/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
int BalanceMyLink(CtrlType *ctrl, GraphType *graph, idxtype *home, int me,
int you, float *flows, float maxdiff, float *diff_cost, float *diff_lbavg,
float avgvwgt)
{
int h, i, ii, j, k;
int nvtxs, ncon;
int nqueues, minval, maxval, higain, vtx, edge, totalv;
int from, to, qnum, index, nchanges, cut, tmp;
int pass, nswaps, nmoves, multiplier;
idxtype *xadj, *vsize, *adjncy, *adjwgt, *where, *ed, *id;
idxtype *hval, *nvpq, *inq, *map, *rmap, *ptr, *myqueue, *changes;
float *nvwgt, lbvec[MAXNCON], pwgts[MAXNCON*2], tpwgts[MAXNCON*2], my_wgt[MAXNCON];
float newgain, oldgain = 0.0;
float lbavg, bestflow, mycost;
float ipc_factor, redist_factor, ftmp;
FPQueueType *queues;
int mype;
MPI_Comm_rank(MPI_COMM_WORLD, &mype);
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
vsize = graph->vsize;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
ipc_factor = ctrl->ipc_factor;
redist_factor = ctrl->redist_factor;
hval = idxmalloc(nvtxs*7, "hval");
id = hval + nvtxs;
ed = hval + nvtxs*2;
map = hval + nvtxs*3;
rmap = hval + nvtxs*4;
myqueue = hval + nvtxs*5;
changes = hval + nvtxs*6;
sset(ncon*2, 0.0, pwgts);
for (h=0; h<ncon; h++) {
tpwgts[h] = -1.0 * flows[h];
tpwgts[ncon+h] = flows[h];
}
for (i=0; i<nvtxs; i++) {
if (where[i] == me) {
for (h=0; h<ncon; h++) {
tpwgts[h] += nvwgt[i*ncon+h];
pwgts[h] += nvwgt[i*ncon+h];
}
}
else {
ASSERTS(where[i] == you);
for (h=0; h<ncon; h++) {
tpwgts[ncon+h] += nvwgt[i*ncon+h];
pwgts[ncon+h] += nvwgt[i*ncon+h];
}
}
}
/* we don't want any tpwgts to be less than zero */
for (h=0; h<ncon; h++) {
if (tpwgts[h] < 0.0) {
tpwgts[ncon+h] += tpwgts[h];
tpwgts[h] = 0.0;
}
if (tpwgts[ncon+h] < 0.0) {
tpwgts[h] += tpwgts[ncon+h];
tpwgts[ncon+h] = 0.0;
}
}
/*******************************/
/* insert vertices into queues */
/*******************************/
minval = maxval = 0;
multiplier = 1;
for (i=0; i<ncon; i++) {
multiplier *= (i+1);
maxval += i*multiplier;
minval += (ncon-1-i)*multiplier;
}
nqueues = maxval-minval+1;
nvpq = idxsmalloc(nqueues, 0, "nvpq");
ptr = idxmalloc(nqueues+1, "ptr");
inq = idxmalloc(nqueues*2, "inq");
queues = (FPQueueType *)(GKmalloc(sizeof(FPQueueType)*nqueues*2, "queues"));
for (i=0; i<nvtxs; i++)
hval[i] = Moc_HashVwgts(ncon, nvwgt+i*ncon) - minval;
for (i=0; i<nvtxs; i++)
nvpq[hval[i]]++;
ptr[0] = 0;
for (i=0; i<nqueues; i++)
ptr[i+1] = ptr[i] + nvpq[i];
for (i=0; i<nvtxs; i++) {
map[i] = ptr[hval[i]];
rmap[ptr[hval[i]]++] = i;
}
for (i=nqueues-1; i>0; i--)
ptr[i] = ptr[i-1];
ptr[0] = 0;
/* initialize queues */
for (i=0; i<nqueues; i++)
if (nvpq[i] > 0) {
FPQueueInit(queues+i, nvpq[i]);
FPQueueInit(queues+i+nqueues, nvpq[i]);
}
/* compute internal/external degrees */
idxset(nvtxs, 0, id);
idxset(nvtxs, 0, ed);
for (j=0; j<nvtxs; j++)
for (k=xadj[j]; k<xadj[j+1]; k++)
if (where[adjncy[k]] == where[j])
id[j] += adjwgt[k];
else
ed[j] += adjwgt[k];
nswaps = 0;
for (pass=0; pass<N_MOC_BAL_PASSES; pass++) {
idxset(nvtxs, -1, myqueue);
idxset(nqueues*2, 0, inq);
/* insert vertices into correct queues */
for (j=0; j<nvtxs; j++) {
index = (where[j] == me) ? 0 : nqueues;
newgain = ipc_factor*(float)(ed[j]-id[j]);
if (home[j] == me || home[j] == you) {
if (where[j] == home[j])
newgain -= redist_factor*(float)vsize[j];
else
newgain += redist_factor*(float)vsize[j];
}
FPQueueInsert(queues+hval[j]+index, map[j]-ptr[hval[j]], newgain);
myqueue[j] = (where[j] == me) ? 0 : 1;
inq[hval[j]+index]++;
}
/* bestflow = sfavg(ncon, flows); */
for (j=0, h=0; h<ncon; h++)
if (fabs(flows[h]) > fabs(flows[j])) j = h;
bestflow = fabs(flows[j]);
nchanges = nmoves = 0;
for (ii=0; ii<nvtxs/2; ii++) {
from = -1;
Moc_DynamicSelectQueue(nqueues, ncon, me, you, inq, flows, &from,
&qnum, minval, avgvwgt, maxdiff);
/* can't find a vertex in one subdomain, try the other */
if (from != -1 && qnum == -1) {
from = (from == me) ? you : me;
if (from == me) {
for (j=0; j<ncon; j++)
if (flows[j] > avgvwgt)
break;
}
else {
for (j=0; j<ncon; j++)
if (flows[j] < -1.0*avgvwgt)
break;
}
if (j != ncon)
Moc_DynamicSelectQueue(nqueues, ncon, me, you, inq, flows, &from,
&qnum, minval, avgvwgt, maxdiff);
}
if (qnum == -1)
break;
to = (from == me) ? you : me;
index = (from == me) ? 0 : nqueues;
higain = FPQueueGetMax(queues+qnum+index);
inq[qnum+index]--;
ASSERTS(higain != -1);
/*****************/
/* make the swap */
/*****************/
vtx = rmap[higain+ptr[qnum]];
myqueue[vtx] = -1;
where[vtx] = to;
nswaps++;
nmoves++;
/* update the flows */
for (j=0; j<ncon; j++)
flows[j] += (to == me) ? nvwgt[vtx*ncon+j] : -1.0*nvwgt[vtx*ncon+j];
/* ftmp = sfavg(ncon, flows); */
for (j=0, h=0; h<ncon; h++)
if (fabs(flows[h]) > fabs(flows[j])) j = h;
ftmp = fabs(flows[j]);
if (ftmp < bestflow) {
bestflow = ftmp;
nchanges = 0;
}
else {
changes[nchanges++] = vtx;
}
SWAP(id[vtx], ed[vtx], tmp);
for (j=xadj[vtx]; j<xadj[vtx+1]; j++) {
edge = adjncy[j];
/* must compute oldgain before changing id/ed */
if (myqueue[edge] != -1) {
oldgain = ipc_factor*(float)(ed[edge]-id[edge]);
if (home[edge] == me || home[edge] == you) {
if (where[edge] == home[edge])
oldgain -= redist_factor*(float)vsize[edge];
else
oldgain += redist_factor*(float)vsize[edge];
}
}
tmp = (to == where[edge] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[edge], ed[edge], tmp);
if (myqueue[edge] != -1) {
newgain = ipc_factor*(float)(ed[edge]-id[edge]);
if (home[edge] == me || home[edge] == you) {
if (where[edge] == home[edge])
newgain -= redist_factor*(float)vsize[edge];
else
newgain += redist_factor*(float)vsize[edge];
}
FPQueueUpdate(queues+hval[edge]+(nqueues*myqueue[edge]),
map[edge]-ptr[hval[edge]], oldgain, newgain);
}
}
}
/****************************/
/* now go back to best flow */
/****************************/
nswaps -= nchanges;
nmoves -= nchanges;
for (i=0; i<nchanges; i++) {
vtx = changes[i];
from = where[vtx];
where[vtx] = to = (from == me) ? you : me;
SWAP(id[vtx], ed[vtx], tmp);
for (j=xadj[vtx]; j<xadj[vtx+1]; j++) {
edge = adjncy[j];
tmp = (to == where[edge] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[edge], ed[edge], tmp);
}
}
for (i=0; i<nqueues; i++) {
if (nvpq[i] > 0) {
FPQueueReset(queues+i);
FPQueueReset(queues+i+nqueues);
}
}
if (nmoves == 0)
break;
}
/***************************/
/* compute 2-way imbalance */
/***************************/
sset(ncon, 0.0, my_wgt);
for (i=0; i<nvtxs; i++)
if (where[i] == me)
for (h=0; h<ncon; h++)
my_wgt[h] += nvwgt[i*ncon+h];
for (i=0; i<ncon; i++) {
ftmp = (pwgts[i]+pwgts[ncon+i])/2.0;
if (ftmp != 0.0)
lbvec[i] = fabs(my_wgt[i]-tpwgts[i]) / ftmp;
else
lbvec[i] = 0.0;
}
lbavg = savg(ncon, lbvec);
*diff_lbavg = lbavg;
/****************/
/* compute cost */
/****************/
cut = totalv = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
totalv += vsize[i];
for (j=xadj[i]; j<xadj[i+1]; j++)
if (where[adjncy[j]] != where[i])
cut += adjwgt[j];
}
cut /= 2;
mycost = cut*ipc_factor + totalv*redist_factor;
*diff_cost = mycost;
/* free memory */
for (i=0; i<nqueues; i++)
if (nvpq[i] > 0) {
FPQueueFree(queues+i);
FPQueueFree(queues+i+nqueues);
}
GKfree((void **)&hval, (void **)&nvpq, (void **)&ptr, (void **)&inq, (void **)&queues, LTERM);
return nswaps;
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Random_KWayEdgeRefineMConn(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor, int npasses, int ffactor)
{
int i, ii, iii, j, jj, k, l, pass, nvtxs, nmoves, nbnd, tvwgt, myndegrees;
int from, me, to, oldcut, vwgt, gain;
int maxndoms, nadd;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where, *pwgts, *perm, *bndptr, *bndind, *minwgt, *maxwgt, *itpwgts;
idxtype *phtable, *pmat, *pmatptr, *ndoms;
EDegreeType *myedegrees;
RInfoType *myrinfo;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bndptr = graph->bndptr;
bndind = graph->bndind;
where = graph->where;
pwgts = graph->pwgts;
pmat = ctrl->wspace.pmat;
phtable = idxwspacemalloc(ctrl, nparts);
ndoms = idxwspacemalloc(ctrl, nparts);
ComputeSubDomainGraph(graph, nparts, pmat, ndoms);
/* Setup the weight intervals of the various subdomains */
minwgt = idxwspacemalloc(ctrl, nparts);
maxwgt = idxwspacemalloc(ctrl, nparts);
itpwgts = idxwspacemalloc(ctrl, nparts);
tvwgt = idxsum(nparts, pwgts);
ASSERT(tvwgt == idxsum(nvtxs, graph->vwgt));
for (i=0; i<nparts; i++) {
itpwgts[i] = tpwgts[i]*tvwgt;
maxwgt[i] = tpwgts[i]*tvwgt*ubfactor;
minwgt[i] = tpwgts[i]*tvwgt*(1.0/ubfactor);
}
perm = idxwspacemalloc(ctrl, nvtxs);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions: [%6d %6d]-[%6d %6d], Balance: %5.3f, Nv-Nb[%6d %6d]. Cut: %6d\n",
pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], minwgt[0], maxwgt[0],
1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nvtxs, graph->nbnd,
graph->mincut));
for (pass=0; pass<npasses; pass++) {
ASSERT(ComputeCut(graph, where) == graph->mincut);
maxndoms = ndoms[idxamax(nparts, ndoms)];
oldcut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (nmoves=iii=0; iii<graph->nbnd; iii++) {
ii = perm[iii];
if (ii >= nbnd)
continue;
i = bndind[ii];
myrinfo = graph->rinfo+i;
if (myrinfo->ed >= myrinfo->id) { /* Total ED is too high */
from = where[i];
vwgt = graph->vwgt[i];
if (myrinfo->id > 0 && pwgts[from]-vwgt < minwgt[from])
continue; /* This cannot be moved! */
myedegrees = myrinfo->edegrees;
myndegrees = myrinfo->ndegrees;
/* Determine the valid domains */
for (j=0; j<myndegrees; j++) {
to = myedegrees[j].pid;
phtable[to] = 1;
pmatptr = pmat + to*nparts;
for (nadd=0, k=0; k<myndegrees; k++) {
if (k == j)
continue;
l = myedegrees[k].pid;
if (pmatptr[l] == 0) {
if (ndoms[l] > maxndoms-1) {
phtable[to] = 0;
nadd = maxndoms;
break;
}
nadd++;
}
}
if (ndoms[to]+nadd > maxndoms)
phtable[to] = 0;
if (nadd == 0)
phtable[to] = 2;
}
/* Find the first valid move */
j = myrinfo->id;
for (k=0; k<myndegrees; k++) {
to = myedegrees[k].pid;
if (!phtable[to])
continue;
gain = myedegrees[k].ed-j; /* j = myrinfo->id. Allow good nodes to move */
if (pwgts[to]+vwgt <= maxwgt[to]+ffactor*gain && gain >= 0)
break;
}
if (k == myndegrees)
continue; /* break out if you did not find a candidate */
for (j=k+1; j<myndegrees; j++) {
to = myedegrees[j].pid;
if (!phtable[to])
continue;
if ((myedegrees[j].ed > myedegrees[k].ed && pwgts[to]+vwgt <= maxwgt[to]) ||
(myedegrees[j].ed == myedegrees[k].ed &&
itpwgts[myedegrees[k].pid]*pwgts[to] < itpwgts[to]*pwgts[myedegrees[k].pid]))
k = j;
}
to = myedegrees[k].pid;
j = 0;
if (myedegrees[k].ed-myrinfo->id > 0)
j = 1;
else if (myedegrees[k].ed-myrinfo->id == 0) {
if (/*(iii&7) == 0 ||*/ phtable[myedegrees[k].pid] == 2 || pwgts[from] >= maxwgt[from] || itpwgts[from]*(pwgts[to]+vwgt) < itpwgts[to]*pwgts[from])
j = 1;
}
if (j == 0)
continue;
/*=====================================================================
* If we got here, we can now move the vertex from 'from' to 'to'
*======================================================================*/
graph->mincut -= myedegrees[k].ed-myrinfo->id;
IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut));
/* Update pmat to reflect the move of 'i' */
pmat[from*nparts+to] += (myrinfo->id-myedegrees[k].ed);
pmat[to*nparts+from] += (myrinfo->id-myedegrees[k].ed);
if (pmat[from*nparts+to] == 0) {
ndoms[from]--;
if (ndoms[from]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
if (pmat[to*nparts+from] == 0) {
ndoms[to]--;
if (ndoms[to]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
/* Update where, weight, and ID/ED information of the vertex you moved */
where[i] = to;
INC_DEC(pwgts[to], pwgts[from], vwgt);
myrinfo->ed += myrinfo->id-myedegrees[k].ed;
SWAP(myrinfo->id, myedegrees[k].ed, j);
if (myedegrees[k].ed == 0)
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].pid = from;
if (myrinfo->ed-myrinfo->id < 0)
BNDDelete(nbnd, bndind, bndptr, i);
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
me = where[ii];
myrinfo = graph->rinfo+ii;
if (myrinfo->edegrees == NULL) {
myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii];
}
myedegrees = myrinfo->edegrees;
ASSERT(CheckRInfo(myrinfo));
if (me == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
if (myrinfo->ed-myrinfo->id >= 0 && bndptr[ii] == -1)
BNDInsert(nbnd, bndind, bndptr, ii);
}
else if (me == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
if (myrinfo->ed-myrinfo->id < 0 && bndptr[ii] != -1)
BNDDelete(nbnd, bndind, bndptr, ii);
}
/* Remove contribution from the .ed of 'from' */
if (me != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == from) {
if (myedegrees[k].ed == adjwgt[j])
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].ed -= adjwgt[j];
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (me != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == to) {
myedegrees[k].ed += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
myedegrees[myrinfo->ndegrees].pid = to;
myedegrees[myrinfo->ndegrees++].ed = adjwgt[j];
}
}
/* Update pmat to reflect the move of 'i' for domains other than 'from' and 'to' */
if (me != from && me != to) {
pmat[me*nparts+from] -= adjwgt[j];
pmat[from*nparts+me] -= adjwgt[j];
if (pmat[me*nparts+from] == 0) {
ndoms[me]--;
if (ndoms[me]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
if (pmat[from*nparts+me] == 0) {
ndoms[from]--;
if (ndoms[from]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
if (pmat[me*nparts+to] == 0) {
ndoms[me]++;
if (ndoms[me] > maxndoms) {
printf("You just increased the maxndoms: %d %d\n", ndoms[me], maxndoms);
maxndoms = ndoms[me];
}
}
if (pmat[to*nparts+me] == 0) {
ndoms[to]++;
if (ndoms[to] > maxndoms) {
printf("You just increased the maxndoms: %d %d\n", ndoms[to], maxndoms);
maxndoms = ndoms[to];
}
}
pmat[me*nparts+to] += adjwgt[j];
pmat[to*nparts+me] += adjwgt[j];
}
ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]);
ASSERT(CheckRInfo(myrinfo));
}
nmoves++;
}
}
graph->nbnd = nbnd;
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\t[%6d %6d], Balance: %5.3f, Nb: %6d. Nmoves: %5d, Cut: %5d, Vol: %5d, %d\n",
pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)],
1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nbnd, nmoves,
graph->mincut, ComputeVolume(graph, where), idxsum(nparts, ndoms)));
if (graph->mincut == oldcut)
break;
}
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function computes the subdomain graph
**************************************************************************/
void EliminateSubDomainEdges(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts)
{
int i, ii, j, k, me, other, nvtxs, total, max, avg, totalout, nind, ncand, ncand2, target, target2, nadd;
int min, move, cpwgt, tvwgt;
idxtype *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *maxpwgt, *pmat, *ndoms, *mypmat, *otherpmat, *ind;
KeyValueType *cand, *cand2;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
adjwgt = graph->adjwgt;
where = graph->where;
pwgts = graph->pwgts; /* We assume that this is properly initialized */
maxpwgt = idxwspacemalloc(ctrl, nparts);
ndoms = idxwspacemalloc(ctrl, nparts);
otherpmat = idxwspacemalloc(ctrl, nparts);
ind = idxwspacemalloc(ctrl, nvtxs);
pmat = ctrl->wspace.pmat;
cand = (KeyValueType *)GKmalloc(nparts*sizeof(KeyValueType), "EliminateSubDomainEdges: cand");
cand2 = (KeyValueType *)GKmalloc(nparts*sizeof(KeyValueType), "EliminateSubDomainEdges: cand");
/* Compute the pmat matrix and ndoms */
ComputeSubDomainGraph(graph, nparts, pmat, ndoms);
/* Compute the maximum allowed weight for each domain */
tvwgt = idxsum(nparts, pwgts);
for (i=0; i<nparts; i++)
maxpwgt[i] = 1.25*tpwgts[i]*tvwgt;
/* Get into the loop eliminating subdomain connections */
for (;;) {
total = idxsum(nparts, ndoms);
avg = total/nparts;
max = ndoms[idxamax(nparts, ndoms)];
/* printf("Adjacent Subdomain Stats: Total: %3d, Max: %3d, Avg: %3d [%5d]\n", total, max, avg, idxsum(nparts*nparts, pmat)); */
if (max < 1.4*avg)
break;
me = idxamax(nparts, ndoms);
mypmat = pmat + me*nparts;
totalout = idxsum(nparts, mypmat);
/*printf("Me: %d, TotalOut: %d,\n", me, totalout);*/
/* Sort the connections according to their cut */
for (ncand2=0, i=0; i<nparts; i++) {
if (mypmat[i] > 0) {
cand2[ncand2].key = mypmat[i];
cand2[ncand2++].val = i;
}
}
ikeysort(ncand2, cand2);
move = 0;
for (min=0; min<ncand2; min++) {
if (cand2[min].key > totalout/(2*ndoms[me]))
break;
other = cand2[min].val;
/*printf("\tMinOut: %d to %d\n", mypmat[other], other);*/
idxset(nparts, 0, otherpmat);
/* Go and find the vertices in 'other' that are connected in 'me' */
for (nind=0, i=0; i<nvtxs; i++) {
if (where[i] == other) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (where[adjncy[j]] == me) {
ind[nind++] = i;
break;
}
}
}
}
/* Go and construct the otherpmat to see where these nind vertices are connected to */
for (cpwgt=0, ii=0; ii<nind; ii++) {
i = ind[ii];
cpwgt += vwgt[i];
for (j=xadj[i]; j<xadj[i+1]; j++)
otherpmat[where[adjncy[j]]] += adjwgt[j];
}
otherpmat[other] = 0;
for (ncand=0, i=0; i<nparts; i++) {
if (otherpmat[i] > 0) {
cand[ncand].key = -otherpmat[i];
cand[ncand++].val = i;
}
}
ikeysort(ncand, cand);
/*
* Go through and the select the first domain that is common with 'me', and
* does not increase the ndoms[target] higher than my ndoms, subject to the
* maxpwgt constraint. Traversal is done from the mostly connected to the least.
