/***********************************************************************************
* This function is the entry point of the parallel ordering algorithm.
* This function assumes that the graph is already nice partitioned among the
* processors and then proceeds to perform recursive bisection.
************************************************************************************/
void ParMETIS_V3_PartGeom(idxtype *vtxdist, int *ndims, float *xyz, idxtype *part, MPI_Comm *comm)
{
int i, npes, mype, nvtxs, firstvtx, dbglvl;
idxtype *xadj, *adjncy;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
int zeroflg = 0;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
if (npes == 1) {
idxset(vtxdist[mype+1]-vtxdist[mype], 0, part);
return;
}
/* Setup a fake graph to allow the rest of the code to work unchanged */
dbglvl = 0;
nvtxs = vtxdist[mype+1]-vtxdist[mype];
firstvtx = vtxdist[mype];
xadj = idxmalloc(nvtxs+1, "ParMETIS_PartGeom: xadj");
adjncy = idxmalloc(nvtxs, "ParMETIS_PartGeom: adjncy");
for (i=0; i<nvtxs; i++) {
xadj[i] = i;
adjncy[i] = firstvtx + (i+1)%nvtxs;
}
xadj[nvtxs] = nvtxs;
/* Proceed with the rest of the code */
SetUpCtrl(&ctrl, npes, dbglvl, *comm);
ctrl.seed = mype;
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*npes);
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, NULL, adjncy, NULL, &zeroflg);
PreAllocateMemory(&ctrl, graph, &wspace);
/*=======================================================
* Compute the initial geometric partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Coordinate_Partition(&ctrl, graph, *ndims, xyz, 0, &wspace);
idxcopy(graph->nvtxs, graph->where, part);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
FreeInitialGraphAndRemap(graph, 0);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
GKfree((void **)&xadj, (void **)&adjncy, 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.
************************************************************************************/
void ParMETIS_RepartLDiffusion(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, realtype *adjwgt, int *wgtflag, int *numflag, int *options,
int *edgecut, idxtype *part, MPI_Comm *comm)
{
int npes, mype;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
if (npes == 1) { /* Take care the npes = 1 case */
idxset(vtxdist[1], 0, part);
*edgecut = 0;
return;
}
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
SetUpCtrl(&ctrl, npes, options, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 70*npes);
graph = SetUpGraph(&ctrl, vtxdist, xadj, vwgt, adjncy, adjwgt, *wgtflag);
graph->vsize = idxsmalloc(graph->nvtxs, 1, "Par_KMetis: vsize");
PreAllocateMemory(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TRACK, printf("%d ParMETIS_RepartLDiffusion about to call AdaptiveUndirected_Partition\n",mype));
AdaptiveUndirected_Partition(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TRACK, printf("%d ParMETIS_RepartLDiffusion about to call ReMapGraph\n",mype));
ReMapGraph(&ctrl, graph, 0, &wspace);
idxcopy(graph->nvtxs, graph->where, part);
*edgecut = graph->mincut;
IMfree((void**)&graph->vsize, LTERM);
FreeInitialGraphAndRemap(graph, *wgtflag);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
}
/***********************************************************************************
* This function creates the fused-element-graph and returns the partition
************************************************************************************/
void ParMETIS_FusedElementGraph(idxtype *vtxdist, idxtype *xadj, realtype *vvol,
