void METIS_NodeRefine(int nvtxs, idxtype *xadj, idxtype *vwgt, idxtype *adjncy,
idxtype *adjwgt, idxtype *where, idxtype *hmarker, float ubfactor)
{
GraphType *graph;
CtrlType ctrl;
ctrl.dbglvl = ONMETIS_DBGLVL;
ctrl.optype = OP_ONMETIS;
graph = CreateGraph();
SetUpGraph(graph, OP_ONMETIS, nvtxs, 1, xadj, adjncy, vwgt, adjwgt, 3);
AllocateWorkSpace(&ctrl, graph, 2);
Allocate2WayNodePartitionMemory(&ctrl, graph);
idxcopy(nvtxs, where, graph->where);
Compute2WayNodePartitionParams(&ctrl, graph);
FM_2WayNodeRefine_OneSidedP(&ctrl, graph, hmarker, ubfactor, 10);
/* FM_2WayNodeRefine_TwoSidedP(&ctrl, graph, hmarker, ubfactor, 10); */
FreeWorkSpace(&ctrl, graph);
idxcopy(nvtxs, graph->where, where);
FreeGraph(graph);
}
/*************************************************************************
* This function takes a bisection and constructs a minimum weight vertex
* separator out of it. It uses the node-based separator refinement for it.
**************************************************************************/
void ConstructSeparator(CtrlType *ctrl, GraphType *graph, float ubfactor)
{
int i, j, k, nvtxs, nbnd;
idxtype *xadj, *where, *bndind;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
nbnd = graph->nbnd;
bndind = graph->bndind;
where = idxcopy(nvtxs, graph->where, idxwspacemalloc(ctrl, nvtxs));
/* Put the nodes in the boundary into the separator */
for (i=0; i<nbnd; i++) {
j = bndind[i];
if (xadj[j+1]-xadj[j] > 0) /* Ignore islands */
where[j] = 2;
}
GKfree(&graph->rdata, LTERM);
Allocate2WayNodePartitionMemory(ctrl, graph);
idxcopy(nvtxs, where, graph->where);
idxwspacefree(ctrl, nvtxs);
ASSERT(IsSeparable(graph));
Compute2WayNodePartitionParams(ctrl, graph);
ASSERT(CheckNodePartitionParams(graph));
FM_2WayNodeRefine(ctrl, graph, ubfactor, 8);
ASSERT(IsSeparable(graph));
}
/******************************************************************************
* This function takes a partition vector that is distributed and reads in
* the original graph and computes the edgecut
*******************************************************************************/
int ComputeRealCut2(idxtype *vtxdist, idxtype *mvtxdist, idxtype *part, idxtype *mpart, char *filename, MPI_Comm comm)
{
int i, j, nvtxs, mype, npes, cut;
idxtype *xadj, *adjncy, *gpart, *gmpart, *perm, *sizes;
MPI_Status status;
MPI_Comm_size(comm, &npes);
MPI_Comm_rank(comm, &mype);
if (mype != 0) {
MPI_Send((void *)part, vtxdist[mype+1]-vtxdist[mype], IDX_DATATYPE, 0, 1, comm);
MPI_Send((void *)mpart, mvtxdist[mype+1]-mvtxdist[mype], IDX_DATATYPE, 0, 1, comm);
}
else { /* Processor 0 does all the rest */
gpart = idxmalloc(vtxdist[npes], "ComputeRealCut: gpart");
idxcopy(vtxdist[1], part, gpart);
gmpart = idxmalloc(mvtxdist[npes], "ComputeRealCut: gmpart");
idxcopy(mvtxdist[1], mpart, gmpart);
for (i=1; i<npes; i++) {
MPI_Recv((void *)(gpart+vtxdist[i]), vtxdist[i+1]-vtxdist[i], IDX_DATATYPE, i, 1, comm, &status);
MPI_Recv((void *)(gmpart+mvtxdist[i]), mvtxdist[i+1]-mvtxdist[i], IDX_DATATYPE, i, 1, comm, &status);
}
/* OK, now go and reconstruct the permutation to go from the graph to mgraph */
perm = idxmalloc(vtxdist[npes], "ComputeRealCut: perm");
sizes = idxsmalloc(npes+1, 0, "ComputeRealCut: sizes");
for (i=0; i<vtxdist[npes]; i++)
sizes[gpart[i]]++;
MAKECSR(i, npes, sizes);
for (i=0; i<vtxdist[npes]; i++)
perm[i] = sizes[gpart[i]]++;
/* Ok, now read the graph from the file */
ReadMetisGraph(filename, &nvtxs, &xadj, &adjncy);
/* OK, now compute the cut */
for (cut=0, i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (gmpart[perm[i]] != gmpart[perm[adjncy[j]]])
cut++;
}
}
cut = cut/2;
GKfree(&gpart, &gmpart, &perm, &sizes, &xadj, &adjncy, LTERM);
return cut;
}
return 0;
}
