/*************************************************************************
* This function computes the subdomain graph
**************************************************************************/
void EliminateSubDomainEdges(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts)
{
int i, ii, j, k, me, other, nvtxs, total, max, avg, totalout, nind, ncand, ncand2, target, target2, nadd;
int min, move, cpwgt, tvwgt;
idxtype *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *maxpwgt, *pmat, *ndoms, *mypmat, *otherpmat, *ind;
KeyValueType *cand, *cand2;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
adjwgt = graph->adjwgt;
where = graph->where;
pwgts = graph->pwgts; /* We assume that this is properly initialized */
maxpwgt = idxwspacemalloc(ctrl, nparts);
ndoms = idxwspacemalloc(ctrl, nparts);
otherpmat = idxwspacemalloc(ctrl, nparts);
ind = idxwspacemalloc(ctrl, nvtxs);
pmat = ctrl->wspace.pmat;
cand = (KeyValueType *)GKmalloc(nparts*sizeof(KeyValueType), "EliminateSubDomainEdges: cand");
cand2 = (KeyValueType *)GKmalloc(nparts*sizeof(KeyValueType), "EliminateSubDomainEdges: cand");
/* Compute the pmat matrix and ndoms */
ComputeSubDomainGraph(graph, nparts, pmat, ndoms);
/* Compute the maximum allowed weight for each domain */
tvwgt = idxsum(nparts, pwgts);
for (i=0; i<nparts; i++)
maxpwgt[i] = 1.25*tpwgts[i]*tvwgt;
/* Get into the loop eliminating subdomain connections */
for (;;) {
total = idxsum(nparts, ndoms);
avg = total/nparts;
max = ndoms[idxamax(nparts, ndoms)];
/* printf("Adjacent Subdomain Stats: Total: %3d, Max: %3d, Avg: %3d [%5d]\n", total, max, avg, idxsum(nparts*nparts, pmat)); */
if (max < 1.4*avg)
break;
me = idxamax(nparts, ndoms);
mypmat = pmat + me*nparts;
totalout = idxsum(nparts, mypmat);
/*printf("Me: %d, TotalOut: %d,\n", me, totalout);*/
/* Sort the connections according to their cut */
for (ncand2=0, i=0; i<nparts; i++) {
if (mypmat[i] > 0) {
cand2[ncand2].key = mypmat[i];
cand2[ncand2++].val = i;
}
}
ikeysort(ncand2, cand2);
move = 0;
for (min=0; min<ncand2; min++) {
if (cand2[min].key > totalout/(2*ndoms[me]))
break;
other = cand2[min].val;
/*printf("\tMinOut: %d to %d\n", mypmat[other], other);*/
idxset(nparts, 0, otherpmat);
/* Go and find the vertices in 'other' that are connected in 'me' */
for (nind=0, i=0; i<nvtxs; i++) {
if (where[i] == other) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (where[adjncy[j]] == me) {
ind[nind++] = i;
break;
}
}
}
}
/* Go and construct the otherpmat to see where these nind vertices are connected to */
for (cpwgt=0, ii=0; ii<nind; ii++) {
i = ind[ii];
cpwgt += vwgt[i];
for (j=xadj[i]; j<xadj[i+1]; j++)
otherpmat[where[adjncy[j]]] += adjwgt[j];
}
otherpmat[other] = 0;
for (ncand=0, i=0; i<nparts; i++) {
if (otherpmat[i] > 0) {
cand[ncand].key = -otherpmat[i];
cand[ncand++].val = i;
}
}
ikeysort(ncand, cand);
/*
* Go through and the select the first domain that is common with 'me', and
* does not increase the ndoms[target] higher than my ndoms, subject to the
* maxpwgt constraint. Traversal is done from the mostly connected to the least.
*/
target = target2 = -1;
for (i=0; i<ncand; i++) {
k = cand[i].val;
if (mypmat[k] > 0) {
if (pwgts[k] + cpwgt > maxpwgt[k]) /* Check if balance will go off */
continue;
for (j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && ndoms[j] >= ndoms[me]-1 && pmat[nparts*j+k] == 0)
break;
}
if (j == nparts) { /* No bad second level effects */
for (nadd=0, j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && pmat[nparts*k+j] == 0)
nadd++;
}
/*printf("\t\tto=%d, nadd=%d, %d\n", k, nadd, ndoms[k]);*/
if (target2 == -1 && ndoms[k]+nadd < ndoms[me]) {
target2 = k;
}
if (nadd == 0) {
target = k;
break;
}
}
}
}
if (target == -1 && target2 != -1)
target = target2;
if (target == -1) {
/* printf("\t\tCould not make the move\n");*/
continue;
}
/*printf("\t\tMoving to %d\n", target);*/
/* Update the partition weights */
INC_DEC(pwgts[target], pwgts[other], cpwgt);
MoveGroupMConn(ctrl, graph, ndoms, pmat, nparts, target, nind, ind);
move = 1;
break;
}
if (move == 0)
break;
}
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
GKfree((void**)&cand, &cand2, LTERM);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Random_KWayEdgeRefine(CtrlType *ctrl, GraphType *graph, int nparts, floattype *tpwgts, floattype 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;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where, *pwgts, *perm, *bndptr, *bndind, *minwgt, *maxwgt, *itpwgts;
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;
/* 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);
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;
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]+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 ((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 < 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;
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\t[%6d %6d], Balance: %5.3f, Nb: %6d. Nmoves: %5d, Cut: %6d, Vol: %6d\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)));
if (graph->mincut == oldcut)
break;
}
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function projects a partition, and at the same time computes the
* parameters for refinement.