realtype *vsurf, idxtype *adjncy, idxtype *vwgt, realtype *adjwgt,
int *wgtflag, int *numflag, int *nparts, int *options,
idxtype *part, MPI_Comm *comm)
{
int npes, mype, nvtxs;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
nvtxs = vtxdist[mype+1]-vtxdist[mype];
/* IFSET(options[OPTION_DBGLVL], DBG_TRACK, printf("%d ParMETIS_FEG npes=%d\n",mype, npes)); */
SetUpCtrl(&ctrl, *nparts, options, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*amax(npes, *nparts));
graph = SetUpGraph(&ctrl, vtxdist, xadj, vwgt, adjncy, adjwgt, *wgtflag);
graph->where = part;
PreAllocateMemory(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
CreateFusedElementGraph(&ctrl, graph, &wspace, numflag);
idxcopy(nvtxs, graph->where, part);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
if (((*wgtflag)&2) == 0)
IMfree((void**)&graph->vwgt, LTERM);
IMfree((void**)&graph->lperm, &graph->peind, &graph->pexadj, &graph->peadjncy,
&graph->peadjloc, &graph->recvptr, &graph->recvind, &graph->sendptr,
&graph->imap, &graph->sendind, &graph, LTERM);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
}
/******************************************************************************
* This function takes a graph and its partition vector and creates a new
* graph corresponding to the one after the movement
*******************************************************************************/
void TestMoveGraph(GraphType *ograph, GraphType *omgraph, idxtype *part, MPI_Comm comm)
{
int npes, mype;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph, *mgraph;
int options[5] = {0, 0, 1, 0, 0};
MPI_Comm_size(comm, &npes);
MPI_Comm_rank(comm, &mype);
SetUpCtrl(&ctrl, npes, 0, comm);
ctrl.CoarsenTo = 1; /* Needed by SetUpGraph, otherwise we can FP errors */
graph = SetUpGraph(&ctrl, ograph->vtxdist, ograph->xadj, NULL, ograph->adjncy, NULL, 0);
AllocateWSpace(&ctrl, graph, &wspace);
SetUp(&ctrl, graph, &wspace);
graph->where = part;
graph->ncon = 1;
mgraph = Mc_MoveGraph(&ctrl, graph, &wspace);
omgraph->gnvtxs = mgraph->gnvtxs;
omgraph->nvtxs = mgraph->nvtxs;
omgraph->nedges = mgraph->nedges;
omgraph->vtxdist = mgraph->vtxdist;
omgraph->xadj = mgraph->xadj;
omgraph->adjncy = mgraph->adjncy;
mgraph->vtxdist = NULL;
mgraph->xadj = NULL;
mgraph->adjncy = NULL;
FreeGraph(mgraph);
graph->where = NULL;
FreeInitialGraphAndRemap(graph, 0, 1);
FreeWSpace(&wspace);
}
/***********************************************************************************
* This function is the entry point of the parallel ordering algorithm.
* This function assumes that the graph is already nice partitioned among the
* processors and then proceeds to perform recursive bisection.
************************************************************************************/
void ParMETIS_V3_NodeND(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy, int *numflag,
int *options, idxtype *order, idxtype *sizes, MPI_Comm *comm)
{
int i, j;
int ltvwgts[MAXNCON];
int nparts, npes, mype, wgtflag = 0, seed = GLOBAL_SEED;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph, *mgraph;
idxtype *morder;
int minnvtxs;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
nparts = npes;
if (!ispow2(npes)) {
if (mype == 0)
printf("Error: The number of processors must be a power of 2!\n");
return;
}
if (vtxdist[npes] < (int)((float)(npes*npes)*1.2)) {
if (mype == 0)
printf("Error: Too many processors for this many vertices.\n");
return;
}
minnvtxs = vtxdist[1]-vtxdist[0];
for (i=0; i<npes; i++)
