/*************************************************************************
* This function is the entry point of the initial partition algorithm
* that does recursive bissection.
* This algorithm assembles the graph to all the processors and preceeds
* by parallelizing the recursive bisection step.
**************************************************************************/
void InitPartition(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, j, ncon, mype, npes, gnvtxs, ngroups;
idx_t *xadj, *adjncy, *adjwgt, *vwgt;
idx_t *part, *gwhere0, *gwhere1;
idx_t *tmpwhere, *tmpvwgt, *tmpxadj, *tmpadjncy, *tmpadjwgt;
graph_t *agraph;
idx_t lnparts, fpart, fpe, lnpes;
idx_t twoparts=2, moptions[METIS_NOPTIONS], edgecut, max_cut;
real_t *tpwgts, *tpwgts2, *lbvec, lbsum, min_lbsum, wsum;
MPI_Comm ipcomm;
struct {
double sum;
int rank;
} lpesum, gpesum;
WCOREPUSH;
ncon = graph->ncon;
ngroups = gk_max(gk_min(RIP_SPLIT_FACTOR, ctrl->npes), 1);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
lbvec = rwspacemalloc(ctrl, ncon);
/* assemble the graph to all the processors */
agraph = AssembleAdaptiveGraph(ctrl, graph);
gnvtxs = agraph->nvtxs;
/* make a copy of the graph's structure for later */
xadj = icopy(gnvtxs+1, agraph->xadj, iwspacemalloc(ctrl, gnvtxs+1));
vwgt = icopy(gnvtxs*ncon, agraph->vwgt, iwspacemalloc(ctrl, gnvtxs*ncon));
adjncy = icopy(agraph->nedges, agraph->adjncy, iwspacemalloc(ctrl, agraph->nedges));
adjwgt = icopy(agraph->nedges, agraph->adjwgt, iwspacemalloc(ctrl, agraph->nedges));
part = iwspacemalloc(ctrl, gnvtxs);
/* create different processor groups */
gkMPI_Comm_split(ctrl->gcomm, ctrl->mype % ngroups, 0, &ipcomm);
gkMPI_Comm_rank(ipcomm, &mype);
gkMPI_Comm_size(ipcomm, &npes);
/* Go into the recursive bisection */
METIS_SetDefaultOptions(moptions);
moptions[METIS_OPTION_SEED] = ctrl->sync + (ctrl->mype % ngroups) + 1;
tpwgts = ctrl->tpwgts;
tpwgts2 = rwspacemalloc(ctrl, 2*ncon);
lnparts = ctrl->nparts;
fpart = fpe = 0;
lnpes = npes;
while (lnpes > 1 && lnparts > 1) {
/* determine the weights of the two partitions as a function of the
weight of the target partition weights */
for (j=(lnparts>>1), i=0; i<ncon; i++) {
tpwgts2[i] = rsum(j, tpwgts+fpart*ncon+i, ncon);
tpwgts2[ncon+i] = rsum(lnparts-j, tpwgts+(fpart+j)*ncon+i, ncon);
wsum = 1.0/(tpwgts2[i] + tpwgts2[ncon+i]);
tpwgts2[i] *= wsum;
tpwgts2[ncon+i] *= wsum;
}
METIS_PartGraphRecursive(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, NULL, agraph->adjwgt, &twoparts, tpwgts2, NULL, moptions,
&edgecut, part);
/* pick one of the branches */
if (mype < fpe+lnpes/2) {
KeepPart(ctrl, agraph, part, 0);
lnpes = lnpes/2;
lnparts = lnparts/2;
}
else {
KeepPart(ctrl, agraph, part, 1);
fpart = fpart + lnparts/2;
fpe = fpe + lnpes/2;
lnpes = lnpes - lnpes/2;
lnparts = lnparts - lnparts/2;
}
}
gwhere0 = iset(gnvtxs, 0, iwspacemalloc(ctrl, gnvtxs));
gwhere1 = iwspacemalloc(ctrl, gnvtxs);
if (lnparts == 1) { /* Case npes is greater than or equal to nparts */
/* Only the first process will assign labels (for the reduction to work) */
if (mype == fpe) {