Matrix* setupCanonicalMatrix(int nvtxs, int nedges, idxtype* xadj,
idxtype* adjncy, idxtype* adjwgt, int ncutify)
{
int i,j;
Matrix* ret;
if ( ncutify )
ret=allocMatrix(nvtxs,nedges,1,0,0);
else
ret=allocMatrix(nvtxs,nedges,0,0,0);
idxcopy(nvtxs+1, xadj, ret->xadj);
idxcopy(nedges, adjncy, ret->adjncy);
if ( adjwgt != NULL )
{
if ( ncutify )
{
for(i=0;i<ret->nvtxs;i++)
{
ret->adjwgtsum[i]=0;
for(j=ret->xadj[i];j<ret->xadj[i+1];j++)
{
ret->adjwgt[j]=(wgttype)adjwgt[j];
ret->adjwgtsum[i]+=ret->adjwgt[j];
}
}
//ncutifyWeights(ret,1,ncutify); //YK removed
}
else
{
for(i=0;i<nedges;i++)
ret->adjwgt[i]=(wgttype)adjwgt[i];
}
normalizeColumns(ret,1,0);
}
else
{
if ( ncutify )
ncutifyWeights(ret,0,ncutify);
normalizeColumns(ret,0,0);
}
// sort each column in ascending order. This is necessary for
// getDprAdjMatrix.
for(i=0;i<nvtxs;i++)
{
ParallelQSort(ret->adjncy,ret->adjwgt,ret->xadj[i],ret->xadj[i+1]-1);
}
return ret;
}
void AllocateNodePartitionParams(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int nparts, nvtxs;
idxtype *vwgt;
NRInfoType *rinfo, *myrinfo;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->KWayInitTmr));
nvtxs = graph->nvtxs;
nparts = ctrl->nparts;
graph->nrinfo = (NRInfoType *)GKmalloc(sizeof(NRInfoType)*nvtxs, "AllocateNodePartitionParams: rinfo");
graph->lpwgts = idxmalloc(2*nparts, "AllocateNodePartitionParams: lpwgts");
graph->gpwgts = idxmalloc(2*nparts, "AllocateNodePartitionParams: gpwgts");
graph->sepind = idxmalloc(nvtxs, "AllocateNodePartitionParams: sepind");
graph->hmarker = idxmalloc(nvtxs, "AllocateNodePartitionParams: hmarker");
/* Allocate additional memory for graph->vwgt in order to store the weights
of the remote vertices */
vwgt = graph->vwgt;
graph->vwgt = idxmalloc(nvtxs+graph->nrecv, "AllocateNodePartitionParams: graph->vwgt");
idxcopy(nvtxs, vwgt, graph->vwgt);
GKfree((void **)&vwgt, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->KWayInitTmr));
}
/*************************************************************************
* This function takes a graph and produces a bisection of it
**************************************************************************/
int MlevelKWayPartitioning(CtrlType *ctrl, GraphType *graph, int nparts, idxtype *part, float *tpwgts, float ubfactor)
{
int i, j, nvtxs, tvwgt, tpwgts2[2];
GraphType *cgraph;
int wgtflag=3, numflag=0, options[10], edgecut;
cgraph = Coarsen2Way(ctrl, graph);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
AllocateKWayPartitionMemory(ctrl, cgraph, nparts);
options[0] = 1;
options[OPTION_CTYPE] = MATCH_SHEMKWAY;
options[OPTION_ITYPE] = IPART_GGPKL;
options[OPTION_RTYPE] = RTYPE_FM;
options[OPTION_DBGLVL] = 0;
METIS_WPartGraphRecursive(&cgraph->nvtxs, cgraph->xadj, cgraph->adjncy, cgraph->vwgt,
cgraph->adjwgt, &wgtflag, &numflag, &nparts, tpwgts, options,
&edgecut, cgraph->where);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
IFSET(ctrl->dbglvl, DBG_IPART, printf("Initial %d-way partitioning cut: %d\n", nparts, edgecut));
IFSET(ctrl->dbglvl, DBG_KWAYPINFO, ComputePartitionInfo(cgraph, nparts, cgraph->where));
RefineKWay(ctrl, graph, cgraph, nparts, tpwgts, ubfactor);
idxcopy(graph->nvtxs, graph->where, part);
GKfree(&graph->gdata, &graph->rdata, LTERM);
return graph->mincut;
}
/***********************************************************************************
* 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 re-adjusts the amount of memory that was allocated if
* it will lead to significant savings
**************************************************************************/
void ReAdjustMemory(GraphType *graph, GraphType *cgraph, int dovsize)
{
if (cgraph->nedges > 100000 && graph->nedges < 0.7*graph->nedges) {
idxcopy(cgraph->nedges, cgraph->adjwgt, cgraph->adjncy+cgraph->nedges);