minnvtxs = (minnvtxs < vtxdist[i+1]-vtxdist[i]) ? minnvtxs : vtxdist[i+1]-vtxdist[i];
if (minnvtxs < (int)((float)npes*1.1)) {
if (mype == 0)
printf("Error: vertices are not distributed equally.\n");
return;
}
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 1);
SetUpCtrl(&ctrl, nparts, options[PMV3_OPTION_DBGLVL], *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*npes);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*amax(npes, nparts));
ctrl.seed = mype;
ctrl.sync = seed;
ctrl.partType = STATIC_PARTITION;
ctrl.ps_relation = -1;
ctrl.tpwgts = fsmalloc(nparts, 1.0/(float)(nparts), "tpwgts");
ctrl.ubvec[0] = 1.03;
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, NULL, adjncy, NULL, &wgtflag);
PreAllocateMemory(&ctrl, graph, &wspace);
/*=======================================================
* Compute the initial k-way partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Moc_Global_Partition(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
/*=======================================================
* Move the graph according to the partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.MoveTmr));
MALLOC_CHECK(NULL);
graph->ncon = 1;
mgraph = Moc_MoveGraph(&ctrl, graph, &wspace);
MALLOC_CHECK(NULL);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.MoveTmr));
/*=======================================================
* Now compute an ordering of the moved graph
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
FreeWSpace(&wspace);
PreAllocateMemory(&ctrl, mgraph, &wspace);
ctrl.ipart = ISEP_NODE;
ctrl.CoarsenTo = amin(vtxdist[npes]+1, amax(20*npes, 1000));
/* compute tvwgts */
for (j=0; j<mgraph->ncon; j++)
ltvwgts[j] = 0;
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<mgraph->ncon; j++)
ltvwgts[j] += mgraph->vwgt[i*mgraph->ncon+j];
for (j=0; j<mgraph->ncon; j++)
ctrl.tvwgts[j] = GlobalSESum(&ctrl, ltvwgts[j]);
mgraph->nvwgt = fmalloc(mgraph->nvtxs*mgraph->ncon, "mgraph->nvwgt");
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<mgraph->ncon; j++)
mgraph->nvwgt[i*mgraph->ncon+j] = (float)(mgraph->vwgt[i*mgraph->ncon+j]) / (float)(ctrl.tvwgts[j]);
morder = idxmalloc(mgraph->nvtxs, "PAROMETIS: morder");
MultilevelOrder(&ctrl, mgraph, morder, sizes, &wspace);
MALLOC_CHECK(NULL);
/* Invert the ordering back to the original graph */
ProjectInfoBack(&ctrl, graph, order, morder, &wspace);
MALLOC_CHECK(NULL);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
free(ctrl.tpwgts);
free(morder);
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph, 0);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 0);
MALLOC_CHECK(NULL);
}
/***********************************************************************************
* This function is the entry point of the parallel kmetis algorithm that uses
* coordinates to compute an initial graph distribution.
************************************************************************************/
void ParMETIS_V3_PartGeomKway(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, int *wgtflag, int *numflag, int *ndims,
float *xyz, int *ncon, int *nparts, float *tpwgts, float *ubvec,
int *options, int *edgecut, idxtype *part, MPI_Comm *comm)
{
int h, i, j;
int nvtxs = -1, npes, mype;
int uwgtflag, cut, gcut, maxnvtxs;
int ltvwgts[MAXNCON];
int moptions[10];
CtrlType ctrl;
idxtype *uvwgt;
WorkSpaceType wspace;
GraphType *graph, *mgraph;
float avg, maximb, balance, *mytpwgts;
int seed, dbglvl = 0;
int iwgtflag, inumflag, incon, inparts, ioptions[10];
float *itpwgts, iubvec[MAXNCON];
