/*************************************************************************
* 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 CreateCoarseGraph_Global(ctrl_t *ctrl, graph_t *graph, idx_t cnvtxs)
{
idx_t h, i, j, k, l, ii, jj, ll, nnbrs, nvtxs, nedges, ncon;
idx_t firstvtx, lastvtx, cfirstvtx, clastvtx, otherlastvtx;
idx_t npes=ctrl->npes, mype=ctrl->mype;
idx_t cnedges, nsend, nrecv, nkeepsize, nrecvsize, nsendsize, v, u;
idx_t *xadj, *adjncy, *adjwgt, *vwgt, *vsize, *vtxdist, *home, *where;
idx_t *match, *cmap;
idx_t *cxadj, *cadjncy, *cadjwgt, *cvwgt, *cvsize = NULL, *chome = NULL,
*cwhere = NULL, *cvtxdist;
idx_t *rsizes, *ssizes, *rlens, *slens, *rgraph, *sgraph, *perm;
idx_t *peind, *recvptr, *recvind;
real_t *nvwgt, *cnvwgt;
graph_t *cgraph;
ikv_t *scand, *rcand;
idx_t htable[LHTSIZE], htableidx[LHTSIZE];
WCOREPUSH;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
vwgt = graph->vwgt;
vsize = graph->vsize;
nvwgt = graph->nvwgt;
home = graph->home;
where = graph->where;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
match = graph->match;
vtxdist = graph->vtxdist;
firstvtx = vtxdist[mype];
lastvtx = vtxdist[mype+1];
cmap = graph->cmap = ismalloc(nvtxs+graph->nrecv, -1, "Global_CreateCoarseGraph: cmap");
nnbrs = graph->nnbrs;
peind = graph->peind;
recvind = graph->recvind;
recvptr = graph->recvptr;
/* Initialize the coarser graph */
cgraph = CreateGraph();
cgraph->nvtxs = cnvtxs;
cgraph->ncon = ncon;
cgraph->level = graph->level+1;
cgraph->finer = graph;
graph->coarser = cgraph;
/*************************************************************
* Obtain the vtxdist of the coarser graph
**************************************************************/
cvtxdist = cgraph->vtxdist = imalloc(npes+1, "Global_CreateCoarseGraph: cvtxdist");
cvtxdist[npes] = cnvtxs; /* Use last position in the cvtxdist as a temp buffer */
gkMPI_Allgather((void *)(cvtxdist+npes), 1, IDX_T, (void *)cvtxdist, 1,
IDX_T, ctrl->comm);
MAKECSR(i, npes, cvtxdist);
cgraph->gnvtxs = cvtxdist[npes];
#ifdef DEBUG_CONTRACT
PrintVector(ctrl, npes+1, 0, cvtxdist, "cvtxdist");
#endif
/*************************************************************
* Construct the cmap vector
**************************************************************/
cfirstvtx = cvtxdist[mype];
clastvtx = cvtxdist[mype+1];
/* Create the cmap of what you know so far locally. */
for (cnvtxs=0, i=0; i<nvtxs; i++) {
if (match[i] >= KEEP_BIT) {
k = match[i] - KEEP_BIT;
if (k>=firstvtx && k<firstvtx+i)
continue; /* Both (i,k) are local and i has been matched via the (k,i) side */
cmap[i] = cfirstvtx + cnvtxs++;
if (k != firstvtx+i && (k>=firstvtx && k<lastvtx)) { /* I'm matched locally */
cmap[k-firstvtx] = cmap[i];
match[k-firstvtx] += KEEP_BIT; /* Add the KEEP_BIT to simplify coding */
}
}
}
PASSERT(ctrl, cnvtxs == clastvtx-cfirstvtx);
CommInterfaceData(ctrl, graph, cmap, cmap+nvtxs);
/* Update the cmap of the locally stored vertices that will go away.
