/*************************************************************************
* This function implements a simple graph adaption strategy.
**************************************************************************/
void AdaptGraph(graph_t *graph, idx_t afactor, MPI_Comm comm)
{
idx_t i, nvtxs, nadapt, firstvtx, lastvtx;
idx_t npes, mype, mypwgt, max, min, sum;
idx_t *vwgt, *xadj, *adjncy, *adjwgt, *perm;
gkMPI_Comm_size(comm, &npes);
gkMPI_Comm_rank(comm, &mype);
srand(mype*afactor);
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
if (graph->adjwgt == NULL)
adjwgt = graph->adjwgt = ismalloc(graph->nedges, 1, "AdaptGraph: adjwgt");
else
adjwgt = graph->adjwgt;
vwgt = graph->vwgt;
firstvtx = graph->vtxdist[mype];
lastvtx = graph->vtxdist[mype+1];
perm = imalloc(nvtxs, "AdaptGraph: perm");
FastRandomPermute(nvtxs, perm, 1);
nadapt = RandomInRange(nvtxs);
nadapt = RandomInRange(nvtxs);
nadapt = RandomInRange(nvtxs);
for (i=0; i<nadapt; i++)
vwgt[perm[i]] = afactor*vwgt[perm[i]];
/*
for (i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (k >= firstvtx && k < lastvtx) {
adjwgt[j] = (int)pow(1.0*(gk_min(vwgt[i],vwgt[k-firstvtx])), .6667);
if (adjwgt[j] == 0)
adjwgt[j] = 1;
}
}
}
*/
mypwgt = isum(nvtxs, vwgt, 1);
MPI_Allreduce((void *)&mypwgt, (void *)&max, 1, IDX_T, MPI_MAX, comm);
MPI_Allreduce((void *)&mypwgt, (void *)&min, 1, IDX_T, MPI_MIN, comm);
MPI_Allreduce((void *)&mypwgt, (void *)&sum, 1, IDX_T, MPI_SUM, comm);
if (mype == 0)
printf("Initial Load Imbalance: %5.4"PRREAL", [%5"PRIDX" %5"PRIDX" %5"PRIDX"] for afactor: %"PRIDX"\n", (1.0*max*npes)/(1.0*sum), min, max, sum, afactor);
free(perm);
}
/*************************************************************************
* 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 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;
}
void Match_Local(ctrl_t *ctrl, graph_t *graph)
{
idx_t h, i, ii, j, k;
idx_t nvtxs, ncon, cnvtxs, firstvtx, maxi, maxidx, edge;
idx_t *xadj, *adjncy, *adjwgt, *vtxdist, *home, *myhome;
idx_t *perm, *match;
real_t maxnvwgt, *nvwgt;
WCOREPUSH;
maxnvwgt = 0.75/((real_t)ctrl->CoarsenTo);
graph->match_type = PARMETIS_MTYPE_LOCAL;
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
home = graph->home;
vtxdist = graph->vtxdist;
firstvtx = vtxdist[ctrl->mype];
match = graph->match = imalloc(nvtxs+graph->nrecv, "HEM_Match: match");
/* wspacemalloc'ed arrays */
myhome = iset(nvtxs+graph->nrecv, UNMATCHED, iwspacemalloc(ctrl, nvtxs+graph->nrecv));
perm = iwspacemalloc(ctrl, nvtxs);
/* if coasening for adaptive/repartition, exchange home information */
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) {
icopy(nvtxs, home, myhome);
CommInterfaceData(ctrl, graph, myhome, myhome+nvtxs);
}
/*************************************************************
* Go now and find a local matching
*************************************************************/
iset(nvtxs, UNMATCHED, match);
iset(graph->nrecv, 0, match+nvtxs); /* Easy way to handle remote vertices */
FastRandomPermute(nvtxs, perm, 1);
for (cnvtxs=0, ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) {
maxidx = maxi = -1;
/* Find a heavy-edge matching, if the weight of the vertex is OK */
for (h=0; h<ncon; h++)
if (nvwgt[i*ncon+h] > maxnvwgt)
break;
if (h == ncon && ii<nvtxs) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
edge = adjncy[j];
/* match only with local vertices */