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
/********************************/
/* Try and take care bad inputs */
/********************************/
if (options != NULL && options[0] == 1)
dbglvl = options[PMV3_OPTION_DBGLVL];
CheckInputs(STATIC_PARTITION, npes, dbglvl, wgtflag, &iwgtflag, numflag, &inumflag,
ncon, &incon, nparts, &inparts, tpwgts, &itpwgts, ubvec, iubvec,
NULL, NULL, options, ioptions, part, comm);
/*********************************/
/* Take care the nparts = 1 case */
/*********************************/
if (inparts <= 1) {
idxset(vtxdist[mype+1]-vtxdist[mype], 0, part);
*edgecut = 0;
return;
}
/******************************/
/* Take care of npes = 1 case */
/******************************/
if (npes == 1 && inparts > 1) {
moptions[0] = 0;
nvtxs = vtxdist[1];
if (incon == 1) {
METIS_WPartGraphKway(&nvtxs, xadj, adjncy, vwgt, adjwgt, &iwgtflag, &inumflag,
&inparts, itpwgts, moptions, edgecut, part);
}
else {
/* ADD: this is because METIS does not support tpwgts for all constraints */
mytpwgts = fmalloc(inparts, "mytpwgts");
for (i=0; i<inparts; i++)
mytpwgts[i] = itpwgts[i*incon];
moptions[7] = -1;
METIS_mCPartGraphRecursive2(&nvtxs, &incon, xadj, adjncy, vwgt, adjwgt, &iwgtflag,
&inumflag, &inparts, mytpwgts, moptions, edgecut, part);
free(mytpwgts);
}
return;
}
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
/*****************************/
/* Set up control structures */
/*****************************/
if (ioptions[0] == 1) {
dbglvl = ioptions[PMV3_OPTION_DBGLVL];
seed = ioptions[PMV3_OPTION_SEED];
}
else {
dbglvl = GLOBAL_DBGLVL;
seed = GLOBAL_SEED;
}
SetUpCtrl(&ctrl, npes, dbglvl, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*incon*amax(npes, inparts));
ctrl.seed = (seed == 0) ? mype : seed*mype;
ctrl.sync = GlobalSEMax(&ctrl, seed);
ctrl.partType = STATIC_PARTITION;
ctrl.ps_relation = -1;
ctrl.tpwgts = itpwgts;
scopy(incon, iubvec, ctrl.ubvec);
uwgtflag = iwgtflag|2;
uvwgt = idxsmalloc(vtxdist[mype+1]-vtxdist[mype], 1, "uvwgt");
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, uvwgt, adjncy, adjwgt, &uwgtflag);
free(graph->nvwgt); graph->nvwgt = NULL;
PreAllocateMemory(&ctrl, graph, &wspace);
/*=================================================================
* Compute the initial npes-way partitioning geometric partitioning
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Coordinate_Partition(&ctrl, graph, *ndims, xyz, 1, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
/*=================================================================
* Move the graph according to the partitioning
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.MoveTmr));
free(uvwgt);
graph->vwgt = ((iwgtflag&2) != 0) ? vwgt : idxsmalloc(graph->nvtxs*incon, 1, "vwgt");
graph->ncon = incon;
j = ctrl.nparts;
ctrl.nparts = ctrl.npes;
mgraph = Moc_MoveGraph(&ctrl, graph, &wspace);
ctrl.nparts = j;
/**********************************************************/
/* Do the same functionality as Moc_SetUpGraph for mgraph */
/**********************************************************/
/* compute tvwgts */
for (j=0; j<incon; j++)
ltvwgts[j] = 0;
for (i=0; i<graph->nvtxs; i++)
for (j=0; j<incon; j++)
ltvwgts[j] += mgraph->vwgt[i*incon+j];
for (j=0; j<incon; j++)
ctrl.tvwgts[j] = GlobalSESum(&ctrl, ltvwgts[j]);
/* check for zero wgt constraints */
for (i=0; i<incon; i++) {
/* ADD: take care of the case in which tvwgts is zero */
if (ctrl.tvwgts[i] == 0) {
if (ctrl.mype == 0) printf("ERROR: sum weight for constraint %d is zero\n", i);