*/
target = target2 = -1;
for (i=0; i<ncand; i++) {
k = cand[i].val;
if (mypmat[k] > 0) {
if (pwgts[k] + cpwgt > maxpwgt[k]) /* Check if balance will go off */
continue;
for (j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && ndoms[j] >= ndoms[me]-1 && pmat[nparts*j+k] == 0)
break;
}
if (j == nparts) { /* No bad second level effects */
for (nadd=0, j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && pmat[nparts*k+j] == 0)
nadd++;
}
/*printf("\t\tto=%d, nadd=%d, %d\n", k, nadd, ndoms[k]);*/
if (target2 == -1 && ndoms[k]+nadd < ndoms[me]) {
target2 = k;
}
if (nadd == 0) {
target = k;
break;
}
}
}
}
if (target == -1 && target2 != -1)
target = target2;
if (target == -1) {
/* printf("\t\tCould not make the move\n");*/
continue;
}
/*printf("\t\tMoving to %d\n", target);*/
/* Update the partition weights */
INC_DEC(pwgts[target], pwgts[other], cpwgt);
MoveGroupMConn(ctrl, graph, ndoms, pmat, nparts, target, nind, ind);
move = 1;
break;
}
if (move == 0)
break;
}
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
GKfree(&cand, &cand2, LTERM);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void MCRandom_KWayEdgeRefineHorizontal(CtrlType *ctrl, GraphType *graph, int nparts,
float *orgubvec, int npasses)
{
int i, ii, iii, j, /*jj,*/ k, /*l,*/ pass, nvtxs, ncon, nmoves, nbnd, myndegrees, same;
int from, me, to, oldcut, gain;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where, *perm, *bndptr, *bndind;
EDegreeType *myedegrees;
RInfoType *myrinfo;
float *npwgts, *nvwgt, *minwgt, *maxwgt, maxlb, minlb, ubvec[MAXNCON], tvec[MAXNCON];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bndptr = graph->bndptr;
bndind = graph->bndind;
where = graph->where;
npwgts = graph->npwgts;
/* Setup the weight intervals of the various subdomains */
minwgt = fwspacemalloc(ctrl, nparts*ncon);
maxwgt = fwspacemalloc(ctrl, nparts*ncon);
/* See if the orgubvec consists of identical constraints */
maxlb = minlb = orgubvec[0];
for (i=1; i<ncon; i++) {
minlb = (orgubvec[i] < minlb ? orgubvec[i] : minlb);
maxlb = (orgubvec[i] > maxlb ? orgubvec[i] : maxlb);
}
same = (fabs(maxlb-minlb) < .01 ? 1 : 0);
/* Let's not get very optimistic. Let Balancing do the work */
ComputeHKWayLoadImbalance(ncon, nparts, npwgts, ubvec);
for (i=0; i<ncon; i++)
ubvec[i] = amax(ubvec[i], orgubvec[i]);
if (!same) {
for (i=0; i<nparts; i++) {
for (j=0; j<ncon; j++) {
maxwgt[i*ncon+j] = ubvec[j]/nparts;
minwgt[i*ncon+j] = 1.0/(ubvec[j]*nparts);
}
}
}
else {
maxlb = ubvec[0];
for (i=1; i<ncon; i++)
maxlb = (ubvec[i] > maxlb ? ubvec[i] : maxlb);
for (i=0; i<nparts; i++) {
for (j=0; j<ncon; j++) {
maxwgt[i*ncon+j] = maxlb/nparts;
minwgt[i*ncon+j] = 1.0/(maxlb*nparts);
}
}
}
perm = idxwspacemalloc(ctrl, nvtxs);
if (ctrl->dbglvl&DBG_REFINE) {
printf("Partitions: [%5.4f %5.4f], Nv-Nb[%6d %6d]. Cut: %6d, LB: ",
npwgts[samin(ncon*nparts, npwgts)], npwgts[samax(ncon*nparts, npwgts)],
graph->nvtxs, graph->nbnd, graph->mincut);
ComputeHKWayLoadImbalance(ncon, nparts, npwgts, tvec);
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
printf("\n");
}
for (pass=0; pass<npasses; pass++) {
ASSERT(ComputeCut(graph, where) == graph->mincut);
oldcut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (nmoves=iii=0; iii<graph->nbnd; iii++) {
ii = perm[iii];
if (ii >= nbnd)
continue;
i = bndind[ii];
myrinfo = graph->rinfo+i;
if (myrinfo->ed >= myrinfo->id) { /* Total ED is too high */
from = where[i];
nvwgt = graph->nvwgt+i*ncon;
if (myrinfo->id > 0 && AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, -1.0, nvwgt, minwgt+from*ncon))
continue; /* This cannot be moved! */
myedegrees = myrinfo->edegrees;
myndegrees = myrinfo->ndegrees;
for (k=0; k<myndegrees; k++) {
to = myedegrees[k].pid;
gain = myedegrees[k].ed - myrinfo->id;
if (gain >= 0 &&
(AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon) ||
IsHBalanceBetterFT(ncon, nparts, npwgts+from*ncon, npwgts+to*ncon, nvwgt, ubvec)))
break;
}
if (k == myndegrees)
continue; /* break out if you did not find a candidate */
for (j=k+1; j<myndegrees; j++) {
to = myedegrees[j].pid;
if ((myedegrees[j].ed > myedegrees[k].ed &&
(AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon) ||
IsHBalanceBetterFT(ncon, nparts, npwgts+from*ncon, npwgts+to*ncon, nvwgt, ubvec))) ||
(myedegrees[j].ed == myedegrees[k].ed &&
IsHBalanceBetterTT(ncon, nparts, npwgts+myedegrees[k].pid*ncon, npwgts+to*ncon, nvwgt, ubvec)))
k = j;
}
to = myedegrees[k].pid;
if (myedegrees[k].ed-myrinfo->id == 0
&& !IsHBalanceBetterFT(ncon, nparts, npwgts+from*ncon, npwgts+to*ncon, nvwgt, ubvec)
&& AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, npwgts+from*ncon, maxwgt+from*ncon))
continue;
/*=====================================================================
* If we got here, we can now move the vertex from 'from' to 'to'
*======================================================================*/
graph->mincut -= myedegrees[k].ed-myrinfo->id;
IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut));
/* Update where, weight, and ID/ED information of the vertex you moved */
saxpy(ncon, 1.0, nvwgt, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt, 1, npwgts+from*ncon, 1);
where[i] = to;
myrinfo->ed += myrinfo->id-myedegrees[k].ed;
SWAP(myrinfo->id, myedegrees[k].ed, j);
if (myedegrees[k].ed == 0)
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].pid = from;
if (myrinfo->ed-myrinfo->id < 0)
BNDDelete(nbnd, bndind, bndptr, i);
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
me = where[ii];
myrinfo = graph->rinfo+ii;
if (myrinfo->edegrees == NULL) {
myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii];
}
myedegrees = myrinfo->edegrees;
ASSERT(CheckRInfo(myrinfo));
if (me == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
if (myrinfo->ed-myrinfo->id >= 0 && bndptr[ii] == -1)
BNDInsert(nbnd, bndind, bndptr, ii);
}
else if (me == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
if (myrinfo->ed-myrinfo->id < 0 && bndptr[ii] != -1)
BNDDelete(nbnd, bndind, bndptr, ii);
}
/* Remove contribution from the .ed of 'from' */
if (me != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == from) {
if (myedegrees[k].ed == adjwgt[j])
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].ed -= adjwgt[j];
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (me != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == to) {
myedegrees[k].ed += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
myedegrees[myrinfo->ndegrees].pid = to;
myedegrees[myrinfo->ndegrees++].ed = adjwgt[j];
}
}
ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]);
ASSERT(CheckRInfo(myrinfo));
}
nmoves++;
}
}
graph->nbnd = nbnd;
if (ctrl->dbglvl&DBG_REFINE) {
printf("\t [%5.4f %5.4f], Nb: %6d, Nmoves: %5d, Cut: %6d, LB: ",
npwgts[samin(ncon*nparts, npwgts)], npwgts[samax(ncon*nparts, npwgts)],
nbnd, nmoves, graph->mincut);
ComputeHKWayLoadImbalance(ncon, nparts, npwgts, tvec);
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
printf("\n");
}
if (graph->mincut == oldcut)
break;
}
fwspacefree(ctrl, ncon*nparts);
fwspacefree(ctrl, ncon*nparts);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Moc_KWayFM(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace, int npasses)
{
int h, i, ii, iii, j, k, c;
int pass, nvtxs, nedges, ncon;
int nmoves, nmoved, nswaps, nzgswaps;
/* int gnswaps, gnzgswaps; */
int me, firstvtx, lastvtx, yourlastvtx;
int from, to = -1, oldto, oldcut, mydomain, yourdomain, imbalanced, overweight;
int npes = ctrl->npes, mype = ctrl->mype, nparts = ctrl->nparts;
int nlupd, nsupd, nnbrs, nchanged;
idxtype *xadj, *ladjncy, *adjwgt, *vtxdist;
idxtype *where, *tmp_where, *moved;
floattype *lnpwgts, *gnpwgts, *ognpwgts, *pgnpwgts, *movewgts, *overfill;
idxtype *update, *supdate, *rupdate, *pe_updates;
idxtype *changed, *perm, *pperm, *htable;
idxtype *peind, *recvptr, *sendptr;
KeyValueType *swchanges, *rwchanges;
RInfoType *rinfo, *myrinfo, *tmp_myrinfo, *tmp_rinfo;
EdgeType *tmp_edegrees, *my_edegrees, *your_edegrees;
floattype lbvec[MAXNCON], *nvwgt, *badmaxpwgt, *ubvec, *tpwgts, lbavg, ubavg;
int *nupds_pe;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->KWayTmr));
/*************************/
/* set up common aliases */
/*************************/
nvtxs = graph->nvtxs;
nedges = graph->nedges;
ncon = graph->ncon;
vtxdist = graph->vtxdist;
xadj = graph->xadj;
ladjncy = graph->adjncy;
adjwgt = graph->adjwgt;
firstvtx = vtxdist[mype];
lastvtx = vtxdist[mype+1];
where = graph->where;
rinfo = graph->rinfo;
lnpwgts = graph->lnpwgts;
gnpwgts = graph->gnpwgts;
ubvec = ctrl->ubvec;
tpwgts = ctrl->tpwgts;
nnbrs = graph->nnbrs;
peind = graph->peind;
recvptr = graph->recvptr;
sendptr = graph->sendptr;
changed = idxmalloc(nvtxs, "KWR: changed");
rwchanges = wspace->pairs;
swchanges = rwchanges + recvptr[nnbrs];
/************************************/
/* set up important data structures */
/************************************/
perm = idxmalloc(nvtxs, "KWR: perm");
pperm = idxmalloc(nparts, "KWR: pperm");
update = idxmalloc(nvtxs, "KWR: update");
supdate = wspace->indices;
rupdate = supdate + recvptr[nnbrs];
nupds_pe = imalloc(npes, "KWR: nupds_pe");
htable = idxsmalloc(nvtxs+graph->nrecv, 0, "KWR: lhtable");
badmaxpwgt = fmalloc(nparts*ncon, "badmaxpwgt");
for (i=0; i<nparts; i++) {
for (h=0; h<ncon; h++) {
badmaxpwgt[i*ncon+h] = ubvec[h]*tpwgts[i*ncon+h];
}
}
movewgts = fmalloc(nparts*ncon, "KWR: movewgts");
ognpwgts = fmalloc(nparts*ncon, "KWR: ognpwgts");
pgnpwgts = fmalloc(nparts*ncon, "KWR: pgnpwgts");
overfill = fmalloc(nparts*ncon, "KWR: overfill");
moved = idxmalloc(nvtxs, "KWR: moved");
tmp_where = idxmalloc(nvtxs+graph->nrecv, "KWR: tmp_where");
tmp_rinfo = (RInfoType *)GKmalloc(sizeof(RInfoType)*nvtxs, "KWR: tmp_rinfo");
tmp_edegrees = (EdgeType *)GKmalloc(sizeof(EdgeType)*nedges, "KWR: tmp_edegrees");
idxcopy(nvtxs+graph->nrecv, where, tmp_where);
for (i=0; i<nvtxs; i++) {
tmp_rinfo[i].id = rinfo[i].id;
tmp_rinfo[i].ed = rinfo[i].ed;
tmp_rinfo[i].ndegrees = rinfo[i].ndegrees;
tmp_rinfo[i].degrees = tmp_edegrees+xadj[i];
for (j=0; j<rinfo[i].ndegrees; j++) {
tmp_rinfo[i].degrees[j].edge = rinfo[i].degrees[j].edge;
tmp_rinfo[i].degrees[j].ewgt = rinfo[i].degrees[j].ewgt;
}
}
nswaps = nzgswaps = 0;
/*********************************************************/
/* perform a small number of passes through the vertices */
/*********************************************************/
for (pass=0; pass<npasses; pass++) {
if (mype == 0)
RandomPermute(nparts, pperm, 1);
MPI_Bcast((void *)pperm, nparts, IDX_DATATYPE, 0, ctrl->comm);
FastRandomPermute(nvtxs, perm, 1);
oldcut = graph->mincut;
/* check to see if the partitioning is imbalanced */
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
ubavg = savg(ncon, ubvec);
lbavg = savg(ncon, lbvec);
imbalanced = (lbavg > ubavg) ? 1 : 0;
for (c=0; c<2; c++) {
scopy(ncon*nparts, gnpwgts, ognpwgts);
sset(ncon*nparts, 0.0, movewgts);
nmoved = 0;
/**********************************************/
/* PASS ONE -- record stats for desired moves */
/**********************************************/
for (iii=0; iii<nvtxs; iii++) {
i = perm[iii];
from = tmp_where[i];
nvwgt = graph->nvwgt+i*ncon;
for (h=0; h<ncon; h++)
if (fabs(nvwgt[h]-gnpwgts[from*ncon+h]) < SMALLFLOAT)
break;
if (h < ncon) {
continue;
}
/* check for a potential improvement */
if (tmp_rinfo[i].ed >= tmp_rinfo[i].id) {
my_edegrees = tmp_rinfo[i].degrees;
for (k=0; k<tmp_rinfo[i].ndegrees; k++) {
to = my_edegrees[k].edge;
if (ProperSide(c, pperm[from], pperm[to])) {
for (h=0; h<ncon; h++)
if (gnpwgts[to*ncon+h]+nvwgt[h] > badmaxpwgt[to*ncon+h] && nvwgt[h] > 0.0)
break;
if (h == ncon)
break;
}
}
oldto = to;
/* check if a subdomain was found that fits */
if (k < tmp_rinfo[i].ndegrees) {
for (j=k+1; j<tmp_rinfo[i].ndegrees; j++) {
to = my_edegrees[j].edge;
if (ProperSide(c, pperm[from], pperm[to])) {
for (h=0; h<ncon; h++)
if (gnpwgts[to*ncon+h]+nvwgt[h] > badmaxpwgt[to*ncon+h] && nvwgt[h] > 0.0)
break;
if (h == ncon) {
if (my_edegrees[j].ewgt > my_edegrees[k].ewgt ||
(my_edegrees[j].ewgt == my_edegrees[k].ewgt &&
IsHBalanceBetterTT(ncon,gnpwgts+oldto*ncon,gnpwgts+to*ncon,nvwgt,ubvec))){
k = j;
oldto = my_edegrees[k].edge;
}
}
}
}
to = oldto;
if (my_edegrees[k].ewgt > tmp_rinfo[i].id ||
(my_edegrees[k].ewgt == tmp_rinfo[i].id &&
(imbalanced || graph->level > 3 || iii % 8 == 0) &&
IsHBalanceBetterFT(ncon,gnpwgts+from*ncon,gnpwgts+to*ncon,nvwgt,ubvec))){
/****************************************/
/* Update tmp arrays of the moved vertex */
/****************************************/
tmp_where[i] = to;
moved[nmoved++] = i;
for (h=0; h<ncon; h++) {
lnpwgts[to*ncon+h] += nvwgt[h];
lnpwgts[from*ncon+h] -= nvwgt[h];
gnpwgts[to*ncon+h] += nvwgt[h];
gnpwgts[from*ncon+h] -= nvwgt[h];
movewgts[to*ncon+h] += nvwgt[h];
movewgts[from*ncon+h] -= nvwgt[h];
}
tmp_rinfo[i].ed += tmp_rinfo[i].id-my_edegrees[k].ewgt;
SWAP(tmp_rinfo[i].id, my_edegrees[k].ewgt, j);
if (my_edegrees[k].ewgt == 0) {
tmp_rinfo[i].ndegrees--;
my_edegrees[k].edge = my_edegrees[tmp_rinfo[i].ndegrees].edge;
my_edegrees[k].ewgt = my_edegrees[tmp_rinfo[i].ndegrees].ewgt;
}
else {
my_edegrees[k].edge = from;
}
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
/* no need to bother about vertices on different pe's */
if (ladjncy[j] >= nvtxs)
continue;
me = ladjncy[j];
mydomain = tmp_where[me];
myrinfo = tmp_rinfo+me;
your_edegrees = myrinfo->degrees;
if (mydomain == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
}
else {
if (mydomain == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
}
}
/* Remove contribution from the .ed of 'from' */
if (mydomain != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == from) {
if (your_edegrees[k].ewgt == adjwgt[j]) {
myrinfo->ndegrees--;
your_edegrees[k].edge = your_edegrees[myrinfo->ndegrees].edge;
your_edegrees[k].ewgt = your_edegrees[myrinfo->ndegrees].ewgt;
}
else {
your_edegrees[k].ewgt -= adjwgt[j];
}
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (mydomain != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == to) {
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
your_edegrees[myrinfo->ndegrees].edge = to;
your_edegrees[myrinfo->ndegrees++].ewgt = adjwgt[j];
}
}
}
}
}
}
}
/******************************************/
/* Let processors know the subdomain wgts */
/* if all proposed moves commit. */
/******************************************/
MPI_Allreduce((void *)lnpwgts, (void *)pgnpwgts, nparts*ncon,
MPI_DOUBLE, MPI_SUM, ctrl->comm);
/**************************/
/* compute overfill array */
/**************************/
overweight = 0;
for (j=0; j<nparts; j++) {
for (h=0; h<ncon; h++) {
if (pgnpwgts[j*ncon+h] > ognpwgts[j*ncon+h]) {
overfill[j*ncon+h] =
(pgnpwgts[j*ncon+h]-badmaxpwgt[j*ncon+h]) /
(pgnpwgts[j*ncon+h]-ognpwgts[j*ncon+h]);
}
else {
overfill[j*ncon+h] = 0.0;
}
overfill[j*ncon+h] = amax(overfill[j*ncon+h], 0.0);
overfill[j*ncon+h] *= movewgts[j*ncon+h];
if (overfill[j*ncon+h] > 0.0)
overweight = 1;
ASSERTP(ctrl, ognpwgts[j*ncon+h] <= badmaxpwgt[j*ncon+h] ||
pgnpwgts[j*ncon+h] <= ognpwgts[j*ncon+h],
(ctrl, "%.4f %.4f %.4f\n", ognpwgts[j*ncon+h],