* The remote processor assigned cmap for them */
for (i=0; i<nvtxs; i++) {
if (match[i] < KEEP_BIT) { /* Only vertices that go away satisfy this*/
cmap[i] = cmap[nvtxs+BSearch(graph->nrecv, recvind, match[i])];
}
}
CommInterfaceData(ctrl, graph, cmap, cmap+nvtxs);
#ifndef NDEBUG
for (i=0; i<nvtxs+graph->nrecv; i++) {
if (cmap[i] == -1)
errexit("cmap[%"PRIDX"] == -1\n", i);
}
#endif
#ifdef DEBUG_CONTRACT
PrintVector(ctrl, nvtxs, firstvtx, cmap, "Cmap");
#endif
/*************************************************************
* Determine how many adjcency lists you need to send/receive.
**************************************************************/
/* first pass: determine sizes */
for (nsend=0, nrecv=0, i=0; i<nvtxs; i++) {
if (match[i] < KEEP_BIT) /* This is going away */
nsend++;
else {
k = match[i]-KEEP_BIT;
if (k<firstvtx || k>=lastvtx) /* This is comming from afar */
nrecv++;
}
}
scand = ikvwspacemalloc(ctrl, nsend);
rcand = graph->rcand = ikvmalloc(nrecv, "CreateCoarseGraph: rcand");
/* second pass: place them in the appropriate arrays */
nkeepsize = nsend = nrecv = 0;
for (i=0; i<nvtxs; i++) {
if (match[i] < KEEP_BIT) { /* This is going away */
scand[nsend].key = match[i];
scand[nsend].val = i;
nsend++;
}
else {
nkeepsize += (xadj[i+1]-xadj[i]);
k = match[i]-KEEP_BIT;
if (k<firstvtx || k>=lastvtx) { /* This is comming from afar */
rcand[nrecv].key = k;
rcand[nrecv].val = cmap[i] - cfirstvtx; /* Set it for use during the partition projection */
PASSERT(ctrl, rcand[nrecv].val>=0 && rcand[nrecv].val<cnvtxs);
nrecv++;
}
}
}
#ifdef DEBUG_CONTRACT
PrintPairs(ctrl, nsend, scand, "scand");
PrintPairs(ctrl, nrecv, rcand, "rcand");
#endif
/***************************************************************
* Determine how many lists and their sizes you will send and
* received for each of the neighboring PEs
****************************************************************/
rlens = graph->rlens = imalloc(nnbrs+1, "CreateCoarseGraph: graph->rlens");
slens = graph->slens = imalloc(nnbrs+1, "CreateCoarseGraph: graph->slens");
rsizes = iset(nnbrs, 0, iwspacemalloc(ctrl, nnbrs));
ssizes = iset(nnbrs, 0, iwspacemalloc(ctrl, nnbrs));
/* Take care the sending data first */
ikvsortii(nsend, scand);
slens[0] = 0;
for (k=i=0; i<nnbrs; i++) {
otherlastvtx = vtxdist[peind[i]+1];
for (; k<nsend && scand[k].key < otherlastvtx; k++)
ssizes[i] += (xadj[scand[k].val+1]-xadj[scand[k].val]);
slens[i+1] = k;
}
/* Take care the receiving data next. You cannot yet determine the rsizes[] */
ikvsortii(nrecv, rcand);
rlens[0] = 0;
for (k=i=0; i<nnbrs; i++) {
otherlastvtx = vtxdist[peind[i]+1];
for (; k<nrecv && rcand[k].key < otherlastvtx; k++);
rlens[i+1] = k;
}
#ifdef DEBUG_CONTRACT
PrintVector(ctrl, nnbrs+1, 0, slens, "slens");
PrintVector(ctrl, nnbrs+1, 0, rlens, "rlens");
#endif
/***************************************************************