MPI_Finalize();
exit(-1);
}
}
/* compute nvwgt */
mgraph->nvwgt = fmalloc(mgraph->nvtxs*incon, "mgraph->nvwgt");
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<incon; j++)
mgraph->nvwgt[i*incon+j] = (float)(mgraph->vwgt[i*incon+j]) / (float)(ctrl.tvwgts[j]);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.MoveTmr));
if (ctrl.dbglvl&DBG_INFO) {
cut = 0;
for (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 = (float)(maxnvtxs)/((float)(graph->gnvtxs)/(float)(npes));
rprintf(&ctrl, "XYZ Cut: %6d \tBalance: %6.3f [%d %d %d]\n",
gcut, balance, maxnvtxs, graph->gnvtxs, npes);
}
/*=================================================================
* Set up the newly moved graph
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
ctrl.nparts = inparts;
FreeWSpace(&wspace);
PreAllocateMemory(&ctrl, mgraph, &wspace);
/*=======================================================
* Now compute the partition of the moved graph
=======================================================*/
if (vtxdist[npes] < SMALLGRAPH || vtxdist[npes] < npes*20 || GlobalSESum(&ctrl, mgraph->nedges) == 0) {
IFSET(ctrl.dbglvl, DBG_INFO, rprintf(&ctrl, "Partitioning a graph of size %d serially\n", vtxdist[npes]));
PartitionSmallGraph(&ctrl, mgraph, &wspace);
}
else {
Moc_Global_Partition(&ctrl, mgraph, &wspace);
}
ParallelReMapGraph(&ctrl, mgraph, &wspace);
/* Invert the ordering back to the original graph */
ctrl.nparts = npes;
ProjectInfoBack(&ctrl, graph, part, mgraph->where, &wspace);
*edgecut = mgraph->mincut;
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
/*******************/
/* Print out stats */
/*******************/
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
if (ctrl.dbglvl&DBG_INFO) {
rprintf(&ctrl, "Final %d-way CUT: %6d \tBalance: ", inparts, mgraph->mincut);
avg = 0.0;
for (h=0; h<incon; h++) {
maximb = 0.0;
for (i=0; i<inparts; i++)
maximb = amax(maximb, mgraph->gnpwgts[i*incon+h]/itpwgts[i*incon+h]);
avg += maximb;
rprintf(&ctrl, "%.3f ", maximb);
}
rprintf(&ctrl, " avg: %.3f\n", avg/(float)incon);
}
GKfree((void **)&itpwgts, LTERM);
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph, iwgtflag);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
}
/*************************************************************************
* This function converts a mesh into a dual graph
**************************************************************************/
void ParMETIS_V3_Mesh2Dual(idxtype *elmdist, idxtype *eptr, idxtype *eind,
int *numflag, int *ncommonnodes, idxtype **xadj,
idxtype **adjncy, MPI_Comm *comm)
{
int i, j, jj, k, kk, m;
int npes, mype, pe, count, mask, pass;
int nelms, lnns, my_nns, node;
int firstelm, firstnode, lnode, nrecv, nsend;
int *scounts, *rcounts, *sdispl, *rdispl;
idxtype *nodedist, *nmap, *auxarray;
idxtype *gnptr, *gnind, *nptr, *nind, *myxadj, *myadjncy = NULL;
idxtype *sbuffer, *rbuffer, *htable;
KeyValueType *nodelist, *recvbuffer;
idxtype ind[200], wgt[200];
int gmaxnode, gminnode;
CtrlType ctrl;
SetUpCtrl(&ctrl, -1, 0, *comm);
npes = ctrl.npes;
mype = ctrl.mype;
nelms = elmdist[mype+1]-elmdist[mype];
if (*numflag == 1)
ChangeNumberingMesh2(elmdist, eptr, eind, NULL, NULL, NULL, npes, mype, 1);
mask = (1<<11)-1;
/*****************************/
/* Determine number of nodes */
/*****************************/
gminnode = GlobalSEMin(&ctrl, eind[idxamin(eptr[nelms], eind)]);
for (i=0; i<eptr[nelms]; i++)
eind[i] -= gminnode;
gmaxnode = GlobalSEMax(&ctrl, eind[idxamax(eptr[nelms], eind)]);
