->id; if (gain >= 0 && (AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon) || 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 ((myedegrees[j].ed > myedegrees[k].ed && (AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon) || IsHBalanceBetterFT(ncon, nparts, npwgts+from*ncon, npwgts+to*ncon, nvwgt, ubvec))) || (myedegrees[j].ed == myedegrees[k].ed && IsHBalanceBetterTT(ncon, nparts, npwgts+myedegrees[k].pid*ncon, npwgts+to*ncon, nvwgt, ubvec))) k = j; } to = myedegrees[k].pid; if (myedegrees[k].ed-myrinfo->id == 0 && !IsHBalanceBetterFT(ncon, nparts, npwgts+from*ncon, npwgts+to*ncon, nvwgt, ubvec) && AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, npwgts+from*ncon, maxwgt+from*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-myrinfo->id < 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)); 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]; } } 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 (graph->mincut == oldcut) break; } fwspacefree(ctrl, ncon*nparts); fwspacefree(ctrl, ncon*nparts); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * This function performs an edge-based FM refinement **************************************************************************/ void MocGeneral2WayBalance(CtrlType *ctrl, GraphType *graph, float *tpwgts, float lbfactor) { int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, me, limit, tmp, cnum; idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind; idxtype *moved, *swaps, *perm, *qnum; float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal; PQueueType parts[MAXNCON][2]; int higain, oldgain, mincut, newcut, mincutorder; int qsizes[MAXNCON][2]; 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); /* 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); qsizes[i][0] = qsizes[i][1] = 0; } for (i=0; i<nvtxs; i++) { qnum[i] = samax(ncon, nvwgt+i*ncon); qsizes[qnum[i]][where[i]]++; } /* printf("Weight Distribution: \t"); for (i=0; i<ncon; i++) printf(" [%d %d]", qsizes[i][0], qsizes[i][1]); printf("\n"); */ for (from=0; from<2; from++) { for (j=0; j<ncon; j++) { if (qsizes[j][from] == 0) { for (i=0; i<nvtxs; i++) { if (where[i] != from) continue; k = samax2(ncon, nvwgt+i*ncon); if (k == j && qsizes[qnum[i]][from] > qsizes[j][from] && nvwgt[i*ncon+qnum[i]] < 1.3*nvwgt[i*ncon+j]) { qsizes[qnum[i]][from]--; qsizes[j][from]++; qnum[i] = j; } } } } } /* printf("Weight Distribution (after):\t "); for (i=0; i<ncon; i++) printf(" [%d %d]", qsizes[i][0], qsizes[i][1]); printf("\n"); */ for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[0]-npwgts[i]); minbal = origbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); 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: %.3f [B]\n", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut, origbal); } 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 (minbal < lbfactor) break; SelectQueue(ncon, npwgts, tpwgts, &from, &cnum, parts); 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]); newbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts); if (newbal < minbal || (newbal == minbal && (newcut < mincut || (newcut == mincut && BetterBalance(ncon, npwgts, tpwgts, mindiff))))) { mincut = newcut; minbal = newbal; mincutorder = nswaps; for (i=0; i<ncon; i++) mindiff[i] = fabs(tpwgts[0]-npwgts[i]); } 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 (l=0; l<ncon; l++) printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]); printf(", %.3f LB: %.3f\n", minbal, newbal); } /************************************************************** * 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 (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 (l=0; l<ncon; l++) printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]); printf("], LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts)); } 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); }
