/*************************************************************************
* This function performs a k-way directed diffusion
**************************************************************************/
real_t WavefrontDiffusion(ctrl_t *ctrl, graph_t *graph, idx_t *home)
{
idx_t ii, i, j, k, l, nvtxs, nedges, nparts;
idx_t from, to, edge, done, nswaps, noswaps, totalv, wsize;
idx_t npasses, first, second, third, mind, maxd;
idx_t *xadj, *adjncy, *adjwgt, *where, *perm;
idx_t *rowptr, *colind, *ed, *psize;
real_t *transfer, *tmpvec;
real_t balance = -1.0, *load, *solution, *workspace;
real_t *nvwgt, *npwgts, flowFactor, cost, ubfactor;
matrix_t matrix;
ikv_t *cand;
idx_t ndirty, nclean, dptr, clean;
nvtxs = graph->nvtxs;
nedges = graph->nedges;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
nparts = ctrl->nparts;
ubfactor = ctrl->ubvec[0];
matrix.nrows = nparts;
flowFactor = 0.35;
flowFactor = (ctrl->mype == 2) ? 0.50 : flowFactor;
flowFactor = (ctrl->mype == 3) ? 0.75 : flowFactor;
flowFactor = (ctrl->mype == 4) ? 1.00 : flowFactor;
/* allocate memory */
solution = rmalloc(4*nparts+2*nedges, "WavefrontDiffusion: solution");
tmpvec = solution + nparts;
npwgts = solution + 2*nparts;
load = solution + 3*nparts;
matrix.values = solution + 4*nparts;
transfer = matrix.transfer = solution + 4*nparts + nedges;
perm = imalloc(2*nvtxs+2*nparts+nedges+1, "WavefrontDiffusion: perm");
ed = perm + nvtxs;
psize = perm + 2*nvtxs;
rowptr = matrix.rowptr = perm + 2*nvtxs + nparts;
colind = matrix.colind = perm + 2*nvtxs + 2*nparts + 1;
/*GKTODO - Potential problem with this malloc */
wsize = gk_max(sizeof(real_t)*nparts*6, sizeof(idx_t)*(nvtxs+nparts*2+1));
workspace = (real_t *)gk_malloc(wsize, "WavefrontDiffusion: workspace");
cand = ikvmalloc(nvtxs, "WavefrontDiffusion: cand");
/*****************************/
/* Populate empty subdomains */
/*****************************/
iset(nparts, 0, psize);
for (i=0; i<nvtxs; i++)
psize[where[i]]++;
mind = iargmin(nparts, psize);
maxd = iargmax(nparts, psize);
if (psize[mind] == 0) {
for (i=0; i<nvtxs; i++) {
k = (RandomInRange(nvtxs)+i)%nvtxs;
if (where[k] == maxd) {
where[k] = mind;
psize[mind]++;
psize[maxd]--;
break;
}
}
}
iset(nvtxs, 0, ed);
rset(nparts, 0.0, npwgts);
for (i=0; i<nvtxs; i++) {
npwgts[where[i]] += nvwgt[i];
for (j=xadj[i]; j<xadj[i+1]; j++)
ed[i] += (where[i] != where[adjncy[j]] ? adjwgt[j] : 0);
}
ComputeLoad(graph, nparts, load, ctrl->tpwgts, 0);
done = 0;
/* zero out the tmpvec array */
rset(nparts, 0.0, tmpvec);
npasses = gk_min(nparts/2, NGD_PASSES);
for (l=0; l<npasses; l++) {
/* Set-up and solve the diffusion equation */
nswaps = 0;
/************************/
/* Solve flow equations */
/************************/
SetUpConnectGraph(graph, &matrix, (idx_t *)workspace);
/* check for disconnected subdomains */
for(i=0; i<matrix.nrows; i++) {
if (matrix.rowptr[i]+1 == matrix.rowptr[i+1]) {
cost = (real_t)(ctrl->mype);
goto CleanUpAndExit;
}
}
ConjGrad2(&matrix, load, solution, 0.001, workspace);
ComputeTransferVector(1, &matrix, solution, transfer, 0);
GetThreeMax(nparts, load, &first, &second, &third);
if (l%3 == 0) {
FastRandomPermute(nvtxs, perm, 1);
}
else {
/*****************************/
/* move dirty vertices first */
/*****************************/
ndirty = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