if (graph->ncon == 1) {
if (dovsize) {
cgraph->gdata = realloc(cgraph->gdata, (5*cgraph->nvtxs+1 + 2*cgraph->nedges)*sizeof(idxtype));
/* Do this, in case everything was copied into new space */
cgraph->xadj = cgraph->gdata;
cgraph->vwgt = cgraph->gdata + cgraph->nvtxs+1;
cgraph->vsize = cgraph->gdata + 2*cgraph->nvtxs+1;
cgraph->adjwgtsum = cgraph->gdata + 3*cgraph->nvtxs+1;
cgraph->cmap = cgraph->gdata + 4*cgraph->nvtxs+1;
cgraph->adjncy = cgraph->gdata + 5*cgraph->nvtxs+1;
cgraph->adjwgt = cgraph->gdata + 5*cgraph->nvtxs+1 + cgraph->nedges;
}
else {
cgraph->gdata = realloc(cgraph->gdata, (4*cgraph->nvtxs+1 + 2*cgraph->nedges)*sizeof(idxtype));
/* Do this, in case everything was copied into new space */
cgraph->xadj = cgraph->gdata;
cgraph->vwgt = cgraph->gdata + cgraph->nvtxs+1;
cgraph->adjwgtsum = cgraph->gdata + 2*cgraph->nvtxs+1;
cgraph->cmap = cgraph->gdata + 3*cgraph->nvtxs+1;
cgraph->adjncy = cgraph->gdata + 4*cgraph->nvtxs+1;
cgraph->adjwgt = cgraph->gdata + 4*cgraph->nvtxs+1 + cgraph->nedges;
}
}
else {
if (dovsize) {
cgraph->gdata = realloc(cgraph->gdata, (4*cgraph->nvtxs+1 + 2*cgraph->nedges)*sizeof(idxtype));
/* Do this, in case everything was copied into new space */
cgraph->xadj = cgraph->gdata;
cgraph->vsize = cgraph->gdata + cgraph->nvtxs+1;
cgraph->adjwgtsum = cgraph->gdata + 2*cgraph->nvtxs+1;
cgraph->cmap = cgraph->gdata + 3*cgraph->nvtxs+1;
cgraph->adjncy = cgraph->gdata + 4*cgraph->nvtxs+1;
cgraph->adjwgt = cgraph->gdata + 4*cgraph->nvtxs+1 + cgraph->nedges;
}
else {
cgraph->gdata = realloc(cgraph->gdata, (3*cgraph->nvtxs+1 + 2*cgraph->nedges)*sizeof(idxtype));
/* Do this, in case everything was copied into new space */
cgraph->xadj = cgraph->gdata;
cgraph->adjwgtsum = cgraph->gdata + cgraph->nvtxs+1;
cgraph->cmap = cgraph->gdata + 2*cgraph->nvtxs+1;
cgraph->adjncy = cgraph->gdata + 3*cgraph->nvtxs+1;
cgraph->adjwgt = cgraph->gdata + 3*cgraph->nvtxs+1 + cgraph->nedges;
}
}
}
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void MocGrowBisection(CtrlType *ctrl, GraphType *graph, float *tpwgts, float ubfactor)
{
int i, j, k, nvtxs, ncon, from, bestcut, mincut, nbfs;
idxtype *bestwhere, *where;
nvtxs = graph->nvtxs;
MocAllocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(graph->nedges, graph->adjwgt);
for (; nbfs>0; nbfs--) {
idxset(nvtxs, 1, where);
where[RandomInRange(nvtxs)] = 0;
MocCompute2WayPartitionParams(ctrl, graph);
MocInit2WayBalance(ctrl, graph, tpwgts);
MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
MocBalance2Way(ctrl, graph, tpwgts, 1.02);
MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
if (bestcut >= graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
/*GKfree(&bestwhere, LTERM);*/
GKfree1((void**)&bestwhere);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void MocGrowBisection2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
int /*i, j, k,*/ nvtxs, /*ncon, from,*/ bestcut, /*mincut,*/ nbfs;
idxtype *bestwhere, *where;
nvtxs = graph->nvtxs;
MocAllocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(graph->nedges, graph->adjwgt);
for (; nbfs>0; nbfs--) {
idxset(nvtxs, 1, where);
where[RandomInRange(nvtxs)] = 0;
MocCompute2WayPartitionParams(ctrl, graph);
MocBalance2Way2(ctrl, graph, tpwgts, ubvec);
MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4);
MocBalance2Way2(ctrl, graph, tpwgts, ubvec);
MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4);
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
GKfree((void**)&bestwhere, LTERM);
}
/*************************************************************************
* This function is the entry point for ONWMETIS. It requires weights on the
* vertices. It is for the case that the matrix has been pre-compressed.
