/*************************************************************************
* 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;
}
graph_t *CompressGraph(ctrl_t *ctrl, idx_t nvtxs, idx_t *xadj, idx_t *adjncy,
idx_t *vwgt, idx_t *cptr, idx_t *cind)
{
idx_t i, ii, iii, j, jj, k, l, cnvtxs, cnedges;
idx_t *cxadj, *cadjncy, *cvwgt, *mark, *map;
ikv_t *keys;
graph_t *graph=NULL;
mark = ismalloc(nvtxs, -1, "CompressGraph: mark");
map = ismalloc(nvtxs, -1, "CompressGraph: map");
keys = ikvmalloc(nvtxs, "CompressGraph: keys");
/* Compute a key for each adjacency list */
for (i=0; i<nvtxs; i++) {
k = 0;
for (j=xadj[i]; j<xadj[i+1]; j++)
k += adjncy[j];
keys[i].key = k+i; /* Add the diagonal entry as well */
keys[i].val = i;
}
ikvsorti(nvtxs, keys);
l = cptr[0] = 0;
for (cnvtxs=i=0; i<nvtxs; i++) {
ii = keys[i].val;
if (map[ii] == -1) {
mark[ii] = i; /* Add the diagonal entry */
for (j=xadj[ii]; j<xadj[ii+1]; j++)
mark[adjncy[j]] = i;
map[ii] = cnvtxs;
cind[l++] = ii;
for (j=i+1; j<nvtxs; j++) {
iii = keys[j].val;
if (keys[i].key != keys[j].key || xadj[ii+1]-xadj[ii] != xadj[iii+1]-xadj[iii])
break; /* Break if keys or degrees are different */
if (map[iii] == -1) { /* Do a comparison if iii has not been mapped */
for (jj=xadj[iii]; jj<xadj[iii+1]; jj++) {
if (mark[adjncy[jj]] != i)
break;
}
if (jj == xadj[iii+1]) { /* Identical adjacency structure */
map[iii] = cnvtxs;
cind[l++] = iii;
}
}
}
cptr[++cnvtxs] = l;
}
}
IFSET(ctrl->dbglvl, METIS_DBG_INFO,
printf(" Compression: reduction in # of vertices: %"PRIDX".\n", nvtxs-cnvtxs));
if (cnvtxs < COMPRESSION_FRACTION*nvtxs) {
/* Sufficient compression is possible, so go ahead and create the
compressed graph */
graph = CreateGraph();
cnedges = 0;
for (i=0; i<cnvtxs; i++) {
ii = cind[cptr[i]];
cnedges += xadj[ii+1]-xadj[ii];
}
/* Allocate memory for the compressed graph */
cxadj = graph->xadj = imalloc(cnvtxs+1, "CompressGraph: xadj");
cvwgt = graph->vwgt = ismalloc(cnvtxs, 0, "CompressGraph: vwgt");
cadjncy = graph->adjncy = imalloc(cnedges, "CompressGraph: adjncy");
graph->adjwgt = ismalloc(cnedges, 1, "CompressGraph: adjwgt");
/* Now go and compress the graph */
iset(nvtxs, -1, mark);
l = cxadj[0] = 0;
for (i=0; i<cnvtxs; i++) {
mark[i] = i; /* Remove any dioganal entries in the compressed graph */
for (j=cptr[i]; j<cptr[i+1]; j++) {
ii = cind[j];
/* accumulate the vertex weights of the consistuent vertices */
cvwgt[i] += (vwgt == NULL ? 1 : vwgt[ii]);
/* generate the combined adjancency list */
for (jj=xadj[ii]; jj<xadj[ii+1]; jj++) {
k = map[adjncy[jj]];
if (mark[k] != i) {
mark[k] = i;
cadjncy[l++] = k;
}
}
}
cxadj[i+1] = l;
}
graph->nvtxs = cnvtxs;
graph->nedges = l;
graph->ncon = 1;
SetupGraph_tvwgt(graph);
SetupGraph_label(graph);
}
gk_free((void **)&keys, &map, &mark, LTERM);
return graph;
}
void EliminateSubDomainEdges(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, ii, j, k, ncon, nparts, scheme, pid_from, pid_to, me, other, nvtxs,