ndirty++;
}
dptr = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
perm[dptr++] = i;
else
perm[ndirty++] = i;
}
PASSERT(ctrl, ndirty == nvtxs);
ndirty = dptr;
nclean = nvtxs-dptr;
FastRandomPermute(ndirty, perm, 0);
FastRandomPermute(nclean, perm+ndirty, 0);
}
if (ctrl->mype == 0) {
for (j=nvtxs, k=0, ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (ed[i] != 0) {
cand[k].key = -ed[i];
cand[k++].val = i;
}
else {
cand[--j].key = 0;
cand[j].val = i;
}
}
ikvsorti(k, cand);
}
for (ii=0; ii<nvtxs/3; ii++) {
i = (ctrl->mype == 0) ? cand[ii].val : perm[ii];
from = where[i];
/* don't move out the last vertex in a subdomain */
if (psize[from] == 1)
continue;
clean = (from == home[i]) ? 1 : 0;
/* only move from top three or dirty vertices */
if (from != first && from != second && from != third && clean)
continue;
/* Scatter the sparse transfer row into the dense tmpvec row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++)
tmpvec[colind[j]] = transfer[j];
for (j=xadj[i]; j<xadj[i+1]; j++) {
to = where[adjncy[j]];
if (from != to) {
if (tmpvec[to] > (flowFactor * nvwgt[i])) {
tmpvec[to] -= nvwgt[i];
INC_DEC(psize[to], psize[from], 1);
INC_DEC(npwgts[to], npwgts[from], nvwgt[i]);
INC_DEC(load[to], load[from], nvwgt[i]);
where[i] = to;
nswaps++;
/* Update external degrees */
ed[i] = 0;
for (k=xadj[i]; k<xadj[i+1]; k++) {
edge = adjncy[k];
ed[i] += (to != where[edge] ? adjwgt[k] : 0);
if (where[edge] == from)
ed[edge] += adjwgt[k];
if (where[edge] == to)
ed[edge] -= adjwgt[k];
}
break;
}
}
}
/* Gather the dense tmpvec row into the sparse transfer row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++) {
transfer[j] = tmpvec[colind[j]];
tmpvec[colind[j]] = 0.0;
}
ASSERT(fabs(rsum(nparts, tmpvec, 1)) < .0001)
}
if (l % 2 == 1) {
balance = rmax(nparts, npwgts)*nparts;
if (balance < ubfactor + 0.035)
done = 1;
if (GlobalSESum(ctrl, done) > 0)
break;
noswaps = (nswaps > 0) ? 0 : 1;
if (GlobalSESum(ctrl, noswaps) > ctrl->npes/2)
break;
}
}
graph->mincut = ComputeSerialEdgeCut(graph);
totalv = Mc_ComputeSerialTotalV(graph, home);
cost = ctrl->ipc_factor * (real_t)graph->mincut + ctrl->redist_factor * (real_t)totalv;
CleanUpAndExit:
gk_free((void **)&solution, (void **)&perm, (void **)&workspace, (void **)&cand, LTERM);
return cost;
}
/*************************************************************************
* This function is the entry point of the initial partitioning algorithm.
* This algorithm assembles the graph to all the processors and preceed
* serially.
**************************************************************************/
idx_t Mc_Diffusion(ctrl_t *ctrl, graph_t *graph, idx_t *vtxdist, idx_t *where,
idx_t *home, idx_t npasses)
{
idx_t h, i, j;
idx_t nvtxs, nedges, ncon, pass, iter, domain, processor;
idx_t nparts, mype, npes, nlinks, me, you, wsize;
idx_t nvisited, nswaps = -1, tnswaps, done, alldone = -1;
idx_t *rowptr, *colind, *diff_where, *sr_where, *ehome, *map, *rmap;
idx_t *pack, *unpack, *match, *proc2sub, *sub2proc;
idx_t *visited, *gvisited;
real_t *transfer, *npwgts, maxdiff, minflow, maxflow;
real_t lbavg, oldlbavg, ubavg, *lbvec;
real_t *diff_flows, *sr_flows;
real_t diff_lbavg, sr_lbavg, diff_cost, sr_cost;
idx_t *rbuffer, *sbuffer;
idx_t *rcount, *rdispl;
real_t *solution, *load, *workspace;
matrix_t matrix;
graph_t *egraph;
if (graph->ncon > 3)