badmaxpwgt[j*ncon+h], pgnpwgts[j*ncon+h]));
}
}
/****************************************************/
/* select moves to undo according to overfill array */
/****************************************************/
if (overweight == 1) {
for (iii=0; iii<nmoved; iii++) {
i = moved[iii];
oldto = tmp_where[i];
nvwgt = graph->nvwgt+i*ncon;
my_edegrees = tmp_rinfo[i].degrees;
for (k=0; k<tmp_rinfo[i].ndegrees; k++)
if (my_edegrees[k].edge == where[i])
break;
for (h=0; h<ncon; h++)
if (nvwgt[h] > 0.0 && overfill[oldto*ncon+h] > nvwgt[h]/4.0)
break;
/**********************************/
/* nullify this move if necessary */
/**********************************/
if (k != tmp_rinfo[i].ndegrees && h != ncon) {
moved[iii] = -1;
from = oldto;
to = where[i];
for (h=0; h<ncon; h++) {
overfill[oldto*ncon+h] = amax(overfill[oldto*ncon+h]-nvwgt[h], 0.0);
}
tmp_where[i] = to;
tmp_rinfo[i].ed += tmp_rinfo[i].id-my_edegrees[k].ewgt;
SWAP(tmp_rinfo[i].id, my_edegrees[k].ewgt, j);
if (my_edegrees[k].ewgt == 0) {
tmp_rinfo[i].ndegrees--;
my_edegrees[k].edge = my_edegrees[tmp_rinfo[i].ndegrees].edge;
my_edegrees[k].ewgt = my_edegrees[tmp_rinfo[i].ndegrees].ewgt;
}
else {
my_edegrees[k].edge = from;
}
for (h=0; h<ncon; h++) {
lnpwgts[to*ncon+h] += nvwgt[h];
lnpwgts[from*ncon+h] -= nvwgt[h];
}
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
/* no need to bother about vertices on different pe's */
if (ladjncy[j] >= nvtxs)
continue;
me = ladjncy[j];
mydomain = tmp_where[me];
myrinfo = tmp_rinfo+me;
your_edegrees = myrinfo->degrees;
if (mydomain == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
}
else {
if (mydomain == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
}
}
/* Remove contribution from the .ed of 'from' */
if (mydomain != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == from) {
if (your_edegrees[k].ewgt == adjwgt[j]) {
myrinfo->ndegrees--;
your_edegrees[k].edge = your_edegrees[myrinfo->ndegrees].edge;
your_edegrees[k].ewgt = your_edegrees[myrinfo->ndegrees].ewgt;
}
else {
your_edegrees[k].ewgt -= adjwgt[j];
}
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (mydomain != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == to) {
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
your_edegrees[myrinfo->ndegrees].edge = to;
your_edegrees[myrinfo->ndegrees++].ewgt = adjwgt[j];
}
}
}
}
}
}
/*************************************************/
/* PASS TWO -- commit the remainder of the moves */
/*************************************************/
nlupd = nsupd = nmoves = nchanged = 0;
for (iii=0; iii<nmoved; iii++) {
i = moved[iii];
if (i == -1)
continue;
where[i] = tmp_where[i];
/* Make sure to update the vertex information */
if (htable[i] == 0) {
/* make sure you do the update */
htable[i] = 1;
update[nlupd++] = i;
}
/* Put the vertices adjacent to i into the update array */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = ladjncy[j];
if (htable[k] == 0) {
htable[k] = 1;
if (k<nvtxs)
update[nlupd++] = k;
else
supdate[nsupd++] = k;
}
}
nmoves++;
nswaps++;
/* check number of zero-gain moves */
for (k=0; k<rinfo[i].ndegrees; k++)
if (rinfo[i].degrees[k].edge == to)
break;
if (rinfo[i].id == rinfo[i].degrees[k].ewgt)
nzgswaps++;
if (graph->pexadj[i+1]-graph->pexadj[i] > 0)
changed[nchanged++] = i;
}
/* Tell interested pe's the new where[] info for the interface vertices */
CommChangedInterfaceData(ctrl, graph, nchanged, changed, where,
swchanges, rwchanges, wspace->pv4);
IFSET(ctrl->dbglvl, DBG_RMOVEINFO,
rprintf(ctrl, "\t[%d %d], [%.4f], [%d %d %d]\n",
pass, c, badmaxpwgt[0],
GlobalSESum(ctrl, nmoves),
GlobalSESum(ctrl, nsupd),
GlobalSESum(ctrl, nlupd)));
/*-------------------------------------------------------------
/ Time to communicate with processors to send the vertices
/ whose degrees need to be update.
/-------------------------------------------------------------*/
/* Issue the receives first */
for (i=0; i<nnbrs; i++) {
MPI_Irecv((void *)(rupdate+sendptr[i]), sendptr[i+1]-sendptr[i], IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. This needs some preporcessing */
for (i=0; i<nsupd; i++) {
htable[supdate[i]] = 0;
supdate[i] = graph->imap[supdate[i]];
}
iidxsort(nsupd, supdate);
for (j=i=0; i<nnbrs; i++) {
yourlastvtx = vtxdist[peind[i]+1];
for (k=j; k<nsupd && supdate[k] < yourlastvtx; k++);
MPI_Isend((void *)(supdate+j), k-j, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i);
j = k;
}
/* OK, now get into the loop waiting for the send/recv operations to finish */
MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++)
MPI_Get_count(ctrl->statuses+i, IDX_DATATYPE, nupds_pe+i);
MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/*-------------------------------------------------------------
/ Place the recieved to-be updated vertices into update[]
/-------------------------------------------------------------*/
for (i=0; i<nnbrs; i++) {
pe_updates = rupdate+sendptr[i];
for (j=0; j<nupds_pe[i]; j++) {
k = pe_updates[j];
if (htable[k-firstvtx] == 0) {
htable[k-firstvtx] = 1;
update[nlupd++] = k-firstvtx;
}
}
}
/*-------------------------------------------------------------
/ Update the rinfo of the vertices in the update[] array
/-------------------------------------------------------------*/
for (ii=0; ii<nlupd; ii++) {
i = update[ii];
ASSERT(ctrl, htable[i] == 1);
htable[i] = 0;
mydomain = where[i];
myrinfo = rinfo+i;
tmp_myrinfo = tmp_rinfo+i;
my_edegrees = myrinfo->degrees;
your_edegrees = tmp_myrinfo->degrees;
graph->lmincut -= myrinfo->ed;
myrinfo->ndegrees = 0;
myrinfo->id = 0;
myrinfo->ed = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
yourdomain = where[ladjncy[j]];
if (mydomain != yourdomain) {
myrinfo->ed += adjwgt[j];
for (k=0; k<myrinfo->ndegrees; k++) {
if (my_edegrees[k].edge == yourdomain) {
my_edegrees[k].ewgt += adjwgt[j];
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
my_edegrees[k].edge = yourdomain;
my_edegrees[k].ewgt = adjwgt[j];
your_edegrees[k].edge = yourdomain;
your_edegrees[k].ewgt = adjwgt[j];
myrinfo->ndegrees++;
}
ASSERT(ctrl, myrinfo->ndegrees <= xadj[i+1]-xadj[i]);
ASSERT(ctrl, tmp_myrinfo->ndegrees <= xadj[i+1]-xadj[i]);
}
else {
myrinfo->id += adjwgt[j];
}
}
graph->lmincut += myrinfo->ed;
tmp_myrinfo->id = myrinfo->id;
tmp_myrinfo->ed = myrinfo->ed;
tmp_myrinfo->ndegrees = myrinfo->ndegrees;
}
/* finally, sum-up the partition weights */
MPI_Allreduce((void *)lnpwgts, (void *)gnpwgts, nparts*ncon,
MPI_DOUBLE, MPI_SUM, ctrl->comm);
}
graph->mincut = GlobalSESum(ctrl, graph->lmincut)/2;
if (graph->mincut == oldcut)
break;
}
/*
gnswaps = GlobalSESum(ctrl, nswaps);
gnzgswaps = GlobalSESum(ctrl, nzgswaps);
if (mype == 0)
printf("niters: %d, nswaps: %d, nzgswaps: %d\n", pass+1, gnswaps, gnzgswaps);
*/
GKfree((void **)&badmaxpwgt, (void **)&update, (void **)&nupds_pe, (void **)&htable, LTERM);
GKfree((void **)&changed, (void **)&pperm, (void **)&perm, (void **)&moved, LTERM);
GKfree((void **)&pgnpwgts, (void **)&ognpwgts, (void **)&overfill, (void **)&movewgts, LTERM);
GKfree((void **)&tmp_where, (void **)&tmp_rinfo, (void **)&tmp_edegrees, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->KWayTmr));
}
/*************************************************************************
* This function balances two partitions by moving the highest gain
* (including negative gain) vertices to the other domain.
* It is used only when tha unbalance is due to non contigous
* subdomains. That is, the are no boundary vertices.
* It moves vertices from the domain that is overweight to the one that
* is underweight.
**************************************************************************/
void MocInit2WayBalance(CtrlType *ctrl, GraphType *graph, float *tpwgts)
{
int i, ii, j, k, l, kwgt, nvtxs, nbnd, ncon, nswaps, from, to, pass, me, cnum, tmp;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *perm, *qnum;
float *nvwgt, *npwgts;
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
nvwgt = graph->nvwgt;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
npwgts = graph->npwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
perm = idxwspacemalloc(ctrl, nvtxs);
qnum = idxwspacemalloc(ctrl, nvtxs);
/* This is called for initial partitioning so we know from where to pick nodes */
from = 1;
to = (from+1)%2;
if (ctrl->dbglvl&DBG_REFINE) {
printf("Parts: [");
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: %.3f [B]\n", tpwgts[0], tpwgts[1],
graph->nvtxs, graph->nbnd, graph->mincut,
Compute2WayHLoadImbalance(ncon, npwgts, tpwgts));
}
for (i=0; i<ncon; i++) {
PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1);
PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1);
}
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
ASSERT(CheckGraph(graph));
/* Compute the queues in which each vertex will be assigned to */
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
/* Insert the nodes of the proper partition in the appropriate priority queue */
RandomPermute(nvtxs, perm, 1);
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (where[i] == from) {
if (ed[i] > 0)
PQueueInsert(&parts[qnum[i]][0], i, ed[i]-id[i]);
else
PQueueInsert(&parts[qnum[i]][1], i, ed[i]-id[i]);
}
}
mincut = graph->mincut;
nbnd = graph->nbnd;
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (AreAnyVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, tpwgts[from]))
break;
if ((cnum = SelectQueueOneWay(ncon, npwgts, tpwgts, from, parts)) == -1)
break;
if ((higain = PQueueGetMax(&parts[cnum][0])) == -1)
higain = PQueueGetMax(&parts[cnum][1]);
mincut -= (ed[higain]-id[higain]);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
where[higain] = to;
if (ctrl->dbglvl&DBG_MOVEINFO) {
printf("Moved %6d from %d(%d). [%5d] %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], mincut);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf(", LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts));
if (ed[higain] == 0 && id[higain] > 0)
printf("\t Pulled from the interior!\n");
}
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (where[k] == from) {
if (ed[k] > 0 && bndptr[k] == -1) { /* It moves in boundary */
PQueueDelete(&parts[qnum[k]][1], k, oldgain);
PQueueInsert(&parts[qnum[k]][0], k, ed[k]-id[k]);
}
else { /* It must be in the boundary already */
if (bndptr[k] == -1)
printf("What you thought was wrong!\n");
PQueueUpdate(&parts[qnum[k]][0], k, oldgain, ed[k]-id[k]);
}
}
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
ASSERTP(ComputeCut(graph, where) == mincut, ("%d != %d\n", ComputeCut(graph, where), mincut));
}
if (ctrl->dbglvl&DBG_REFINE) {
printf("\tMincut: %6d, NBND: %6d, NPwgts: ", mincut, nbnd);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf(", LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts));
}
graph->mincut = mincut;
graph->nbnd = nbnd;
for (i=0; i<ncon; i++) {
PQueueFree(ctrl, &parts[i][0]);
PQueueFree(ctrl, &parts[i][1]);
}
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
void FM_Mc2WayCutRefine(ctrl_t *ctrl, graph_t *graph, real_t *ntpwgts, idx_t niter)
{
idx_t i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass,
me, limit, tmp, cnum;
idx_t *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *id, *ed,
*bndptr, *bndind;
idx_t *moved, *swaps, *perm, *qnum;
idx_t higain, mincut, initcut, newcut, mincutorder;
real_t *invtvwgt, *ubfactors, *minbalv, *newbalv;
real_t origbal, minbal, newbal, rgain, ffactor;
rpq_t **queues;
WCOREPUSH;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
invtvwgt = graph->invtvwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
pwgts = graph->pwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = iwspacemalloc(ctrl, nvtxs);
swaps = iwspacemalloc(ctrl, nvtxs);
perm = iwspacemalloc(ctrl, nvtxs);
qnum = iwspacemalloc(ctrl, nvtxs);
ubfactors = rwspacemalloc(ctrl, ncon);
newbalv = rwspacemalloc(ctrl, ncon);
minbalv = rwspacemalloc(ctrl, ncon);
limit = gk_min(gk_max(0.01*nvtxs, 25), 150);
/* Determine a fudge factor to allow the refinement routines to get out
of tight balancing constraints. */
ffactor = .5/gk_max(20, nvtxs);
/* Initialize the queues */
queues = (rpq_t **)wspacemalloc(ctrl, 2*ncon*sizeof(rpq_t *));
for (i=0; i<2*ncon; i++)
queues[i] = rpqCreate(nvtxs);
for (i=0; i<nvtxs; i++)
qnum[i] = iargmax_nrm(ncon, vwgt+i*ncon, invtvwgt);
/* Determine the unbalance tolerance for each constraint. The tolerance is
equal to the maximum of the original load imbalance and the user-supplied
allowed tolerance. The rationale behind this approach is to allow the
refinement routine to improve the cut, without having to worry about fixing
load imbalance problems. The load imbalance is addressed by the balancing
routines. */
origbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ctrl->ubfactors, ubfactors);
for (i=0; i<ncon; i++)
ubfactors[i] = (ubfactors[i] > 0 ? ctrl->ubfactors[i]+ubfactors[i] : ctrl->ubfactors[i]);
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, origbal, -2));
iset(nvtxs, -1, moved);
for (pass=0; pass<niter; pass++) { /* Do a number of passes */
for (i=0; i<2*ncon; i++)
rpqReset(queues[i]);
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
minbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ubfactors, minbalv);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
irandArrayPermute(nbnd, perm, nbnd/5, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(ed[i] > 0 || id[i] == 0);
ASSERT(bndptr[i] != -1);
//rgain = 1.0*(ed[i]-id[i])/sqrt(vwgt[i*ncon+qnum[i]]+1);
//rgain = (ed[i]-id[i] > 0 ? 1.0*(ed[i]-id[i])/sqrt(vwgt[i*ncon+qnum[i]]+1) : ed[i]-id[i]);
rgain = ed[i]-id[i];
rpqInsert(queues[2*qnum[i]+where[i]], i, rgain);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
SelectQueue(graph, ctrl->pijbm, ubfactors, queues, &from, &cnum);
to = (from+1)%2;
if (from == -1 || (higain = rpqGetTop(queues[2*cnum+from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
newcut -= (ed[higain]-id[higain]);
iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1);
iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+from*ncon, 1);
newbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ubfactors, newbalv);
if ((newcut < mincut && newbal <= ffactor) ||
(newcut == mincut && (newbal < minbal ||
(newbal == minbal && BetterBalance2Way(ncon, minbalv, newbalv))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
rcopy(ncon, newbalv, minbalv);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+from*ncon, 1);
iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
if (ctrl->dbglvl&METIS_DBG_MOVEINFO) {
printf("Moved%6"PRIDX" from %"PRIDX"(%"PRIDX") Gain:%5"PRIDX", "
"Cut:%5"PRIDX", NPwgts:", higain, from, cnum, ed[higain]-id[higain], newcut);
for (l=0; l<ncon; l++)
printf("(%.3"PRREAL" %.3"PRREAL")", pwgts[l]*invtvwgt[l], pwgts[ncon+l]*invtvwgt[l]);
printf(" %+.3"PRREAL" LB: %.3"PRREAL"(%+.3"PRREAL")\n",
minbal, ComputeLoadImbalance(graph, 2, ctrl->pijbm), newbal);
}
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
rpqDelete(queues[2*qnum[k]+where[k]], k);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1) {
//rgain = 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1);
//rgain = (ed[k]-id[k] > 0 ?