**************************************************************************/
void ProjectVolKWayPartition(CtrlType *ctrl, GraphType *graph, int nparts)
{
int i, j, k, nvtxs, me, other, istart, iend, ndegrees;
idxtype *xadj, *adjncy, *adjwgt, *adjwgtsum;
idxtype *cmap, *where;
idxtype *cwhere;
GraphType *cgraph;
VRInfoType *crinfo, *rinfo, *myrinfo;
VEDegreeType *myedegrees;
idxtype *htable;
cgraph = graph->coarser;
cwhere = cgraph->where;
crinfo = cgraph->vrinfo;
nvtxs = graph->nvtxs;
cmap = graph->cmap;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
AllocateVolKWayPartitionMemory(ctrl, graph, nparts);
where = graph->where;
rinfo = graph->vrinfo;
/* Go through and project partition and compute id/ed for the nodes */
for (i=0; i<nvtxs; i++) {
k = cmap[i];
where[i] = cwhere[k];
cmap[i] = crinfo[k].ed; /* For optimization */
}
htable = idxset(nparts, -1, idxwspacemalloc(ctrl, nparts));
ctrl->wspace.cdegree = 0;
for (i=0; i<nvtxs; i++) {
me = where[i];
myrinfo = rinfo+i;
myrinfo->id = myrinfo->ed = myrinfo->nid = myrinfo->ndegrees = 0;
myrinfo->edegrees = NULL;
myrinfo->id = adjwgtsum[i];
myrinfo->nid = xadj[i+1]-xadj[i];
if (cmap[i] > 0) { /* If it is an interface node. Note cmap[i] = crinfo[cmap[i]].ed */
istart = xadj[i];
iend = xadj[i+1];
myedegrees = myrinfo->edegrees = ctrl->wspace.vedegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += iend-istart;
ndegrees = 0;
for (j=istart; j<iend; j++) {
other = where[adjncy[j]];
if (me != other) {
myrinfo->ed += adjwgt[j];
myrinfo->nid--;
if ((k = htable[other]) == -1) {
htable[other] = ndegrees;
myedegrees[ndegrees].gv = 0;
myedegrees[ndegrees].pid = other;
myedegrees[ndegrees].ed = adjwgt[j];
myedegrees[ndegrees++].ned = 1;
}
else {
myedegrees[k].ed += adjwgt[j];
myedegrees[k].ned++;
}
}
}
myrinfo->id -= myrinfo->ed;
/* Remove space for edegrees if it was interior */
if (myrinfo->ed == 0) {
myrinfo->edegrees = NULL;
ctrl->wspace.cdegree -= iend-istart;
}
else {
myrinfo->ndegrees = ndegrees;
for (j=0; j<ndegrees; j++)
htable[myedegrees[j].pid] = -1;
}
}
}
ComputeKWayVolGains(ctrl, graph, nparts);
idxcopy(nparts, cgraph->pwgts, graph->pwgts);
graph->mincut = cgraph->mincut;
FreeGraph(graph->coarser);
graph->coarser = NULL;
idxwspacefree(ctrl, nparts);
}
/*************************************************************************
* 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, 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);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void MCGreedy_KWayEdgeBalanceHorizontal(CtrlType *ctrl, GraphType *graph, int nparts,
float *ubvec, int npasses)
{
int i, ii, j, k, 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 creates the coarser graph
**************************************************************************/
void CreateCoarseGraph_NVW(CtrlType *ctrl, GraphType *graph, int cnvtxs, idxtype *match, idxtype *perm)
{
int i, j, jj, k, kk, l, m, istart, iend, nvtxs, nedges, ncon, cnedges, v, u, mask;
idxtype *xadj, *adjncy, *adjwgtsum, *auxadj;
idxtype *cmap, *htable;
idxtype *cxadj, *cvwgt, *cadjncy, *cadjwgt, *cadjwgtsum;
float *nvwgt, *cnvwgt;
GraphType *cgraph;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgtsum = graph->adjwgtsum;
cmap = graph->cmap;
/* Initialize the coarser graph */
cgraph = SetUpCoarseGraph(graph, cnvtxs, 0);
cxadj = cgraph->xadj;
cvwgt = cgraph->vwgt;
cnvwgt = cgraph->nvwgt;
cadjwgtsum = cgraph->adjwgtsum;
cadjncy = cgraph->adjncy;
cadjwgt = cgraph->adjwgt;
iend = xadj[nvtxs];
auxadj = ctrl->wspace.auxcore;
memcpy(auxadj, adjncy, iend*sizeof(idxtype));
for (i=0; i<iend; i++)
auxadj[i] = cmap[auxadj[i]];
mask = HTLENGTH;
htable = idxset(mask+1, -1, idxwspacemalloc(ctrl, mask+1));
cxadj[0] = cnvtxs = cnedges = 0;
for (i=0; i<nvtxs; i++) {
v = perm[i];
if (cmap[v] != cnvtxs)
continue;
u = match[v];
cvwgt[cnvtxs] = 1;
cadjwgtsum[cnvtxs] = adjwgtsum[v];
nedges = 0;
istart = xadj[v];
iend = xadj[v+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
kk = k&mask;
if ((m = htable[kk]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = 1;
htable[kk] = nedges++;
}
else if (cadjncy[m] == k) {
cadjwgt[m]++;
}
else {
for (jj=0; jj<nedges; jj++) {
if (cadjncy[jj] == k) {
cadjwgt[jj]++;
break;
}
}
if (jj == nedges) {
cadjncy[nedges] = k;
cadjwgt[nedges++] = 1;
}
}
}
if (v != u) {
cvwgt[cnvtxs]++;
cadjwgtsum[cnvtxs] += adjwgtsum[u];
istart = xadj[u];
iend = xadj[u+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
kk = k&mask;
if ((m = htable[kk]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = 1;
htable[kk] = nedges++;
}
else if (cadjncy[m] == k) {
cadjwgt[m]++;
}
else {
for (jj=0; jj<nedges; jj++) {
if (cadjncy[jj] == k) {
cadjwgt[jj]++;
break;
}
}
if (jj == nedges) {
cadjncy[nedges] = k;
cadjwgt[nedges++] = 1;
}
}
}
/* Remove the contracted adjacency weight */
jj = htable[cnvtxs&mask];
if (jj >= 0 && cadjncy[jj] != cnvtxs) {
for (jj=0; jj<nedges; jj++) {
if (cadjncy[jj] == cnvtxs)
break;
}
}
if (jj >= 0 && cadjncy[jj] == cnvtxs) { /* This 2nd check is needed for non-adjacent matchings */
cadjwgtsum[cnvtxs] -= cadjwgt[jj];
cadjncy[jj] = cadjncy[--nedges];
cadjwgt[jj] = cadjwgt[nedges];
}
}
ASSERTP(cadjwgtsum[cnvtxs] == idxsum(nedges, cadjwgt), ("%d %d %d %d %d\n", cnvtxs, cadjwgtsum[cnvtxs], idxsum(nedges, cadjwgt), adjwgtsum[u], adjwgtsum[v]));
for (j=0; j<nedges; j++)
htable[cadjncy[j]&mask] = -1; /* Zero out the htable */
htable[cnvtxs&mask] = -1;
cnedges += nedges;
cxadj[++cnvtxs] = cnedges;
cadjncy += nedges;
cadjwgt += nedges;
}
cgraph->nedges = cnedges;
ReAdjustMemory(graph, cgraph, 0);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
idxwspacefree(ctrl, mask+1);
}
/*************************************************************************
* This function projects a partition, and at the same time computes the
* parameters for refinement.
