/*************************************************************************
* 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 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 is the driver for the adaptive refinement mode of ParMETIS
**************************************************************************/
void Adaptive_Partition(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int i;
int tewgt, tvsize;
floattype gtewgt, gtvsize;
floattype ubavg, lbavg, lbvec[MAXNCON];
/************************************/
/* Set up important data structures */
/************************************/
SetUp(ctrl, graph, wspace);
ubavg = savg(graph->ncon, ctrl->ubvec);
tewgt = idxsum(graph->nedges, graph->adjwgt);
tvsize = idxsum(graph->nvtxs, graph->vsize);
gtewgt = (floattype) GlobalSESum(ctrl, tewgt) + 1.0; /* The +1 were added to remove any FPE */
gtvsize = (floattype) GlobalSESum(ctrl, tvsize) + 1.0;
ctrl->redist_factor = ctrl->redist_base * ((gtewgt/gtvsize)/ ctrl->edge_size_ratio);
IFSET(ctrl->dbglvl, DBG_PROGRESS, rprintf(ctrl, "[%6d %8d %5d %5d][%d]\n",
graph->gnvtxs, GlobalSESum(ctrl, graph->nedges), GlobalSEMin(ctrl, graph->nvtxs), GlobalSEMax(ctrl, graph->nvtxs), ctrl->CoarsenTo));
if (graph->gnvtxs < 1.3*ctrl->CoarsenTo ||
(graph->finer != NULL && graph->gnvtxs > graph->finer->gnvtxs*COARSEN_FRACTION)) {
/***********************************************/
/* Balance the partition on the coarsest graph */
/***********************************************/
graph->where = idxsmalloc(graph->nvtxs+graph->nrecv, -1, "graph->where");
idxcopy(graph->nvtxs, graph->home, graph->where);
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
lbavg = savg(graph->ncon, lbvec);
if (lbavg > ubavg + 0.035 && ctrl->partType != REFINE_PARTITION)
Balance_Partition(ctrl, graph, wspace);
if (ctrl->dbglvl&DBG_PROGRESS) {
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
rprintf(ctrl, "nvtxs: %10d, balance: ", graph->gnvtxs);
for (i=0; i<graph->ncon; i++)
rprintf(ctrl, "%.3f ", lbvec[i]);
rprintf(ctrl, "\n");
}
/* check if no coarsening took place */
if (graph->finer == NULL) {
Moc_ComputePartitionParams(ctrl, graph, wspace);
Moc_KWayBalance(ctrl, graph, wspace, graph->ncon);
Moc_KWayAdaptiveRefine(ctrl, graph, wspace, NGR_PASSES);
}
}
else {
/*******************************/
/* Coarsen it and partition it */
/*******************************/
switch (ctrl->ps_relation) {
case COUPLED:
Mc_LocalMatch_HEM(ctrl, graph, wspace);
break;
case DISCOUPLED:
default:
Moc_GlobalMatch_Balance(ctrl, graph, wspace);
break;
}
Adaptive_Partition(ctrl, graph->coarser, wspace);
/********************************/
/* project partition and refine */
/********************************/
Moc_ProjectPartition(ctrl, graph, wspace);
Moc_ComputePartitionParams(ctrl, graph, wspace);
if (graph->ncon > 1 && graph->level < 4) {
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
lbavg = savg(graph->ncon, lbvec);
if (lbavg > ubavg + 0.025) {
Moc_KWayBalance(ctrl, graph, wspace, graph->ncon);
}
}
Moc_KWayAdaptiveRefine(ctrl, graph, wspace, NGR_PASSES);
if (ctrl->dbglvl&DBG_PROGRESS) {
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
rprintf(ctrl, "nvtxs: %10d, cut: %8d, balance: ", graph->gnvtxs, graph->mincut);
for (i=0; i<graph->ncon; i++)
rprintf(ctrl, "%.3f ", lbvec[i]);
rprintf(ctrl, "\n");
}
}
}
/*************************************************************************
* This function is the driver to the multi-constraint partitioning algorithm.