/**************************/
/* Check for input errors */
/**************************/
ASSERTS(nelms > 0);
/* construct node distribution array */
nodedist = idxsmalloc(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 = (KeyValueType *)GKmalloc(eptr[nelms]*sizeof(KeyValueType), "nodelist");
auxarray = idxmalloc(eptr[nelms], "auxarray");
htable = idxsmalloc(amax(my_nns, mask+1), -1, "htable");
scounts = imalloc(4*npes+2, "scounts");
rcounts = scounts+npes;
sdispl = scounts+2*npes;
rdispl = scounts+3*npes+1;
/*********************************************/
/* 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 */
}
}
ikeysort(eptr[nelms], nodelist);
for (count=1, i=1; i<eptr[nelms]; i++) {
if (nodelist[i].key > nodelist[i-1].key)
count++;
}
lnns = count;
nmap = idxmalloc(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 */
}
MPI_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;
}
ASSERTS(pe < npes);
MPI_Alltoall((void *)scounts, 1, MPI_INT, (void *)rcounts, 1, MPI_INT, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
icopy(npes, rcounts, rdispl);
MAKECSR(i, npes, rdispl);
ASSERTS(sdispl[npes] == eptr[nelms]*2);
nrecv = rdispl[npes]/2;
recvbuffer = (KeyValueType *)GKmalloc(amax(1, nrecv)*sizeof(KeyValueType), "recvbuffer");
MPI_Alltoallv((void *)nodelist, scounts, sdispl, IDX_DATATYPE, (void *)recvbuffer,
rcounts, rdispl, IDX_DATATYPE, *comm);
/**************************************/
/* construct global node-element list */
/**************************************/
gnptr = idxsmalloc(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;
ASSERTS(lnode >= 0 && lnode < my_nns)
gnptr[lnode]++;
}
}
MAKECSR(i, my_nns, gnptr);
gnind = idxmalloc(amax(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;
}
}
MPI_Alltoall((void *)scounts, 1, MPI_INT, (void *)rcounts, 1, MPI_INT, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
/* create the send buffer */
nsend = sdispl[npes];
sbuffer = (idxtype *)realloc(nodelist, sizeof(idxtype)*amax(1, nsend));
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;
}
}
ASSERTS(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 = (idxtype *)realloc(recvbuffer, sizeof(idxtype)*amax(1, nrecv));
MPI_Alltoallv((void *)sbuffer, scounts, sdispl, IDX_DATATYPE, (void *)rbuffer,
rcounts, rdispl, IDX_DATATYPE, *comm);
k = -1;
nptr = idxsmalloc(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);
ASSERTS(k+1 == lnns);
ASSERTS(nptr[lnns] == nrecv)
myxadj = *xadj = idxsmalloc(nelms+1, 0, "xadj");
idxset(mask+1, -1, htable);
firstelm = elmdist[mype];
/* Two passes -- in first pass, simply find out the memory requirements */
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;
}
}
}
}
}
for (j=0; j<count; j++) {
htable[(ind[j]&mask)] = -1;
if (wgt[j] >= *ncommonnodes) {
if (pass == 0)
myxadj[i]++;
else
myadjncy[myxadj[i]++] = ind[j];
}
}
}
if (pass == 0) {
MAKECSR(i, nelms, myxadj);
myadjncy = *adjncy = idxmalloc(myxadj[nelms], "adjncy");
}
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)
ChangeNumberingMesh2(elmdist, eptr, eind, myxadj, myadjncy, NULL, npes, mype, 0);
/* do not free nodelist, recvbuffer, rbuffer */
GKfree((void **)&scounts, (void **)&nodedist, (void **)&nmap, (void **)&sbuffer,
(void **)&htable, (void **)&nptr, (void **)&nind, (void **)&gnptr,
(void **)&gnind, (void **)&auxarray, LTERM);
FreeCtrl(&ctrl);
return;
}
/***********************************************************************************
* 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.