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart;
}
}
else { /* Case in which npes is smaller than nparts */
/* create the normalized tpwgts for the lnparts from ctrl->tpwgts */
tpwgts = rwspacemalloc(ctrl, lnparts*ncon);
for (j=0; j<ncon; j++) {
for (wsum=0.0, i=0; i<lnparts; i++) {
tpwgts[i*ncon+j] = ctrl->tpwgts[(fpart+i)*ncon+j];
wsum += tpwgts[i*ncon+j];
}
for (wsum=1.0/wsum, i=0; i<lnparts; i++)
tpwgts[i*ncon+j] *= wsum;
}
METIS_PartGraphKway(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, NULL, agraph->adjwgt, &lnparts, tpwgts, NULL, moptions,
&edgecut, part);
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart + part[i];
}
gkMPI_Allreduce((void *)gwhere0, (void *)gwhere1, gnvtxs, IDX_T, MPI_SUM, ipcomm);
if (ngroups > 1) {
tmpxadj = agraph->xadj;
tmpadjncy = agraph->adjncy;
tmpadjwgt = agraph->adjwgt;
tmpvwgt = agraph->vwgt;
tmpwhere = agraph->where;
agraph->xadj = xadj;
agraph->adjncy = adjncy;
agraph->adjwgt = adjwgt;
agraph->vwgt = vwgt;
agraph->where = gwhere1;
agraph->vwgt = vwgt;
agraph->nvtxs = gnvtxs;
edgecut = ComputeSerialEdgeCut(agraph);
ComputeSerialBalance(ctrl, agraph, gwhere1, lbvec);
lbsum = rsum(ncon, lbvec, 1);
gkMPI_Allreduce((void *)&edgecut, (void *)&max_cut, 1, IDX_T, MPI_MAX, ctrl->gcomm);
gkMPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, REAL_T, MPI_MIN, ctrl->gcomm);
lpesum.sum = lbsum;
if (min_lbsum < UNBALANCE_FRACTION*ncon) {
if (lbsum < UNBALANCE_FRACTION*ncon)
lpesum.sum = edgecut;
else
lpesum.sum = max_cut;
}
lpesum.rank = ctrl->mype;
gkMPI_Allreduce((void *)&lpesum, (void *)&gpesum, 1, MPI_DOUBLE_INT,
MPI_MINLOC, ctrl->gcomm);
gkMPI_Bcast((void *)gwhere1, gnvtxs, IDX_T, gpesum.rank, ctrl->gcomm);
agraph->xadj = tmpxadj;
agraph->adjncy = tmpadjncy;
agraph->adjwgt = tmpadjwgt;
agraph->vwgt = tmpvwgt;
agraph->where = tmpwhere;
}
icopy(graph->nvtxs, gwhere1+graph->vtxdist[ctrl->mype], graph->where);
FreeGraph(agraph);
gkMPI_Comm_free(&ipcomm);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
WCOREPOP;
}
void TestParMetis_GPart(char *filename, char *xyzfile, MPI_Comm comm)
{
idx_t ncon, nparts, npes, mype, opt2, realcut;
graph_t graph, mgraph;
idx_t *part, *mpart, *savepart, *order, *sizes;
idx_t numflag=0, wgtflag=0, options[10], edgecut, ndims;
real_t ipc2redist, *xyz=NULL, *tpwgts = NULL, ubvec[MAXNCON];
gkMPI_Comm_size(comm, &npes);
gkMPI_Comm_rank(comm, &mype);
ParallelReadGraph(&graph, filename, comm);
if (xyzfile)
xyz = ReadTestCoordinates(&graph, xyzfile, &ndims, comm);
gkMPI_Barrier(comm);
part = imalloc(graph.nvtxs, "TestParMetis_V3: part");
tpwgts = rmalloc(MAXNCON*npes*2, "TestParMetis_V3: tpwgts");
rset(MAXNCON, 1.05, ubvec);
graph.vwgt = ismalloc(graph.nvtxs*5, 1, "TestParMetis_GPart: vwgt");
/*======================================================================
/ ParMETIS_V3_PartKway
/=======================================================================*/
options[0] = 1;
options[1] = 3;
options[2] = 1;
wgtflag = 2;
numflag = 0;
edgecut = 0;
for (nparts=2*npes; nparts>=npes/2 && nparts > 0; nparts = nparts/2) {
for (ncon=1; ncon<=NCON; ncon++) {