**************************************************************************/
void METIS_EdgeComputeSeparator(int *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *adjwgt, int *options, int *sepsize, idxtype *part)
{
int i, j, tvwgt, tpwgts[2];
GraphType graph;
CtrlType ctrl;
SetUpGraph(&graph, OP_ONMETIS, *nvtxs, 1, xadj, adjncy, vwgt, adjwgt, 3);
tvwgt = idxsum(*nvtxs, graph.vwgt);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = ONMETIS_CTYPE;
ctrl.IType = ONMETIS_ITYPE;
ctrl.RType = ONMETIS_RTYPE;
ctrl.dbglvl = ONMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.oflags = 0;
ctrl.pfactor = 0;
ctrl.nseps = 5;
ctrl.optype = OP_OEMETIS;
ctrl.CoarsenTo = amin(100, *nvtxs-1);
ctrl.maxvwgt = 1.5*tvwgt/ctrl.CoarsenTo;
InitRandom(options[7]);
AllocateWorkSpace(&ctrl, &graph, 2);
/*============================================================
* Perform the bisection
*============================================================*/
tpwgts[0] = tvwgt/2;
tpwgts[1] = tvwgt-tpwgts[0];
MlevelEdgeBisection(&ctrl, &graph, tpwgts, 1.05);
ConstructMinCoverSeparator(&ctrl, &graph, 1.05);
*sepsize = graph.pwgts[2];
idxcopy(*nvtxs, graph.where, part);
GKfree((void**)&graph.gdata, &graph.rdata, &graph.label, LTERM);
FreeWorkSpace(&ctrl, &graph);
}
/*************************************************************************
* This function is the entry point for ONWMETIS. It requires weights on the
* vertices. It is for the case that the matrix has been pre-compressed.
**************************************************************************/
void METIS_NodeComputeSeparator(int *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *adjwgt, float *ubfactor, int *options, int *sepsize, idxtype *part)
{
int i, j, tvwgt, tpwgts[2];
GraphType graph;
CtrlType ctrl;
SetUpGraph(&graph, OP_ONMETIS, *nvtxs, 1, xadj, adjncy, vwgt, adjwgt, 3);
tvwgt = idxsum(*nvtxs, graph.vwgt);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = ONMETIS_CTYPE;
ctrl.IType = ONMETIS_ITYPE;
ctrl.RType = ONMETIS_RTYPE;
ctrl.dbglvl = ONMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.oflags = OFLAG_COMPRESS; /* For by-passing the pre-coarsening for multiple runs */
ctrl.RType = 2; /* Standard 1-sided node refinement code */
ctrl.pfactor = 0;
ctrl.nseps = 5; /* This should match NUM_INIT_MSECTIONS in ParMETISLib/defs.h */
ctrl.optype = OP_ONMETIS;
InitRandom(options[7]);
AllocateWorkSpace(&ctrl, &graph, 2);
/*============================================================
* Perform the bisection
*============================================================*/
tpwgts[0] = tvwgt/2;
tpwgts[1] = tvwgt-tpwgts[0];
MlevelNodeBisectionMultiple(&ctrl, &graph, tpwgts, *ubfactor*.95);
*sepsize = graph.pwgts[2];
idxcopy(*nvtxs, graph.where, part);
GKfree((void **)&graph.gdata, &graph.rdata, &graph.label, LTERM);
FreeWorkSpace(&ctrl, &graph);
}
/*************************************************************************
* This function is the entry point for ONWMETIS. It requires weights on the
* vertices. It is for the case that the matrix has been pre-compressed.