total, max, avg, totalout, nind=0, ncand=0, ncand2, target, target2,
nadd, bestnadd=0;
idx_t min, move, *cpwgt;
idx_t *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *maxpwgt,
*mypmat, *otherpmat, *kpmat, *ind;
idx_t *nads, **adids, **adwgts;
ikv_t *cand, *cand2;
ipq_t queue;
real_t *tpwgts, badfactor=1.4;
idx_t *pptr, *pind;
idx_t *vmarker=NULL, *pmarker=NULL, *modind=NULL; /* volume specific work arrays */
WCOREPUSH;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
adjwgt = (ctrl->objtype == METIS_OBJTYPE_VOL ? NULL : graph->adjwgt);
where = graph->where;
pwgts = graph->pwgts; /* We assume that this is properly initialized */
nparts = ctrl->nparts;
tpwgts = ctrl->tpwgts;
cpwgt = iwspacemalloc(ctrl, ncon);
maxpwgt = iwspacemalloc(ctrl, nparts*ncon);
ind = iwspacemalloc(ctrl, nvtxs);
otherpmat = iset(nparts, 0, iwspacemalloc(ctrl, nparts));
cand = ikvwspacemalloc(ctrl, nparts);
cand2 = ikvwspacemalloc(ctrl, nparts);
pptr = iwspacemalloc(ctrl, nparts+1);
pind = iwspacemalloc(ctrl, nvtxs);
iarray2csr(nvtxs, nparts, where, pptr, pind);
if (ctrl->objtype == METIS_OBJTYPE_VOL) {
/* Vol-refinement specific working arrays */
modind = iwspacemalloc(ctrl, nvtxs);
vmarker = iset(nvtxs, 0, iwspacemalloc(ctrl, nvtxs));
pmarker = iset(nparts, -1, iwspacemalloc(ctrl, nparts));
}
/* Compute the pmat matrix and ndoms */
ComputeSubDomainGraph(ctrl, graph);
nads = ctrl->nads;
adids = ctrl->adids;
adwgts = ctrl->adwgts;
mypmat = iset(nparts, 0, ctrl->pvec1);
kpmat = iset(nparts, 0, ctrl->pvec2);
/* Compute the maximum allowed weight for each domain */
for (i=0; i<nparts; i++) {
for (j=0; j<ncon; j++)
maxpwgt[i*ncon+j] =
(ncon == 1 ? 1.25 : 1.025)*tpwgts[i]*graph->tvwgt[j]*ctrl->ubfactors[j];
}
ipqInit(&queue, nparts);
/* Get into the loop eliminating subdomain connections */
while (1) {
total = isum(nparts, nads, 1);
avg = total/nparts;
max = nads[iargmax(nparts, nads)];
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("Adjacent Subdomain Stats: Total: %3"PRIDX", "
"Max: %3"PRIDX"[%zu], Avg: %3"PRIDX"\n",
total, max, iargmax(nparts, nads), avg));
if (max < badfactor*avg)
break;
/* Add the subdomains that you will try to reduce their connectivity */
ipqReset(&queue);
for (i=0; i<nparts; i++) {
if (nads[i] >= avg + (max-avg)/2)
ipqInsert(&queue, i, nads[i]);
}
move = 0;
while ((me = ipqGetTop(&queue)) != -1) {
totalout = isum(nads[me], adwgts[me], 1);
for (ncand2=0, i=0; i<nads[me]; i++) {
mypmat[adids[me][i]] = adwgts[me][i];
/* keep track of the weakly connected adjacent subdomains */
if (2*nads[me]*adwgts[me][i] < totalout) {
cand2[ncand2].val = adids[me][i];
cand2[ncand2++].key = adwgts[me][i];
}
}
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("Me: %"PRIDX", Degree: %4"PRIDX", TotalOut: %"PRIDX",\n",
me, nads[me], totalout));
/* Sort the connections according to their cut */
ikvsorti(ncand2, cand2);
/* Two schemes are used for eliminating subdomain edges.