if (ncon > 1 && nparts > 1)
Mc_AdaptGraph(&graph, part, ncon, nparts, comm);
else
iset(graph.nvtxs, 1, graph.vwgt);
if (mype == 0)
printf("\nTesting ParMETIS_V3_PartKway with ncon: %"PRIDX", nparts: %"PRIDX"\n", ncon, nparts);
rset(nparts*ncon, 1.0/(real_t)nparts, tpwgts);
ParMETIS_V3_PartKway(graph.vtxdist, graph.xadj, graph.adjncy, graph.vwgt,
NULL, &wgtflag, &numflag, &ncon, &nparts, tpwgts, ubvec, options,
&edgecut, part, &comm);
realcut = ComputeRealCut(graph.vtxdist, part, filename, comm);
if (mype == 0) {
printf("ParMETIS_V3_PartKway reported a cut of %"PRIDX" [%s:%"PRIDX"]\n", edgecut,
(edgecut == realcut ? "OK" : "ERROR"), realcut);
}
if (mype == 0)
printf("\nTesting ParMETIS_V3_RefineKway with ncon: %"PRIDX", nparts: %"PRIDX"\n", ncon, nparts);
options[3] = PARMETIS_PSR_UNCOUPLED;
ParMETIS_V3_RefineKway(graph.vtxdist, graph.xadj, graph.adjncy, graph.vwgt,
NULL, &wgtflag, &numflag, &ncon, &nparts, tpwgts, ubvec, options,
&edgecut, part, &comm);
realcut = ComputeRealCut(graph.vtxdist, part, filename, comm);
if (mype == 0) {
printf("ParMETIS_V3_RefineKway reported a cut of %"PRIDX" [%s:%"PRIDX"]\n", edgecut,
(edgecut == realcut ? "OK" : "ERROR"), realcut);
}
}
}
/*======================================================================
/ ParMETIS_V3_PartGeomKway
/=======================================================================*/
if (xyzfile != NULL) {
options[0] = 1;
options[1] = 3;
options[2] = 1;
wgtflag = 2;
numflag = 0;
for (nparts=2*npes; nparts>=npes/2 && nparts > 0; nparts = nparts/2) {
for (ncon=1; ncon<=NCON; ncon++) {
if (ncon > 1)
Mc_AdaptGraph(&graph, part, ncon, nparts, comm);
else
iset(graph.nvtxs, 1, graph.vwgt);
if (mype == 0)
printf("\nTesting ParMETIS_V3_PartGeomKway with ncon: %"PRIDX", nparts: %"PRIDX"\n", ncon, nparts);
rset(nparts*ncon, 1.0/(real_t)nparts, tpwgts);
ParMETIS_V3_PartGeomKway(graph.vtxdist, graph.xadj, graph.adjncy, graph.vwgt,
NULL, &wgtflag, &numflag, &ndims, xyz, &ncon, &nparts, tpwgts, ubvec,
options, &edgecut, part, &comm);
realcut = ComputeRealCut(graph.vtxdist, part, filename, comm);
if (mype == 0)
printf("ParMETIS_V3_PartGeomKway reported a cut of %"PRIDX" [%s:%"PRIDX"]\n", edgecut,
(edgecut == realcut ? "OK" : "ERROR"), realcut);
}
}
}
/*======================================================================
/ ParMETIS_V3_PartGeom
/=======================================================================*/
if (xyz != NULL) {
wgtflag = 0;
numflag = 0;
if (mype == 0)
printf("\nTesting ParMETIS_V3_PartGeom\n");
ParMETIS_V3_PartGeom(graph.vtxdist, &ndims, xyz, part, &comm);
realcut = ComputeRealCut(graph.vtxdist, part, filename, comm);
if (mype == 0)
printf("ParMETIS_V3_PartGeom reported a cut of %"PRIDX"\n", realcut);
}
/*======================================================================
/ Coupled ParMETIS_V3_RefineKway
/=======================================================================*/
options[0] = 1;
options[1] = 3;
options[2] = 1;
options[3] = PARMETIS_PSR_COUPLED;
nparts = npes;
wgtflag = 0;
numflag = 0;
ncon = 1;
rset(nparts*ncon, 1.0/(real_t)nparts, tpwgts);