**************************************************************************/
void METIS_NodeComputeSeparator(idxtype *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *adjwgt, idxtype *options, idxtype *sepsize, idxtype *part)
{
idxtype i, j, tvwgt, tpwgts[2];
GraphType graph;
CtrlType ctrl;
SetUpGraph(&graph, OP_ONMETIS, *nvtxs, 1, xadj, adjncy, vwgt, adjwgt, 3);
tvwgt = idxsum(*nvtxs, graph.vwgt, 1);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = ONMETIS_CTYPE;
ctrl.IType = ONMETIS_ITYPE;
ctrl.RType = ONMETIS_RTYPE;
ctrl.dbglvl = ONMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.oflags = 0;
ctrl.pfactor = 0;
ctrl.nseps = 3;
ctrl.optype = OP_ONMETIS;
ctrl.CoarsenTo = amin(100, *nvtxs-1);
ctrl.maxvwgt = 1.5*tvwgt/ctrl.CoarsenTo;
InitRandom(options[7]);
AllocateWorkSpace(&ctrl, &graph, 2);
/*============================================================
* Perform the bisection
*============================================================*/
tpwgts[0] = tvwgt/2;
tpwgts[1] = tvwgt-tpwgts[0];
MlevelNodeBisectionMultiple(&ctrl, &graph, tpwgts, 1.02);
*sepsize = graph.pwgts[2];
idxcopy(*nvtxs, graph.where, part);
FreeGraph(&graph, 0);
FreeWorkSpace(&ctrl, &graph);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void MocGrowBisectionNew2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
idxtype i, j, k, nvtxs, ncon, from, bestcut, mincut, nbfs, inbfs;
idxtype *bestwhere, *where;
nvtxs = graph->nvtxs;
MocAllocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
for (inbfs=0; inbfs<nbfs; inbfs++) {
idxset(nvtxs, 1, where);
where[RandomInRange(nvtxs)] = 0;
MocCompute2WayPartitionParams(ctrl, graph);
MocInit2WayBalance2(ctrl, graph, tpwgts, ubvec);
MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4);
if (inbfs == 0 || bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
gk_free((void **)&bestwhere, LTERM);
}
/*************************************************************************
* This function takes a graph and produces a bisection of it
**************************************************************************/
int MCMlevelKWayPartitioning(CtrlType *ctrl, GraphType *graph, int nparts, idxtype *part,
float *rubvec)
{
int i, j, nvtxs;
GraphType *cgraph;
int options[10], edgecut;
cgraph = MCCoarsen2Way(ctrl, graph);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
MocAllocateKWayPartitionMemory(ctrl, cgraph, nparts);
options[0] = 1;
options[OPTION_CTYPE] = MATCH_SBHEM_INFNORM;
options[OPTION_ITYPE] = IPART_RANDOM;
options[OPTION_RTYPE] = RTYPE_FM;
options[OPTION_DBGLVL] = 0;
/* Determine what you will use as the initial partitioner, based on tolerances */
for (i=0; i<graph->ncon; i++) {
if (rubvec[i] > 1.2)
break;
}
if (i == graph->ncon)
METIS_mCPartGraphRecursiveInternal(&cgraph->nvtxs, &cgraph->ncon,
cgraph->xadj, cgraph->adjncy, cgraph->nvwgt, cgraph->adjwgt, &nparts,
options, &edgecut, cgraph->where);
else
METIS_mCHPartGraphRecursiveInternal(&cgraph->nvtxs, &cgraph->ncon,
cgraph->xadj, cgraph->adjncy, cgraph->nvwgt, cgraph->adjwgt, &nparts,
rubvec, options, &edgecut, cgraph->where);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
IFSET(ctrl->dbglvl, DBG_IPART, printf("Initial %d-way partitioning cut: %d\n", nparts, edgecut));
IFSET(ctrl->dbglvl, DBG_KWAYPINFO, ComputePartitionInfo(cgraph, nparts, cgraph->where));
MocRefineKWayHorizontal(ctrl, graph, cgraph, nparts, rubvec);
idxcopy(graph->nvtxs, graph->where, part);
GKfree(&graph->nvwgt, &graph->npwgts, &graph->gdata, &graph->rdata, LTERM);
return graph->mincut;
}
/***********************************************************************************
* 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 is the entry point for detecting contacts between
* bounding boxes and surface nodes
**************************************************************************/
void METIS_FindContacts(void *raw_cinfo, idxtype *nboxes, double *boxcoords, idxtype *nparts,