if (myhome[edge] != myhome[i] || edge >= nvtxs)
continue;
for (h=0; h<ncon; h++)
if (nvwgt[edge*ncon+h] > maxnvwgt)
break;
if (h == ncon) {
if (match[edge] == UNMATCHED &&
(maxi == -1 ||
adjwgt[maxi] < adjwgt[j] ||
(adjwgt[maxi] == adjwgt[j] &&
BetterVBalance(ncon,nvwgt+i*ncon,nvwgt+maxidx*ncon,nvwgt+edge*ncon) >= 0))) {
maxi = j;
maxidx = edge;
}
}
}
}
if (maxi != -1) {
k = adjncy[maxi];
match[i] = firstvtx+k + (i <= k ? KEEP_BIT : 0);
match[k] = firstvtx+i + (i > k ? KEEP_BIT : 0);
}
else {
match[i] = firstvtx+i + KEEP_BIT;
}
cnvtxs++;
}
}
CommInterfaceData(ctrl, graph, match, match+nvtxs);
#ifdef DEBUG_MATCH
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match1");
#endif
if (ctrl->dbglvl&DBG_MATCHINFO) {
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match");
myprintf(ctrl, "Cnvtxs: %"PRIDX"\n", cnvtxs);
rprintf(ctrl, "Done with matching...\n");
}
WCOREPOP;
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
CreateCoarseGraph_Local(ctrl, graph, cnvtxs);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
}
void Match_Global(ctrl_t *ctrl, graph_t *graph)
{
idx_t h, i, ii, j, k;
idx_t nnbrs, nvtxs, ncon, cnvtxs, firstvtx, lastvtx, maxi, maxidx, nkept;
idx_t otherlastvtx, nrequests, nchanged, pass, nmatched, wside;
idx_t *xadj, *adjncy, *adjwgt, *vtxdist, *home, *myhome;
idx_t *match;
idx_t *peind, *sendptr, *recvptr;
idx_t *perm, *iperm, *nperm, *changed;
real_t *nvwgt, maxnvwgt;
idx_t *nreqs_pe;
ikv_t *match_requests, *match_granted, *pe_requests;
idx_t last_unmatched;
WCOREPUSH;
maxnvwgt = 0.75/((real_t)(ctrl->CoarsenTo));
graph->match_type = PARMETIS_MTYPE_GLOBAL;
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
home = graph->home;
nvwgt = graph->nvwgt;
vtxdist = graph->vtxdist;
firstvtx = vtxdist[ctrl->mype];
lastvtx = vtxdist[ctrl->mype+1];
nnbrs = graph->nnbrs;
peind = graph->peind;
sendptr = graph->sendptr;
recvptr = graph->recvptr;
match = graph->match = ismalloc(nvtxs+graph->nrecv, UNMATCHED, "GlobalMatch: match");
/* wspacemalloc'ed arrays */
myhome = iset(nvtxs+graph->nrecv, UNMATCHED, iwspacemalloc(ctrl, nvtxs+graph->nrecv));
nreqs_pe = iset(nnbrs, 0, iwspacemalloc(ctrl, nnbrs));
perm = iwspacemalloc(ctrl, nvtxs);
iperm = iwspacemalloc(ctrl, nvtxs);
nperm = iwspacemalloc(ctrl, nnbrs);
changed = iwspacemalloc(ctrl, nvtxs);
match_requests = ikvwspacemalloc(ctrl, graph->nsend);
match_granted = ikvwspacemalloc(ctrl, graph->nrecv);
/* create the traversal order */
FastRandomPermute(nvtxs, perm, 1);
for (i=0; i<nvtxs; i++)
iperm[perm[i]] = i;
iincset(nnbrs, 0, nperm);
/* if coasening for adaptive/repartition, exchange home information */
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) {
PASSERT(ctrl, home != NULL);
icopy(nvtxs, home, myhome);
CommInterfaceData(ctrl, graph, myhome, myhome+nvtxs);
}
/* if coarsening for ordering, replace home with where information */
if (ctrl->partType == ORDER_PARTITION) {
PASSERT(ctrl, graph->where != NULL);
icopy(nvtxs, graph->where, myhome);
CommInterfaceData(ctrl, graph, myhome, myhome+nvtxs);
}
/* mark all heavy vertices as TOO_HEAVY so they will not be matched */
for (nchanged=i=0; i<nvtxs; i++) {
for (h=0; h<ncon; h++) {
if (nvwgt[i*ncon+h] > maxnvwgt) {
match[i] = TOO_HEAVY;
nchanged++;
break;
}
}
}
/* If no heavy vertices, pick one at random and treat it as such so that
at the end of the matching each partition will still have one vertex.