**************************************************************************/
void MocProjectKWayPartition(CtrlType *ctrl, GraphType *graph, int nparts)
{
int i, j, k, nvtxs, nbnd, me, other, istart, iend, ndegrees;
idxtype *xadj, *adjncy, *adjwgt, *adjwgtsum;
idxtype *cmap, *where, *bndptr, *bndind;
idxtype *cwhere;
GraphType *cgraph;
RInfoType *crinfo, *rinfo, *myrinfo;
EDegreeType *myedegrees;
idxtype *htable;
cgraph = graph->coarser;
cwhere = cgraph->where;
crinfo = cgraph->rinfo;
nvtxs = graph->nvtxs;
cmap = graph->cmap;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
MocAllocateKWayPartitionMemory(ctrl, graph, nparts);
where = graph->where;
rinfo = graph->rinfo;
bndind = graph->bndind;
bndptr = idxset(nvtxs, -1, graph->bndptr);
/* Go through and project partition and compute id/ed for the nodes */
for (i=0; i<nvtxs; i++) {
k = cmap[i];
where[i] = cwhere[k];
cmap[i] = crinfo[k].ed; /* For optimization */
}
htable = idxset(nparts, -1, idxwspacemalloc(ctrl, nparts));
ctrl->wspace.cdegree = 0;
for (nbnd=0, i=0; i<nvtxs; i++) {
me = where[i];
myrinfo = rinfo+i;
myrinfo->id = myrinfo->ed = myrinfo->ndegrees = 0;
myrinfo->edegrees = NULL;
myrinfo->id = adjwgtsum[i];
if (cmap[i] > 0) { /* If it is an interface node. Note cmap[i] = crinfo[cmap[i]].ed */
istart = xadj[i];
iend = xadj[i+1];
myedegrees = myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += iend-istart;
ndegrees = 0;
for (j=istart; j<iend; j++) {
other = where[adjncy[j]];
if (me != other) {
myrinfo->ed += adjwgt[j];
if ((k = htable[other]) == -1) {
htable[other] = ndegrees;
myedegrees[ndegrees].pid = other;
myedegrees[ndegrees++].ed = adjwgt[j];
}
else {
myedegrees[k].ed += adjwgt[j];
}
}
}
myrinfo->id -= myrinfo->ed;
/* Remove space for edegrees if it was interior */
if (myrinfo->ed == 0) {
myrinfo->edegrees = NULL;
ctrl->wspace.cdegree -= iend-istart;
}
else {
if (myrinfo->ed-myrinfo->id >= 0)
BNDInsert(nbnd, bndind, bndptr, i);
myrinfo->ndegrees = ndegrees;
for (j=0; j<ndegrees; j++)
htable[myedegrees[j].pid] = -1;
}
}
}
scopy(graph->ncon*nparts, cgraph->npwgts, graph->npwgts);
graph->mincut = cgraph->mincut;
graph->nbnd = nbnd;
FreeGraph(graph->coarser);
graph->coarser = NULL;
idxwspacefree(ctrl, nparts);
ASSERT(CheckBnd2(graph));
}
/*************************************************************************
* This function takes a graph and a bisection and splits it into two graphs.
* It relies on the fact that adjwgt is all set to 1.
**************************************************************************/
int SplitGraphOrderCC(CtrlType *ctrl, GraphType *graph, GraphType *sgraphs, int ncmps, idxtype *cptr, idxtype *cind)
{
int i, ii, iii, j, k, l, istart, iend, mypart, nvtxs, snvtxs, snedges;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *adjwgtsum, *label, *where, *bndptr, *bndind;
idxtype *sxadj, *svwgt, *sadjncy, *sadjwgt, *sadjwgtsum, *slabel;
idxtype *rename;
idxtype *auxadjncy, *auxadjwgt;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->SplitTmr));
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
label = graph->label;
where = graph->where;
bndptr = graph->bndptr;
bndind = graph->bndind;
ASSERT(bndptr != NULL);
/* Go and use bndptr to also mark the boundary nodes in the two partitions */
for (ii=0; ii<graph->nbnd; ii++) {
i = bndind[ii];
for (j=xadj[i]; j<xadj[i+1]; j++)
bndptr[adjncy[j]] = 1;
}
rename = idxwspacemalloc(ctrl, nvtxs);
/* Go and split the graph a component at a time */
for (iii=0; iii<ncmps; iii++) {
RandomPermute(cptr[iii+1]-cptr[iii], cind+cptr[iii], 0);
snvtxs = snedges = 0;
for (j=cptr[iii]; j<cptr[iii+1]; j++) {
i = cind[j];
rename[i] = snvtxs++;
snedges += xadj[i+1]-xadj[i];
}
SetUpSplitGraph(graph, sgraphs+iii, snvtxs, snedges);
sxadj = sgraphs[iii].xadj;
svwgt = sgraphs[iii].vwgt;
sadjwgtsum = sgraphs[iii].adjwgtsum;
sadjncy = sgraphs[iii].adjncy;
sadjwgt = sgraphs[iii].adjwgt;
slabel = sgraphs[iii].label;
snvtxs = snedges = sxadj[0] = 0;
for (ii=cptr[iii]; ii<cptr[iii+1]; ii++) {
i = cind[ii];
istart = xadj[i];
iend = xadj[i+1];
if (bndptr[i] == -1) { /* This is an interior vertex */
auxadjncy = sadjncy + snedges - istart;
auxadjwgt = sadjwgt + snedges - istart;
for(j=istart; j<iend; j++)
auxadjncy[j] = adjncy[j];
snedges += iend-istart;
}
else {
l = snedges;
for (j=istart; j<iend; j++) {
k = adjncy[j];
if (where[k] != 2)
sadjncy[l++] = k;
}
snedges = l;
}
svwgt[snvtxs] = vwgt[i];
sadjwgtsum[snvtxs] = snedges-sxadj[snvtxs];
slabel[snvtxs] = label[i];
sxadj[++snvtxs] = snedges;
}
idxset(snedges, 1, sadjwgt);
for (i=0; i<snedges; i++)
sadjncy[i] = rename[sadjncy[i]];
sgraphs[iii].nvtxs = snvtxs;
sgraphs[iii].nedges = snedges;
sgraphs[iii].ncon = 1;
if (snvtxs < MMDSWITCH)
sgraphs[iii].adjwgt = NULL; /* A marker to call MMD on the driver */
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->SplitTmr));
idxwspacefree(ctrl, nvtxs);
return ncmps;
}
/*************************************************************************
* This function performs a node-based FM refinement. This is the
* one-way version
**************************************************************************/
void FM_2WayNodeRefine_OneSided(CtrlType *ctrl, GraphType *graph, 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, *swaps, *perm;
PQueueType parts;
NRInfoType *rinfo;
int higain, oldgain, mincut, initcut, mincutorder;
int pass, to, other, limit;
int 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);