if (mype == 0)
printf("\nTesting coupled ParMETIS_V3_RefineKway with default options (before move)\n");
ParMETIS_V3_RefineKway(graph.vtxdist, graph.xadj, graph.adjncy, NULL, NULL,
&wgtflag, &numflag, &ncon, &nparts, tpwgts, ubvec, options, &edgecut,
part, &comm);
/* Compute a good partition and move the graph. Do so quietly! */
options[0] = 0;
nparts = npes;
wgtflag = 0;
numflag = 0;
ncon = 1;
rset(nparts*ncon, 1.0/(real_t)nparts, tpwgts);
ParMETIS_V3_PartKway(graph.vtxdist, graph.xadj, graph.adjncy, NULL, NULL,
&wgtflag, &numflag, &ncon, &npes, tpwgts, ubvec, options, &edgecut,
part, &comm);
TestMoveGraph(&graph, &mgraph, part, comm);
gk_free((void **)&(graph.vwgt), LTERM);
mpart = ismalloc(mgraph.nvtxs, mype, "TestParMetis_V3: mpart");
savepart = imalloc(mgraph.nvtxs, "TestParMetis_V3: savepart");
/*======================================================================
/ Coupled ParMETIS_V3_RefineKway
/=======================================================================*/
options[0] = 1;
options[1] = 3;
options[2] = 1;
options[3] = PARMETIS_PSR_COUPLED;
nparts = npes;
wgtflag = 0;
numflag = 0;
for (ncon=1; ncon<=NCON; ncon++) {
if (mype == 0)
printf("\nTesting coupled ParMETIS_V3_RefineKway with ncon: %"PRIDX", nparts: %"PRIDX"\n", ncon, nparts);
rset(nparts*ncon, 1.0/(real_t)nparts, tpwgts);
ParMETIS_V3_RefineKway(mgraph.vtxdist, mgraph.xadj, mgraph.adjncy, NULL, NULL,
&wgtflag, &numflag, &ncon, &nparts, tpwgts, ubvec, options, &edgecut,
mpart, &comm);
realcut = ComputeRealCutFromMoved(graph.vtxdist, mgraph.vtxdist, part, mpart,
filename, comm);
if (mype == 0)
printf("ParMETIS_V3_RefineKway reported a cut of %"PRIDX" [%s:%"PRIDX"]\n", edgecut,
(edgecut == realcut ? "OK" : "ERROR"), realcut);
}
/*ADAPTIVE:*/
/*======================================================================
/ ParMETIS_V3_AdaptiveRepart
/=======================================================================*/
mgraph.vwgt = ismalloc(mgraph.nvtxs*NCON, 1, "TestParMetis_V3: mgraph.vwgt");
mgraph.vsize = ismalloc(mgraph.nvtxs, 1, "TestParMetis_V3: mgraph.vsize");
AdaptGraph(&mgraph, 4, comm);
options[0] = 1;
options[1] = 7;
options[2] = 1;
options[3] = PARMETIS_PSR_COUPLED;
wgtflag = 2;
numflag = 0;
for (nparts=2*npes; nparts>=npes/2; nparts = nparts/2) {
options[0] = 0;
ncon = 1;
wgtflag = 0;
rset(nparts*ncon, 1.0/(real_t)nparts, tpwgts);
ParMETIS_V3_PartKway(mgraph.vtxdist, mgraph.xadj, mgraph.adjncy, NULL, NULL,
&wgtflag, &numflag, &ncon, &nparts, tpwgts, ubvec, options, &edgecut,
savepart, &comm);
options[0] = 1;
wgtflag = 2;
for (ncon=1; ncon<=NCON; ncon++) {
rset(nparts*ncon, 1.0/(real_t)nparts, tpwgts);
if (ncon > 1)
Mc_AdaptGraph(&mgraph, savepart, ncon, nparts, comm);
else
AdaptGraph(&mgraph, 4, comm);
for (ipc2redist=1000.0; ipc2redist>=0.001; ipc2redist/=1000.0) {
icopy(mgraph.nvtxs, savepart, mpart);
if (mype == 0)
printf("\nTesting ParMETIS_V3_AdaptiveRepart with ipc2redist: %.3"PRREAL", ncon: %"PRIDX", nparts: %"PRIDX"\n",
ipc2redist, ncon, nparts);
ParMETIS_V3_AdaptiveRepart(mgraph.vtxdist, mgraph.xadj, mgraph.adjncy,