* Exchange size information
****************************************************************/
/* Issue the receives first. */
for (i=0; i<nnbrs; i++) {
if (rlens[i+1]-rlens[i] > 0) /* Issue a receive only if you are getting something */
gkMPI_Irecv((void *)(rsizes+i), 1, IDX_T, peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Take care the sending data next */
for (i=0; i<nnbrs; i++) {
if (slens[i+1]-slens[i] > 0) /* Issue a send only if you are sending something */
gkMPI_Isend((void *)(ssizes+i), 1, IDX_T, peind[i], 1, ctrl->comm, ctrl->sreq+i);
}
/* OK, now get into the loop waiting for the operations to finish */
for (i=0; i<nnbrs; i++) {
if (rlens[i+1]-rlens[i] > 0)
gkMPI_Wait(ctrl->rreq+i, &ctrl->status);
}
for (i=0; i<nnbrs; i++) {
if (slens[i+1]-slens[i] > 0)
gkMPI_Wait(ctrl->sreq+i, &ctrl->status);
}
#ifdef DEBUG_CONTRACT
PrintVector(ctrl, nnbrs, 0, rsizes, "rsizes");
PrintVector(ctrl, nnbrs, 0, ssizes, "ssizes");
#endif
/*************************************************************
* Allocate memory for received/sent graphs and start sending
* and receiving data.
* rgraph and sgraph is a different data structure than CSR
* to facilitate single message exchange.
**************************************************************/
nrecvsize = isum(nnbrs, rsizes, 1);
nsendsize = isum(nnbrs, ssizes, 1);
rgraph = iwspacemalloc(ctrl, (4+ncon)*nrecv+2*nrecvsize);
WCOREPUSH; /* for freeing sgraph right away */
sgraph = iwspacemalloc(ctrl, (4+ncon)*nsend+2*nsendsize);
/* Deal with the received portion first */
for (l=i=0; i<nnbrs; i++) {
/* Issue a receive only if you are getting something */
if (rlens[i+1]-rlens[i] > 0) {
gkMPI_Irecv((void *)(rgraph+l), (4+ncon)*(rlens[i+1]-rlens[i])+2*rsizes[i],
IDX_T, peind[i], 1, ctrl->comm, ctrl->rreq+i);
l += (4+ncon)*(rlens[i+1]-rlens[i])+2*rsizes[i];
}
}
/* Deal with the sent portion now */
for (ll=l=i=0; i<nnbrs; i++) {
if (slens[i+1]-slens[i] > 0) { /* Issue a send only if you are sending something */
for (k=slens[i]; k<slens[i+1]; k++) {
ii = scand[k].val;
sgraph[ll++] = firstvtx+ii;
sgraph[ll++] = xadj[ii+1]-xadj[ii];
for (h=0; h<ncon; h++)
sgraph[ll++] = vwgt[ii*ncon+h];
sgraph[ll++] = (ctrl->partType == STATIC_PARTITION || ctrl->partType == ORDER_PARTITION
? -1 : vsize[ii]);
sgraph[ll++] = (ctrl->partType == STATIC_PARTITION || ctrl->partType == ORDER_PARTITION
? -1 : home[ii]);
for (jj=xadj[ii]; jj<xadj[ii+1]; jj++) {
sgraph[ll++] = cmap[adjncy[jj]];
sgraph[ll++] = adjwgt[jj];
}
}
PASSERT(ctrl, ll-l == (4+ncon)*(slens[i+1]-slens[i])+2*ssizes[i]);
/*myprintf(ctrl, "Sending to pe:%"PRIDX", %"PRIDX" lists of size %"PRIDX"\n", peind[i], slens[i+1]-slens[i], ssizes[i]); */
gkMPI_Isend((void *)(sgraph+l), ll-l, IDX_T, peind[i], 1, ctrl->comm, ctrl->sreq+i);
l = ll;
}
}
/* OK, now get into the loop waiting for the operations to finish */
for (i=0; i<nnbrs; i++) {
if (rlens[i+1]-rlens[i] > 0)