idxtype **r_cntptr, idxtype **r_cntind)
{
idxtype i, ncnts, tncnts, maxtncnts;
idxtype *cntptr, *cntind, *auxcntind, *stack, *marker;
ContactInfoType *cinfo;
cinfo = (ContactInfoType *)raw_cinfo;
maxtncnts = 6*(*nboxes);
cntptr = idxsmalloc(*nboxes+1, 0, "METIS_FindContacts: cntptr");
cntind = idxmalloc(maxtncnts, "METIS_FindContacts: cntind");
auxcntind = idxmalloc(*nparts, "METIS_FindContacts: auxcntind");
stack = idxmalloc(cinfo->nnodes, "METIS_FindContacts: stack");
marker = idxsmalloc(*nparts, 0, "METIS_FindContacts: marker");
/* Go through each box and determine its contacting partitions */
for (tncnts=0, i=0; i<*nboxes; i++) {
ncnts = FindBoxContacts(cinfo, boxcoords+i*6, stack, auxcntind, marker);
if (ncnts == 0)
mprintf("CSearchError: Box has no contacts!\n");
if (ncnts + tncnts >= maxtncnts) {
maxtncnts += (tncnts+ncnts)*(*nboxes-i)/i;
if ((cntind = (idxtype *)realloc(cntind, maxtncnts*sizeof(idxtype))) == NULL)
errexit("Realloc failed! of %d words!\n", maxtncnts);
}
cntptr[i] = ncnts;
idxcopy(ncnts, auxcntind, cntind+tncnts);
tncnts += ncnts;
}
MAKECSR(i, *nboxes, cntptr);
*r_cntptr = cntptr;
*r_cntind = cntind;
gk_free((void **)&auxcntind, &stack, &marker, LTERM);
}
/******************************************************************************
* This function takes a partition vector that is distributed and reads in
* the original graph and computes the edgecut
*******************************************************************************/
int ComputeRealCut(idxtype *vtxdist, idxtype *part, char *filename, MPI_Comm comm)
{
int i, j, nvtxs, mype, npes, cut;
idxtype *xadj, *adjncy, *gpart;
MPI_Status status;
MPI_Comm_size(comm, &npes);
MPI_Comm_rank(comm, &mype);
if (mype != 0) {
MPI_Send((void *)part, vtxdist[mype+1]-vtxdist[mype], IDX_DATATYPE, 0, 1, comm);
}
else { /* Processor 0 does all the rest */
gpart = idxmalloc(vtxdist[npes], "ComputeRealCut: gpart");
idxcopy(vtxdist[1], part, gpart);
for (i=1; i<npes; i++)
MPI_Recv((void *)(gpart+vtxdist[i]), vtxdist[i+1]-vtxdist[i], IDX_DATATYPE, i, 1, comm, &status);
ReadMetisGraph(filename, &nvtxs, &xadj, &adjncy);
/* OK, now compute the cut */
for (cut=0, i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (gpart[i] != gpart[adjncy[j]])
cut++;
}
}
cut = cut/2;
GKfree(&gpart, &xadj, &adjncy, LTERM);
return cut;
}
return 0;
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Moc_KWayFM(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace, int npasses)
{
int h, i, ii, iii, j, k, c;
int pass, nvtxs, nedges, ncon;
int nmoves, nmoved, nswaps, nzgswaps;
/* int gnswaps, gnzgswaps; */
int me, firstvtx, lastvtx, yourlastvtx;
int from, to = -1, oldto, oldcut, mydomain, yourdomain, imbalanced, overweight;
int npes = ctrl->npes, mype = ctrl->mype, nparts = ctrl->nparts;
int nlupd, nsupd, nnbrs, nchanged;
idxtype *xadj, *ladjncy, *adjwgt, *vtxdist;
idxtype *where, *tmp_where, *moved;
floattype *lnpwgts, *gnpwgts, *ognpwgts, *pgnpwgts, *movewgts, *overfill;
idxtype *update, *supdate, *rupdate, *pe_updates;
idxtype *changed, *perm, *pperm, *htable;
idxtype *peind, *recvptr, *sendptr;
KeyValueType *swchanges, *rwchanges;
RInfoType *rinfo, *myrinfo, *tmp_myrinfo, *tmp_rinfo;
EdgeType *tmp_edegrees, *my_edegrees, *your_edegrees;
floattype lbvec[MAXNCON], *nvwgt, *badmaxpwgt, *ubvec, *tpwgts, lbavg, ubavg;
int *nupds_pe;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->KWayTmr));
/*************************/
/* set up common aliases */
/*************************/
nvtxs = graph->nvtxs;
nedges = graph->nedges;
ncon = graph->ncon;
vtxdist = graph->vtxdist;
xadj = graph->xadj;
ladjncy = graph->adjncy;
adjwgt = graph->adjwgt;
firstvtx = vtxdist[mype];
lastvtx = vtxdist[mype+1];
where = graph->where;
rinfo = graph->rinfo;
lnpwgts = graph->lnpwgts;