/************************************************************************* * This function performs k-way refinement **************************************************************************/ void Random_KWayEdgeRefineMConn(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor, int npasses, int ffactor) { int i, ii, iii, j, jj, k, l, pass, nvtxs, nmoves, nbnd, tvwgt, myndegrees; int from, me, to, oldcut, vwgt, gain; int maxndoms, nadd; idxtype *xadj, *adjncy, *adjwgt; idxtype *where, *pwgts, *perm, *bndptr, *bndind, *minwgt, *maxwgt, *itpwgts; idxtype *phtable, *pmat, *pmatptr, *ndoms; EDegreeType *myedegrees; RInfoType *myrinfo; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; bndptr = graph->bndptr; bndind = graph->bndind; 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); 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); maxndoms = ndoms[idxamax(nparts, ndoms)]; oldcut = graph->mincut; nbnd = graph->nbnd; RandomPermute(nbnd, perm, 1); for (nmoves=iii=0; iii<graph->nbnd; iii++) { ii = perm[iii]; if (ii >= nbnd) continue; i = bndind[ii]; myrinfo = graph->rinfo+i; if (myrinfo->ed >= myrinfo->id) { /* Total ED is too high */ from = where[i]; vwgt = graph->vwgt[i]; if (myrinfo->id > 0 && 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; if (nadd == 0) phtable[to] = 2; } /* Find the first valid move */ j = myrinfo->id; for (k=0; k<myndegrees; k++) { to = myedegrees[k].pid; if (!phtable[to]) continue; gain = myedegrees[k].ed-j; /* j = myrinfo->id. Allow good nodes to move */ if (pwgts[to]+vwgt <= maxwgt[to]+ffactor*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 (!phtable[to]) continue; 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 ||*/ phtable[myedegrees[k].pid] == 2 || 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 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-myrinfo->id < 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)); 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 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]; } 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: %5d, Vol: %5d, %d\n", pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], 1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nbnd, nmoves, graph->mincut, ComputeVolume(graph, where), idxsum(nparts, ndoms))); if (graph->mincut == oldcut) break; } idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nparts); idxwspacefree(ctrl, nvtxs); }
/************************************************************************* * 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 moves a collection of vertices and updates their rinfo **************************************************************************/ void MoveGroupMConn(CtrlType *ctrl, GraphType *graph, idxtype *ndoms, idxtype *pmat, int nparts, int to, int nind, idxtype *ind) { int i, ii, iii, j, jj, k, l, nvtxs, nbnd, myndegrees; int from, me; idxtype *xadj, *adjncy, *adjwgt; idxtype *where, *bndptr, *bndind; EDegreeType *myedegrees; RInfoType *myrinfo; nvtxs = graph->nvtxs; xadj = graph->xadj; adjncy = graph->adjncy; adjwgt = graph->adjwgt; where = graph->where; bndptr = graph->bndptr; bndind = graph->bndind; nbnd = graph->nbnd; for (iii=0; iii<nind; iii++) { i = ind[iii]; from = where[i]; myrinfo = graph->rinfo+i; if (myrinfo->edegrees == NULL) { myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree; ctrl->wspace.cdegree += xadj[i+1]-xadj[i]; myrinfo->ndegrees = 0; } myedegrees = myrinfo->edegrees; /* find the location of 'to' in myrinfo or create it if it is not there */ for (k=0; k<myrinfo->ndegrees; k++) { if (myedegrees[k].pid == to) break; } if (k == myrinfo->ndegrees) { myedegrees[k].pid = to; myedegrees[k].ed = 0; myrinfo->ndegrees++; } graph->mincut -= myedegrees[k].ed-myrinfo->id; /* 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 (pmat[to*nparts+from] == 0) ndoms[to]--; /* Update where, weight, and ID/ED information of the vertex you moved */ 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-myrinfo->id < 0 && bndptr[i] != -1) 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)); 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 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 (pmat[from*nparts+me] == 0) ndoms[from]--; if (pmat[me*nparts+to] == 0) ndoms[me]++; if (pmat[to*nparts+me] == 0) ndoms[to]++; pmat[me*nparts+to] += adjwgt[j]; pmat[to*nparts+me] += adjwgt[j]; } ASSERT(CheckRInfo(myrinfo)); } ASSERT(CheckRInfo(graph->rinfo+i)); } graph->nbnd = nbnd; }