gkMPI_Wait(ctrl->rreq+i, &ctrl->status);
}
for (i=0; i<nnbrs; i++) {
if (slens[i+1]-slens[i] > 0)
gkMPI_Wait(ctrl->sreq+i, &ctrl->status);
}
#ifdef DEBUG_CONTRACT
rprintf(ctrl, "Graphs were sent!\n");
PrintTransferedGraphs(ctrl, nnbrs, peind, slens, rlens, sgraph, rgraph);
#endif
WCOREPOP; /* free sgraph */
/*************************************************************
* Setup the mapping from indices returned by BSearch to
* those that are actually stored
**************************************************************/
perm = iset(graph->nrecv, -1, iwspacemalloc(ctrl, graph->nrecv));
for (j=i=0; i<nrecv; i++) {
perm[BSearch(graph->nrecv, recvind, rgraph[j])] = j+1;
j += (4+ncon)+2*rgraph[j+1];
}
/*************************************************************
* Finally, create the coarser graph
**************************************************************/
/* Allocate memory for the coarser graph, and fire up coarsening */
cxadj = cgraph->xadj = imalloc(cnvtxs+1, "CreateCoarserGraph: cxadj");
cvwgt = cgraph->vwgt = imalloc(cnvtxs*ncon, "CreateCoarserGraph: cvwgt");
cnvwgt = cgraph->nvwgt = rmalloc(cnvtxs*ncon, "CreateCoarserGraph: cnvwgt");
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) {
cvsize = cgraph->vsize = imalloc(cnvtxs, "CreateCoarserGraph: cvsize");
chome = cgraph->home = imalloc(cnvtxs, "CreateCoarserGraph: chome");
}
if (where != NULL)
cwhere = cgraph->where = imalloc(cnvtxs, "CreateCoarserGraph: cwhere");
/* these are just upper bound estimates for now */
cadjncy = iwspacemalloc(ctrl, nkeepsize+nrecvsize);
cadjwgt = iwspacemalloc(ctrl, nkeepsize+nrecvsize);
iset(LHTSIZE, -1, htable);
cxadj[0] = cnvtxs = cnedges = 0;
for (i=0; i<nvtxs; i++) {
if (match[i] >= KEEP_BIT) {
v = firstvtx+i;
u = match[i]-KEEP_BIT;
if (u>=firstvtx && u<lastvtx && v > u)
continue; /* I have already collapsed it as (u,v) */
/* Collapse the v vertex first, which you know is local */
for (h=0; h<ncon; h++)
cvwgt[cnvtxs*ncon+h] = vwgt[i*ncon+h];
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) {
cvsize[cnvtxs] = vsize[i];
chome[cnvtxs] = home[i];
}
if (where != NULL)
cwhere[cnvtxs] = where[i];
nedges = 0;
/* Collapse the v (i) vertex first */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = cmap[adjncy[j]];
if (k < 0)
printf("k=%d\n", (int)k);
if (k != cfirstvtx+cnvtxs) { /* If this is not an internal edge */
l = k&MASK;
if (htable[l] == -1) { /* Seeing this for first time */
htable[l] = k;
htableidx[l] = cnedges+nedges;
cadjncy[cnedges+nedges] = k;
cadjwgt[cnedges+nedges++] = adjwgt[j];
}
else if (htable[l] == k) {
cadjwgt[htableidx[l]] += adjwgt[j];
}
else { /* Now you have to go and do a search. Expensive case */
for (l=0; l<nedges; l++) {
if (cadjncy[cnedges+l] == k)
break;
}
if (l < nedges) {
cadjwgt[cnedges+l] += adjwgt[j];
}
else {
cadjncy[cnedges+nedges] = k;
cadjwgt[cnedges+nedges++] = adjwgt[j];
}
}
}
}