mgraph.vwgt, mgraph.vsize, NULL, &wgtflag, &numflag, &ncon, &nparts,
tpwgts, ubvec, &ipc2redist, options, &edgecut, mpart, &comm);
realcut = ComputeRealCutFromMoved(graph.vtxdist, mgraph.vtxdist, part, mpart,
filename, comm);
if (mype == 0)
printf("ParMETIS_V3_AdaptiveRepart reported a cut of %"PRIDX" [%s:%"PRIDX"]\n", edgecut,
(edgecut == realcut ? "OK" : "ERROR"), realcut);
}
}
}
gk_free((void **)&tpwgts, &part, &mpart, &savepart, &xyz, &mgraph.vwgt, &mgraph.vsize, LTERM);
}
/***********************************************************************************
* This function is the entry point of the parallel multilevel local diffusion
* algorithm. It uses parallel undirected diffusion followed by adaptive k-way
* refinement. This function utilizes local coarsening.
************************************************************************************/
int ParMETIS_V3_RefineKway(idx_t *vtxdist, idx_t *xadj, idx_t *adjncy, idx_t *vwgt,
idx_t *adjwgt, idx_t *wgtflag, idx_t *numflag, idx_t *ncon, idx_t *nparts,
real_t *tpwgts, real_t *ubvec, idx_t *options, idx_t *edgecut, idx_t *part,
MPI_Comm *comm)
{
idx_t npes, mype, status;
ctrl_t *ctrl=NULL;
graph_t *graph=NULL;
size_t curmem;
/* Check the input parameters and return if an error */
status = CheckInputsPartKway(vtxdist, xadj, adjncy, vwgt, adjwgt, wgtflag,
numflag, ncon, nparts, tpwgts, ubvec, options, edgecut, part, comm);
if (GlobalSEMinComm(*comm, status) == 0)
return METIS_ERROR;
status = METIS_OK;
gk_malloc_init();
curmem = gk_GetCurMemoryUsed();
/* Setup ctrl */
ctrl = SetupCtrl(PARMETIS_OP_RMETIS, options, *ncon, *nparts, tpwgts, ubvec, *comm);
npes = ctrl->npes;
mype = ctrl->mype;
/* Take care the nparts == 1 case */
if (*nparts == 1) {
iset(vtxdist[mype+1]-vtxdist[mype], (*numflag == 0 ? 0 : 1), part);
*edgecut = 0;
goto DONE;
}
/* setup the graph */
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
graph = SetupGraph(ctrl, *ncon, vtxdist, xadj, vwgt, NULL, adjncy, adjwgt, *wgtflag);
if (ctrl->ps_relation == PARMETIS_PSR_COUPLED)
iset(graph->nvtxs, mype, graph->home);
else
icopy(graph->nvtxs, part, graph->home);
/* Allocate workspace */
AllocateWSpace(ctrl, 10*graph->nvtxs);
/* Partition and Remap */
STARTTIMER(ctrl, ctrl->TotalTmr);
ctrl->CoarsenTo = gk_min(vtxdist[npes]+1, 50*(*ncon)*gk_max(npes, *nparts));
Adaptive_Partition(ctrl, graph);
ParallelReMapGraph(ctrl, graph);
icopy(graph->nvtxs, graph->where, part);
*edgecut = graph->mincut;
STOPTIMER(ctrl, ctrl->TotalTmr);
/* Take care of output */
IFSET(ctrl->dbglvl, DBG_TIME, PrintTimingInfo(ctrl));
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->gcomm));
IFSET(ctrl->dbglvl, DBG_INFO, PrintPostPartInfo(ctrl, graph, 1));
FreeInitialGraphAndRemap(graph);
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
DONE:
FreeCtrl(&ctrl);
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 (int)status;
}
/***********************************************************************************
* This function is the entry point of the parallel kmetis algorithm that uses
* coordinates to compute an initial graph distribution.