This is to eliminate the cases in which once a matching has been
computed, a processor ends up having no vertices */
if (nchanged == 0)
match[RandomInRange(nvtxs)] = TOO_HEAVY;
CommInterfaceData(ctrl, graph, match, match+nvtxs);
/* set initial value of nkept based on how over/under weight the
partition is to begin with */
nkept = graph->gnvtxs/ctrl->npes - nvtxs;
/* Find a matching by doing multiple iterations */
for (nmatched=pass=0; pass<NMATCH_PASSES; pass++) {
wside = (graph->level+pass)%2;
nchanged = nrequests = 0;
for (last_unmatched=ii=nmatched; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
maxidx = i;
maxi = -1;
/* Deal with islands. Find another vertex and match it with */
if (xadj[i] == xadj[i+1]) {
last_unmatched = gk_max(ii, last_unmatched)+1;
for (; last_unmatched<nvtxs; last_unmatched++) {
k = perm[last_unmatched];
if (match[k] == UNMATCHED && myhome[i] == myhome[k]) {
match[i] = firstvtx+k + (i <= k ? KEEP_BIT : 0);
match[k] = firstvtx+i + (i > k ? KEEP_BIT : 0);
changed[nchanged++] = i;
changed[nchanged++] = k;
break;
}
}
continue;
}
/* Find a heavy-edge matching. */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (match[k] == UNMATCHED && myhome[k] == myhome[i]) {
if (ncon == 1) {
if (maxi == -1 || adjwgt[maxi] < adjwgt[j] ||
(adjwgt[maxi] == adjwgt[j] && RandomInRange(xadj[i+1]-xadj[i]) == 0)) {
maxi = j;
maxidx = k;
}
}
else {
if (maxi == -1 || adjwgt[maxi] < adjwgt[j] ||
(adjwgt[maxi] == adjwgt[j] && maxidx < nvtxs && k < nvtxs &&
BetterVBalance(ncon,nvwgt+i*ncon,nvwgt+maxidx*ncon,nvwgt+k*ncon) >= 0)) {
maxi = j;
maxidx = k;
}
}
}
}
if (maxi != -1) {
k = adjncy[maxi];
if (k < nvtxs) { /* Take care the local vertices first */
/* Here we give preference the local matching by granting it right away */
match[i] = firstvtx+k + (i <= k ? KEEP_BIT : 0);
match[k] = firstvtx+i + (i > k ? KEEP_BIT : 0);
changed[nchanged++] = i;
changed[nchanged++] = k;
}
else { /* Take care any remote boundary vertices */
match[k] = MAYBE_MATCHED;
/* Alternate among which vertices will issue the requests */
if ((wside == 0 && firstvtx+i < graph->imap[k]) ||
(wside == 1 && firstvtx+i > graph->imap[k])) {
match[i] = MAYBE_MATCHED;
match_requests[nrequests].key = graph->imap[k];
match_requests[nrequests].val = firstvtx+i;
nrequests++;
}
}
}
}
}
#ifdef DEBUG_MATCH
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match1");
myprintf(ctrl, "[c: %2"PRIDX"] Nlocal: %"PRIDX", Nrequests: %"PRIDX"\n", c, nlocal, nrequests);
#endif
/***********************************************************
* Exchange the match_requests, requests for me are stored in
* match_granted
************************************************************/
/* Issue the receives first. Note that from each PE can receive a maximum
of the interface node that it needs to send it in the case of a mat-vec */
for (i=0; i<nnbrs; i++) {
gkMPI_Irecv((void *)(match_granted+recvptr[i]), 2*(recvptr[i+1]-recvptr[i]), IDX_T,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. This needs some work */