gnpwgts = graph->gnpwgts;
ubvec = ctrl->ubvec;
tpwgts = ctrl->tpwgts;
nnbrs = graph->nnbrs;
peind = graph->peind;
recvptr = graph->recvptr;
sendptr = graph->sendptr;
changed = idxmalloc(nvtxs, "KWR: changed");
rwchanges = wspace->pairs;
swchanges = rwchanges + recvptr[nnbrs];
/************************************/
/* set up important data structures */
/************************************/
perm = idxmalloc(nvtxs, "KWR: perm");
pperm = idxmalloc(nparts, "KWR: pperm");
update = idxmalloc(nvtxs, "KWR: update");
supdate = wspace->indices;
rupdate = supdate + recvptr[nnbrs];
nupds_pe = imalloc(npes, "KWR: nupds_pe");
htable = idxsmalloc(nvtxs+graph->nrecv, 0, "KWR: lhtable");
badmaxpwgt = fmalloc(nparts*ncon, "badmaxpwgt");
for (i=0; i<nparts; i++) {
for (h=0; h<ncon; h++) {
badmaxpwgt[i*ncon+h] = ubvec[h]*tpwgts[i*ncon+h];
}
}
movewgts = fmalloc(nparts*ncon, "KWR: movewgts");
ognpwgts = fmalloc(nparts*ncon, "KWR: ognpwgts");
pgnpwgts = fmalloc(nparts*ncon, "KWR: pgnpwgts");
overfill = fmalloc(nparts*ncon, "KWR: overfill");
moved = idxmalloc(nvtxs, "KWR: moved");
tmp_where = idxmalloc(nvtxs+graph->nrecv, "KWR: tmp_where");
tmp_rinfo = (RInfoType *)GKmalloc(sizeof(RInfoType)*nvtxs, "KWR: tmp_rinfo");
tmp_edegrees = (EdgeType *)GKmalloc(sizeof(EdgeType)*nedges, "KWR: tmp_edegrees");
idxcopy(nvtxs+graph->nrecv, where, tmp_where);
for (i=0; i<nvtxs; i++) {
tmp_rinfo[i].id = rinfo[i].id;
tmp_rinfo[i].ed = rinfo[i].ed;
tmp_rinfo[i].ndegrees = rinfo[i].ndegrees;
tmp_rinfo[i].degrees = tmp_edegrees+xadj[i];
for (j=0; j<rinfo[i].ndegrees; j++) {
tmp_rinfo[i].degrees[j].edge = rinfo[i].degrees[j].edge;
tmp_rinfo[i].degrees[j].ewgt = rinfo[i].degrees[j].ewgt;
}
}
nswaps = nzgswaps = 0;
/*********************************************************/
/* perform a small number of passes through the vertices */
/*********************************************************/
for (pass=0; pass<npasses; pass++) {
if (mype == 0)
RandomPermute(nparts, pperm, 1);
MPI_Bcast((void *)pperm, nparts, IDX_DATATYPE, 0, ctrl->comm);
FastRandomPermute(nvtxs, perm, 1);
oldcut = graph->mincut;
/* check to see if the partitioning is imbalanced */
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
ubavg = savg(ncon, ubvec);
lbavg = savg(ncon, lbvec);
imbalanced = (lbavg > ubavg) ? 1 : 0;
for (c=0; c<2; c++) {
scopy(ncon*nparts, gnpwgts, ognpwgts);
sset(ncon*nparts, 0.0, movewgts);
nmoved = 0;
/**********************************************/
/* PASS ONE -- record stats for desired moves */
/**********************************************/
for (iii=0; iii<nvtxs; iii++) {
i = perm[iii];
from = tmp_where[i];
nvwgt = graph->nvwgt+i*ncon;
for (h=0; h<ncon; h++)
if (fabs(nvwgt[h]-gnpwgts[from*ncon+h]) < SMALLFLOAT)
break;
if (h < ncon) {
continue;
}
/* check for a potential improvement */
if (tmp_rinfo[i].ed >= tmp_rinfo[i].id) {
my_edegrees = tmp_rinfo[i].degrees;
for (k=0; k<tmp_rinfo[i].ndegrees; k++) {
to = my_edegrees[k].edge;
if (ProperSide(c, pperm[from], pperm[to])) {
for (h=0; h<ncon; h++)
if (gnpwgts[to*ncon+h]+nvwgt[h] > badmaxpwgt[to*ncon+h] && nvwgt[h] > 0.0)
break;
if (h == ncon)
break;
}
}
oldto = to;
/* check if a subdomain was found that fits */
if (k < tmp_rinfo[i].ndegrees) {
for (j=k+1; j<tmp_rinfo[i].ndegrees; j++) {
to = my_edegrees[j].edge;
if (ProperSide(c, pperm[from], pperm[to])) {
for (h=0; h<ncon; h++)
if (gnpwgts[to*ncon+h]+nvwgt[h] > badmaxpwgt[to*ncon+h] && nvwgt[h] > 0.0)
break;
if (h == ncon) {
if (my_edegrees[j].ewgt > my_edegrees[k].ewgt ||
(my_edegrees[j].ewgt == my_edegrees[k].ewgt &&
IsHBalanceBetterTT(ncon,gnpwgts+oldto*ncon,gnpwgts+to*ncon,nvwgt,ubvec))){
k = j;
oldto = my_edegrees[k].edge;
}
}
}
}
to = oldto;
if (my_edegrees[k].ewgt > tmp_rinfo[i].id ||
(my_edegrees[k].ewgt == tmp_rinfo[i].id &&
(imbalanced || graph->level > 3 || iii % 8 == 0) &&