The first, tries to eliminate subdomain edges by moving remote groups
of vertices to subdomains that 'me' is already connected to.
The second, tries to eliminate subdomain edges by moving entire sets of
my vertices that connect to the 'other' subdomain to a subdomain that
I'm already connected to.
These two schemes are applied in sequence. */
target = target2 = -1;
for (scheme=0; scheme<2; scheme++) {
for (min=0; min<ncand2; min++) {
other = cand2[min].val;
/* pid_from is the subdomain from where the vertices will be removed.
pid_to is the adjacent subdomain to pid_from that defines the
(me, other) subdomain edge that needs to be removed */
if (scheme == 0) {
pid_from = other;
pid_to = me;
}
else {
pid_from = me;
pid_to = other;
}
/* Go and find the vertices in 'other' that are connected in 'me' */
for (nind=0, ii=pptr[pid_from]; ii<pptr[pid_from+1]; ii++) {
i = pind[ii];
ASSERT(where[i] == pid_from);
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (where[adjncy[j]] == pid_to) {
ind[nind++] = i;
break;
}
}
}
/* Go and construct the otherpmat to see where these nind vertices are
connected to */
iset(ncon, 0, cpwgt);
for (ncand=0, ii=0; ii<nind; ii++) {
i = ind[ii];
iaxpy(ncon, 1, vwgt+i*ncon, 1, cpwgt, 1);
for (j=xadj[i]; j<xadj[i+1]; j++) {
if ((k = where[adjncy[j]]) == pid_from)
continue;
if (otherpmat[k] == 0)
cand[ncand++].val = k;
otherpmat[k] += (adjwgt ? adjwgt[j] : 1);
}
}
for (i=0; i<ncand; i++) {
cand[i].key = otherpmat[cand[i].val];
ASSERT(cand[i].key > 0);
}
ikvsortd(ncand, cand);
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("\tMinOut: %4"PRIDX", to: %3"PRIDX", TtlWgt: %5"PRIDX"[#:%"PRIDX"]\n",
mypmat[other], other, isum(ncon, cpwgt, 1), nind));
/* Go through and select the first domain that is common with 'me', and does
not increase the nads[target] higher than nads[me], subject to the maxpwgt
constraint. Traversal is done from the mostly connected to the least. */
for (i=0; i<ncand; i++) {
k = cand[i].val;
if (mypmat[k] > 0) {
/* Check if balance will go off */
if (!ivecaxpylez(ncon, 1, cpwgt, pwgts+k*ncon, maxpwgt+k*ncon))
continue;
/* get a dense vector out of k's connectivity */
for (j=0; j<nads[k]; j++)
kpmat[adids[k][j]] = adwgts[k][j];
/* Check if the move to domain k will increase the nads of another
subdomain j that the set of vertices being moved are connected
to but domain k is not connected to. */
for (j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && kpmat[j] == 0 && nads[j]+1 >= nads[me])
break;
}
/* There were no bad second level effects. See if you can find a
subdomain to move to. */
if (j == nparts) {
for (nadd=0, j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && kpmat[j] == 0)
nadd++;
}
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("\t\tto=%"PRIDX", nadd=%"PRIDX", %"PRIDX"\n", k, nadd, nads[k]));
if (nads[k]+nadd < nads[me]) {
if (target2 == -1 || nads[target2]+bestnadd > nads[k]+nadd ||
(nads[target2]+bestnadd == nads[k]+nadd && bestnadd > nadd)) {
target2 = k;
bestnadd = nadd;
}
}
if (nadd == 0)
target = k;
}
/* reset kpmat for the next iteration */
for (j=0; j<nads[k]; j++)
kpmat[adids[k][j]] = 0;