ikvsorti(nrequests, match_requests);
for (j=i=0; i<nnbrs; i++) {
otherlastvtx = vtxdist[peind[i]+1];
for (k=j; k<nrequests && match_requests[k].key < otherlastvtx; k++);
gkMPI_Isend((void *)(match_requests+j), 2*(k-j), IDX_T, peind[i], 1,
ctrl->comm, ctrl->sreq+i);
j = k;
}
/* OK, now get into the loop waiting for the operations to finish */
gkMPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++) {
gkMPI_Get_count(ctrl->statuses+i, IDX_T, nreqs_pe+i);
nreqs_pe[i] = nreqs_pe[i]/2; /* Adjust for pairs of IDX_T */
}
gkMPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/***********************************************************
* Now, go and service the requests that you received in
* match_granted
************************************************************/
RandomPermute(nnbrs, nperm, 0);
for (ii=0; ii<nnbrs; ii++) {
i = nperm[ii];
pe_requests = match_granted+recvptr[i];
for (j=0; j<nreqs_pe[i]; j++) {
k = pe_requests[j].key;
PASSERTP(ctrl, k >= firstvtx && k < lastvtx, (ctrl, "%"PRIDX" %"PRIDX" %"PRIDX" %"PRIDX" %"PRIDX"\n", firstvtx, lastvtx, k, j, peind[i]));
/* myprintf(ctrl, "Requesting a match %"PRIDX" %"PRIDX"\n", pe_requests[j].key, pe_requests[j].val); */
if (match[k-firstvtx] == UNMATCHED) { /* Bingo, lets grant this request */
changed[nchanged++] = k-firstvtx;
if (nkept >= 0) { /* decide who to keep it based on local balance */
match[k-firstvtx] = pe_requests[j].val + KEEP_BIT;
nkept--;
}
else {
match[k-firstvtx] = pe_requests[j].val;
pe_requests[j].key += KEEP_BIT;
nkept++;
}
/* myprintf(ctrl, "Request from pe:%"PRIDX" (%"PRIDX" %"PRIDX") granted!\n", peind[i], pe_requests[j].val, pe_requests[j].key); */
}
else { /* We are not granting the request */
/* myprintf(ctrl, "Request from pe:%"PRIDX" (%"PRIDX" %"PRIDX") not granted!\n", peind[i], pe_requests[j].val, pe_requests[j].key); */
pe_requests[j].key = UNMATCHED;
}
}
}
/***********************************************************
* Exchange the match_granted information. It is stored in
* match_requests
************************************************************/
/* Issue the receives first. Note that from each PE can receive a maximum
of the interface node that it needs to send during the case of a mat-vec */
for (i=0; i<nnbrs; i++) {
gkMPI_Irecv((void *)(match_requests+sendptr[i]), 2*(sendptr[i+1]-sendptr[i]), IDX_T,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. */
for (i=0; i<nnbrs; i++) {
gkMPI_Isend((void *)(match_granted+recvptr[i]), 2*nreqs_pe[i], IDX_T,
peind[i], 1, ctrl->comm, ctrl->sreq+i);
}
/* OK, now get into the loop waiting for the operations to finish */
gkMPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++) {
gkMPI_Get_count(ctrl->statuses+i, IDX_T, nreqs_pe+i);
nreqs_pe[i] = nreqs_pe[i]/2; /* Adjust for pairs of IDX_T */
}
gkMPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/***********************************************************
* Now, go and through the match_requests and update local
* match information for the matchings that were granted.