************************************************************************************/
int ParMETIS_V3_PartGeomKway(idx_t *vtxdist, idx_t *xadj, idx_t *adjncy,
idx_t *vwgt, idx_t *adjwgt, idx_t *wgtflag, idx_t *numflag, idx_t *ndims,
real_t *xyz, idx_t *ncon, idx_t *nparts, real_t *tpwgts, real_t *ubvec,
idx_t *options, idx_t *edgecut, idx_t *part, MPI_Comm *comm)
{
idx_t h, i, j, npes, mype, status, nvtxs, seed, dbglvl;
idx_t cut, gcut, maxnvtxs;
idx_t moptions[METIS_NOPTIONS];
ctrl_t *ctrl;
graph_t *graph, *mgraph;
real_t balance;
size_t curmem;
/* Check the input parameters and return if an error */
status = CheckInputsPartGeomKway(vtxdist, xadj, adjncy, vwgt, adjwgt, wgtflag,
numflag, ndims, xyz, ncon, nparts, tpwgts, ubvec, options,
edgecut, part, comm);
if (GlobalSEMinComm(*comm, status) == 0)
return METIS_ERROR;
status = METIS_OK;
gk_malloc_init();
curmem = gk_GetCurMemoryUsed();
/* Setup the ctrl */
ctrl = SetupCtrl(PARMETIS_OP_GKMETIS, options, *ncon, *nparts, tpwgts, ubvec, *comm);
npes = ctrl->npes;
mype = ctrl->mype;
/* Take care the nparts == 1 case */
if (*nparts == 1) {
iset(vtxdist[mype+1]-vtxdist[mype], (*numflag == 0 ? 0 : 1), part);
*edgecut = 0;
goto DONE;
}
/* Take care of npes == 1 case */
if (npes == 1) {
nvtxs = vtxdist[1] - vtxdist[0]; /* subtraction is required when numflag==1 */
METIS_SetDefaultOptions(moptions);
moptions[METIS_OPTION_NUMBERING] = *numflag;
status = METIS_PartGraphKway(&nvtxs, ncon, xadj, adjncy, vwgt, NULL, adjwgt,
nparts, tpwgts, ubvec, moptions, edgecut, part);
goto DONE;
}
/* Setup the graph */
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
graph = SetupGraph(ctrl, *ncon, vtxdist, xadj, vwgt, NULL, adjncy, adjwgt, *wgtflag);
gk_free((void **)&graph->nvwgt, LTERM);
/* Allocate the workspace */
AllocateWSpace(ctrl, 10*graph->nvtxs);
/* Compute the initial npes-way partitioning geometric partitioning */
STARTTIMER(ctrl, ctrl->TotalTmr);
Coordinate_Partition(ctrl, graph, *ndims, xyz, 1);
STOPTIMER(ctrl, ctrl->TotalTmr);
/* Move the graph according to the partitioning */
STARTTIMER(ctrl, ctrl->MoveTmr);
ctrl->nparts = npes;
mgraph = MoveGraph(ctrl, graph);
ctrl->nparts = *nparts;
SetupGraph_nvwgts(ctrl, mgraph); /* compute nvwgts for the moved graph */
if (ctrl->dbglvl&DBG_INFO) {
CommInterfaceData(ctrl, graph, graph->where, graph->where+graph->nvtxs);
for (cut=0, i=0; i<graph->nvtxs; i++) {
for (j=graph->xadj[i]; j<graph->xadj[i+1]; j++) {
if (graph->where[i] != graph->where[graph->adjncy[j]])
cut += graph->adjwgt[j];
}
}
gcut = GlobalSESum(ctrl, cut)/2;
maxnvtxs = GlobalSEMax(ctrl, mgraph->nvtxs);
balance = (real_t)(maxnvtxs)/((real_t)(graph->gnvtxs)/(real_t)(npes));
rprintf(ctrl, "XYZ Cut: %6"PRIDX" \tBalance: %6.3"PRREAL" [%"PRIDX" %"PRIDX" %"PRIDX"]\n",
gcut, balance, maxnvtxs, graph->gnvtxs, npes);
}
STOPTIMER(ctrl, ctrl->MoveTmr);
/* Compute the partition of the moved graph */
STARTTIMER(ctrl, ctrl->TotalTmr);
ctrl->CoarsenTo = gk_min(vtxdist[npes]+1, 25*(*ncon)*gk_max(npes, *nparts));
if (vtxdist[npes] < SMALLGRAPH
|| vtxdist[npes] < npes*20
|| GlobalSESum(ctrl, mgraph->nedges) == 0) { /* serially */