}
if (target != -1)
break;
}
/* reset the otherpmat for the next iteration */
for (i=0; i<ncand; i++)
otherpmat[cand[i].val] = 0;
if (target == -1 && target2 != -1)
target = target2;
if (target != -1) {
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("\t\tScheme: %"PRIDX". Moving to %"PRIDX"\n", scheme, target));
move = 1;
break;
}
}
if (target != -1)
break; /* A move was found. No need to try the other scheme */
}
/* reset the mypmat for next iteration */
for (i=0; i<nads[me]; i++)
mypmat[adids[me][i]] = 0;
/* Note that once a target is found the above loops exit right away. So the
following variables are valid */
if (target != -1) {
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
MoveGroupMinConnForCut(ctrl, graph, target, nind, ind);
break;
case METIS_OBJTYPE_VOL:
MoveGroupMinConnForVol(ctrl, graph, target, nind, ind, vmarker,
pmarker, modind);
break;
default:
gk_errexit(SIGERR, "Unknown objtype of %d\n", ctrl->objtype);
}
/* Update the csr representation of the partitioning vector */
iarray2csr(nvtxs, nparts, where, pptr, pind);
}
}
if (move == 0)
break;
}
ipqFree(&queue);
WCOREPOP;
}
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 converts a mesh into a dual graph
**************************************************************************/
int ParMETIS_V3_Mesh2Dual(idx_t *elmdist, idx_t *eptr, idx_t *eind,
idx_t *numflag, idx_t *ncommon, idx_t **r_xadj,
idx_t **r_adjncy, MPI_Comm *comm)
{
idx_t i, j, jj, k, kk, m;
idx_t npes, mype, pe, count, mask, pass;
idx_t nelms, lnns, my_nns, node;
idx_t firstelm, firstnode, lnode, nrecv, nsend;
idx_t *scounts, *rcounts, *sdispl, *rdispl;
idx_t *nodedist, *nmap, *auxarray;
idx_t *gnptr, *gnind, *nptr, *nind, *myxadj=NULL, *myadjncy = NULL;
idx_t *sbuffer, *rbuffer, *htable;
ikv_t *nodelist, *recvbuffer;
idx_t maxcount, *ind, *wgt;
idx_t gmaxnode, gminnode;
size_t curmem;
gk_malloc_init();
curmem = gk_GetCurMemoryUsed();
/* Get basic comm info */
gkMPI_Comm_size(*comm, &npes);
gkMPI_Comm_rank(*comm, &mype);
nelms = elmdist[mype+1]-elmdist[mype];
if (*numflag > 0)
ChangeNumberingMesh(elmdist, eptr, eind, NULL, NULL, NULL, npes, mype, 1);
mask = (1<<11)-1;
/*****************************/
/* Determine number of nodes */
/*****************************/
gminnode = GlobalSEMinComm(*comm, imin(eptr[nelms], eind));
for (i=0; i<eptr[nelms]; i++)
eind[i] -= gminnode;
gmaxnode = GlobalSEMaxComm(*comm, imax(eptr[nelms], eind));
/**************************/
/* Check for input errors */
/**************************/
ASSERT(nelms > 0);
/* construct node distribution array */
nodedist = ismalloc(npes+1, 0, "nodedist");
for (nodedist[0]=0, i=0,j=gmaxnode+1; i<npes; i++) {
k = j/(npes-i);
nodedist[i+1] = nodedist[i]+k;
j -= k;
}
my_nns = nodedist[mype+1]-nodedist[mype];
firstnode = nodedist[mype];
nodelist = ikvmalloc(eptr[nelms], "nodelist");
auxarray = imalloc(eptr[nelms], "auxarray");
htable = ismalloc(gk_max(my_nns, mask+1), -1, "htable");
scounts = imalloc(npes, "scounts");
rcounts = imalloc(npes, "rcounts");
sdispl = imalloc(npes+1, "sdispl");
rdispl = imalloc(npes+1, "rdispl");
/*********************************************/