return 0;
WCOREPUSH;
nvtxs = graph->nvtxs;
nedges = graph->nedges;
ncon = graph->ncon;
nparts = ctrl->nparts;
mype = ctrl->mype;
npes = ctrl->npes;
ubavg = ravg(ncon, ctrl->ubvec);
/* initialize variables and allocate memory */
lbvec = rwspacemalloc(ctrl, ncon);
diff_flows = rwspacemalloc(ctrl, ncon);
sr_flows = rwspacemalloc(ctrl, ncon);
load = rwspacemalloc(ctrl, nparts);
solution = rwspacemalloc(ctrl, nparts);
npwgts = graph->gnpwgts = rwspacemalloc(ctrl, ncon*nparts);
matrix.values = rwspacemalloc(ctrl, nedges);
transfer = matrix.transfer = rwspacemalloc(ctrl, ncon*nedges);
proc2sub = iwspacemalloc(ctrl, gk_max(nparts, npes*2));
sub2proc = iwspacemalloc(ctrl, nparts);
match = iwspacemalloc(ctrl, nparts);
rowptr = matrix.rowptr = iwspacemalloc(ctrl, nparts+1);
colind = matrix.colind = iwspacemalloc(ctrl, nedges);
rcount = iwspacemalloc(ctrl, npes);
rdispl = iwspacemalloc(ctrl, npes+1);
pack = iwspacemalloc(ctrl, nvtxs);
unpack = iwspacemalloc(ctrl, nvtxs);
rbuffer = iwspacemalloc(ctrl, nvtxs);
sbuffer = iwspacemalloc(ctrl, nvtxs);
map = iwspacemalloc(ctrl, nvtxs);
rmap = iwspacemalloc(ctrl, nvtxs);
diff_where = iwspacemalloc(ctrl, nvtxs);
ehome = iwspacemalloc(ctrl, nvtxs);
wsize = gk_max(sizeof(real_t)*nparts*6, sizeof(idx_t)*(nvtxs+nparts*2+1));
workspace = (real_t *)gk_malloc(wsize, "Mc_Diffusion: workspace");
graph->ckrinfo = (ckrinfo_t *)gk_malloc(nvtxs*sizeof(ckrinfo_t), "Mc_Diffusion: rinfo");
/* construct subdomain connectivity graph */
matrix.nrows = nparts;
SetUpConnectGraph(graph, &matrix, (idx_t *)workspace);
nlinks = (matrix.nnzs-nparts) / 2;
visited = iwspacemalloc(ctrl, matrix.nnzs);
gvisited = iwspacemalloc(ctrl, matrix.nnzs);
for (pass=0; pass<npasses; pass++) {
rset(matrix.nnzs*ncon, 0.0, transfer);
iset(matrix.nnzs, 0, gvisited);
iset(matrix.nnzs, 0, visited);
iter = nvisited = 0;
/* compute ncon flow solutions */
for (h=0; h<ncon; h++) {
rset(nparts, 0.0, solution);
ComputeLoad(graph, nparts, load, ctrl->tpwgts, h);
lbvec[h] = (rmax(nparts, load)+1.0/nparts) * (real_t)nparts;
ConjGrad2(&matrix, load, solution, 0.001, workspace);
ComputeTransferVector(ncon, &matrix, solution, transfer, h);
}
oldlbavg = ravg(ncon, lbvec);
tnswaps = 0;
maxdiff = 0.0;
for (i=0; i<nparts; i++) {
for (j=rowptr[i]; j<rowptr[i+1]; j++) {
maxflow = rmax(ncon, transfer+j*ncon);
minflow = rmin(ncon, transfer+j*ncon);
maxdiff = (maxflow - minflow > maxdiff) ? maxflow - minflow : maxdiff;
}
}
while (nvisited < nlinks) {
/* compute independent sets of subdomains */
iset(gk_max(nparts, npes*2), UNMATCHED, proc2sub);
CSR_Match_SHEM(&matrix, match, proc2sub, gvisited, ncon);
/* set up the packing arrays */
iset(nparts, UNMATCHED, sub2proc);
for (i=0; i<npes*2; i++) {
if (proc2sub[i] == UNMATCHED)
break;
sub2proc[proc2sub[i]] = i/2;
}
iset(npes, 0, rcount);
for (i=0; i<nvtxs; i++) {
domain = where[i];
processor = sub2proc[domain];
if (processor != UNMATCHED)
rcount[processor]++;
}
rdispl[0] = 0;
for (i=1; i<npes+1; i++)
rdispl[i] = rdispl[i-1] + rcount[i-1];
iset(nvtxs, UNMATCHED, unpack);
for (i=0; i<nvtxs; i++) {
domain = where[i];
processor = sub2proc[domain];
if (processor != UNMATCHED)
unpack[rdispl[processor]++] = i;
}
SHIFTCSR(i, npes, rdispl);
iset(nvtxs, UNMATCHED, pack);
for (i=0; i<rdispl[npes]; i++) {
ASSERT(unpack[i] != UNMATCHED);
domain = where[unpack[i]];