mind = idxwspacemalloc(ctrl, nvtxs+1);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("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 = 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 = 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,
printf("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,
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 (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut))
break;
}
PQueueFree(ctrl, &parts);
idxwspacefree(ctrl, nvtxs+1);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs a node-based FM refinement
**************************************************************************/
void FM_2WayNodeBalance(CtrlType *ctrl, GraphType *graph, float ubfactor)
{
int 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;
int higain, oldgain;
int 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 (abs(pwgts[0]-pwgts[1]) < (int)((ubfactor-1.0)*(pwgts[0]+pwgts[1])))
return;
if (abs(pwgts[0]-pwgts[1]) < 3*idxsum(nvtxs, vwgt)/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,
printf("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,
printf("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,
printf("\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 performs a node-based FM refinement
**************************************************************************/
void FM_2WayNodeRefineEqWgt(CtrlType *ctrl, GraphType *graph, 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 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));
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 = 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 = 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,
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 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 = 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 takes a bisection and constructs a minimum weight vertex
* separator out of it. It uses an unweighted minimum-cover algorithm
* followed by node-based separator refinement.
**************************************************************************/
void ConstructMinCoverSeparator0(CtrlType *ctrl, GraphType *graph, float ubfactor)
{
int i, ii, j, jj, k, l, nvtxs, nbnd, bnvtxs[3], bnedges[2], csize;
idxtype *xadj, *adjncy, *bxadj, *badjncy;
idxtype *where, *bndind, *bndptr, *vmap, *ivmap, *cover;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
nbnd = graph->nbnd;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
vmap = idxwspacemalloc(ctrl, nvtxs);
ivmap = idxwspacemalloc(ctrl, nbnd);
cover = idxwspacemalloc(ctrl, nbnd);
if (nbnd > 0) {
/* Go through the boundary and determine the sizes of the bipartite graph */
bnvtxs[0] = bnvtxs[1] = bnedges[0] = bnedges[1] = 0;
for (i=0; i<nbnd; i++) {
j = bndind[i];
k = where[j];
if (xadj[j+1]-xadj[j] > 0) {
bnvtxs[k]++;
bnedges[k] += xadj[j+1]-xadj[j];
}
}
bnvtxs[2] = bnvtxs[0]+bnvtxs[1];
bnvtxs[1] = bnvtxs[0];
bnvtxs[0] = 0;
bxadj = idxmalloc(bnvtxs[2]+1, "ConstructMinCoverSeparator: bxadj");
badjncy = idxmalloc(bnedges[0]+bnedges[1]+1, "ConstructMinCoverSeparator: badjncy");
/* Construct the ivmap and vmap */
ASSERT(idxset(nvtxs, -1, vmap) == vmap);
for (i=0; i<nbnd; i++) {
j = bndind[i];
k = where[j];
if (xadj[j+1]-xadj[j] > 0) {
vmap[j] = bnvtxs[k];
ivmap[bnvtxs[k]++] = j;
}
}
/* OK, go through and put the vertices of each part starting from 0 */
bnvtxs[1] = bnvtxs[0];
bnvtxs[0] = 0;
bxadj[0] = l = 0;
for (k=0; k<2; k++) {
for (ii=0; ii<nbnd; ii++) {
i = bndind[ii];
if (where[i] == k && xadj[i] < xadj[i+1]) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
jj = adjncy[j];
if (where[jj] != k) {
ASSERT(bndptr[jj] != -1);
ASSERTP(vmap[jj] != -1, ("%d %d %d\n", jj, vmap[jj], graph->bndptr[jj]));
badjncy[l++] = vmap[jj];
}
}
bxadj[++bnvtxs[k]] = l;
}
}
}
ASSERT(l <= bnedges[0]+bnedges[1]);
MinCover(bxadj, badjncy, bnvtxs[0], bnvtxs[1], cover, &csize);
IFSET(ctrl->dbglvl, DBG_SEPINFO,
printf("Nvtxs: %6d, [%5d %5d], Cut: %6d, SS: [%6d %6d], Cover: %6d\n", nvtxs, graph->pwgts[0], graph->pwgts[1], graph->mincut, bnvtxs[0], bnvtxs[1]-bnvtxs[0], csize));
for (i=0; i<csize; i++) {
j = ivmap[cover[i]];
where[j] = 2;
}
GKfree(&bxadj, &badjncy, LTERM);
for (i=0; i<nbnd; i++)
bndptr[bndind[i]] = -1;
for (nbnd=i=0; i<nvtxs; i++) {
if (where[i] == 2) {
bndind[nbnd] = i;
bndptr[i] = nbnd++;
}
}
}
else {
IFSET(ctrl->dbglvl, DBG_SEPINFO,
printf("Nvtxs: %6d, [%5d %5d], Cut: %6d, SS: [%6d %6d], Cover: %6d\n", nvtxs, graph->pwgts[0], graph->pwgts[1], graph->mincut, 0, 0, 0));
}
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, graph->nbnd);
idxwspacefree(ctrl, graph->nbnd);
graph->nbnd = nbnd;
ASSERT(IsSeparable(graph));
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Greedy_KWayEdgeBalance(CtrlType *ctrl, GraphType *graph, int nparts, floattype *tpwgts, floattype 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;
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 [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;
}
nmoves = 0;
for (;;) {
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;
for (k=0; k<myndegrees; k++) {
to = myedegrees[k].pid;
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 (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 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 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\n",
pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)],
1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nbnd, nmoves, graph->mincut));
}
PQueueFree(ctrl, &queue);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function takes a graph and a bisection and splits it into two graphs.