**************************************************************************/
void Moc_Global_Partition(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int i, ncon, nparts;
floattype ftmp, ubavg, lbavg, lbvec[MAXNCON];
ncon = graph->ncon;
nparts = ctrl->nparts;
ubavg = savg(graph->ncon, ctrl->ubvec);
SetUp(ctrl, graph, wspace);
if (ctrl->dbglvl&DBG_PROGRESS) {
rprintf(ctrl, "[%6d %8d %5d %5d] [%d] [", graph->gnvtxs, GlobalSESum(ctrl, graph->nedges),
GlobalSEMin(ctrl, graph->nvtxs), GlobalSEMax(ctrl, graph->nvtxs), ctrl->CoarsenTo);
for (i=0; i<ncon; i++)
rprintf(ctrl, " %.3f", GlobalSEMinFloat(ctrl,graph->nvwgt[samin_strd(graph->nvtxs, graph->nvwgt+i, ncon)*ncon+i]));
rprintf(ctrl, "] [");
for (i=0; i<ncon; i++)
rprintf(ctrl, " %.3f", GlobalSEMaxFloat(ctrl, graph->nvwgt[samax_strd(graph->nvtxs, graph->nvwgt+i, ncon)*ncon+i]));
rprintf(ctrl, "]\n");
}
if (graph->gnvtxs < 1.3*ctrl->CoarsenTo ||
(graph->finer != NULL &&
graph->gnvtxs > graph->finer->gnvtxs*COARSEN_FRACTION)) {
/* Done with coarsening. Find a partition */
graph->where = idxmalloc(graph->nvtxs+graph->nrecv, "graph->where");
Moc_InitPartition_RB(ctrl, graph, wspace);
if (ctrl->dbglvl&DBG_PROGRESS) {
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
rprintf(ctrl, "nvtxs: %10d, balance: ", graph->gnvtxs);
for (i=0; i<graph->ncon; i++)
rprintf(ctrl, "%.3f ", lbvec[i]);
rprintf(ctrl, "\n");
}
/* In case no coarsening took place */
if (graph->finer == NULL) {
Moc_ComputePartitionParams(ctrl, graph, wspace);
Moc_KWayFM(ctrl, graph, wspace, NGR_PASSES);
}
}
else {
Moc_GlobalMatch_Balance(ctrl, graph, wspace);
Moc_Global_Partition(ctrl, graph->coarser, wspace);
Moc_ProjectPartition(ctrl, graph, wspace);
Moc_ComputePartitionParams(ctrl, graph, wspace);
if (graph->ncon > 1 && graph->level < 3) {
for (i=0; i<ncon; i++) {
ftmp = ssum_strd(nparts, graph->gnpwgts+i, ncon);
if (ftmp != 0.0)
lbvec[i] = (floattype)(nparts) *
graph->gnpwgts[samax_strd(nparts, graph->gnpwgts+i, ncon)*ncon+i]/ftmp;
else
lbvec[i] = 1.0;
}
lbavg = savg(graph->ncon, lbvec);
if (lbavg > ubavg + 0.035) {
if (ctrl->dbglvl&DBG_PROGRESS) {
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
rprintf(ctrl, "nvtxs: %10d, cut: %8d, balance: ", graph->gnvtxs, graph->mincut);
for (i=0; i<graph->ncon; i++)
rprintf(ctrl, "%.3f ", lbvec[i]);
rprintf(ctrl, "\n");
}
Moc_KWayBalance(ctrl, graph, wspace, graph->ncon);
}
}
Moc_KWayFM(ctrl, graph, wspace, NGR_PASSES);
if (ctrl->dbglvl&DBG_PROGRESS) {
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
rprintf(ctrl, "nvtxs: %10d, cut: %8d, balance: ", graph->gnvtxs, graph->mincut);
for (i=0; i<graph->ncon; i++)
rprintf(ctrl, "%.3f ", lbvec[i]);
rprintf(ctrl, "\n");
}
if (graph->level != 0)
GKfree((void **)&graph->lnpwgts, (void **)&graph->gnpwgts, LTERM);
}
return;
}