/* Collapse the u vertex next */
if (v != u) {
if (u>=firstvtx && u<lastvtx) { /* Local vertex */
u -= firstvtx;
for (h=0; h<ncon; h++)
cvwgt[cnvtxs*ncon+h] += vwgt[u*ncon+h];
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION)
cvsize[cnvtxs] += vsize[u];
for (j=xadj[u]; j<xadj[u+1]; j++) {
k = cmap[adjncy[j]];
if (k != cfirstvtx+cnvtxs) { /* If this is not an internal edge */
l = k&MASK;
if (htable[l] == -1) { /* Seeing this for first time */
htable[l] = k;
htableidx[l] = cnedges+nedges;
cadjncy[cnedges+nedges] = k;
cadjwgt[cnedges+nedges] = adjwgt[j];
nedges++;
}
else if (htable[l] == k) {
cadjwgt[htableidx[l]] += adjwgt[j];
}
else { /* Now you have to go and do a search. Expensive case */
for (l=0; l<nedges; l++) {
if (cadjncy[cnedges+l] == k)
break;
}
if (l < nedges) {
cadjwgt[cnedges+l] += adjwgt[j];
}
else {
cadjncy[cnedges+nedges] = k;
cadjwgt[cnedges+nedges] = adjwgt[j];
nedges++;
}
}
}
}
}
else { /* Remote vertex */
u = perm[BSearch(graph->nrecv, recvind, u)];
for (h=0; h<ncon; h++)
/* Remember that the +1 stores the vertex weight */
cvwgt[cnvtxs*ncon+h] += rgraph[(u+1)+h];
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) {
cvsize[cnvtxs] += rgraph[u+1+ncon];
chome[cnvtxs] = rgraph[u+2+ncon];
}
for (j=0; j<rgraph[u]; j++) {
k = rgraph[u+3+ncon+2*j];
if (k != cfirstvtx+cnvtxs) { /* If this is not an internal edge */
l = k&MASK;
if (htable[l] == -1) { /* Seeing this for first time */
htable[l] = k;
htableidx[l] = cnedges+nedges;
cadjncy[cnedges+nedges] = k;
cadjwgt[cnedges+nedges] = rgraph[u+3+ncon+2*j+1];
nedges++;
}
else if (htable[l] == k) {
cadjwgt[htableidx[l]] += rgraph[u+3+ncon+2*j+1];
}
else { /* Now you have to go and do a search. Expensive case */
for (l=0; l<nedges; l++) {
if (cadjncy[cnedges+l] == k)
break;
}
if (l < nedges) {
cadjwgt[cnedges+l] += rgraph[u+3+ncon+2*j+1];
}
else {
cadjncy[cnedges+nedges] = k;
cadjwgt[cnedges+nedges] = rgraph[u+3+ncon+2*j+1];
nedges++;
}
}
}
}
}
}
cnedges += nedges;
for (j=cxadj[cnvtxs]; j<cnedges; j++)
htable[cadjncy[j]&MASK] = -1; /* reset the htable */
cxadj[++cnvtxs] = cnedges;
}
}
cgraph->nedges = cnedges;
/* ADD: In order to keep from having to change this too much */
/* ADD: I kept vwgt array and recomputed nvwgt for each coarser graph */
for (j=0; j<cnvtxs; j++) {
for (h=0; h<ncon; h++)
cgraph->nvwgt[j*ncon+h] = ctrl->invtvwgts[h]*cvwgt[j*ncon+h];
}
cgraph->adjncy = imalloc(cnedges, "CreateCoarserGraph: cadjncy");
cgraph->adjwgt = imalloc(cnedges, "CreateCoarserGraph: cadjwgt");
icopy(cnedges, cadjncy, cgraph->adjncy);
icopy(cnedges, cadjwgt, cgraph->adjwgt);
WCOREPOP;
/* Note that graph->where works fine even if it is NULL */
gk_free((void **)&graph->where, LTERM);
}
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;
}
/*************************************************************************
* 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;
}