**************************************************************************/
void SplitGraphPart(CtrlType *ctrl, GraphType *graph, GraphType *lgraph, GraphType *rgraph)
{
int i, j, k, kk, l, istart, iend, mypart, nvtxs, ncon, snvtxs[2], snedges[2], sum;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *adjwgtsum, *label, *where, *bndptr;
idxtype *sxadj[2], *svwgt[2], *sadjncy[2], *sadjwgt[2], *sadjwgtsum[2], *slabel[2];
idxtype *rename;
idxtype *auxadjncy, *auxadjwgt;
floattype *nvwgt, *snvwgt[2], *npwgts;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->SplitTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
vwgt = graph->vwgt;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
label = graph->label;
where = graph->where;
bndptr = graph->bndptr;
npwgts = graph->npwgts;
ASSERT(bndptr != NULL);
rename = idxwspacemalloc(ctrl, nvtxs);
snvtxs[0] = snvtxs[1] = snedges[0] = snedges[1] = 0;
for (i=0; i<nvtxs; i++) {
k = where[i];
rename[i] = snvtxs[k]++;
snedges[k] += xadj[i+1]-xadj[i];
}
SetUpSplitGraph(graph, lgraph, snvtxs[0], snedges[0]);
sxadj[0] = lgraph->xadj;
svwgt[0] = lgraph->vwgt;
snvwgt[0] = lgraph->nvwgt;
sadjwgtsum[0] = lgraph->adjwgtsum;
sadjncy[0] = lgraph->adjncy;
sadjwgt[0] = lgraph->adjwgt;
slabel[0] = lgraph->label;
SetUpSplitGraph(graph, rgraph, snvtxs[1], snedges[1]);
sxadj[1] = rgraph->xadj;
svwgt[1] = rgraph->vwgt;
snvwgt[1] = rgraph->nvwgt;
sadjwgtsum[1] = rgraph->adjwgtsum;
sadjncy[1] = rgraph->adjncy;
sadjwgt[1] = rgraph->adjwgt;
slabel[1] = rgraph->label;
snvtxs[0] = snvtxs[1] = snedges[0] = snedges[1] = 0;
sxadj[0][0] = sxadj[1][0] = 0;
for (i=0; i<nvtxs; i++) {
mypart = where[i];
sum = adjwgtsum[i];
istart = xadj[i];
iend = xadj[i+1];
if (bndptr[i] == -1) { /* This is an interior vertex */
auxadjncy = sadjncy[mypart] + snedges[mypart] - istart;
auxadjwgt = sadjwgt[mypart] + snedges[mypart] - istart;
for(j=istart; j<iend; j++) {
auxadjncy[j] = adjncy[j];
auxadjwgt[j] = adjwgt[j];
}
snedges[mypart] += iend-istart;
}
else {
auxadjncy = sadjncy[mypart];
auxadjwgt = sadjwgt[mypart];
l = snedges[mypart];
for (j=istart; j<iend; j++) {
k = adjncy[j];
if (where[k] == mypart) {
auxadjncy[l] = k;
auxadjwgt[l++] = adjwgt[j];
}
else {
sum -= adjwgt[j];
}
}
snedges[mypart] = l;
}
if (ncon == 1)
svwgt[mypart][snvtxs[mypart]] = vwgt[i];
else {
for (kk=0; kk<ncon; kk++)
snvwgt[mypart][snvtxs[mypart]*ncon+kk] = nvwgt[i*ncon+kk]/npwgts[mypart*ncon+kk];
}
sadjwgtsum[mypart][snvtxs[mypart]] = sum;
slabel[mypart][snvtxs[mypart]] = label[i];
sxadj[mypart][++snvtxs[mypart]] = snedges[mypart];
}
for (mypart=0; mypart<2; mypart++) {
iend = sxadj[mypart][snvtxs[mypart]];
auxadjncy = sadjncy[mypart];
for (i=0; i<iend; i++)
auxadjncy[i] = rename[auxadjncy[i]];
}
lgraph->nedges = snedges[0];
rgraph->nedges = snedges[1];
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->SplitTmr));
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function takes a graph and a bisection and splits it into two graphs.
* This function relies on the fact that adjwgt is all equal to 1.
**************************************************************************/
void SplitGraphOrder(CtrlType *ctrl, GraphType *graph, GraphType *lgraph, GraphType *rgraph)
{
int i, ii, j, k, l, istart, iend, mypart, nvtxs, snvtxs[3], snedges[3];
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *adjwgtsum, *label, *where, *bndptr, *bndind;
idxtype *sxadj[2], *svwgt[2], *sadjncy[2], *sadjwgt[2], *sadjwgtsum[2], *slabel[2];
idxtype *rename;
idxtype *auxadjncy, *auxadjwgt;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->SplitTmr));
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
label = graph->label;
where = graph->where;
bndptr = graph->bndptr;
bndind = graph->bndind;
ASSERT(bndptr != NULL);
rename = idxwspacemalloc(ctrl, nvtxs);
snvtxs[0] = snvtxs[1] = snvtxs[2] = snedges[0] = snedges[1] = snedges[2] = 0;
for (i=0; i<nvtxs; i++) {
k = where[i];
rename[i] = snvtxs[k]++;
snedges[k] += xadj[i+1]-xadj[i];
}
SetUpSplitGraph(graph, lgraph, snvtxs[0], snedges[0]);
sxadj[0] = lgraph->xadj;
svwgt[0] = lgraph->vwgt;
sadjwgtsum[0] = lgraph->adjwgtsum;
sadjncy[0] = lgraph->adjncy;
sadjwgt[0] = lgraph->adjwgt;
slabel[0] = lgraph->label;
SetUpSplitGraph(graph, rgraph, snvtxs[1], snedges[1]);
sxadj[1] = rgraph->xadj;
svwgt[1] = rgraph->vwgt;
sadjwgtsum[1] = rgraph->adjwgtsum;
sadjncy[1] = rgraph->adjncy;
sadjwgt[1] = rgraph->adjwgt;
slabel[1] = rgraph->label;
/* Go and use bndptr to also mark the boundary nodes in the two partitions */
for (ii=0; ii<graph->nbnd; ii++) {
i = bndind[ii];
for (j=xadj[i]; j<xadj[i+1]; j++)
bndptr[adjncy[j]] = 1;
}
snvtxs[0] = snvtxs[1] = snedges[0] = snedges[1] = 0;
sxadj[0][0] = sxadj[1][0] = 0;
for (i=0; i<nvtxs; i++) {
if ((mypart = where[i]) == 2)
continue;
istart = xadj[i];
iend = xadj[i+1];
if (bndptr[i] == -1) { /* This is an interior vertex */
auxadjncy = sadjncy[mypart] + snedges[mypart] - istart;
for(j=istart; j<iend; j++)
auxadjncy[j] = adjncy[j];
snedges[mypart] += iend-istart;
}
else {
auxadjncy = sadjncy[mypart];
l = snedges[mypart];
for (j=istart; j<iend; j++) {
k = adjncy[j];
if (where[k] == mypart)
auxadjncy[l++] = k;
}
snedges[mypart] = l;
}
svwgt[mypart][snvtxs[mypart]] = vwgt[i];
sadjwgtsum[mypart][snvtxs[mypart]] = snedges[mypart]-sxadj[mypart][snvtxs[mypart]];
slabel[mypart][snvtxs[mypart]] = label[i];
sxadj[mypart][++snvtxs[mypart]] = snedges[mypart];
}
for (mypart=0; mypart<2; mypart++) {
iend = snedges[mypart];