// 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1) : ed[k]-id[k]);
rgain = ed[k]-id[k];
rpqUpdate(queues[2*qnum[k]+where[k]], k, rgain);
}
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1) {
//rgain = 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1);
//rgain = (ed[k]-id[k] > 0 ?
// 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1) : ed[k]-id[k]);
rgain = ed[k]-id[k];
rpqInsert(queues[2*qnum[k]+where[k]], k, rgain);
}
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1);
iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->nbnd = nbnd;
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, minbal, mincutorder));
if (mincutorder <= 0 || mincut == initcut)
break;
}
for (i=0; i<2*ncon; i++)
rpqDestroy(queues[i]);
WCOREPOP;
}
/*************************************************************************
* This function performs a k-way directed diffusion
**************************************************************************/
real_t WavefrontDiffusion(ctrl_t *ctrl, graph_t *graph, idx_t *home)
{
idx_t ii, i, j, k, l, nvtxs, nedges, nparts;
idx_t from, to, edge, done, nswaps, noswaps, totalv, wsize;
idx_t npasses, first, second, third, mind, maxd;
idx_t *xadj, *adjncy, *adjwgt, *where, *perm;
idx_t *rowptr, *colind, *ed, *psize;
real_t *transfer, *tmpvec;
real_t balance = -1.0, *load, *solution, *workspace;
real_t *nvwgt, *npwgts, flowFactor, cost, ubfactor;
matrix_t matrix;
ikv_t *cand;
idx_t ndirty, nclean, dptr, clean;
nvtxs = graph->nvtxs;
nedges = graph->nedges;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
nparts = ctrl->nparts;
ubfactor = ctrl->ubvec[0];
matrix.nrows = nparts;
flowFactor = 0.35;
flowFactor = (ctrl->mype == 2) ? 0.50 : flowFactor;
flowFactor = (ctrl->mype == 3) ? 0.75 : flowFactor;
flowFactor = (ctrl->mype == 4) ? 1.00 : flowFactor;
/* allocate memory */
solution = rmalloc(4*nparts+2*nedges, "WavefrontDiffusion: solution");
tmpvec = solution + nparts;
npwgts = solution + 2*nparts;
load = solution + 3*nparts;
matrix.values = solution + 4*nparts;
transfer = matrix.transfer = solution + 4*nparts + nedges;
perm = imalloc(2*nvtxs+2*nparts+nedges+1, "WavefrontDiffusion: perm");
ed = perm + nvtxs;
psize = perm + 2*nvtxs;
rowptr = matrix.rowptr = perm + 2*nvtxs + nparts;
colind = matrix.colind = perm + 2*nvtxs + 2*nparts + 1;
/*GKTODO - Potential problem with this malloc */
wsize = gk_max(sizeof(real_t)*nparts*6, sizeof(idx_t)*(nvtxs+nparts*2+1));
workspace = (real_t *)gk_malloc(wsize, "WavefrontDiffusion: workspace");
cand = ikvmalloc(nvtxs, "WavefrontDiffusion: cand");
/*****************************/
/* Populate empty subdomains */
/*****************************/
iset(nparts, 0, psize);
for (i=0; i<nvtxs; i++)
psize[where[i]]++;
mind = iargmin(nparts, psize);
maxd = iargmax(nparts, psize);
if (psize[mind] == 0) {
for (i=0; i<nvtxs; i++) {
k = (RandomInRange(nvtxs)+i)%nvtxs;
if (where[k] == maxd) {
where[k] = mind;
psize[mind]++;
psize[maxd]--;
break;
}
}
}
iset(nvtxs, 0, ed);
rset(nparts, 0.0, npwgts);
for (i=0; i<nvtxs; i++) {
npwgts[where[i]] += nvwgt[i];
for (j=xadj[i]; j<xadj[i+1]; j++)
ed[i] += (where[i] != where[adjncy[j]] ? adjwgt[j] : 0);
}
ComputeLoad(graph, nparts, load, ctrl->tpwgts, 0);
done = 0;
/* zero out the tmpvec array */
rset(nparts, 0.0, tmpvec);
npasses = gk_min(nparts/2, NGD_PASSES);
for (l=0; l<npasses; l++) {
/* Set-up and solve the diffusion equation */
nswaps = 0;
/************************/
/* Solve flow equations */
/************************/
SetUpConnectGraph(graph, &matrix, (idx_t *)workspace);
/* check for disconnected subdomains */
for(i=0; i<matrix.nrows; i++) {
if (matrix.rowptr[i]+1 == matrix.rowptr[i+1]) {
cost = (real_t)(ctrl->mype);
goto CleanUpAndExit;
}
}
ConjGrad2(&matrix, load, solution, 0.001, workspace);
ComputeTransferVector(1, &matrix, solution, transfer, 0);
GetThreeMax(nparts, load, &first, &second, &third);
if (l%3 == 0) {
FastRandomPermute(nvtxs, perm, 1);
}
else {
/*****************************/
/* move dirty vertices first */
/*****************************/
ndirty = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
ndirty++;
}
dptr = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
perm[dptr++] = i;
else
perm[ndirty++] = i;
}
PASSERT(ctrl, ndirty == nvtxs);
ndirty = dptr;
nclean = nvtxs-dptr;
FastRandomPermute(ndirty, perm, 0);
FastRandomPermute(nclean, perm+ndirty, 0);
}
if (ctrl->mype == 0) {
for (j=nvtxs, k=0, ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (ed[i] != 0) {
cand[k].key = -ed[i];
cand[k++].val = i;
}
else {
cand[--j].key = 0;
cand[j].val = i;
}
}
ikvsorti(k, cand);
}
for (ii=0; ii<nvtxs/3; ii++) {
i = (ctrl->mype == 0) ? cand[ii].val : perm[ii];
from = where[i];
/* don't move out the last vertex in a subdomain */
if (psize[from] == 1)
continue;
clean = (from == home[i]) ? 1 : 0;
/* only move from top three or dirty vertices */
if (from != first && from != second && from != third && clean)
continue;
/* Scatter the sparse transfer row into the dense tmpvec row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++)
tmpvec[colind[j]] = transfer[j];
for (j=xadj[i]; j<xadj[i+1]; j++) {
to = where[adjncy[j]];
if (from != to) {
if (tmpvec[to] > (flowFactor * nvwgt[i])) {
tmpvec[to] -= nvwgt[i];
INC_DEC(psize[to], psize[from], 1);
INC_DEC(npwgts[to], npwgts[from], nvwgt[i]);
INC_DEC(load[to], load[from], nvwgt[i]);
where[i] = to;
nswaps++;
/* Update external degrees */
ed[i] = 0;
for (k=xadj[i]; k<xadj[i+1]; k++) {
edge = adjncy[k];
ed[i] += (to != where[edge] ? adjwgt[k] : 0);
if (where[edge] == from)
ed[edge] += adjwgt[k];
if (where[edge] == to)
ed[edge] -= adjwgt[k];
}
break;
}
}
}
/* Gather the dense tmpvec row into the sparse transfer row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++) {
transfer[j] = tmpvec[colind[j]];
tmpvec[colind[j]] = 0.0;
}
ASSERT(fabs(rsum(nparts, tmpvec, 1)) < .0001)
}
if (l % 2 == 1) {
balance = rmax(nparts, npwgts)*nparts;
if (balance < ubfactor + 0.035)
done = 1;
if (GlobalSESum(ctrl, done) > 0)
break;
noswaps = (nswaps > 0) ? 0 : 1;
if (GlobalSESum(ctrl, noswaps) > ctrl->npes/2)
break;
}
}
graph->mincut = ComputeSerialEdgeCut(graph);
totalv = Mc_ComputeSerialTotalV(graph, home);
cost = ctrl->ipc_factor * (real_t)graph->mincut + ctrl->redist_factor * (real_t)totalv;
CleanUpAndExit:
gk_free((void **)&solution, (void **)&perm, (void **)&workspace, (void **)&cand, LTERM);
return cost;
}
/*************************************************************************
* This function balances two partitions by moving the highest gain
* (including negative gain) vertices to the other domain.
* It is used only when tha unbalance is due to non contigous
* subdomains. That is, the are no boundary vertices.
* It moves vertices from the domain that is overweight to the one that
* is underweight.
**************************************************************************/
void Mc_Serial_Init2WayBalance(GraphType *graph, float *tpwgts)
{
int i, ii, j, k;
int kwgt, nvtxs, nbnd, ncon, nswaps, from, to, cnum, tmp;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *qnum;
float *nvwgt, *npwgts;
FPQueueType parts[MAXNCON][2];
int higain, oldgain, mincut;
KeyValueType *cand;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
nvwgt = graph->nvwgt;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->sendind;
ed = graph->recvind;
npwgts = graph->gnpwgts;
bndptr = graph->sendptr;
bndind = graph->recvptr;
qnum = idxmalloc(nvtxs, "qnum");
cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand");
/* This is called for initial partitioning so we know from where to pick nodes */
from = 1;
to = (from+1)%2;
for (i=0; i<ncon; i++) {
FPQueueInit(&parts[i][0], nvtxs);
FPQueueInit(&parts[i][1], nvtxs);
}
/* Compute the queues in which each vertex will be assigned to */
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
for (i=0; i<nvtxs; i++) {
cand[i].key = id[i]-ed[i];
cand[i].val = i;
}
ikeysort(nvtxs, cand);
/* Insert the nodes of the proper partition in the appropriate priority queue */
for (ii=0; ii<nvtxs; ii++) {
i = cand[ii].val;
if (where[i] == from) {
if (ed[i] > 0)
FPQueueInsert(&parts[qnum[i]][0], i, (float)(ed[i]-id[i]));
else
FPQueueInsert(&parts[qnum[i]][1], i, (float)(ed[i]-id[i]));
}
}
mincut = graph->mincut;
nbnd = graph->gnvtxs;
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (Serial_AreAnyVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, tpwgts+from*ncon))
break;
if ((cnum = Serial_SelectQueueOneWay(ncon, npwgts, tpwgts, from, parts)) == -1)
break;
if ((higain = FPQueueGetMax(&parts[cnum][0])) == -1)
higain = FPQueueGetMax(&parts[cnum][1]);
mincut -= (ed[higain]-id[higain]);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
where[higain] = to;
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (where[k] == from) {
if (ed[k] > 0 && bndptr[k] == -1) { /* It moves in boundary */
FPQueueDelete(&parts[qnum[k]][1], k);
FPQueueInsert(&parts[qnum[k]][0], k, (float)(ed[k]-id[k]));
}
else { /* It must be in the boundary already */
FPQueueUpdate(&parts[qnum[k]][0], k, (float)(oldgain), (float)(ed[k]-id[k]));
}
}
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->gnvtxs = nbnd;
for (i=0; i<ncon; i++) {
FPQueueFree(&parts[i][0]);
FPQueueFree(&parts[i][1]);
}
GKfree((void **)&cand, (void **)&qnum, LTERM);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void MocFM_2WayEdgeRefine(CtrlType *ctrl, GraphType *graph, float *tpwgts, int npasses)
{
int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, me, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *perm, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, initcut, newcut, mincutorder;
float rtpwgts[2];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
npwgts = graph->npwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
qnum = idxwspacemalloc(ctrl, nvtxs);
limit = amin(amax(0.01*nvtxs, 25), 150);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1);
PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1);
}
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
origbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
rtpwgts[0] = origbal*tpwgts[0];
rtpwgts[1] = origbal*tpwgts[1];
/*
if (ctrl->dbglvl&DBG_REFINE) {
printf("Parts: [");
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: %.3f\n", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut, origbal);
}
*/
idxset(nvtxs, -1, moved);
for (pass=0; pass<npasses; pass++) { /* Do a number of passes */
for (i=0; i<ncon; i++) {
PQueueReset(&parts[i][0]);
PQueueReset(&parts[i][1]);
}
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
minbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(ed[i] > 0 || id[i] == 0);
ASSERT(bndptr[i] != -1);
PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
SelectQueue(ncon, npwgts, rtpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if ((newcut < mincut && newbal-origbal <= .00001) ||
(newcut == mincut && (newbal < minbal ||
(newbal == minbal && BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/*
if (ctrl->dbglvl&DBG_MOVEINFO) {
printf("Moved %6d from %d(%d). Gain: %5d, Cut: %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf(", %.3f LB: %.3f\n", minbal, newbal);
}
*/
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
PQueueDelete(&parts[qnum[k]][where[k]], k, oldgain);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]);
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
PQueueInsert(&parts[qnum[k]][where[k]], k, ed[k]-id[k]);
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/*
if (ctrl->dbglvl&DBG_REFINE) {
printf("\tMincut: %6d at %5d, NBND: %6d, NPwgts: [", mincut, mincutorder, nbnd);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("], LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts));
}
*/
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut == initcut)
break;
}
for (i=0; i<ncon; i++) {
PQueueFree(ctrl, &parts[i][0]);
PQueueFree(ctrl, &parts[i][1]);
}
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void FM_2WayEdgeRefine(CtrlType *ctrl, GraphType *graph, int *tpwgts, int npasses)
{
int i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, limit, tmp;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts;
idxtype *moved, *swaps, *perm;
PQueueType parts[2];
int higain, oldgain, mincut, mindiff, origdiff, initcut, newcut, mincutorder, avgvwgt;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
pwgts = graph->pwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
limit = (int) amin(amax(0.01*nvtxs, 15), 100);
avgvwgt = amin((pwgts[0]+pwgts[1])/20, 2*(pwgts[0]+pwgts[1])/nvtxs);
tmp = graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)];
PQueueInit(ctrl, &parts[0], nvtxs, tmp);
PQueueInit(ctrl, &parts[1], nvtxs, tmp);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions: [%6d %6d] T[%6d %6d], Nv-Nb[%6d %6d]. ICut: %6d\n",
pwgts[0], pwgts[1], tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
origdiff = abs(tpwgts[0]-pwgts[0]);
idxset(nvtxs, -1, moved);
for (pass=0; pass<npasses; pass++) { /* Do a number of passes */
PQueueReset(&parts[0]);
PQueueReset(&parts[1]);
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
mindiff = abs(tpwgts[0]-pwgts[0]);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = perm[ii];
ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0);
ASSERT(bndptr[bndind[i]] != -1);
PQueueInsert(&parts[where[bndind[i]]], bndind[i], ed[bndind[i]]-id[bndind[i]]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
from = (tpwgts[0]-pwgts[0] < tpwgts[1]-pwgts[1] ? 0 : 1);
to = (from+1)%2;
if ((higain = PQueueGetMax(&parts[from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
newcut -= (ed[higain]-id[higain]);
INC_DEC(pwgts[to], pwgts[from], vwgt[higain]);
if ((newcut < mincut && abs(tpwgts[0]-pwgts[0]) <= origdiff+avgvwgt) ||
(newcut == mincut && abs(tpwgts[0]-pwgts[0]) < mindiff)) {
mincut = newcut;
mindiff = abs(tpwgts[0]-pwgts[0]);
mincutorder = nswaps;
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
INC_DEC(pwgts[from], pwgts[to], vwgt[higain]);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
printf("Moved %6d from %d. [%3d %3d] %5d [%4d %4d]\n", higain, from, ed[higain]-id[higain], vwgt[higain], newcut, pwgts[0], pwgts[1]));
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
PQueueDelete(&parts[where[k]], k, oldgain);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
PQueueUpdate(&parts[where[k]], k, oldgain, ed[k]-id[k]);
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
PQueueInsert(&parts[where[k]], k, ed[k]-id[k]);
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
INC_DEC(pwgts[to], pwgts[(to+1)%2], vwgt[higain]);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\tMinimum cut: %6d at %5d, PWGTS: [%6d %6d], NBND: %6d\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut == initcut)
break;
}
PQueueFree(ctrl, &parts[0]);
PQueueFree(ctrl, &parts[1]);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
void FM_2WayNodeRefine_TwoSidedP(CtrlType *ctrl, GraphType *graph,
idxtype *hmarker, float ubfactor, int npasses)
{
int i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind;
idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idxtype *mptr, *mind, *moved, *swaps, *perm;
PQueueType parts[2];
NRInfoType *rinfo;
int higain, oldgain, mincut, initcut, mincutorder;
int pass, to, other, limit;
int badmaxpwgt, mindiff, newdiff;
int u[2], g[2];
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
i = ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt);
PQueueInit(ctrl, &parts[0], nvtxs, i);
PQueueInit(ctrl, &parts[1], nvtxs, i);
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
mptr = idxwspacemalloc(ctrl, nvtxs+1);
mind = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions: [%6d %6d] Nv-Nb[%6d %6d]. ISep: %6d\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
badmaxpwgt = (int)(ubfactor*amax(pwgts[0], pwgts[1]));
for (pass=0; pass<npasses; pass++) {
idxset(nvtxs, -1, moved);
PQueueReset(&parts[0]);
PQueueReset(&parts[1]);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(where[i] == 2);
if (hmarker[i] == -1) {
PQueueInsert(&parts[0], i, vwgt[i]-rinfo[i].edegrees[1]);
PQueueInsert(&parts[1], i, vwgt[i]-rinfo[i].edegrees[0]);
moved[i] = -5;
}
else if (hmarker[i] != 2) {
PQueueInsert(&parts[hmarker[i]], i, vwgt[i]-rinfo[i].edegrees[(hmarker[i]+1)%2]);
moved[i] = -(10+hmarker[i]);
}
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = nbnd;
/******************************************************
* Get into the FM loop
*******************************************************/
mptr[0] = nmind = 0;
mindiff = abs(pwgts[0]-pwgts[1]);
to = (pwgts[0] < pwgts[1] ? 0 : 1);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
u[0] = PQueueSeeMax(&parts[0]);
u[1] = PQueueSeeMax(&parts[1]);
if (u[0] != -1 && u[1] != -1) {
g[0] = vwgt[u[0]]-rinfo[u[0]].edegrees[1];
g[1] = vwgt[u[1]]-rinfo[u[1]].edegrees[0];
to = (g[0] > g[1] ? 0 : (g[0] < g[1] ? 1 : pass%2));
if (pwgts[to]+vwgt[u[to]] > badmaxpwgt)
to = (to+1)%2;
}
else if (u[0] == -1 && u[1] == -1) {
break;
}
else if (u[0] != -1 && pwgts[0]+vwgt[u[0]] <= badmaxpwgt) {
to = 0;
}
else if (u[1] != -1 && pwgts[1]+vwgt[u[1]] <= badmaxpwgt) {
to = 1;
}
else
break;
other = (to+1)%2;
higain = PQueueGetMax(&parts[to]);
/* Delete its matching entry in the other queue */