************************************************************/
for (i=0; i<nnbrs; i++) {
pe_requests = match_requests+sendptr[i];
for (j=0; j<nreqs_pe[i]; j++) {
match[pe_requests[j].val-firstvtx] = pe_requests[j].key;
if (pe_requests[j].key != UNMATCHED)
changed[nchanged++] = pe_requests[j].val-firstvtx;
}
}
for (i=0; i<nchanged; i++) {
ii = iperm[changed[i]];
perm[ii] = perm[nmatched];
iperm[perm[nmatched]] = ii;
nmatched++;
}
CommChangedInterfaceData(ctrl, graph, nchanged, changed, match,
match_requests, match_granted);
}
/* Traverse the vertices and those that were unmatched, match them with themselves */
cnvtxs = 0;
for (i=0; i<nvtxs; i++) {
if (match[i] == UNMATCHED || match[i] == TOO_HEAVY) {
match[i] = (firstvtx+i) + KEEP_BIT;
cnvtxs++;
}
else if (match[i] >= KEEP_BIT) { /* A matched vertex which I get to keep */
cnvtxs++;
}
}
if (ctrl->dbglvl&DBG_MATCHINFO) {
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match");
myprintf(ctrl, "Cnvtxs: %"PRIDX"\n", cnvtxs);
rprintf(ctrl, "Done with matching...\n");
}
WCOREPOP;
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
CreateCoarseGraph_Global(ctrl, graph, cnvtxs);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
}
/*************************************************************************
* This function finds a matching
**************************************************************************/
void Moc_GlobalMatch_Balance(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int h, i, ii, j, k;
int nnbrs, nvtxs, ncon, cnvtxs, firstvtx, lastvtx, maxi, maxidx, nkept;
int otherlastvtx, nrequests, nchanged, pass, nmatched, wside;
idxtype *xadj, *ladjncy, *adjwgt, *vtxdist, *home, *myhome, *shome, *rhome;
idxtype *match, *rmatch, *smatch;
idxtype *peind, *sendptr, *recvptr;
idxtype *perm, *iperm, *nperm, *changed;
floattype *nvwgt, maxnvwgt;
int *nreqs_pe;
KeyValueType *match_requests, *match_granted, *pe_requests;
maxnvwgt = 1.0/((floattype)(ctrl->nparts)*MAXNVWGT_FACTOR);
graph->match_type = MATCH_GLOBAL;
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
ladjncy = graph->adjncy;
adjwgt = graph->adjwgt;
home = graph->home;
nvwgt = graph->nvwgt;
vtxdist = graph->vtxdist;
firstvtx = vtxdist[ctrl->mype];
lastvtx = vtxdist[ctrl->mype+1];
match = graph->match = idxsmalloc(nvtxs+graph->nrecv, UNMATCHED, "HEM_Match: match");
myhome = idxsmalloc(nvtxs+graph->nrecv, UNMATCHED, "HEM_Match: myhome");
/*------------------------------------------------------------
/ Send/Receive the home information of interface vertices
/------------------------------------------------------------*/
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) {
idxcopy(nvtxs, home, myhome);
shome = wspace->indices;
rhome = myhome + nvtxs;
CommInterfaceData(ctrl, graph, myhome, shome, rhome);
}
nnbrs = graph->nnbrs;
peind = graph->peind;
sendptr = graph->sendptr;
recvptr = graph->recvptr;
/* Use wspace->indices as the tmp space for matching info of the boundary
* vertices that are sent and received */
rmatch = match + nvtxs;
smatch = wspace->indices;
changed = smatch+graph->nsend;
/* Use wspace->indices as the tmp space for match requests of the boundary
* vertices that are sent and received */
match_requests = wspace->pairs;
match_granted = match_requests + graph->nsend;
nreqs_pe = ismalloc(nnbrs, 0, "Match_HEM: nreqs_pe");
nkept = graph->gnvtxs/ctrl->npes - nvtxs;
perm = (idxtype *)wspace->degrees;
iperm = perm + nvtxs;
FastRandomPermute(nvtxs, perm, 1);