IFSET(ctrl->dbglvl, DBG_INFO,
rprintf(ctrl, "Partitioning a graph of size %"PRIDX" serially\n", vtxdist[npes]));
PartitionSmallGraph(ctrl, mgraph);
}
else { /* in parallel */
Global_Partition(ctrl, mgraph);
}
ParallelReMapGraph(ctrl, mgraph);
/* Invert the ordering back to the original graph */
ctrl->nparts = npes;
ProjectInfoBack(ctrl, graph, part, mgraph->where);
ctrl->nparts = *nparts;
*edgecut = mgraph->mincut;
STOPTIMER(ctrl, ctrl->TotalTmr);
/* Print some stats */
IFSET(ctrl->dbglvl, DBG_TIME, PrintTimingInfo(ctrl));
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->gcomm));
IFSET(ctrl->dbglvl, DBG_INFO, PrintPostPartInfo(ctrl, mgraph, 0));
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph);
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
DONE:
FreeCtrl(&ctrl);
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 (int)status;
}
/***********************************************************************************
* This function is the entry point of the serial ordering algorithm.
************************************************************************************/
int ParMETIS_SerialNodeND(idx_t *vtxdist, idx_t *xadj, idx_t *adjncy,
idx_t *numflag, idx_t *options, idx_t *order, idx_t *sizes,
MPI_Comm *comm)
{
idx_t i, npes, mype;
ctrl_t *ctrl=NULL;
graph_t *agraph=NULL;
idx_t *perm=NULL, *iperm=NULL;
idx_t *sendcount, *displs;
/* Setup the ctrl */
ctrl = SetupCtrl(PARMETIS_OP_OMETIS, options, 1, 1, NULL, NULL, *comm);
npes = ctrl->npes;
mype = ctrl->mype;
if (!ispow2(npes)) {
if (mype == 0)
printf("Error: The number of processors must be a power of 2!\n");
FreeCtrl(&ctrl);
return METIS_ERROR;
}
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 1);
STARTTIMER(ctrl, ctrl->TotalTmr);
STARTTIMER(ctrl, ctrl->MoveTmr);
agraph = AssembleEntireGraph(ctrl, vtxdist, xadj, adjncy);
STOPTIMER(ctrl, ctrl->MoveTmr);
if (mype == 0) {
perm = imalloc(agraph->nvtxs, "PAROMETISS: perm");
iperm = imalloc(agraph->nvtxs, "PAROMETISS: iperm");
METIS_NodeNDP(agraph->nvtxs, agraph->xadj, agraph->adjncy,
agraph->vwgt, npes, NULL, perm, iperm, sizes);
}
STARTTIMER(ctrl, ctrl->MoveTmr);
/* Broadcast the sizes array */
gkMPI_Bcast((void *)sizes, 2*npes, IDX_T, 0, ctrl->gcomm);
/* Scatter the iperm */
sendcount = imalloc(npes, "PAROMETISS: sendcount");
displs = imalloc(npes, "PAROMETISS: displs");
for (i=0; i<npes; i++) {
sendcount[i] = vtxdist[i+1]-vtxdist[i];
displs[i] = vtxdist[i];
}
gkMPI_Scatterv((void *)iperm, sendcount, displs, IDX_T, (void *)order,
vtxdist[mype+1]-vtxdist[mype], IDX_T, 0, ctrl->gcomm);
STOPTIMER(ctrl, ctrl->MoveTmr);
STOPTIMER(ctrl, ctrl->TotalTmr);
IFSET(ctrl->dbglvl, DBG_TIME, PrintTimingInfo(ctrl));
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->gcomm));
gk_free((void **)&agraph->xadj, &agraph->adjncy, &perm, &iperm,
&sendcount, &displs, &agraph, LTERM);
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 0);
goto DONE;
DONE:
FreeCtrl(&ctrl);
return METIS_OK;
}
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 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;
}