idxset(iend, 1, sadjwgt[mypart]);
auxadjncy = sadjncy[mypart];
for (i=0; i<iend; i++)
auxadjncy[i] = rename[auxadjncy[i]];
}
lgraph->nvtxs = snvtxs[0];
lgraph->nedges = snedges[0];
rgraph->nvtxs = snvtxs[1];
rgraph->nedges = snedges[1];
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->SplitTmr));
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 = (int) 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 MCRandom_KWayEdgeRefineHorizontal(CtrlType *ctrl, GraphType *graph, int nparts,
float *orgubvec, int npasses)
{
int i, ii, iii, j, k, 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 MocFM_2WayEdgeRefine2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *orgubvec,
int npasses)
{
int i, j, ii, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *perm, *qnum;
float *nvwgt, *npwgts, origdiff[MAXNCON], origbal[MAXNCON], minbal[MAXNCON];
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, initcut, newcut, mincutorder;
float *maxwgt, *minwgt, ubvec[MAXNCON], 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);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, origbal);
for (i=0; i<ncon; i++) {
origdiff[i] = fabs(tpwgts[0]-npwgts[i]);
ubvec[i] = amax(origbal[i], orgubvec[i]);
}
/* 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);
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("\n");
}
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;
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, minbal);
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++) {
SelectQueue2(ncon, npwgts, tpwgts, &from, &cnum, parts, maxwgt);
to = (from+1)%2;
if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
newcut -= (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);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
if ((newcut < mincut && AreAllBelow(ncon, tvec, ubvec)) ||
(newcut == mincut && IsBetter2wayBalance(ncon, tvec, minbal, ubvec))) {
mincut = newcut;
for (i=0; i<ncon; i++)
minbal[i] = tvec[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 (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
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 && 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: ");
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
printf("\n");
}
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);
fwspacefree(ctrl, 2*ncon);
fwspacefree(ctrl, 2*ncon);
}
/*************************************************************************
* This function creates the coarser graph
**************************************************************************/
void CreateCoarseGraphNoMask(CtrlType *ctrl, GraphType *graph, int cnvtxs, idxtype *match, idxtype *perm)
{
int i, j, k, m, istart, iend, nvtxs, nedges, ncon, cnedges, v, u, dovsize;
idxtype *xadj, *vwgt, *vsize, *adjncy, *adjwgt, *adjwgtsum, *auxadj;
idxtype *cmap, *htable;
idxtype *cxadj, *cvwgt, *cvsize, *cadjncy, *cadjwgt, *cadjwgtsum;
float *nvwgt, *cnvwgt;
GraphType *cgraph;
dovsize = (ctrl->optype == OP_KVMETIS ? 1 : 0);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
vwgt = graph->vwgt;
vsize = graph->vsize;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
cmap = graph->cmap;
/* Initialize the coarser graph */
cgraph = SetUpCoarseGraph(graph, cnvtxs, dovsize);
cxadj = cgraph->xadj;
cvwgt = cgraph->vwgt;
cvsize = cgraph->vsize;
cnvwgt = cgraph->nvwgt;
cadjwgtsum = cgraph->adjwgtsum;
cadjncy = cgraph->adjncy;
cadjwgt = cgraph->adjwgt;
htable = idxset(cnvtxs, -1, idxwspacemalloc(ctrl, cnvtxs));
iend = xadj[nvtxs];
auxadj = ctrl->wspace.auxcore;
memcpy(auxadj, adjncy, iend*sizeof(idxtype));
for (i=0; i<iend; i++)
auxadj[i] = cmap[auxadj[i]];
cxadj[0] = cnvtxs = cnedges = 0;
for (i=0; i<nvtxs; i++) {
v = perm[i];
if (cmap[v] != cnvtxs)
continue;
u = match[v];
if (ncon == 1)
cvwgt[cnvtxs] = vwgt[v];
else
scopy(ncon, nvwgt+v*ncon, cnvwgt+cnvtxs*ncon);
if (dovsize)
cvsize[cnvtxs] = vsize[v];
cadjwgtsum[cnvtxs] = adjwgtsum[v];
nedges = 0;
istart = xadj[v];
iend = xadj[v+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
if ((m = htable[k]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = adjwgt[j];
htable[k] = nedges++;
}
else {
cadjwgt[m] += adjwgt[j];
}
}
if (v != u) {
if (ncon == 1)
cvwgt[cnvtxs] += vwgt[u];
else
saxpy(ncon, 1.0, nvwgt+u*ncon, 1, cnvwgt+cnvtxs*ncon, 1);
if (dovsize)
cvsize[cnvtxs] += vsize[u];
cadjwgtsum[cnvtxs] += adjwgtsum[u];
istart = xadj[u];
iend = xadj[u+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
if ((m = htable[k]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = adjwgt[j];
htable[k] = nedges++;
}
else {
cadjwgt[m] += adjwgt[j];
}
}
/* Remove the contracted adjacency weight */
if ((j = htable[cnvtxs]) != -1) {
ASSERT(cadjncy[j] == cnvtxs);
cadjwgtsum[cnvtxs] -= cadjwgt[j];
cadjncy[j] = cadjncy[--nedges];
cadjwgt[j] = cadjwgt[nedges];
htable[cnvtxs] = -1;
}
}
ASSERTP(cadjwgtsum[cnvtxs] == idxsum(nedges, cadjwgt), ("%d %d\n", cadjwgtsum[cnvtxs], idxsum(nedges, cadjwgt)));
for (j=0; j<nedges; j++)
htable[cadjncy[j]] = -1; /* Zero out the htable */
cnedges += nedges;
cxadj[++cnvtxs] = cnedges;
cadjncy += nedges;
cadjwgt += nedges;
}
cgraph->nedges = cnedges;
ReAdjustMemory(graph, cgraph, dovsize);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
idxwspacefree(ctrl, cnvtxs);
}
/*************************************************************************
* This function 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 MocInit2WayBalance2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