************************************************************************************/
void ParMETIS_V3_RefineKway(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, int *wgtflag, int *numflag, int *ncon,
int *nparts, float *tpwgts, float *ubvec, int *options, int *edgecut,
idxtype *part, MPI_Comm *comm)
{
int h, i;
int npes, mype;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
int tewgt, tvsize, nmoved, maxin, maxout;
float gtewgt, gtvsize, avg, maximb;
int ps_relation, seed, dbglvl = 0;
int iwgtflag, inumflag, incon, inparts, ioptions[10];
float *itpwgts, iubvec[MAXNCON];
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
/********************************/
/* Try and take care bad inputs */
/********************************/
if (options != NULL && options[0] == 1)
dbglvl = options[PMV3_OPTION_DBGLVL];
CheckInputs(REFINE_PARTITION, npes, dbglvl, wgtflag, &iwgtflag, numflag, &inumflag,
ncon, &incon, nparts, &inparts, tpwgts, &itpwgts, ubvec, iubvec,
NULL, NULL, options, ioptions, part, comm);
/* ADD: take care of disconnected graph */
/* ADD: take care of highly unbalanced vtxdist */
/*********************************/
/* Take care the nparts = 1 case */
/*********************************/
if (inparts <= 1) {
idxset(vtxdist[mype+1]-vtxdist[mype], 0, part);
*edgecut = 0;
return;
}
/**************************/
/* Set up data structures */
/**************************/
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
/*****************************/
/* Set up control structures */
/*****************************/
if (ioptions[0] == 1) {
dbglvl = ioptions[PMV3_OPTION_DBGLVL];
seed = ioptions[PMV3_OPTION_SEED];
ps_relation = (npes == inparts) ? ioptions[PMV3_OPTION_PSR] : DISCOUPLED;
}
else {
dbglvl = GLOBAL_DBGLVL;
seed = GLOBAL_SEED;
ps_relation = (npes == inparts) ? COUPLED : DISCOUPLED;
}
SetUpCtrl(&ctrl, inparts, dbglvl, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 50*incon*amax(npes, inparts));
ctrl.ipc_factor = 1000.0;
ctrl.redist_factor = 1.0;
ctrl.redist_base = 1.0;
ctrl.seed = (seed == 0) ? mype : seed*mype;
ctrl.sync = GlobalSEMax(&ctrl, seed);
ctrl.partType = REFINE_PARTITION;
ctrl.ps_relation = ps_relation;
ctrl.tpwgts = itpwgts;
graph = Moc_SetUpGraph(&ctrl, incon, vtxdist, xadj, vwgt, adjncy, adjwgt, &iwgtflag);
graph->vsize = idxsmalloc(graph->nvtxs, 1, "vsize");
graph->home = idxmalloc(graph->nvtxs, "home");
if (ctrl.ps_relation == COUPLED)
idxset(graph->nvtxs, mype, graph->home);
else
idxcopy(graph->nvtxs, part, graph->home);
tewgt = idxsum(graph->nedges, graph->adjwgt);
tvsize = idxsum(graph->nvtxs, graph->vsize);
gtewgt = (float) GlobalSESum(&ctrl, tewgt) + 1.0/graph->gnvtxs;
gtvsize = (float) GlobalSESum(&ctrl, tvsize) + 1.0/graph->gnvtxs;
ctrl.edge_size_ratio = gtewgt/gtvsize;
scopy(incon, iubvec, ctrl.ubvec);
PreAllocateMemory(&ctrl, graph, &wspace);
/***********************/
/* Partition and Remap */
/***********************/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Adaptive_Partition(&ctrl, graph, &wspace);
ParallelReMapGraph(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
idxcopy(graph->nvtxs, graph->where, part);
if (edgecut != NULL)
*edgecut = graph->mincut;
/***********************/
/* Take care of output */
/***********************/
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
if (ctrl.dbglvl&DBG_INFO) {
Mc_ComputeMoveStatistics(&ctrl, graph, &nmoved, &maxin, &maxout);
rprintf(&ctrl, "Final %3d-way Cut: %6d \tBalance: ", inparts, graph->mincut);
avg = 0.0;
for (h=0; h<incon; h++) {
maximb = 0.0;
for (i=0; i<inparts; i++)
maximb = amax(maximb, graph->gnpwgts[i*incon+h]/itpwgts[i*incon+h]);
avg += maximb;
rprintf(&ctrl, "%.3f ", maximb);
}
rprintf(&ctrl, "\nNMoved: %d %d %d %d\n", nmoved, maxin, maxout, maxin+maxout);
}
/*************************************/
/* Free memory, renumber, and return */
/*************************************/
GKfree((void **)&graph->lnpwgts, (void **)&graph->gnpwgts, (void **)&graph->nvwgt, (void **)(&graph->home), (void **)(&graph->vsize), LTERM);
GKfree((void **)&itpwgts, LTERM);
FreeInitialGraphAndRemap(graph, iwgtflag);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
return;
}