void FM_2WayNodeRefine2Sided(ctrl_t *ctrl, graph_t *graph, idx_t niter) { idx_t i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind; idx_t *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr; idx_t *mptr, *mind, *moved, *swaps; rpq_t *queues[2]; nrinfo_t *rinfo; idx_t higain, oldgain, mincut, initcut, mincutorder; idx_t pass, to, other, limit; idx_t badmaxpwgt, mindiff, newdiff; idx_t u[2], g[2]; real_t mult; WCOREPUSH; 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; queues[0] = rpqCreate(nvtxs); queues[1] = rpqCreate(nvtxs); moved = iwspacemalloc(ctrl, nvtxs); swaps = iwspacemalloc(ctrl, nvtxs); mptr = iwspacemalloc(ctrl, nvtxs+1); mind = iwspacemalloc(ctrl, 2*nvtxs); mult = 0.5*ctrl->ubfactors[0]; badmaxpwgt = (idx_t)(mult*(pwgts[0]+pwgts[1]+pwgts[2])); IFSET(ctrl->dbglvl, METIS_DBG_REFINE, printf("Partitions-N2: [%6"PRIDX" %6"PRIDX"] Nv-Nb[%6"PRIDX" %6"PRIDX"]. ISep: %6"PRIDX"\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut)); for (pass=0; pass<niter; pass++) { iset(nvtxs, -1, moved); rpqReset(queues[0]); rpqReset(queues[1]); mincutorder = -1; initcut = mincut = graph->mincut; nbnd = graph->nbnd; /* use the swaps array in place of the traditional perm array to save memory */ my_irandArrayPermute_r(nbnd, swaps, nbnd, 1, &ctrl->curseed); for (ii=0; ii<nbnd; ii++) { i = bndind[swaps[ii]]; ASSERT(where[i] == 2); rpqInsert(queues[0], i, vwgt[i]-rinfo[i].edegrees[1]); rpqInsert(queues[1], i, vwgt[i]-rinfo[i].edegrees[0]); } ASSERT(CheckNodeBnd(graph, nbnd)); ASSERT(CheckNodePartitionParams(graph)); limit = (ctrl->compress ? gk_min(5*nbnd, 400) : gk_min(2*nbnd, 300)); /****************************************************** * Get into the FM loop *******************************************************/ mptr[0] = nmind = 0; mindiff = iabs(pwgts[0]-pwgts[1]); to = (pwgts[0] < pwgts[1] ? 0 : 1); for (nswaps=0; nswaps<nvtxs; nswaps++) { u[0] = rpqSeeTopVal(queues[0]); u[1] = rpqSeeTopVal(queues[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)); if (pwgts[to]+vwgt[u[to]] > badmaxpwgt) to = (to+1)%2; } else if (u[0] == -1 && u[1] == -1) { break; } else if (u[0] != -1 && pwgts[0]+vwgt[u[0]] <= badmaxpwgt) { to = 0; } else if (u[1] != -1 && pwgts[1]+vwgt[u[1]] <= badmaxpwgt) { to = 1; } else break; other = (to+1)%2; higain = rpqGetTop(queues[to]); if (moved[higain] == -1) /* Delete if it was in the separator originally */ rpqDelete(queues[other], higain); ASSERT(bndptr[higain] != -1); /* The following check is to ensure we break out if there is a posibility of over-running the mind array. */ if (nmind + xadj[higain+1]-xadj[higain] >= 2*nvtxs-1) break; pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]); newdiff = iabs(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 > 2*limit || (nswaps - mincutorder > limit && pwgts[2] > 1.10*mincut)) { 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)) rpqUpdate(queues[other], k, oldgain-vwgt[higain]); } else if (where[k] == other) { /* This vertex is pulled into the separator */ ASSERTP(bndptr[k] == -1, ("%"PRIDX" %"PRIDX" %"PRIDX"\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)) rpqUpdate(queues[to], kk, oldgain+vwgt[k]); } } /* Insert the new vertex into the priority queue. Only one side! */ if (moved[k] == -1) { rpqInsert(queues[to], k, vwgt[k]-edegrees[other]); moved[k] = -(2+to); } } } mptr[nswaps+1] = nmind; IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO, printf("Moved %6"PRIDX" to %3"PRIDX", Gain: %5"PRIDX" [%5"PRIDX"] [%4"PRIDX" %4"PRIDX"] \t[%5"PRIDX" %5"PRIDX" %5"PRIDX"]\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, METIS_DBG_REFINE, printf("\tMinimum sep: %6"PRIDX" at %5"PRIDX", PWGTS: [%6"PRIDX" %6"PRIDX"], NBND: %6"PRIDX"\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd)); graph->mincut = mincut; graph->nbnd = nbnd; if (mincutorder == -1 || mincut >= initcut) break; } rpqDestroy(queues[0]); rpqDestroy(queues[1]); WCOREPOP; }