processor = sub2proc[domain];
if (processor != UNMATCHED)
pack[unpack[i]] = i;
}
/* Compute the flows */
if (proc2sub[mype*2] != UNMATCHED) {
me = proc2sub[2*mype];
you = proc2sub[2*mype+1];
ASSERT(me != you);
for (j=rowptr[me]; j<rowptr[me+1]; j++) {
if (colind[j] == you) {
visited[j] = 1;
rcopy(ncon, transfer+j*ncon, diff_flows);
break;
}
}
for (j=rowptr[you]; j<rowptr[you+1]; j++) {
if (colind[j] == me) {
visited[j] = 1;
for (h=0; h<ncon; h++) {
if (transfer[j*ncon+h] > 0.0)
diff_flows[h] = -1.0 * transfer[j*ncon+h];
}
break;
}
}
nswaps = 1;
rcopy(ncon, diff_flows, sr_flows);
iset(nvtxs, 0, sbuffer);
for (i=0; i<nvtxs; i++) {
if (where[i] == me || where[i] == you)
sbuffer[i] = 1;
}
egraph = ExtractGraph(ctrl, graph, sbuffer, map, rmap);
if (egraph != NULL) {
icopy(egraph->nvtxs, egraph->where, diff_where);
for (j=0; j<egraph->nvtxs; j++)
ehome[j] = home[map[j]];
RedoMyLink(ctrl, egraph, ehome, me, you, sr_flows, &sr_cost, &sr_lbavg);
if (ncon <= 4) {
sr_where = egraph->where;
egraph->where = diff_where;
nswaps = BalanceMyLink(ctrl, egraph, ehome, me, you, diff_flows, maxdiff,
&diff_cost, &diff_lbavg, 1.0/(real_t)nvtxs);
if ((sr_lbavg < diff_lbavg &&
(diff_lbavg >= ubavg-1.0 || sr_cost == diff_cost)) ||
(sr_lbavg < ubavg-1.0 && sr_cost < diff_cost)) {
for (i=0; i<egraph->nvtxs; i++)
where[map[i]] = sr_where[i];
}
else {
for (i=0; i<egraph->nvtxs; i++)
where[map[i]] = diff_where[i];
}
}
else {
for (i=0; i<egraph->nvtxs; i++)
where[map[i]] = egraph->where[i];
}
gk_free((void **)&egraph->xadj, &egraph->nvwgt, &egraph->adjncy, &egraph, LTERM);
}
/* Pack the flow data */
iset(nvtxs, UNMATCHED, sbuffer);
for (i=0; i<nvtxs; i++) {
domain = where[i];
if (domain == you || domain == me)
sbuffer[pack[i]] = where[i];
}
}
/* Broadcast the flow data */
gkMPI_Allgatherv((void *)&sbuffer[rdispl[mype]], rcount[mype], IDX_T,
(void *)rbuffer, rcount, rdispl, IDX_T, ctrl->comm);
/* Unpack the flow data */
for (i=0; i<rdispl[npes]; i++) {
if (rbuffer[i] != UNMATCHED)
where[unpack[i]] = rbuffer[i];
}
/* Do other stuff */
gkMPI_Allreduce((void *)visited, (void *)gvisited, matrix.nnzs,
IDX_T, MPI_MAX, ctrl->comm);
nvisited = isum(matrix.nnzs, gvisited, 1)/2;
tnswaps += GlobalSESum(ctrl, nswaps);
if (iter++ == NGD_PASSES)
break;
}
/* perform serial refinement */
Mc_ComputeSerialPartitionParams(ctrl, graph, nparts);
Mc_SerialKWayAdaptRefine(ctrl, graph, nparts, home, ctrl->ubvec, 10);
/* check for early breakout */
for (h=0; h<ncon; h++) {
lbvec[h] = (real_t)(nparts) *
npwgts[rargmax_strd(nparts,npwgts+h,ncon)*ncon+h];
}
lbavg = ravg(ncon, lbvec);
done = 0;
if (tnswaps == 0 || lbavg >= oldlbavg || lbavg <= ubavg + 0.035)
done = 1;
alldone = GlobalSEMax(ctrl, done);
if (alldone == 1)
break;
}
/* ensure that all subdomains have at least one vertex */
/*
iset(nparts, 0, match);
for (i=0; i<nvtxs; i++)
match[where[i]]++;
done = 0;
while (done == 0) {
done = 1;
me = iargmin(nparts, match);
if (match[me] == 0) {
if (ctrl->mype == PE) printf("WARNING: empty subdomain %"PRIDX" in Mc_Diffusion\n", me);
you = iargmax(nparts, match);
for (i=0; i<nvtxs; i++) {
if (where[i] == you) {
where[i] = me;
match[you]--;
match[me]++;
done = 0;
break;
}
}
}
}
*/
/* now free memory and return */
gk_free((void **)&workspace, (void **)&graph->ckrinfo, LTERM);
graph->gnpwgts = NULL;
graph->ckrinfo = NULL;
WCOREPOP;
return 0;
}