for (i=0; i<nvtxs; i++)
iperm[perm[i]] = i;
nperm = iperm + nvtxs;
for (i=0; i<nnbrs; i++)
nperm[i] = i;
/*************************************************************
* Go now and find a matching by doing multiple iterations
*************************************************************/
/* First nullify the heavy vertices */
for (nchanged=i=0; i<nvtxs; i++) {
for (h=0; h<ncon; h++)
if (nvwgt[i*ncon+h] > maxnvwgt) {
break;
}
if (h != ncon) {
match[i] = TOO_HEAVY;
nchanged++;
}
}
if (GlobalSESum(ctrl, nchanged) > 0) {
IFSET(ctrl->dbglvl, DBG_PROGRESS,
rprintf(ctrl, "We found %d heavy vertices!\n", GlobalSESum(ctrl, nchanged)));
CommInterfaceData(ctrl, graph, match, smatch, rmatch);
}
for (nmatched=pass=0; pass<NMATCH_PASSES; pass++) {
wside = (graph->level+pass)%2;
nchanged = nrequests = 0;
for (ii=nmatched; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
maxidx = i;
maxi = -1;
/* Find a heavy-edge matching */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = ladjncy[j];
if (match[k] == UNMATCHED &&
myhome[k] == myhome[i] &&
(maxi == -1 ||
adjwgt[maxi] < adjwgt[j] ||
(maxidx < nvtxs &&
k < nvtxs &&
adjwgt[maxi] == adjwgt[j] &&
BetterVBalance(ncon,nvwgt+i*ncon,nvwgt+maxidx*ncon,nvwgt+k*ncon) >= 0))) {
maxi = j;
maxidx = k;
}
}
if (maxi != -1) {
k = ladjncy[maxi];
if (k < nvtxs) { /* Take care the local vertices first */
/* Here we give preference the local matching by granting it right away */
if (i <= k) {
match[i] = firstvtx+k + KEEP_BIT;
match[k] = firstvtx+i;
}
else {
match[i] = firstvtx+k;
match[k] = firstvtx+i + KEEP_BIT;
}
changed[nchanged++] = i;
changed[nchanged++] = k;
}
else { /* Take care any remote boundary vertices */
match[k] = MAYBE_MATCHED;
/* Alternate among which vertices will issue the requests */
if ((wside ==0 && firstvtx+i < graph->imap[k]) || (wside == 1 && firstvtx+i > graph->imap[k])) {
match[i] = MAYBE_MATCHED;
match_requests[nrequests].key = graph->imap[k];
match_requests[nrequests].val = firstvtx+i;
nrequests++;
}
}
}
}
}
#ifdef DEBUG_MATCH
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match1");
myprintf(ctrl, "[c: %2d] Nlocal: %d, Nrequests: %d\n", c, nlocal, nrequests);
#endif
/***********************************************************
* Exchange the match_requests, requests for me are stored in
* match_granted
************************************************************/
/* Issue the receives first. Note that from each PE can receive a maximum
of the interface node that it needs to send it in the case of a mat-vec */
for (i=0; i<nnbrs; i++) {
MPI_Irecv((void *)(match_granted+recvptr[i]), 2*(recvptr[i+1]-recvptr[i]), IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. This needs some work */
ikeysort(nrequests, match_requests);
for (j=i=0; i<nnbrs; i++) {
otherlastvtx = vtxdist[peind[i]+1];
for (k=j; k<nrequests && match_requests[k].key < otherlastvtx; k++);
MPI_Isend((void *)(match_requests+j), 2*(k-j), IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i);
j = k;
}
/* OK, now get into the loop waiting for the operations to finish */
MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++) {
MPI_Get_count(ctrl->statuses+i, IDX_DATATYPE, nreqs_pe+i);
nreqs_pe[i] = nreqs_pe[i]/2; /* Adjust for pairs of IDX_DATATYPE */
}
MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/***********************************************************
* Now, go and service the requests that you received in
* match_granted