if (moved[higain] == -5)
PQueueDelete(&parts[other], higain, vwgt[higain]-rinfo[higain].edegrees[to]);
ASSERT(bndptr[higain] != -1);
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = abs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > limit) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
oldgain = vwgt[k]-rinfo[k].edegrees[to];
rinfo[k].edegrees[to] += vwgt[higain];
if (moved[k] == -5 || moved[k] == -(10+other))
PQueueUpdate(&parts[other], k, oldgain, oldgain-vwgt[higain]);
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
oldgain = vwgt[kk]-rinfo[kk].edegrees[other];
rinfo[kk].edegrees[other] -= vwgt[k];
if (moved[kk] == -5 || moved[kk] == -(10+to))
PQueueUpdate(&parts[to], kk, oldgain, oldgain+vwgt[k]);
}
}
/* Insert the new vertex into the priority queue (if it has not been moved). */
if (moved[k] == -1 && (hmarker[k] == -1 || hmarker[k] == to)) {
PQueueInsert(&parts[to], k, vwgt[k]-edegrees[other]);
moved[k] = -(10+to);
}
#ifdef FULLMOVES /* this does not work as well as the above partial one */
if (moved[k] == -1) {
if (hmarker[k] == -1) {
PQueueInsert(&parts[0], k, vwgt[k]-edegrees[1]);
PQueueInsert(&parts[1], k, vwgt[k]-edegrees[0]);
moved[k] = -5;
}
else if (hmarker[k] != 2) {
PQueueInsert(&parts[hmarker[k]], k, vwgt[k]-edegrees[(hmarker[k]+1)%2]);
moved[k] = -(10+hmarker[k]);
}
}
#endif
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
printf("Moved %6d to %3d, Gain: %5d [%5d] [%4d %4d] \t[%5d %5d %5d]\n", higain, to, g[to], g[other], vwgt[u[to]], vwgt[u[other]], pwgts[0], pwgts[1], pwgts[2]));
}
/****************************************************************
* Roll back computation
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
to = where[higain];
other = (to+1)%2;
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\tMinimum sep: %6d at %5d, PWGTS: [%6d %6d], NBND: %6d\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut >= initcut)
break;
}
PQueueFree(ctrl, &parts[0]);
PQueueFree(ctrl, &parts[1]);
idxwspacefree(ctrl, nvtxs+1);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
void FM_2WayNodeRefine_OneSidedP(CtrlType *ctrl, GraphType *graph,
idxtype *hmarker, float ubfactor, int npasses)
{
int i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind, nbad, qsize;
idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idxtype *mptr, *mind, *swaps, *perm, *inqueue;
PQueueType parts;
NRInfoType *rinfo;
int higain, oldgain, mincut, initcut, mincutorder;
int pass, from, to, limit;
int badmaxpwgt, mindiff, newdiff;
ASSERT(graph->mincut == graph->pwgts[2]);
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
PQueueInit(ctrl, &parts, nvtxs, ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt));
perm = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
mptr = idxwspacemalloc(ctrl, nvtxs+1);
mind = idxwspacemalloc(ctrl, nvtxs);
inqueue = idxwspacemalloc(ctrl, nvtxs);
idxset(nvtxs, -1, inqueue);
badmaxpwgt = (int)(ubfactor*amax(pwgts[0], pwgts[1]));
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions-N1: [%6d %6d] Nv-Nb[%6d %6d] MaxPwgt[%6d]. ISep: %6d\n",
pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, badmaxpwgt, graph->mincut));
to = (pwgts[0] < pwgts[1] ? 1 : 0);
for (pass=0; pass<npasses; pass++) {
from = to;
to = (from+1)%2;
PQueueReset(&parts);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(where[i] == 2);
if (hmarker[i] == -1 || hmarker[i] == to) {
PQueueInsert(&parts, i, vwgt[i]-rinfo[i].edegrees[from]);
inqueue[i] = pass;
}
}
qsize = parts.nnodes;
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = nbnd;
/******************************************************
* Get into the FM loop
*******************************************************/
mptr[0] = nmind = nbad = 0;
mindiff = abs(pwgts[0]-pwgts[1]);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if ((higain = PQueueGetMax(&parts)) == -1)
break;
inqueue[higain] = -1;
ASSERT(bndptr[higain] != -1);
if (pwgts[to]+vwgt[higain] > badmaxpwgt) { /* Skip this vertex */
if (nbad++ > limit)
break;
else {
nswaps--;
continue;
}
}
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[from]);
newdiff = abs(pwgts[to]+vwgt[higain] - (pwgts[from]-rinfo[higain].edegrees[from]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
nbad = 0;
}
else {
if (nbad++ > limit) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[from]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
rinfo[k].edegrees[to] += vwgt[higain];
}
else if (where[k] == from) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[from] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
oldgain = vwgt[kk]-rinfo[kk].edegrees[from];
rinfo[kk].edegrees[from] -= vwgt[k];
/* Update the gain of this node if it was skipped */
if (inqueue[kk] == pass)
PQueueUpdateUp(&parts, kk, oldgain, oldgain+vwgt[k]);
}
}
/* Insert the new vertex into the priority queue. Safe due to one-sided moves */
if (hmarker[k] == -1 || hmarker[k] == to) {
PQueueInsert(&parts, k, vwgt[k]-edegrees[from]);
inqueue[k] = pass;
}
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
printf("Moved %6d to %3d, Gain: %5d [%5d] \t[%5d %5d %5d] [%3d %2d]\n",
higain, to, (vwgt[higain]-rinfo[higain].edegrees[from]), vwgt[higain], pwgts[0], pwgts[1], pwgts[2], nswaps, limit));
}
/****************************************************************
* Roll back computation
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
ASSERT(where[higain] == to);
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = from;
INC_DEC(pwgts[from], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[from] += vwgt[k];
}
}
}
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\tMinimum sep: %6d at %5d, PWGTS: [%6d %6d], NBND: %6d, QSIZE: %6d\n",
mincut, mincutorder, pwgts[0], pwgts[1], nbnd, qsize));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut))
break;
}
PQueueFree(ctrl, &parts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs+1);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function balances two partitions by moving boundary nodes
* from the domain that is overweight to the one that is underweight.
**************************************************************************/
void Bnd2WayBalance(CtrlType *ctrl, GraphType *graph, int *tpwgts)
{
int i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, tmp;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts;
idxtype *moved, *perm;
PQueueType parts;
int higain, oldgain, mincut, mindiff;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
pwgts = graph->pwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
/* Determine from which domain you will be moving data */
mindiff = abs(tpwgts[0]-pwgts[0]);
from = (pwgts[0] < tpwgts[0] ? 1 : 0);
to = (from+1)%2;
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions: [%6d %6d] T[%6d %6d], Nv-Nb[%6d %6d]. ICut: %6d [B]\n",
pwgts[0], pwgts[1], tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
tmp = graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)];
PQueueInit(ctrl, &parts, nvtxs, tmp);
idxset(nvtxs, -1, moved);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert the boundary nodes of the proper partition whose size is OK in the priority queue */
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = perm[ii];
ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0);
ASSERT(bndptr[bndind[i]] != -1);
if (where[bndind[i]] == from && vwgt[bndind[i]] <= mindiff)
PQueueInsert(&parts, bndind[i], ed[bndind[i]]-id[bndind[i]]);
}
mincut = graph->mincut;
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if ((higain = PQueueGetMax(&parts)) == -1)
break;
ASSERT(bndptr[higain] != -1);
if (pwgts[to]+vwgt[higain] > tpwgts[to])
break;
mincut -= (ed[higain]-id[higain]);
INC_DEC(pwgts[to], pwgts[from], vwgt[higain]);
where[higain] = to;
moved[higain] = nswaps;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
printf("Moved %6d from %d. [%3d %3d] %5d [%4d %4d]\n", higain, from, ed[higain]-id[higain], vwgt[higain], mincut, pwgts[0], pwgts[1]));
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1 && where[k] == from && vwgt[k] <= mindiff) /* Remove it if in the queues */
PQueueDelete(&parts, k, oldgain);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1 && where[k] == from && vwgt[k] <= mindiff)
PQueueUpdate(&parts, k, oldgain, ed[k]-id[k]);
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1 && where[k] == from && vwgt[k] <= mindiff)
PQueueInsert(&parts, k, ed[k]-id[k]);
}
}
}
}
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\tMinimum cut: %6d, PWGTS: [%6d %6d], NBND: %6d\n", mincut, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
PQueueFree(ctrl, &parts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs a node-based FM refinement
**************************************************************************/
void FM_2WayNodeRefineEqWgt(CtrlType *ctrl, GraphType *graph, idxtype npasses)
{
idxtype i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind;
idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idxtype *mptr, *mind, *moved, *swaps, *perm;
PQueueType parts[2];
NRInfoType *rinfo;
idxtype higain, oldgain, mincut, initcut, mincutorder;
idxtype pass, to, other, limit;
idxtype mindiff, newdiff;
idxtype u[2], g[2];
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
i = ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt);
PQueueInit(ctrl, &parts[0], nvtxs, i);
PQueueInit(ctrl, &parts[1], nvtxs, i);
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
mptr = idxwspacemalloc(ctrl, nvtxs+1);
mind = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("Partitions: [%6D %6D] Nv-Nb[%6D %6D]. ISep: %6D\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
for (pass=0; pass<npasses; pass++) {
idxset(nvtxs, -1, moved);
PQueueReset(&parts[0]);
PQueueReset(&parts[1]);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(where[i] == 2);
PQueueInsert(&parts[0], i, vwgt[i]-rinfo[i].edegrees[1]);
PQueueInsert(&parts[1], i, vwgt[i]-rinfo[i].edegrees[0]);
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = (ctrl->oflags&OFLAG_COMPRESS ? amin(5*nbnd, 400) : amin(2*nbnd, 300));
/******************************************************
* Get into the FM loop
*******************************************************/
mptr[0] = nmind = 0;
mindiff = idxtype_abs(pwgts[0]-pwgts[1]);
to = (pwgts[0] < pwgts[1] ? 0 : 1);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
to = (pwgts[0] < pwgts[1] ? 0 : 1);
if (pwgts[0] == pwgts[1]) {
u[0] = PQueueSeeMax(&parts[0]);
u[1] = PQueueSeeMax(&parts[1]);
if (u[0] != -1 && u[1] != -1) {
g[0] = vwgt[u[0]]-rinfo[u[0]].edegrees[1];
g[1] = vwgt[u[1]]-rinfo[u[1]].edegrees[0];
to = (g[0] > g[1] ? 0 : (g[0] < g[1] ? 1 : pass%2));
}
}
other = (to+1)%2;
if ((higain = PQueueGetMax(&parts[to])) == -1)
break;
if (moved[higain] == -1) /* Delete if it was in the separator originally */
PQueueDelete(&parts[other], higain, vwgt[higain]-rinfo[higain].edegrees[to]);
ASSERT(bndptr[higain] != -1);
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = idxtype_abs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > limit) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
oldgain = vwgt[k]-rinfo[k].edegrees[to];
rinfo[k].edegrees[to] += vwgt[higain];
if (moved[k] == -1 || moved[k] == -(2+other))
PQueueUpdate(&parts[other], k, oldgain, oldgain-vwgt[higain]);
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
oldgain = vwgt[kk]-rinfo[kk].edegrees[other];
rinfo[kk].edegrees[other] -= vwgt[k];
if (moved[kk] == -1 || moved[kk] == -(2+to))
PQueueUpdate(&parts[to], kk, oldgain, oldgain+vwgt[k]);
}
}
/* Insert the new vertex into the priority queue. Only one side! */
if (moved[k] == -1) {
PQueueInsert(&parts[to], k, vwgt[k]-edegrees[other]);
moved[k] = -(2+to);
}
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
mprintf("Moved %6D to %3D, Gain: %5D [%5D] [%4D %4D] \t[%5D %5D %5D]\n", higain, to, g[to], g[other], vwgt[u[to]], vwgt[u[other]], pwgts[0], pwgts[1], pwgts[2]));
}
/****************************************************************
* Roll back computation
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
to = where[higain];
other = (to+1)%2;
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("\tMinimum sep: %6D at %5D, PWGTS: [%6D %6D], NBND: %6D\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut >= initcut)
break;
}
PQueueFree(ctrl, &parts[0]);
PQueueFree(ctrl, &parts[1]);
idxwspacefree(ctrl, nvtxs+1);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Mc_SerialKWayAdaptRefine(GraphType *graph, int nparts, idxtype *home,
float *orgubvec, int npasses)
{
int i, ii, iii, j, k;
int nvtxs, ncon, pass, nmoves, myndegrees;
int from, me, myhome, to, oldcut, gain, tmp;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where;
EdgeType *mydegrees;
RInfoType *rinfo, *myrinfo;
float *npwgts, *nvwgt, *minwgt, *maxwgt, ubvec[MAXNCON];
int gain_is_greater, gain_is_same, fit_in_to, fit_in_from, going_home;
int zero_gain, better_balance_ft, better_balance_tt;
KeyValueType *cand;
int mype;
MPI_Comm_rank(MPI_COMM_WORLD, &mype);
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
rinfo = graph->rinfo;
npwgts = graph->gnpwgts;
/* Setup the weight intervals of the various subdomains */
cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand");
minwgt = fmalloc(nparts*ncon, "minwgt");
maxwgt = fmalloc(nparts*ncon, "maxwgt");
ComputeHKWayLoadImbalance(ncon, nparts, npwgts, ubvec);
for (i=0; i<ncon; i++)
ubvec[i] = amax(ubvec[i], orgubvec[i]);
for (i=0; i<nparts; i++) {
for (j=0; j<ncon; j++) {
maxwgt[i*ncon+j] = ubvec[j]/(float)nparts;
minwgt[i*ncon+j] = ubvec[j]*(float)nparts;
}
}
for (pass=0; pass<npasses; pass++) {
oldcut = graph->mincut;
for (i=0; i<nvtxs; i++) {
cand[i].key = rinfo[i].id-rinfo[i].ed;
cand[i].val = i;
}
ikeysort(nvtxs, cand);
nmoves = 0;
for (iii=0; iii<nvtxs; iii++) {
i = cand[iii].val;
myrinfo = rinfo+i;
if (myrinfo->ed >= myrinfo->id) {
from = where[i];
myhome = home[i];
nvwgt = graph->nvwgt+i*ncon;
if (myrinfo->id > 0 &&
AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, -1.0, nvwgt, minwgt+from*ncon))
continue;
mydegrees = myrinfo->degrees;
myndegrees = myrinfo->ndegrees;
for (k=0; k<myndegrees; k++) {
to = mydegrees[k].edge;
gain = mydegrees[k].ewgt - myrinfo->id;
if (gain >= 0 &&
(AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon) ||
IsHBalanceBetterFT(ncon,npwgts+from*ncon,npwgts+to*ncon,nvwgt,ubvec))) {
break;
}
}
/* break out if you did not find a candidate */
if (k == myndegrees)
continue;
for (j=k+1; j<myndegrees; j++) {
to = mydegrees[j].edge;
going_home = (myhome == to);
gain_is_same = (mydegrees[j].ewgt == mydegrees[k].ewgt);
gain_is_greater = (mydegrees[j].ewgt > mydegrees[k].ewgt);
fit_in_to = AreAllHVwgtsBelow(ncon,1.0,npwgts+to*ncon,1.0,nvwgt,maxwgt+to*ncon);
better_balance_ft = IsHBalanceBetterFT(ncon,npwgts+from*ncon,
npwgts+to*ncon,nvwgt,ubvec);
better_balance_tt = IsHBalanceBetterTT(ncon,npwgts+mydegrees[k].edge*ncon,
npwgts+to*ncon,nvwgt,ubvec);
if (
(gain_is_greater &&
(fit_in_to ||
better_balance_ft)
)
||
(gain_is_same &&
(
(fit_in_to &&
going_home)
||
better_balance_tt
)
)
) {
k = j;
}
}
to = mydegrees[k].edge;
going_home = (myhome == to);
zero_gain = (mydegrees[k].ewgt == myrinfo->id);
fit_in_from = AreAllHVwgtsBelow(ncon,1.0,npwgts+from*ncon,0.0,npwgts+from*ncon,
maxwgt+from*ncon);
better_balance_ft = IsHBalanceBetterFT(ncon,npwgts+from*ncon,
npwgts+to*ncon,nvwgt,ubvec);
if (zero_gain &&
!going_home &&
!better_balance_ft &&
fit_in_from)
continue;
/*=====================================================================
* If we got here, we can now move the vertex from 'from' to 'to'
*======================================================================*/
graph->mincut -= mydegrees[k].ewgt-myrinfo->id;
/* Update where, weight, and ID/ED information of the vertex you moved */
saxpy2(ncon, 1.0, nvwgt, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt, 1, npwgts+from*ncon, 1);
where[i] = to;
myrinfo->ed += myrinfo->id-mydegrees[k].ewgt;
SWAP(myrinfo->id, mydegrees[k].ewgt, tmp);
if (mydegrees[k].ewgt == 0) {
myrinfo->ndegrees--;
mydegrees[k].edge = mydegrees[myrinfo->ndegrees].edge;
mydegrees[k].ewgt = mydegrees[myrinfo->ndegrees].ewgt;
}
else
mydegrees[k].edge = from;
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
me = where[ii];
myrinfo = rinfo+ii;
mydegrees = myrinfo->degrees;
if (me == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
}
else {
if (me == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
}
}
/* Remove contribution of the ed from 'from' */
if (me != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (mydegrees[k].edge == from) {
if (mydegrees[k].ewgt == adjwgt[j]) {
myrinfo->ndegrees--;
mydegrees[k].edge = mydegrees[myrinfo->ndegrees].edge;
mydegrees[k].ewgt = mydegrees[myrinfo->ndegrees].ewgt;
}
else
mydegrees[k].ewgt -= adjwgt[j];
break;
}
}
}
/* Add contribution of the ed to 'to' */
if (me != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (mydegrees[k].edge == to) {
mydegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
mydegrees[myrinfo->ndegrees].edge = to;
mydegrees[myrinfo->ndegrees++].ewgt = adjwgt[j];
}
}
}
nmoves++;
}
}
if (graph->mincut == oldcut)
break;
}
GKfree((void **)&minwgt, (void **)&maxwgt, (void **)&cand, LTERM);
return;
}
/*************************************************************************