/* first find a local numbering of the nodes */
/*********************************************/
for (i=0; i<nelms; i++) {
for (j=eptr[i]; j<eptr[i+1]; j++) {
nodelist[j].key = eind[j];
nodelist[j].val = j;
auxarray[j] = i; /* remember the local element ID that uses this node */
}
}
ikvsorti(eptr[nelms], nodelist);
for (count=1, i=1; i<eptr[nelms]; i++) {
if (nodelist[i].key > nodelist[i-1].key)
count++;
}
lnns = count;
nmap = imalloc(lnns, "nmap");
/* renumber the nodes of the elements array */
count = 1;
nmap[0] = nodelist[0].key;
eind[nodelist[0].val] = 0;
nodelist[0].val = auxarray[nodelist[0].val]; /* Store the local element ID */
for (i=1; i<eptr[nelms]; i++) {
if (nodelist[i].key > nodelist[i-1].key) {
nmap[count] = nodelist[i].key;
count++;
}
eind[nodelist[i].val] = count-1;
nodelist[i].val = auxarray[nodelist[i].val]; /* Store the local element ID */
}
gkMPI_Barrier(*comm);
/**********************************************************/
/* perform comms necessary to construct node-element list */
/**********************************************************/
iset(npes, 0, scounts);
for (pe=i=0; i<eptr[nelms]; i++) {
while (nodelist[i].key >= nodedist[pe+1])
pe++;
scounts[pe] += 2;
}
ASSERT(pe < npes);
gkMPI_Alltoall((void *)scounts, 1, IDX_T, (void *)rcounts, 1, IDX_T, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
icopy(npes, rcounts, rdispl);
MAKECSR(i, npes, rdispl);
ASSERT(sdispl[npes] == eptr[nelms]*2);
nrecv = rdispl[npes]/2;
recvbuffer = ikvmalloc(gk_max(1, nrecv), "recvbuffer");
gkMPI_Alltoallv((void *)nodelist, scounts, sdispl, IDX_T, (void *)recvbuffer,
rcounts, rdispl, IDX_T, *comm);
/**************************************/
/* construct global node-element list */
/**************************************/
gnptr = ismalloc(my_nns+1, 0, "gnptr");
for (i=0; i<npes; i++) {
for (j=rdispl[i]/2; j<rdispl[i+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
ASSERT(lnode >= 0 && lnode < my_nns)
gnptr[lnode]++;
}
}
MAKECSR(i, my_nns, gnptr);
gnind = imalloc(gk_max(1, gnptr[my_nns]), "gnind");
for (pe=0; pe<npes; pe++) {
firstelm = elmdist[pe];
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
gnind[gnptr[lnode]++] = recvbuffer[j].val+firstelm;
}
}
SHIFTCSR(i, my_nns, gnptr);
/*********************************************************/
/* send the node-element info to the relevant processors */
/*********************************************************/
iset(npes, 0, scounts);
/* use a hash table to ensure that each node is sent to a proc only once */
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
if (htable[lnode] == -1) {
scounts[pe] += gnptr[lnode+1]-gnptr[lnode];
htable[lnode] = 1;
}
}
/* now reset the hash table */
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
htable[lnode] = -1;
}
}
gkMPI_Alltoall((void *)scounts, 1, IDX_T, (void *)rcounts, 1, IDX_T, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
/* create the send buffer */
nsend = sdispl[npes];
sbuffer = imalloc(gk_max(1, nsend), "sbuffer");
count = 0;
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
if (htable[lnode] == -1) {
for (k=gnptr[lnode]; k<gnptr[lnode+1]; k++) {