int i, ii, j, k, l, kwgt, nvtxs, nbnd, ncon, nswaps, from, to, pass, me, cnum, tmp, imin;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *perm, *qnum;
float *nvwgt, *npwgts, minwgt;
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;
moved = idxwspacemalloc(ctrl, nvtxs);
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, ComputeLoadImbalance(ncon, 2, 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);
}
idxset(nvtxs, -1, moved);
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]);
}
}
/*
for (i=0; i<ncon; i++)
printf("Queue #%d has %d %d\n", i, parts[i][0].nnodes, parts[i][1].nnodes);
*/
/* Determine the termination criterion */
imin = 0;
for (i=1; i<ncon; i++)
imin = (ubvec[i] < ubvec[imin] ? i : imin);
minwgt = .5/ubvec[imin];
mincut = graph->mincut;
nbnd = graph->nbnd;
for (nswaps=0; nswaps<nvtxs; nswaps++) {
/* Exit as soon as the minimum weight crossed over */
if (npwgts[to*ncon+imin] > minwgt)
break;
if ((cnum = SelectQueueOneWay2(ncon, npwgts+to*ncon, parts, ubvec)) == -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;
moved[higain] = nswaps;
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", ComputeLoadImbalance(ncon, 2, 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 (moved[k] == -1 && 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", ComputeLoadImbalance(ncon, 2, 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);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function computes the initial id/ed
**************************************************************************/
void ComputeKWayVolGains(CtrlType *ctrl, GraphType *graph, int nparts)
{
int i, ii, j, k, kk, l, nvtxs, me, other, pid, myndegrees;
idxtype *xadj, *vsize, *adjncy, *adjwgt, *where, *bndind, *bndptr, *ophtable;
VRInfoType *rinfo, *myrinfo, *orinfo;
VEDegreeType *myedegrees, *oedegrees;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vsize = graph->vsize;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
bndind = graph->bndind;
bndptr = idxset(nvtxs, -1, graph->bndptr);
rinfo = graph->vrinfo;
ophtable = idxset(nparts, -1, idxwspacemalloc(ctrl, nparts));
/*------------------------------------------------------------
/ Compute now the iv/ev degrees
/------------------------------------------------------------*/
graph->minvol = graph->nbnd = 0;
for (i=0; i<nvtxs; i++) {
myrinfo = rinfo+i;
myrinfo->gv = -MAXIDX;
if (myrinfo->ndegrees > 0) {
me = where[i];
myedegrees = myrinfo->edegrees;
myndegrees = myrinfo->ndegrees;
graph->minvol += myndegrees*vsize[i];
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
other = where[ii];
orinfo = rinfo+ii;
oedegrees = orinfo->edegrees;
for (k=0; k<orinfo->ndegrees; k++)
ophtable[oedegrees[k].pid] = k;
ophtable[other] = 1; /* this is to simplify coding */
if (me == other) {
/* Find which domains 'i' is connected and 'ii' is not and update their gain */
for (k=0; k<myndegrees; k++) {
if (ophtable[myedegrees[k].pid] == -1)
myedegrees[k].gv -= vsize[ii];
}
}
else {
ASSERT(ophtable[me] != -1);
if (oedegrees[ophtable[me]].ned == 1) { /* I'm the only connection of 'ii' in 'me' */
/* Increase the gains for all the common domains between 'i' and 'ii' */
for (k=0; k<myndegrees; k++) {
if (ophtable[myedegrees[k].pid] != -1)
myedegrees[k].gv += vsize[ii];
}
}
else {
/* Find which domains 'i' is connected and 'ii' is not and update their gain */
for (k=0; k<myndegrees; k++) {
if (ophtable[myedegrees[k].pid] == -1)
myedegrees[k].gv -= vsize[ii];
}
}
}
for (kk=0; kk<orinfo->ndegrees; kk++)
ophtable[oedegrees[kk].pid] = -1;
ophtable[other] = -1;
}
/* Compute the max vgain */
for (k=0; k<myndegrees; k++) {
if (myedegrees[k].gv > myrinfo->gv)
myrinfo->gv = myedegrees[k].gv;
}
}
if (myrinfo->ed > 0 && myrinfo->id == 0)
myrinfo->gv += vsize[i];
if (myrinfo->gv >= 0 || myrinfo->ed-myrinfo->id >= 0)
BNDInsert(graph->nbnd, bndind, bndptr, i);
}
idxwspacefree(ctrl, nparts);
}
/*************************************************************************
* This function finds all the connected components induced by the
* partitioning vector in wgraph->where and tries to push them around to
* remove some of them
**************************************************************************/
void EliminateComponents(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor)
{
int i, ii, j, jj, k, me, nvtxs, tvwgt, first, last, nleft, ncmps, cwgt, other, target, deltawgt;
idxtype *xadj, *adjncy, *vwgt, *adjwgt, *where, *pwgts, *maxpwgt;
idxtype *cpvec, *touched, *perm, *todo, *cind, *cptr, *npcmps;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
adjwgt = graph->adjwgt;
where = graph->where;
pwgts = graph->pwgts;
touched = idxset(nvtxs, 0, idxwspacemalloc(ctrl, nvtxs));
cptr = idxwspacemalloc(ctrl, nvtxs+1);
cind = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
todo = idxwspacemalloc(ctrl, nvtxs);
maxpwgt = idxwspacemalloc(ctrl, nparts);
cpvec = idxwspacemalloc(ctrl, nparts);
npcmps = idxset(nparts, 0, idxwspacemalloc(ctrl, nparts));
for (i=0; i<nvtxs; i++)
perm[i] = todo[i] = i;
/* Find the connected componends induced by the partition */
ncmps = -1;
first = last = 0;
nleft = nvtxs;
while (nleft > 0) {
if (first == last) { /* Find another starting vertex */
cptr[++ncmps] = first;
ASSERT(touched[todo[0]] == 0);
i = todo[0];
cind[last++] = i;
touched[i] = 1;
me = where[i];
npcmps[me]++;
}
i = cind[first++];
k = perm[i];
j = todo[k] = todo[--nleft];