* This function performs a node-based FM refinement. This is the
* one-way version
**************************************************************************/
void FM_2WayNodeRefine_OneSided(CtrlType *ctrl, GraphType *graph, float ubfactor, idxtype npasses)
{
idxtype i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind;
idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idxtype *mptr, *mind, *swaps, *perm;
PQueueType parts;
NRInfoType *rinfo;
idxtype higain, oldgain, mincut, initcut, mincutorder;
idxtype pass, to, other, limit;
idxtype badmaxpwgt, mindiff, newdiff;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
PQueueInit(ctrl, &parts, nvtxs, ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt));
perm = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
mptr = idxwspacemalloc(ctrl, nvtxs+1);
mind = idxwspacemalloc(ctrl, nvtxs);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("Partitions-N1: [%6D %6D] Nv-Nb[%6D %6D]. ISep: %6D\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
badmaxpwgt = (int)(ubfactor*(pwgts[0]+pwgts[1]+pwgts[2])/2);
to = (pwgts[0] < pwgts[1] ? 1 : 0);
for (pass=0; pass<npasses; pass++) {
other = to;
to = (to+1)%2;
PQueueReset(&parts);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(where[i] == 2);
PQueueInsert(&parts, i, vwgt[i]-rinfo[i].edegrees[other]);
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = (ctrl->oflags&OFLAG_COMPRESS ? amin(5*nbnd, 400) : amin(2*nbnd, 300));
/******************************************************
* Get into the FM loop
*******************************************************/
mptr[0] = nmind = 0;
mindiff = idxtype_abs(pwgts[0]-pwgts[1]);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if ((higain = PQueueGetMax(&parts)) == -1)
break;
ASSERT(bndptr[higain] != -1);
if (pwgts[to]+vwgt[higain] > badmaxpwgt)
break; /* No point going any further. Balance will be bad */
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = idxtype_abs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > limit) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
rinfo[k].edegrees[to] += vwgt[higain];
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
oldgain = vwgt[kk]-rinfo[kk].edegrees[other];
rinfo[kk].edegrees[other] -= vwgt[k];
/* Since the moves are one-sided this vertex has not been moved yet */
PQueueUpdateUp(&parts, kk, oldgain, oldgain+vwgt[k]);
}
}
/* Insert the new vertex into the priority queue. Safe due to one-sided moves */
PQueueInsert(&parts, k, vwgt[k]-edegrees[other]);
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
mprintf("Moved %6D to %3D, Gain: %5D [%5D] \t[%5D %5D %5D] [%3D %2D]\n",
higain, to, (vwgt[higain]-rinfo[higain].edegrees[other]), vwgt[higain], pwgts[0], pwgts[1], pwgts[2], nswaps, limit));
}
/****************************************************************
* Roll back computation
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
ASSERT(where[higain] == to);
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("\tMinimum sep: %6D at %5D, PWGTS: [%6D %6D], NBND: %6D\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut))
break;
}
PQueueFree(ctrl, &parts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs+1);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void Mc_Serial_FM_2WayRefine(GraphType *graph, float *tpwgts, int npasses)
{
int i, ii, j, k;
int kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
FPQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, initcut, newcut, mincutorder;
float rtpwgts[MAXNCON*2];
KeyValueType *cand;
int mype;
MPI_Comm_rank(MPI_COMM_WORLD, &mype);
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->sendind;
ed = graph->recvind;
npwgts = graph->gnpwgts;
bndptr = graph->sendptr;
bndind = graph->recvptr;
moved = idxmalloc(nvtxs, "moved");
swaps = idxmalloc(nvtxs, "swaps");
qnum = idxmalloc(nvtxs, "qnum");
cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand");
limit = amin(amax(0.01*nvtxs, 25), 150);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
FPQueueInit(&parts[i][0], nvtxs);
FPQueueInit(&parts[i][1], nvtxs);
}
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
origbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
for (i=0; i<ncon; i++) {
rtpwgts[i] = origbal*tpwgts[i];
rtpwgts[ncon+i] = origbal*tpwgts[ncon+i];
}
idxset(nvtxs, -1, moved);
for (pass=0; pass<npasses; pass++) { /* Do a number of passes */
for (i=0; i<ncon; i++) {
FPQueueReset(&parts[i][0]);
FPQueueReset(&parts[i][1]);
}
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
minbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
/* Insert boundary nodes in the priority queues */
nbnd = graph->gnvtxs;
for (i=0; i<nbnd; i++) {
cand[i].key = id[i]-ed[i];
cand[i].val = i;
}
ikeysort(nbnd, cand);
for (ii=0; ii<nbnd; ii++) {
i = bndind[cand[ii].val];
FPQueueInsert(&parts[qnum[i]][where[i]], i, (float)(ed[i]-id[i]));
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
Serial_SelectQueue(ncon, npwgts, rtpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = FPQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if ((newcut < mincut && newbal-origbal <= .00001) ||
(newcut == mincut && (newbal < minbal ||
(newbal == minbal && Serial_BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
FPQueueDelete(&parts[qnum[k]][where[k]], k);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
FPQueueUpdate(&parts[qnum[k]][where[k]], k, (float)oldgain, (float)(ed[k]-id[k]));
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
FPQueueInsert(&parts[qnum[k]][where[k]], k, (float)(ed[k]-id[k]));
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->gnvtxs = nbnd;
if (mincutorder == -1 || mincut == initcut)
break;
}
for (i=0; i<ncon; i++) {
FPQueueFree(&parts[i][0]);
FPQueueFree(&parts[i][1]);
}
GKfree((void **)&cand, (void **)&qnum, (void **)&moved, (void **)&swaps, LTERM);
return;
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void MocGeneral2WayBalance2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *perm, *qnum;
float *nvwgt, *npwgts, origbal[MAXNCON], minbal[MAXNCON], newbal[MAXNCON];
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, newcut, mincutorder;
float *maxwgt, *minwgt, tvec[MAXNCON];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
npwgts = graph->npwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
qnum = idxwspacemalloc(ctrl, nvtxs);
limit = amin(amax(0.01*nvtxs, 15), 100);
/* Setup the weight intervals of the two subdomains */
minwgt = fwspacemalloc(ctrl, 2*ncon);
maxwgt = fwspacemalloc(ctrl, 2*ncon);
for (i=0; i<2; i++) {
for (j=0; j<ncon; j++) {
maxwgt[i*ncon+j] = tpwgts[i]*ubvec[j];
minwgt[i*ncon+j] = tpwgts[i]*(1.0/ubvec[j]);
}
}
/* Initialize the queues */
for (i=0; i<ncon; i++) {
PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1);
PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1);
}
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, origbal);
for (i=0; i<ncon; i++)
minbal[i] = origbal[i];
newcut = mincut = graph->mincut;
mincutorder = -1;
if (ctrl->dbglvl&DBG_REFINE) {
printf("Parts: [");
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: ", tpwgts[0], tpwgts[1],
graph->nvtxs, graph->nbnd, graph->mincut);
for (i=0; i<ncon; i++)
printf("%.3f ", origbal[i]);
printf("[B]\n");
}
idxset(nvtxs, -1, moved);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert all nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nvtxs, perm, 1);
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (AreAllBelow(ncon, minbal, ubvec))
break;
SelectQueue3(ncon, npwgts, tpwgts, &from, &cnum, parts, maxwgt);
to = (from+1)%2;
if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, newbal);
if (IsBetter2wayBalance(ncon, newbal, minbal, ubvec) ||
(IsBetter2wayBalance(ncon, newbal, origbal, ubvec) && newcut < mincut)) {
mincut = newcut;
for (i=0; i<ncon; i++)
minbal[i] = newbal[i];
mincutorder = nswaps;
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
if (ctrl->dbglvl&DBG_MOVEINFO) {
printf("Moved %6d from %d(%d). Gain: %5d, Cut: %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut);
for (i=0; i<ncon; i++)
printf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
printf(", LB: ");
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
if (mincutorder == nswaps)
printf(" *\n");
else
printf("\n");
}
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (moved[k] == -1)
PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]);
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
if (ctrl->dbglvl&DBG_REFINE) {
printf("\tMincut: %6d at %5d, NBND: %6d, NPwgts: [", mincut, mincutorder, nbnd);
for (i=0; i<ncon; i++)
printf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]);
printf("], LB: ");
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
printf("\n");
}
graph->mincut = mincut;
graph->nbnd = nbnd;
for (i=0; i<ncon; i++) {
PQueueFree(ctrl, &parts[i][0]);
PQueueFree(ctrl, &parts[i][1]);
}
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
fwspacefree(ctrl, 2*ncon);
fwspacefree(ctrl, 2*ncon);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void Mc_Serial_Balance2Way(GraphType *graph, float *tpwgts, float lbfactor)
{
int i, ii, j, k, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
FPQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, newcut, mincutorder;
int qsizes[MAXNCON][2];
KeyValueType *cand;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->sendind;
ed = graph->recvind;
npwgts = graph->gnpwgts;
bndptr = graph->sendptr;
bndind = graph->recvptr;
moved = idxmalloc(nvtxs, "moved");
swaps = idxmalloc(nvtxs, "swaps");
qnum = idxmalloc(nvtxs, "qnum");
cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand");
limit = amin(amax(0.01*nvtxs, 15), 100);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
FPQueueInit(&parts[i][0], nvtxs);
FPQueueInit(&parts[i][1], nvtxs);
qsizes[i][0] = qsizes[i][1] = 0;
}
for (i=0; i<nvtxs; i++) {
qnum[i] = samax(ncon, nvwgt+i*ncon);
qsizes[qnum[i]][where[i]]++;
}
for (from=0; from<2; from++) {
for (j=0; j<ncon; j++) {
if (qsizes[j][from] == 0) {
for (i=0; i<nvtxs; i++) {
if (where[i] != from)
continue;
k = samax2(ncon, nvwgt+i*ncon);
if (k == j &&
qsizes[qnum[i]][from] > qsizes[j][from] &&
nvwgt[i*ncon+qnum[i]] < 1.3*nvwgt[i*ncon+j]) {
qsizes[qnum[i]][from]--;
qsizes[j][from]++;
qnum[i] = j;
}
}
}
}
}
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
minbal = origbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
newcut = mincut = graph->mincut;
mincutorder = -1;
idxset(nvtxs, -1, moved);
/* Insert all nodes in the priority queues */
nbnd = graph->gnvtxs;
for (i=0; i<nvtxs; i++) {
cand[i].key = id[i]-ed[i];
cand[i].val = i;
}
ikeysort(nvtxs, cand);
for (ii=0; ii<nvtxs; ii++) {
i = cand[ii].val;
FPQueueInsert(&parts[qnum[i]][where[i]], i, (float)(ed[i]-id[i]));
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (minbal < lbfactor)
break;
Serial_SelectQueue(ncon, npwgts, tpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = FPQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if (newbal < minbal || (newbal == minbal &&
(newcut < mincut || (newcut == mincut &&
Serial_BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (moved[k] == -1)
FPQueueUpdate(&parts[qnum[k]][where[k]], k, (float)(oldgain), (float)(ed[k]-id[k]));
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->gnvtxs = nbnd;
for (i=0; i<ncon; i++) {
FPQueueFree(&parts[i][0]);
FPQueueFree(&parts[i][1]);
}
GKfree((void **)&cand, (void **)&qnum, (void **)&moved, (void **)&swaps, LTERM);
return;
}
void FM_2WayCutRefine(ctrl_t *ctrl, graph_t *graph, real_t *ntpwgts, idx_t niter)
{
idx_t i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, limit, tmp;
idx_t *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts;
idx_t *moved, *swaps, *perm;
rpq_t *queues[2];
idx_t higain, mincut, mindiff, origdiff, initcut, newcut, mincutorder, avgvwgt;
idx_t tpwgts[2];
WCOREPUSH;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
pwgts = graph->pwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = iwspacemalloc(ctrl, nvtxs);
swaps = iwspacemalloc(ctrl, nvtxs);
perm = iwspacemalloc(ctrl, nvtxs);
tpwgts[0] = graph->tvwgt[0]*ntpwgts[0];
tpwgts[1] = graph->tvwgt[0]-tpwgts[0];
limit = gk_min(gk_max(0.01*nvtxs, 15), 100);
avgvwgt = gk_min((pwgts[0]+pwgts[1])/20, 2*(pwgts[0]+pwgts[1])/nvtxs);
queues[0] = rpqCreate(nvtxs);
queues[1] = rpqCreate(nvtxs);
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, 0, -2));
origdiff = iabs(tpwgts[0]-pwgts[0]);
iset(nvtxs, -1, moved);
for (pass=0; pass<niter; pass++) { /* Do a number of passes */
rpqReset(queues[0]);
rpqReset(queues[1]);
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
mindiff = iabs(tpwgts[0]-pwgts[0]);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
irandArrayPermute(nbnd, perm, nbnd, 1);
for (ii=0; ii<nbnd; ii++) {
i = perm[ii];
ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0);
ASSERT(bndptr[bndind[i]] != -1);
rpqInsert(queues[where[bndind[i]]], bndind[i], ed[bndind[i]]-id[bndind[i]]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
from = (tpwgts[0]-pwgts[0] < tpwgts[1]-pwgts[1] ? 0 : 1);
to = (from+1)%2;
if ((higain = rpqGetTop(queues[from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
newcut -= (ed[higain]-id[higain]);
INC_DEC(pwgts[to], pwgts[from], vwgt[higain]);
if ((newcut < mincut && iabs(tpwgts[0]-pwgts[0]) <= origdiff+avgvwgt) ||
(newcut == mincut && iabs(tpwgts[0]-pwgts[0]) < mindiff)) {
mincut = newcut;
mindiff = iabs(tpwgts[0]-pwgts[0]);
mincutorder = nswaps;
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
INC_DEC(pwgts[from], pwgts[to], vwgt[higain]);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO,
printf("Moved %6"PRIDX" from %"PRIDX". [%3"PRIDX" %3"PRIDX"] %5"PRIDX" [%4"PRIDX" %4"PRIDX"]\n", higain, from, ed[higain]-id[higain], vwgt[higain], newcut, pwgts[0], pwgts[1]));
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
rpqDelete(queues[where[k]], k);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
rpqUpdate(queues[where[k]], k, ed[k]-id[k]);
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
rpqInsert(queues[where[k]], k, ed[k]-id[k]);
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
INC_DEC(pwgts[to], pwgts[(to+1)%2], vwgt[higain]);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->nbnd = nbnd;
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, 0, mincutorder));
if (mincutorder <= 0 || mincut == initcut)
break;
}
rpqDestroy(queues[0]);
rpqDestroy(queues[1]);
WCOREPOP;
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void MocGeneral2WayBalance(CtrlType *ctrl, GraphType *graph, float *tpwgts, float lbfactor)
{
int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, me, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *perm, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, newcut, mincutorder;
int qsizes[MAXNCON][2];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
npwgts = graph->npwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
qnum = idxwspacemalloc(ctrl, nvtxs);
limit = amin(amax(0.01*nvtxs, 15), 100);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1);
PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1);
qsizes[i][0] = qsizes[i][1] = 0;
}
for (i=0; i<nvtxs; i++) {
qnum[i] = samax(ncon, nvwgt+i*ncon);
qsizes[qnum[i]][where[i]]++;
}
/*
printf("Weight Distribution: \t");
for (i=0; i<ncon; i++)
printf(" [%d %d]", qsizes[i][0], qsizes[i][1]);
printf("\n");
*/
for (from=0; from<2; from++) {
for (j=0; j<ncon; j++) {
if (qsizes[j][from] == 0) {
for (i=0; i<nvtxs; i++) {
if (where[i] != from)
continue;
k = samax2(ncon, nvwgt+i*ncon);
if (k == j && qsizes[qnum[i]][from] > qsizes[j][from] && nvwgt[i*ncon+qnum[i]] < 1.3*nvwgt[i*ncon+j]) {
qsizes[qnum[i]][from]--;
qsizes[j][from]++;
qnum[i] = j;
}
}
}
}
}
/*
printf("Weight Distribution (after):\t ");
for (i=0; i<ncon; i++)
printf(" [%d %d]", qsizes[i][0], qsizes[i][1]);
printf("\n");
*/
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
minbal = origbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
newcut = mincut = graph->mincut;
mincutorder = -1;
if (ctrl->dbglvl&DBG_REFINE) {
printf("Parts: [");
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: %.3f [B]\n", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut, origbal);
}
idxset(nvtxs, -1, moved);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert all nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nvtxs, perm, 1);
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (minbal < lbfactor)
break;
SelectQueue(ncon, npwgts, tpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if (newbal < minbal || (newbal == minbal &&
(newcut < mincut || (newcut == mincut && BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
if (ctrl->dbglvl&DBG_MOVEINFO) {
printf("Moved %6d from %d(%d). Gain: %5d, Cut: %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf(", %.3f LB: %.3f\n", minbal, newbal);
}
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (moved[k] == -1)
PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]);
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
if (ctrl->dbglvl&DBG_REFINE) {
printf("\tMincut: %6d at %5d, NBND: %6d, NPwgts: [", mincut, mincutorder, nbnd);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("], LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts));