if (k == gnptr[lnode])
sbuffer[count++] = -1*(gnind[k]+1);
else
sbuffer[count++] = gnind[k];
}
htable[lnode] = 1;
}
}
ASSERT(count == sdispl[pe+1]);
/* now reset the hash table */
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
htable[lnode] = -1;
}
}
icopy(npes, rcounts, rdispl);
MAKECSR(i, npes, rdispl);
nrecv = rdispl[npes];
rbuffer = imalloc(gk_max(1, nrecv), "rbuffer");
gkMPI_Alltoallv((void *)sbuffer, scounts, sdispl, IDX_T, (void *)rbuffer,
rcounts, rdispl, IDX_T, *comm);
k = -1;
nptr = ismalloc(lnns+1, 0, "nptr");
nind = rbuffer;
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]; j<rdispl[pe+1]; j++) {
if (nind[j] < 0) {
k++;
nind[j] = (-1*nind[j])-1;
}
nptr[k]++;
}
}
MAKECSR(i, lnns, nptr);
ASSERT(k+1 == lnns);
ASSERT(nptr[lnns] == nrecv)
myxadj = *r_xadj = (idx_t *)malloc(sizeof(idx_t)*(nelms+1));
if (myxadj == NULL)
gk_errexit(SIGMEM, "Failed to allocate memory for the dual graph's xadj array.\n");
iset(nelms+1, 0, myxadj);
iset(mask+1, -1, htable);
firstelm = elmdist[mype];
/* Two passes -- in first pass, simply find out the memory requirements */
maxcount = 200;
ind = imalloc(maxcount, "ParMETIS_V3_Mesh2Dual: ind");
wgt = imalloc(maxcount, "ParMETIS_V3_Mesh2Dual: wgt");
for (pass=0; pass<2; pass++) {
for (i=0; i<nelms; i++) {
for (count=0, j=eptr[i]; j<eptr[i+1]; j++) {
node = eind[j];
for (k=nptr[node]; k<nptr[node+1]; k++) {
if ((kk=nind[k]) == firstelm+i)
continue;
m = htable[(kk&mask)];
if (m == -1) {
ind[count] = kk;
wgt[count] = 1;
htable[(kk&mask)] = count++;
}
else {
if (ind[m] == kk) {
wgt[m]++;
}
else {
for (jj=0; jj<count; jj++) {
if (ind[jj] == kk) {
wgt[jj]++;
break;
}
}
if (jj == count) {
ind[count] = kk;
wgt[count++] = 1;
}
}
}
/* Adjust the memory.
This will be replaced by a idxrealloc() when GKlib will be incorporated */
if (count == maxcount-1) {
maxcount *= 2;
ind = irealloc(ind, maxcount, "ParMETIS_V3_Mesh2Dual: ind");
wgt = irealloc(wgt, maxcount, "ParMETIS_V3_Mesh2Dual: wgt");
}
}
}
for (j=0; j<count; j++) {
htable[(ind[j]&mask)] = -1;
if (wgt[j] >= *ncommon) {
if (pass == 0)
myxadj[i]++;
else
myadjncy[myxadj[i]++] = ind[j];
}
}
}
if (pass == 0) {
MAKECSR(i, nelms, myxadj);
myadjncy = *r_adjncy = (idx_t *)malloc(sizeof(idx_t)*myxadj[nelms]);
if (myadjncy == NULL)
gk_errexit(SIGMEM, "Failed to allocate memory for dual graph's adjncy array.\n");
}
else {
SHIFTCSR(i, nelms, myxadj);
}
}
/*****************************************/
/* correctly renumber the elements array */
/*****************************************/
for (i=0; i<eptr[nelms]; i++)
eind[i] = nmap[eind[i]] + gminnode;
if (*numflag == 1)
ChangeNumberingMesh(elmdist, eptr, eind, myxadj, myadjncy, NULL, npes, mype, 0);
/* do not free nodelist, recvbuffer, rbuffer */
gk_free((void **)&nodedist, &nodelist, &auxarray, &htable, &scounts, &rcounts,
&sdispl, &rdispl, &nmap, &recvbuffer, &gnptr, &gnind, &sbuffer, &rbuffer,
&nptr, &ind, &wgt, LTERM);
if (gk_GetCurMemoryUsed() - curmem > 0) {
printf("ParMETIS appears to have a memory leak of %zdbytes. Report this.\n",
(ssize_t)(gk_GetCurMemoryUsed() - curmem));
}
gk_malloc_cleanup(0);
return METIS_OK;
}