************************************************************/
RandomPermute(nnbrs, nperm, 0);
for (ii=0; ii<nnbrs; ii++) {
i = nperm[ii];
pe_requests = match_granted+recvptr[i];
for (j=0; j<nreqs_pe[i]; j++) {
k = pe_requests[j].key;
ASSERTP(ctrl, k >= firstvtx && k < lastvtx, (ctrl, "%d %d %d %d %d\n", firstvtx, lastvtx, k, j, peind[i]));
/* myprintf(ctrl, "Requesting a match %d %d\n", pe_requests[j].key, pe_requests[j].val); */
if (match[k-firstvtx] == UNMATCHED) { /* Bingo, lets grant this request */
changed[nchanged++] = k-firstvtx;
if (nkept >= 0) { /* Flip a coin for who gets it */
match[k-firstvtx] = pe_requests[j].val + KEEP_BIT;
nkept--;
}
else {
match[k-firstvtx] = pe_requests[j].val;
pe_requests[j].key += KEEP_BIT;
nkept++;
}
/* myprintf(ctrl, "Request from pe:%d (%d %d) granted!\n", peind[i], pe_requests[j].val, pe_requests[j].key); */
}
else { /* We are not granting the request */
/* myprintf(ctrl, "Request from pe:%d (%d %d) not granted!\n", peind[i], pe_requests[j].val, pe_requests[j].key); */
pe_requests[j].key = UNMATCHED;
}
}
}
/***********************************************************
* Exchange the match_granted information. It is stored in
* match_requests
************************************************************/
/* Issue the receives first. Note that from each PE can receive a maximum
of the interface node that it needs to send during the case of a mat-vec */
for (i=0; i<nnbrs; i++) {
MPI_Irecv((void *)(match_requests+sendptr[i]), 2*(sendptr[i+1]-sendptr[i]), IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. */
for (i=0; i<nnbrs; i++) {
MPI_Isend((void *)(match_granted+recvptr[i]), 2*nreqs_pe[i], IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->sreq+i);
}
/* OK, now get into the loop waiting for the operations to finish */
MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++) {
MPI_Get_count(ctrl->statuses+i, IDX_DATATYPE, nreqs_pe+i);
nreqs_pe[i] = nreqs_pe[i]/2; /* Adjust for pairs of IDX_DATATYPE */
}
MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/***********************************************************
* Now, go and through the match_requests and update local
* match information for the matchings that were granted.
************************************************************/
for (i=0; i<nnbrs; i++) {
pe_requests = match_requests+sendptr[i];
for (j=0; j<nreqs_pe[i]; j++) {
match[pe_requests[j].val-firstvtx] = pe_requests[j].key;
if (pe_requests[j].key != UNMATCHED)
changed[nchanged++] = pe_requests[j].val-firstvtx;
}
}
for (i=0; i<nchanged; i++) {
ii = iperm[changed[i]];
perm[ii] = perm[nmatched];
iperm[perm[nmatched]] = ii;
nmatched++;
}
CommChangedInterfaceData(ctrl, graph, nchanged, changed, match, match_requests, match_granted, wspace->pv4);
}
/* Traverse the vertices and those that were unmatched, match them with themselves */
cnvtxs = 0;
for (i=0; i<nvtxs; i++) {
if (match[i] == UNMATCHED || match[i] == TOO_HEAVY) {
match[i] = (firstvtx+i) + KEEP_BIT;
cnvtxs++;
}
else if (match[i] >= KEEP_BIT) { /* A matched vertex which I get to keep */
cnvtxs++;
}
}
if (ctrl->dbglvl&DBG_MATCHINFO) {
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match");
myprintf(ctrl, "Cnvtxs: %d\n", cnvtxs);
rprintf(ctrl, "Done with matching...\n");
}
GKfree((void **)(&myhome), (void **)(&nreqs_pe), LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
Moc_Global_CreateCoarseGraph(ctrl, graph, wspace, cnvtxs);
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
}