}
graph->mincut = mincut;
graph->nbnd = nbnd;
for (i=0; i<ncon; i++) {
PQueueFree(ctrl, &parts[i][0]);
PQueueFree(ctrl, &parts[i][1]);
}
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void MCGreedy_KWayEdgeBalanceHorizontal(CtrlType *ctrl, GraphType *graph, int nparts,
float *ubvec, int npasses)
{
int i, ii, /*iii,*/ j, /*jj,*/ k, /*l,*/ pass, nvtxs, ncon, nbnd, myndegrees, oldgain, gain, nmoves;
int from, me, to, oldcut;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where, *perm, *bndptr, *bndind, *moved;
EDegreeType *myedegrees;
RInfoType *myrinfo;
PQueueType queue;
float *npwgts, *nvwgt, *minwgt, *maxwgt, tvec[MAXNCON];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
npwgts = graph->npwgts;
/* Setup the weight intervals of the various subdomains */
minwgt = fwspacemalloc(ctrl, ncon*nparts);
maxwgt = fwspacemalloc(ctrl, ncon*nparts);
for (i=0; i<nparts; i++) {
for (j=0; j<ncon; j++) {
maxwgt[i*ncon+j] = ubvec[j]/nparts;
minwgt[i*ncon+j] = 1.0/(ubvec[j]*nparts);
}
}
perm = idxwspacemalloc(ctrl, nvtxs);
moved = idxwspacemalloc(ctrl, nvtxs);
PQueueInit(ctrl, &queue, nvtxs, graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]);
if (ctrl->dbglvl&DBG_REFINE) {
printf("Partitions: [%5.4f %5.4f], Nv-Nb[%6d %6d]. Cut: %6d, LB: ",
npwgts[samin(ncon*nparts, npwgts)], npwgts[samax(ncon*nparts, npwgts)],
graph->nvtxs, graph->nbnd, graph->mincut);
ComputeHKWayLoadImbalance(ncon, nparts, npwgts, tvec);
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
printf("[B]\n");
}
for (pass=0; pass<npasses; pass++) {
ASSERT(ComputeCut(graph, where) == graph->mincut);
/* Check to see if things are out of balance, given the tolerance */
if (MocIsHBalanced(ncon, nparts, npwgts, ubvec))
break;
PQueueReset(&queue);
idxset(nvtxs, -1, moved);
oldcut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
PQueueInsert(&queue, i, graph->rinfo[i].ed - graph->rinfo[i].id);
moved[i] = 2;
}
nmoves = 0;
for (;;) {
if ((i = PQueueGetMax(&queue)) == -1)
break;
moved[i] = 1;
myrinfo = graph->rinfo+i;
from = where[i];
nvwgt = graph->nvwgt+i*ncon;
if (AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, -1.0, nvwgt, minwgt+from*ncon))
continue; /* This cannot be moved! */
myedegrees = myrinfo->edegrees;
myndegrees = myrinfo->ndegrees;
for (k=0; k<myndegrees; k++) {
to = myedegrees[k].pid;
if (IsHBalanceBetterFT(ncon, nparts, npwgts+from*ncon, npwgts+to*ncon, nvwgt, ubvec))
break;
}
if (k == myndegrees)
continue; /* break out if you did not find a candidate */
for (j=k+1; j<myndegrees; j++) {
to = myedegrees[j].pid;
if (IsHBalanceBetterTT(ncon, nparts, npwgts+myedegrees[k].pid*ncon, npwgts+to*ncon, nvwgt, ubvec))
k = j;
}
to = myedegrees[k].pid;
j = 0;
if (!AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, maxwgt+from*ncon))
j++;
if (myedegrees[k].ed-myrinfo->id >= 0)
j++;
if (!AreAllHVwgtsAbove(ncon, 1.0, npwgts+to*ncon, 0.0, nvwgt, minwgt+to*ncon) &&
AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon))
j++;
if (j == 0)
continue;
/* DELETE
if (myedegrees[k].ed-myrinfo->id < 0 &&
AreAllHVwgtsBelow(ncon, 1.0, npwgts+from*ncon, 0.0, nvwgt, maxwgt+from*ncon) &&
AreAllHVwgtsAbove(ncon, 1.0, npwgts+to*ncon, 0.0, nvwgt, minwgt+to*ncon) &&
AreAllHVwgtsBelow(ncon, 1.0, npwgts+to*ncon, 1.0, nvwgt, maxwgt+to*ncon))
continue;
*/
/*=====================================================================
* If we got here, we can now move the vertex from 'from' to 'to'
*======================================================================*/
graph->mincut -= myedegrees[k].ed-myrinfo->id;
IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut));
/* Update where, weight, and ID/ED information of the vertex you moved */
saxpy(ncon, 1.0, nvwgt, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt, 1, npwgts+from*ncon, 1);
where[i] = to;
myrinfo->ed += myrinfo->id-myedegrees[k].ed;
SWAP(myrinfo->id, myedegrees[k].ed, j);
if (myedegrees[k].ed == 0)
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].pid = from;
if (myrinfo->ed == 0)
BNDDelete(nbnd, bndind, bndptr, i);
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
me = where[ii];
myrinfo = graph->rinfo+ii;
if (myrinfo->edegrees == NULL) {
myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii];
}
myedegrees = myrinfo->edegrees;
ASSERT(CheckRInfo(myrinfo));
oldgain = (myrinfo->ed-myrinfo->id);
if (me == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
if (myrinfo->ed > 0 && bndptr[ii] == -1)
BNDInsert(nbnd, bndind, bndptr, ii);
}
else if (me == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
if (myrinfo->ed == 0 && bndptr[ii] != -1)
BNDDelete(nbnd, bndind, bndptr, ii);
}
/* Remove contribution from the .ed of 'from' */
if (me != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == from) {
if (myedegrees[k].ed == adjwgt[j])
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].ed -= adjwgt[j];
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (me != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == to) {
myedegrees[k].ed += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
myedegrees[myrinfo->ndegrees].pid = to;
myedegrees[myrinfo->ndegrees++].ed = adjwgt[j];
}
}
/* Update the queue */
if (me == to || me == from) {
gain = myrinfo->ed-myrinfo->id;
if (moved[ii] == 2) {
if (myrinfo->ed > 0)
PQueueUpdate(&queue, ii, oldgain, gain);
else {
PQueueDelete(&queue, ii, oldgain);
moved[ii] = -1;
}
}
else if (moved[ii] == -1 && myrinfo->ed > 0) {
PQueueInsert(&queue, ii, gain);
moved[ii] = 2;
}
}
ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]);
ASSERT(CheckRInfo(myrinfo));
}
nmoves++;
}
graph->nbnd = nbnd;
if (ctrl->dbglvl&DBG_REFINE) {
printf("\t [%5.4f %5.4f], Nb: %6d, Nmoves: %5d, Cut: %6d, LB: ",
npwgts[samin(ncon*nparts, npwgts)], npwgts[samax(ncon*nparts, npwgts)],
nbnd, nmoves, graph->mincut);
ComputeHKWayLoadImbalance(ncon, nparts, npwgts, tvec);
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
printf("\n");
}
if (nmoves == 0)
break;
}
PQueueFree(ctrl, &queue);
fwspacefree(ctrl, ncon*nparts);
fwspacefree(ctrl, ncon*nparts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Greedy_KWayEdgeRefine(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor, int npasses)
{
int i, ii, iii, j, jj, k, l, pass, nvtxs, nbnd, tvwgt, myndegrees, oldgain, gain;
int from, me, to, oldcut, vwgt;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where, *pwgts, *perm, *bndptr, *bndind, *minwgt, *maxwgt, *moved, *itpwgts;
EDegreeType *myedegrees;
RInfoType *myrinfo;
PQueueType queue;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
/* Setup the weight intervals of the various subdomains */
minwgt = idxwspacemalloc(ctrl, nparts);
maxwgt = idxwspacemalloc(ctrl, nparts);
itpwgts = idxwspacemalloc(ctrl, nparts);
tvwgt = idxsum(nparts, pwgts);
ASSERT(tvwgt == idxsum(nvtxs, graph->vwgt));
for (i=0; i<nparts; i++) {
itpwgts[i] = tpwgts[i]*tvwgt;
maxwgt[i] = tpwgts[i]*tvwgt*ubfactor;
minwgt[i] = tpwgts[i]*tvwgt*(1.0/ubfactor);
}
perm = idxwspacemalloc(ctrl, nvtxs);
moved = idxwspacemalloc(ctrl, nvtxs);
PQueueInit(ctrl, &queue, nvtxs, graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions: [%6d %6d]-[%6d %6d], Balance: %5.3f, Nv-Nb[%6d %6d]. Cut: %6d\n",
pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], minwgt[0], maxwgt[0],
1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nvtxs, graph->nbnd,
graph->mincut));
for (pass=0; pass<npasses; pass++) {
ASSERT(ComputeCut(graph, where) == graph->mincut);
PQueueReset(&queue);
idxset(nvtxs, -1, moved);
oldcut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
PQueueInsert(&queue, i, graph->rinfo[i].ed - graph->rinfo[i].id);
moved[i] = 2;
}
for (iii=0;; iii++) {
if ((i = PQueueGetMax(&queue)) == -1)
break;
moved[i] = 1;
myrinfo = graph->rinfo+i;
from = where[i];
vwgt = graph->vwgt[i];
if (pwgts[from]-vwgt < minwgt[from])
continue; /* This cannot be moved! */
myedegrees = myrinfo->edegrees;
myndegrees = myrinfo->ndegrees;
j = myrinfo->id;
for (k=0; k<myndegrees; k++) {
to = myedegrees[k].pid;
gain = myedegrees[k].ed-j; /* j = myrinfo->id. Allow good nodes to move */
if (pwgts[to]+vwgt <= maxwgt[to]+gain && gain >= 0)
break;
}
if (k == myndegrees)
continue; /* break out if you did not find a candidate */
for (j=k+1; j<myndegrees; j++) {
to = myedegrees[j].pid;
if ((myedegrees[j].ed > myedegrees[k].ed && pwgts[to]+vwgt <= maxwgt[to]) ||
(myedegrees[j].ed == myedegrees[k].ed &&
itpwgts[myedegrees[k].pid]*pwgts[to] < itpwgts[to]*pwgts[myedegrees[k].pid]))
k = j;
}
to = myedegrees[k].pid;
j = 0;
if (myedegrees[k].ed-myrinfo->id > 0)
j = 1;
else if (myedegrees[k].ed-myrinfo->id == 0) {
if ((iii&7) == 0 || pwgts[from] >= maxwgt[from] || itpwgts[from]*(pwgts[to]+vwgt) < itpwgts[to]*pwgts[from])
j = 1;
}
if (j == 0)
continue;
/*=====================================================================
* If we got here, we can now move the vertex from 'from' to 'to'
*======================================================================*/
graph->mincut -= myedegrees[k].ed-myrinfo->id;
IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut));
/* Update where, weight, and ID/ED information of the vertex you moved */
where[i] = to;
INC_DEC(pwgts[to], pwgts[from], vwgt);
myrinfo->ed += myrinfo->id-myedegrees[k].ed;
SWAP(myrinfo->id, myedegrees[k].ed, j);
if (myedegrees[k].ed == 0)
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].pid = from;
if (myrinfo->ed < myrinfo->id)
BNDDelete(nbnd, bndind, bndptr, i);
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
me = where[ii];
myrinfo = graph->rinfo+ii;
if (myrinfo->edegrees == NULL) {
myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii];
}
myedegrees = myrinfo->edegrees;
ASSERT(CheckRInfo(myrinfo));
oldgain = (myrinfo->ed-myrinfo->id);
if (me == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
if (myrinfo->ed-myrinfo->id >= 0 && bndptr[ii] == -1)
BNDInsert(nbnd, bndind, bndptr, ii);
}
else if (me == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
if (myrinfo->ed-myrinfo->id < 0 && bndptr[ii] != -1)
BNDDelete(nbnd, bndind, bndptr, ii);
}
/* Remove contribution from the .ed of 'from' */
if (me != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == from) {
if (myedegrees[k].ed == adjwgt[j])
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].ed -= adjwgt[j];
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (me != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == to) {
myedegrees[k].ed += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
myedegrees[myrinfo->ndegrees].pid = to;
myedegrees[myrinfo->ndegrees++].ed = adjwgt[j];
}
}
/* Update the queue */
if (me == to || me == from) {
gain = myrinfo->ed-myrinfo->id;
if (moved[ii] == 2) {
if (gain >= 0)
PQueueUpdate(&queue, ii, oldgain, gain);
else {
PQueueDelete(&queue, ii, oldgain);
moved[ii] = -1;
}
}
else if (moved[ii] == -1 && gain >= 0) {
PQueueInsert(&queue, ii, gain);
moved[ii] = 2;
}
}
ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]);
ASSERT(CheckRInfo(myrinfo));
}
}
graph->nbnd = nbnd;
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\t[%6d %6d], Balance: %5.3f, Nb: %6d. Cut: %6d\n",
pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)],
1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nbnd, graph->mincut));
if (graph->mincut == oldcut)
break;
}
PQueueFree(ctrl, &queue);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function finds all the connected components induced by the
* partitioning vector in wgraph->where and tries to push them around to
* remove some of them
**************************************************************************/
void EliminateComponents(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor)
{
int i, ii, j, jj, k, me, nvtxs, tvwgt, first, last, nleft, ncmps, cwgt, other, target, deltawgt;
idxtype *xadj, *adjncy, *vwgt, *adjwgt, *where, *pwgts, *maxpwgt;
idxtype *cpvec, *touched, *perm, *todo, *cind, *cptr, *npcmps;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
adjwgt = graph->adjwgt;
where = graph->where;
pwgts = graph->pwgts;
touched = idxset(nvtxs, 0, idxwspacemalloc(ctrl, nvtxs));
cptr = idxwspacemalloc(ctrl, nvtxs);
cind = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
todo = idxwspacemalloc(ctrl, nvtxs);
maxpwgt = idxwspacemalloc(ctrl, nparts);
cpvec = idxwspacemalloc(ctrl, nparts);
npcmps = idxset(nparts, 0, idxwspacemalloc(ctrl, nparts));
for (i=0; i<nvtxs; i++)
perm[i] = todo[i] = i;
/* Find the connected componends induced by the partition */
ncmps = -1;
first = last = 0;
nleft = nvtxs;
while (nleft > 0) {
if (first == last) { /* Find another starting vertex */
cptr[++ncmps] = first;
ASSERT(touched[todo[0]] == 0);
i = todo[0];
cind[last++] = i;
touched[i] = 1;
me = where[i];
npcmps[me]++;
}
i = cind[first++];
k = perm[i];
j = todo[k] = todo[--nleft];
perm[j] = k;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (where[k] == me && !touched[k]) {
cind[last++] = k;
touched[k] = 1;
}
}
}
cptr[++ncmps] = first;
/* printf("I found %d components, for this %d-way partition\n", ncmps, nparts); */
if (ncmps > nparts) { /* There are more components than processors */
/* First determine the max allowed load imbalance */
tvwgt = idxsum(nparts, pwgts);
for (i=0; i<nparts; i++)
maxpwgt[i] = ubfactor*tpwgts[i]*tvwgt;
deltawgt = 5;
for (i=0; i<ncmps; i++) {
me = where[cind[cptr[i]]]; /* Get the domain of this component */
if (npcmps[me] == 1)
continue; /* Skip it because it is contigous */
/*printf("Trying to move %d from %d\n", i, me); */
/* Determine the weight of the block to be moved and abort if too high */
for (cwgt=0, j=cptr[i]; j<cptr[i+1]; j++)
cwgt += vwgt[cind[j]];
if (cwgt > .30*pwgts[me])
continue; /* Skip the component if it is over 30% of the weight */
/* Determine the connectivity */
idxset(nparts, 0, cpvec);
for (j=cptr[i]; j<cptr[i+1]; j++) {
ii = cind[j];
for (jj=xadj[ii]; jj<xadj[ii+1]; jj++)
cpvec[where[adjncy[jj]]] += adjwgt[jj];
}
cpvec[me] = 0;
target = -1;
for (j=0; j<nparts; j++) {
if (cpvec[j] > 0 && (cwgt < deltawgt || pwgts[j] + cwgt < maxpwgt[j])) {
if (target == -1 || cpvec[target] < cpvec[j])
target = j;
}
}
/* printf("\tMoving it to %d [%d]\n", target, cpvec[target]);*/
if (target != -1) {
/* Assign all the vertices of 'me' to 'target' and update data structures */
INC_DEC(pwgts[target], pwgts[me], cwgt);
npcmps[me]--;
MoveGroup(ctrl, graph, nparts, target, i, cptr, cind);
}
}
}
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void GrowBisectionNode(CtrlType *ctrl, GraphType *graph, float ubfactor)
{
int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], tpwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *bndind;
idxtype *queue, *touched, *gain, *bestwhere;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
queue = idxmalloc(nvtxs, "BisectGraph: queue");
touched = idxmalloc(nvtxs, "BisectGraph: touched");
tpwgts[0] = idxsum(nvtxs, vwgt);
tpwgts[1] = tpwgts[0]/2;
tpwgts[0] -= tpwgts[1];
maxpwgt[0] = ubfactor*tpwgts[0];
maxpwgt[1] = ubfactor*tpwgts[1];
minpwgt[0] = (1.0/ubfactor)*tpwgts[0];
minpwgt[1] = (1.0/ubfactor)*tpwgts[1];
/* Allocate memory for graph->rdata. Allocate sufficient memory for both edge and node */
graph->rdata = idxmalloc(5*nvtxs+3, "GrowBisectionNode: graph->rdata");
graph->pwgts = graph->rdata;
graph->where = graph->rdata + 3;
graph->bndptr = graph->rdata + nvtxs + 3;
graph->bndind = graph->rdata + 2*nvtxs + 3;
graph->nrinfo = (NRInfoType *)(graph->rdata + 3*nvtxs + 3);
graph->id = graph->rdata + 3*nvtxs + 3;
graph->ed = graph->rdata + 4*nvtxs + 3;
where = graph->where;
bndind = graph->bndind;
nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = tpwgts[0]+tpwgts[1];
for (nbfs++; nbfs>0; nbfs--) {
idxset(nvtxs, 0, touched);
pwgts[1] = tpwgts[0]+tpwgts[1];
pwgts[0] = 0;
idxset(nvtxs, 1, where);
queue[0] = RandomInRange(nvtxs);
touched[queue[0]] = 1;
first = 0; last = 1;
nleft = nvtxs-1;
drain = 0;
/* Start the BFS from queue to get a partition */
if (nbfs >= 1) {
for (;;) {
if (first == last) { /* Empty. Disconnected graph! */
if (nleft == 0 || drain)
break;
k = RandomInRange(nleft);
for (i=0; i<nvtxs; i++) {
if (touched[i] == 0) {
if (k == 0)
break;
else
k--;
}
}
queue[0] = i;
touched[i] = 1;
first = 0; last = 1;;
nleft--;
}
i = queue[first++];
if (pwgts[1]-vwgt[i] < minpwgt[1]) {
drain = 1;
continue;
}
where[i] = 0;
INC_DEC(pwgts[0], pwgts[1], vwgt[i]);
if (pwgts[1] <= maxpwgt[1])
break;
drain = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (touched[k] == 0) {
queue[last++] = k;
touched[k] = 1;
nleft--;
}
}
}
}
/*************************************************************
* Do some partition refinement
**************************************************************/
Compute2WayPartitionParams(ctrl, graph);
Balance2Way(ctrl, graph, tpwgts, ubfactor);
FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
/* Construct and refine the vertex separator */
for (i=0; i<graph->nbnd; i++)
where[bndind[i]] = 2;
Compute2WayNodePartitionParams(ctrl, graph);
FM_2WayNodeRefine(ctrl, graph, ubfactor, 6);
/* printf("ISep: [%d %d %d] %d\n", graph->pwgts[0], graph->pwgts[1], graph->pwgts[2], bestcut); */
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
Compute2WayNodePartitionParams(ctrl, graph);