perm[j] = k;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (where[k] == me && !touched[k]) {
cind[last++] = k;
touched[k] = 1;
}
}
}
cptr[++ncmps] = first;
/* printf("I found %d components, for this %d-way partition\n", ncmps, nparts); */
if (ncmps > nparts) { /* There are more components than processors */
/* First determine the max allowed load imbalance */
tvwgt = idxsum(nparts, pwgts);
for (i=0; i<nparts; i++)
maxpwgt[i] = ubfactor*tpwgts[i]*tvwgt;
deltawgt = 5;
for (i=0; i<ncmps; i++) {
me = where[cind[cptr[i]]]; /* Get the domain of this component */
if (npcmps[me] == 1)
continue; /* Skip it because it is contigous */
/*printf("Trying to move %d from %d\n", i, me); */
/* Determine the weight of the block to be moved and abort if too high */
for (cwgt=0, j=cptr[i]; j<cptr[i+1]; j++)
cwgt += vwgt[cind[j]];
if (cwgt > .30*pwgts[me])
continue; /* Skip the component if it is over 30% of the weight */
/* Determine the connectivity */
idxset(nparts, 0, cpvec);
for (j=cptr[i]; j<cptr[i+1]; j++) {
ii = cind[j];
for (jj=xadj[ii]; jj<xadj[ii+1]; jj++)
cpvec[where[adjncy[jj]]] += adjwgt[jj];
}
cpvec[me] = 0;
target = -1;
for (j=0; j<nparts; j++) {
if (cpvec[j] > 0 && (cwgt < deltawgt || pwgts[j] + cwgt < maxpwgt[j])) {
if (target == -1 || cpvec[target] < cpvec[j])
target = j;
}
}
/* printf("\tMoving it to %d [%d]\n", target, cpvec[target]);*/
if (target != -1) {
/* Assign all the vertices of 'me' to 'target' and update data structures */
INC_DEC(pwgts[target], pwgts[me], cwgt);
npcmps[me]--;
MoveGroup(ctrl, graph, nparts, target, i, cptr, cind);
}
}
}
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs+1);
}
/*************************************************************************
* This function finds a matching using the HEM heuristic
**************************************************************************/
void MCMatch_SHEM(CtrlType *ctrl, GraphType *graph)
{
int i, ii, j, k, nvtxs, cnvtxs, ncon, maxidx, maxwgt, avgdegree;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *match, *cmap, *degrees, *perm, *tperm;
float *nvwgt;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
cmap = graph->cmap;
match = idxset(nvtxs, UNMATCHED, idxwspacemalloc(ctrl, nvtxs));
perm = idxwspacemalloc(ctrl, nvtxs);
tperm = idxwspacemalloc(ctrl, nvtxs);
degrees = idxwspacemalloc(ctrl, nvtxs);
RandomPermute(nvtxs, tperm, 1);
avgdegree = (int)(0.7*(xadj[nvtxs]/nvtxs));
for (i=0; i<nvtxs; i++)
degrees[i] = (xadj[i+1]-xadj[i] > avgdegree ? avgdegree : xadj[i+1]-xadj[i]);
BucketSortKeysInc(nvtxs, avgdegree, degrees, tperm, perm);
cnvtxs = 0;
/* Take care any islands. Islands are matched with non-islands due to coarsening */
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
if (xadj[i] < xadj[i+1])
break;
maxidx = i;
for (j=nvtxs-1; j>ii; j--) {
k = perm[j];
if (match[k] == UNMATCHED && xadj[k] < xadj[k+1]) {
maxidx = k;
break;
}
}
cmap[i] = cmap[maxidx] = cnvtxs++;
match[i] = maxidx;
match[maxidx] = i;
}
}
/* Continue with normal matching */
for (; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
maxidx = i;
maxwgt = 0;
/* Find a heavy-edge matching, subject to maxvwgt constraints */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (match[k] == UNMATCHED && maxwgt <= adjwgt[j] &&
AreAllVwgtsBelowFast(ncon, nvwgt+i*ncon, nvwgt+k*ncon, ctrl->nmaxvwgt)) {
maxwgt = adjwgt[j];
maxidx = adjncy[j];
}
}
cmap[i] = cmap[maxidx] = cnvtxs++;
match[i] = maxidx;
match[maxidx] = i;
}
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
idxwspacefree(ctrl, nvtxs); /* degrees */
idxwspacefree(ctrl, nvtxs); /* tperm */
CreateCoarseGraph(ctrl, graph, cnvtxs, match, perm);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void MocGeneral2WayBalance2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
idxtype 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, origbal[MAXNCON], minbal[MAXNCON], newbal[MAXNCON];
PQueueType parts[MAXNCON][2];
idxtype 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] = gk_fargmax(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) {
mprintf("Parts: [");
for (l=0; l<ncon; l++)
mprintf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
mprintf("] 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++)
mprintf("%.3f ", origbal[i]);
mprintf("[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;
gk_faxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
gk_faxpy(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]);
gk_faxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
gk_faxpy(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) {
mprintf("Moved %6D from %D(%D). Gain: %5D, Cut: %5D, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut);
for (i=0; i<ncon; i++)
mprintf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
mprintf(", LB: ");
for (i=0; i<ncon; i++)
mprintf("%.3f ", tvec[i]);
if (mincutorder == nswaps)
mprintf(" *\n");
else
mprintf("\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);
gk_faxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
gk_faxpy(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) {
mprintf("\tMincut: %6D at %5D, NBND: %6D, NPwgts: [", mincut, mincutorder, nbnd);
for (i=0; i<ncon; i++)
mprintf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]);
mprintf("], LB: ");
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
for (i=0; i<ncon; i++)
mprintf("%.3f ", tvec[i]);
mprintf("\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);
}