GKfree(&bestwhere, &queue, &touched, LTERM);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Greedy_KWayEdgeBalanceMConn(CtrlType *ctrl, GraphType *graph, int nparts, float *tpwgts, float ubfactor, int npasses)
{
int i, ii, iii, j, jj, k, l, pass, nvtxs, nbnd, tvwgt, myndegrees, oldgain, gain, nmoves;
int from, me, to, oldcut, vwgt, maxndoms, nadd;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where, *pwgts, *perm, *bndptr, *bndind, *minwgt, *maxwgt, *moved, *itpwgts;
idxtype *phtable, *pmat, *pmatptr, *ndoms;
EDegreeType *myedegrees;
RInfoType *myrinfo;
PQueueType queue;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
pmat = ctrl->wspace.pmat;
phtable = idxwspacemalloc(ctrl, nparts);
ndoms = idxwspacemalloc(ctrl, nparts);
ComputeSubDomainGraph(graph, nparts, pmat, ndoms);
/* Setup the weight intervals of the various subdomains */
minwgt = idxwspacemalloc(ctrl, nparts);
maxwgt = idxwspacemalloc(ctrl, nparts);
itpwgts = idxwspacemalloc(ctrl, nparts);
tvwgt = idxsum(nparts, pwgts);
ASSERT(tvwgt == idxsum(nvtxs, graph->vwgt));
for (i=0; i<nparts; i++) {
itpwgts[i] = tpwgts[i]*tvwgt;
maxwgt[i] = tpwgts[i]*tvwgt*ubfactor;
minwgt[i] = tpwgts[i]*tvwgt*(1.0/ubfactor);
}
perm = idxwspacemalloc(ctrl, nvtxs);
moved = idxwspacemalloc(ctrl, nvtxs);
PQueueInit(ctrl, &queue, nvtxs, graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)]);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions: [%6d %6d]-[%6d %6d], Balance: %5.3f, Nv-Nb[%6d %6d]. Cut: %6d [B]\n",
pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)], minwgt[0], maxwgt[0],
1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nvtxs, graph->nbnd,
graph->mincut));
for (pass=0; pass<npasses; pass++) {
ASSERT(ComputeCut(graph, where) == graph->mincut);
/* Check to see if things are out of balance, given the tolerance */
for (i=0; i<nparts; i++) {
if (pwgts[i] > maxwgt[i])
break;
}
if (i == nparts) /* Things are balanced. Return right away */
break;
PQueueReset(&queue);
idxset(nvtxs, -1, moved);
oldcut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
PQueueInsert(&queue, i, graph->rinfo[i].ed - graph->rinfo[i].id);
moved[i] = 2;
}
maxndoms = ndoms[idxamax(nparts, ndoms)];
for (nmoves=0;;) {
if ((i = PQueueGetMax(&queue)) == -1)
break;
moved[i] = 1;
myrinfo = graph->rinfo+i;
from = where[i];
vwgt = graph->vwgt[i];
if (pwgts[from]-vwgt < minwgt[from])
continue; /* This cannot be moved! */
myedegrees = myrinfo->edegrees;
myndegrees = myrinfo->ndegrees;
/* Determine the valid domains */
for (j=0; j<myndegrees; j++) {
to = myedegrees[j].pid;
phtable[to] = 1;
pmatptr = pmat + to*nparts;
for (nadd=0, k=0; k<myndegrees; k++) {
if (k == j)
continue;
l = myedegrees[k].pid;
if (pmatptr[l] == 0) {
if (ndoms[l] > maxndoms-1) {
phtable[to] = 0;
nadd = maxndoms;
break;
}
nadd++;
}
}
if (ndoms[to]+nadd > maxndoms)
phtable[to] = 0;
}
for (k=0; k<myndegrees; k++) {
to = myedegrees[k].pid;
if (!phtable[to])
continue;
if (pwgts[to]+vwgt <= maxwgt[to] || itpwgts[from]*(pwgts[to]+vwgt) <= itpwgts[to]*pwgts[from])
break;
}
if (k == myndegrees)
continue; /* break out if you did not find a candidate */
for (j=k+1; j<myndegrees; j++) {
to = myedegrees[j].pid;
if (!phtable[to])
continue;
if (itpwgts[myedegrees[k].pid]*pwgts[to] < itpwgts[to]*pwgts[myedegrees[k].pid])
k = j;
}
to = myedegrees[k].pid;
if (pwgts[from] < maxwgt[from] && pwgts[to] > minwgt[to] && myedegrees[k].ed-myrinfo->id < 0)
continue;
/*=====================================================================
* If we got here, we can now move the vertex from 'from' to 'to'
*======================================================================*/
graph->mincut -= myedegrees[k].ed-myrinfo->id;
IFSET(ctrl->dbglvl, DBG_MOVEINFO, printf("\t\tMoving %6d to %3d. Gain: %4d. Cut: %6d\n", i, to, myedegrees[k].ed-myrinfo->id, graph->mincut));
/* Update pmat to reflect the move of 'i' */
pmat[from*nparts+to] += (myrinfo->id-myedegrees[k].ed);
pmat[to*nparts+from] += (myrinfo->id-myedegrees[k].ed);
if (pmat[from*nparts+to] == 0) {
ndoms[from]--;
if (ndoms[from]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
if (pmat[to*nparts+from] == 0) {
ndoms[to]--;
if (ndoms[to]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
/* Update where, weight, and ID/ED information of the vertex you moved */
where[i] = to;
INC_DEC(pwgts[to], pwgts[from], vwgt);
myrinfo->ed += myrinfo->id-myedegrees[k].ed;
SWAP(myrinfo->id, myedegrees[k].ed, j);
if (myedegrees[k].ed == 0)
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].pid = from;
if (myrinfo->ed == 0)
BNDDelete(nbnd, bndind, bndptr, i);
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
me = where[ii];
myrinfo = graph->rinfo+ii;
if (myrinfo->edegrees == NULL) {
myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii];
}
myedegrees = myrinfo->edegrees;
ASSERT(CheckRInfo(myrinfo));
oldgain = (myrinfo->ed-myrinfo->id);
if (me == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
if (myrinfo->ed > 0 && bndptr[ii] == -1)
BNDInsert(nbnd, bndind, bndptr, ii);
}
else if (me == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
if (myrinfo->ed == 0 && bndptr[ii] != -1)
BNDDelete(nbnd, bndind, bndptr, ii);
}
/* Remove contribution from the .ed of 'from' */
if (me != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == from) {
if (myedegrees[k].ed == adjwgt[j])
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].ed -= adjwgt[j];
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (me != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == to) {
myedegrees[k].ed += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
myedegrees[myrinfo->ndegrees].pid = to;
myedegrees[myrinfo->ndegrees++].ed = adjwgt[j];
}
}
/* Update pmat to reflect the move of 'i' for domains other than 'from' and 'to' */
if (me != from && me != to) {
pmat[me*nparts+from] -= adjwgt[j];
pmat[from*nparts+me] -= adjwgt[j];
if (pmat[me*nparts+from] == 0) {
ndoms[me]--;
if (ndoms[me]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
if (pmat[from*nparts+me] == 0) {
ndoms[from]--;
if (ndoms[from]+1 == maxndoms)
maxndoms = ndoms[idxamax(nparts, ndoms)];
}
if (pmat[me*nparts+to] == 0) {
ndoms[me]++;
if (ndoms[me] > maxndoms) {
printf("You just increased the maxndoms: %d %d\n", ndoms[me], maxndoms);
maxndoms = ndoms[me];
}
}
if (pmat[to*nparts+me] == 0) {
ndoms[to]++;
if (ndoms[to] > maxndoms) {
printf("You just increased the maxndoms: %d %d\n", ndoms[to], maxndoms);
maxndoms = ndoms[to];
}
}
pmat[me*nparts+to] += adjwgt[j];
pmat[to*nparts+me] += adjwgt[j];
}
/* Update the queue */
if (me == to || me == from) {
gain = myrinfo->ed-myrinfo->id;
if (moved[ii] == 2) {
if (myrinfo->ed > 0)
PQueueUpdate(&queue, ii, oldgain, gain);
else {
PQueueDelete(&queue, ii, oldgain);
moved[ii] = -1;
}
}
else if (moved[ii] == -1 && myrinfo->ed > 0) {
PQueueInsert(&queue, ii, gain);
moved[ii] = 2;
}
}
ASSERT(myrinfo->ndegrees <= xadj[ii+1]-xadj[ii]);
ASSERT(CheckRInfo(myrinfo));
}
nmoves++;
}
graph->nbnd = nbnd;
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\t[%6d %6d], Balance: %5.3f, Nb: %6d. Nmoves: %5d, Cut: %6d, %d\n",
pwgts[idxamin(nparts, pwgts)], pwgts[idxamax(nparts, pwgts)],
1.0*nparts*pwgts[idxamax(nparts, pwgts)]/tvwgt, graph->nbnd, nmoves, graph->mincut,idxsum(nparts, ndoms)));
}
PQueueFree(ctrl, &queue);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nparts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void GrowBisection(CtrlType *ctrl, GraphType *graph, int *tpwgts, float ubfactor)
{
int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where;
idxtype *queue, *touched, *gain, *bestwhere;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
Allocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
queue = idxmalloc(nvtxs, "BisectGraph: queue");
touched = idxmalloc(nvtxs, "BisectGraph: touched");
ASSERTP(tpwgts[0]+tpwgts[1] == idxsum(nvtxs, vwgt), ("%d %d\n", tpwgts[0]+tpwgts[1], idxsum(nvtxs, vwgt)));
maxpwgt[0] = ubfactor*tpwgts[0];
maxpwgt[1] = ubfactor*tpwgts[1];
minpwgt[0] = (1.0/ubfactor)*tpwgts[0];
minpwgt[1] = (1.0/ubfactor)*tpwgts[1];
nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(nvtxs, graph->adjwgtsum)+1; /* The +1 is for the 0 edges case */
for (; nbfs>0; nbfs--) {
idxset(nvtxs, 0, touched);
pwgts[1] = tpwgts[0]+tpwgts[1];
pwgts[0] = 0;
idxset(nvtxs, 1, where);
queue[0] = RandomInRange(nvtxs);
touched[queue[0]] = 1;
first = 0; last = 1;
nleft = nvtxs-1;
drain = 0;
/* Start the BFS from queue to get a partition */
for (;;) {
if (first == last) { /* Empty. Disconnected graph! */
if (nleft == 0 || drain)
break;
k = RandomInRange(nleft);
for (i=0; i<nvtxs; i++) {
if (touched[i] == 0) {
if (k == 0)
break;
else
k--;
}
}
queue[0] = i;
touched[i] = 1;
first = 0; last = 1;;
nleft--;
}
i = queue[first++];
if (pwgts[0] > 0 && pwgts[1]-vwgt[i] < minpwgt[1]) {
drain = 1;
continue;
}
where[i] = 0;
INC_DEC(pwgts[0], pwgts[1], vwgt[i]);
if (pwgts[1] <= maxpwgt[1])
break;
drain = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (touched[k] == 0) {
queue[last++] = k;
touched[k] = 1;
nleft--;
}
}
}
/* Check to see if we hit any bad limiting cases */
if (pwgts[1] == 0) {
i = RandomInRange(nvtxs);
where[i] = 1;
INC_DEC(pwgts[1], pwgts[0], vwgt[i]);
}
/*************************************************************
* Do some partition refinement
**************************************************************/
Compute2WayPartitionParams(ctrl, graph);
/*printf("IPART: %3d [%5d %5d] [%5d %5d] %5d\n", graph->nvtxs, pwgts[0], pwgts[1], graph->pwgts[0], graph->pwgts[1], graph->mincut); */
Balance2Way(ctrl, graph, tpwgts, ubfactor);
/*printf("BPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/
FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
/*printf("RPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
GKfree(&bestwhere, &queue, &touched, LTERM);
}
/*************************************************************************
* This function moves a collection of vertices and updates their rinfo
**************************************************************************/
void MoveGroupMConn(CtrlType *ctrl, GraphType *graph, idxtype *ndoms, idxtype *pmat,
int nparts, int to, int nind, idxtype *ind)
{
int i, ii, iii, j, jj, k, l, nvtxs, nbnd, myndegrees;
int from, me;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *where, *bndptr, *bndind;
EDegreeType *myedegrees;
RInfoType *myrinfo;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
bndptr = graph->bndptr;
bndind = graph->bndind;
nbnd = graph->nbnd;
for (iii=0; iii<nind; iii++) {
i = ind[iii];
from = where[i];
myrinfo = graph->rinfo+i;
if (myrinfo->edegrees == NULL) {
myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[i+1]-xadj[i];
myrinfo->ndegrees = 0;
}
myedegrees = myrinfo->edegrees;
/* find the location of 'to' in myrinfo or create it if it is not there */
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == to)
break;
}
if (k == myrinfo->ndegrees) {
myedegrees[k].pid = to;
myedegrees[k].ed = 0;
myrinfo->ndegrees++;
}
graph->mincut -= myedegrees[k].ed-myrinfo->id;
/* Update pmat to reflect the move of 'i' */
pmat[from*nparts+to] += (myrinfo->id-myedegrees[k].ed);
pmat[to*nparts+from] += (myrinfo->id-myedegrees[k].ed);
if (pmat[from*nparts+to] == 0)
ndoms[from]--;
if (pmat[to*nparts+from] == 0)
ndoms[to]--;
/* Update where, weight, and ID/ED information of the vertex you moved */
where[i] = to;
myrinfo->ed += myrinfo->id-myedegrees[k].ed;
SWAP(myrinfo->id, myedegrees[k].ed, j);
if (myedegrees[k].ed == 0)
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].pid = from;
if (myrinfo->ed-myrinfo->id < 0 && bndptr[i] != -1)
BNDDelete(nbnd, bndind, bndptr, i);
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
me = where[ii];
myrinfo = graph->rinfo+ii;
if (myrinfo->edegrees == NULL) {
myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[ii+1]-xadj[ii];
}
myedegrees = myrinfo->edegrees;
ASSERT(CheckRInfo(myrinfo));
if (me == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
if (myrinfo->ed-myrinfo->id >= 0 && bndptr[ii] == -1)
BNDInsert(nbnd, bndind, bndptr, ii);
}
else if (me == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
if (myrinfo->ed-myrinfo->id < 0 && bndptr[ii] != -1)
BNDDelete(nbnd, bndind, bndptr, ii);
}
/* Remove contribution from the .ed of 'from' */
if (me != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == from) {
if (myedegrees[k].ed == adjwgt[j])
myedegrees[k] = myedegrees[--myrinfo->ndegrees];
else
myedegrees[k].ed -= adjwgt[j];
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (me != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == to) {
myedegrees[k].ed += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
myedegrees[myrinfo->ndegrees].pid = to;
myedegrees[myrinfo->ndegrees++].ed = adjwgt[j];
}
}
/* Update pmat to reflect the move of 'i' for domains other than 'from' and 'to' */
if (me != from && me != to) {
pmat[me*nparts+from] -= adjwgt[j];
pmat[from*nparts+me] -= adjwgt[j];
if (pmat[me*nparts+from] == 0)
ndoms[me]--;
if (pmat[from*nparts+me] == 0)
ndoms[from]--;
if (pmat[me*nparts+to] == 0)
ndoms[me]++;
if (pmat[to*nparts+me] == 0)
ndoms[to]++;
pmat[me*nparts+to] += adjwgt[j];
pmat[to*nparts+me] += adjwgt[j];
}
ASSERT(CheckRInfo(myrinfo));
}
ASSERT(CheckRInfo(graph->rinfo+i));
}
graph->nbnd = nbnd;
}
void FM_2WayNodeRefine1Sided(ctrl_t *ctrl, graph_t *graph, idx_t niter)
{
idx_t i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind, iend;
idx_t *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idx_t *mptr, *mind, *swaps;
rpq_t *queue;
nrinfo_t *rinfo;
idx_t higain, mincut, initcut, mincutorder;
idx_t pass, to, other, limit;
idx_t badmaxpwgt, mindiff, newdiff;
real_t mult;
WCOREPUSH;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
queue = rpqCreate(nvtxs);
swaps = iwspacemalloc(ctrl, nvtxs);
mptr = iwspacemalloc(ctrl, nvtxs+1);
mind = iwspacemalloc(ctrl, 2*nvtxs);
mult = 0.5*ctrl->ubfactors[0];
badmaxpwgt = (idx_t)(mult*(pwgts[0]+pwgts[1]+pwgts[2]));
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
printf("Partitions-N1: [%6"PRIDX" %6"PRIDX"] Nv-Nb[%6"PRIDX" %6"PRIDX"]. ISep: %6"PRIDX"\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
to = (pwgts[0] < pwgts[1] ? 1 : 0);
for (pass=0; pass<2*niter; pass++) { /* the 2*niter is for the two sides */
other = to;
to = (to+1)%2;
rpqReset(queue);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
/* use the swaps array in place of the traditional perm array to save memory */
my_irandArrayPermute_r(nbnd, swaps, nbnd, 1, &ctrl->curseed);
for (ii=0; ii<nbnd; ii++) {
i = bndind[swaps[ii]];
ASSERT(where[i] == 2);
rpqInsert(queue, i, vwgt[i]-rinfo[i].edegrees[other]);
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = (ctrl->compress ? gk_min(5*nbnd, 500) : gk_min(3*nbnd, 300));
/******************************************************
* Get into the FM loop
*******************************************************/
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux3Tmr));
mptr[0] = nmind = 0;
mindiff = iabs(pwgts[0]-pwgts[1]);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if ((higain = rpqGetTop(queue)) == -1)
break;
ASSERT(bndptr[higain] != -1);
/* The following check is to ensure we break out if there is a posibility
of over-running the mind array. */
if (nmind + xadj[higain+1]-xadj[higain] >= 2*nvtxs-1)
break;
if (pwgts[to]+vwgt[higain] > badmaxpwgt)
break; /* No point going any further. Balance will be bad */
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = iabs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > 3*limit ||
(nswaps - mincutorder > limit && pwgts[2] > 1.10*mincut)) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux1Tmr));
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
rinfo[k].edegrees[to] += vwgt[higain];
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%"PRIDX" %"PRIDX" %"PRIDX"\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k], iend=xadj[k+1]; jj<iend; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
rinfo[kk].edegrees[other] -= vwgt[k];
/* Since the moves are one-sided this vertex has not been moved yet */
rpqUpdate(queue, kk, vwgt[kk]-rinfo[kk].edegrees[other]);
}
}
/* Insert the new vertex into the priority queue. Safe due to one-sided moves */
rpqInsert(queue, k, vwgt[k]-edegrees[other]);
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux1Tmr));
IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO,
printf("Moved %6"PRIDX" to %3"PRIDX", Gain: %5"PRIDX" [%5"PRIDX"] \t[%5"PRIDX" %5"PRIDX" %5"PRIDX"] [%3"PRIDX" %2"PRIDX"]\n",
higain, to, (vwgt[higain]-rinfo[higain].edegrees[other]), vwgt[higain],
pwgts[0], pwgts[1], pwgts[2], nswaps, limit));
}
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux3Tmr));
/****************************************************************
* Roll back computation
*****************************************************************/
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux2Tmr));
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
ASSERT(where[higain] == to);
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k], iend=xadj[k+1]; jj<iend; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux2Tmr));
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
printf("\tMinimum sep: %6"PRIDX" at %5"PRIDX", PWGTS: [%6"PRIDX" %6"PRIDX"], NBND: %6"PRIDX"\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut))
break;
}
rpqDestroy(queue);
WCOREPOP;
}