IsHBalanceBetterFT(ncon,gnpwgts+from*ncon,gnpwgts+to*ncon,nvwgt,ubvec))){
/****************************************/
/* Update tmp arrays of the moved vertex */
/****************************************/
tmp_where[i] = to;
moved[nmoved++] = i;
for (h=0; h<ncon; h++) {
lnpwgts[to*ncon+h] += nvwgt[h];
lnpwgts[from*ncon+h] -= nvwgt[h];
gnpwgts[to*ncon+h] += nvwgt[h];
gnpwgts[from*ncon+h] -= nvwgt[h];
movewgts[to*ncon+h] += nvwgt[h];
movewgts[from*ncon+h] -= nvwgt[h];
}
tmp_rinfo[i].ed += tmp_rinfo[i].id-my_edegrees[k].ewgt;
SWAP(tmp_rinfo[i].id, my_edegrees[k].ewgt, j);
if (my_edegrees[k].ewgt == 0) {
tmp_rinfo[i].ndegrees--;
my_edegrees[k].edge = my_edegrees[tmp_rinfo[i].ndegrees].edge;
my_edegrees[k].ewgt = my_edegrees[tmp_rinfo[i].ndegrees].ewgt;
}
else {
my_edegrees[k].edge = from;
}
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
/* no need to bother about vertices on different pe's */
if (ladjncy[j] >= nvtxs)
continue;
me = ladjncy[j];
mydomain = tmp_where[me];
myrinfo = tmp_rinfo+me;
your_edegrees = myrinfo->degrees;
if (mydomain == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
}
else {
if (mydomain == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
}
}
/* Remove contribution from the .ed of 'from' */
if (mydomain != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == from) {
if (your_edegrees[k].ewgt == adjwgt[j]) {
myrinfo->ndegrees--;
your_edegrees[k].edge = your_edegrees[myrinfo->ndegrees].edge;
your_edegrees[k].ewgt = your_edegrees[myrinfo->ndegrees].ewgt;
}
else {
your_edegrees[k].ewgt -= adjwgt[j];
}
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (mydomain != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == to) {
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
your_edegrees[myrinfo->ndegrees].edge = to;
your_edegrees[myrinfo->ndegrees++].ewgt = adjwgt[j];
}
}
}
}
}
}
}
/******************************************/
/* Let processors know the subdomain wgts */
/* if all proposed moves commit. */
/******************************************/
MPI_Allreduce((void *)lnpwgts, (void *)pgnpwgts, nparts*ncon,
MPI_DOUBLE, MPI_SUM, ctrl->comm);
/**************************/
/* compute overfill array */
/**************************/
overweight = 0;
for (j=0; j<nparts; j++) {
for (h=0; h<ncon; h++) {
if (pgnpwgts[j*ncon+h] > ognpwgts[j*ncon+h]) {
overfill[j*ncon+h] =
(pgnpwgts[j*ncon+h]-badmaxpwgt[j*ncon+h]) /
(pgnpwgts[j*ncon+h]-ognpwgts[j*ncon+h]);
}
else {
overfill[j*ncon+h] = 0.0;
}
overfill[j*ncon+h] = amax(overfill[j*ncon+h], 0.0);
overfill[j*ncon+h] *= movewgts[j*ncon+h];
if (overfill[j*ncon+h] > 0.0)
overweight = 1;
ASSERTP(ctrl, ognpwgts[j*ncon+h] <= badmaxpwgt[j*ncon+h] ||
pgnpwgts[j*ncon+h] <= ognpwgts[j*ncon+h],
(ctrl, "%.4f %.4f %.4f\n", ognpwgts[j*ncon+h],
badmaxpwgt[j*ncon+h], pgnpwgts[j*ncon+h]));
}
}
/****************************************************/
/* select moves to undo according to overfill array */
/****************************************************/
if (overweight == 1) {
for (iii=0; iii<nmoved; iii++) {
i = moved[iii];
oldto = tmp_where[i];
nvwgt = graph->nvwgt+i*ncon;
my_edegrees = tmp_rinfo[i].degrees;
for (k=0; k<tmp_rinfo[i].ndegrees; k++)
if (my_edegrees[k].edge == where[i])
break;
for (h=0; h<ncon; h++)
if (nvwgt[h] > 0.0 && overfill[oldto*ncon+h] > nvwgt[h]/4.0)
break;
/**********************************/
/* nullify this move if necessary */
/**********************************/
if (k != tmp_rinfo[i].ndegrees && h != ncon) {
moved[iii] = -1;
from = oldto;
to = where[i];
for (h=0; h<ncon; h++) {
overfill[oldto*ncon+h] = amax(overfill[oldto*ncon+h]-nvwgt[h], 0.0);
}
tmp_where[i] = to;
tmp_rinfo[i].ed += tmp_rinfo[i].id-my_edegrees[k].ewgt;
SWAP(tmp_rinfo[i].id, my_edegrees[k].ewgt, j);
if (my_edegrees[k].ewgt == 0) {
tmp_rinfo[i].ndegrees--;
my_edegrees[k].edge = my_edegrees[tmp_rinfo[i].ndegrees].edge;
my_edegrees[k].ewgt = my_edegrees[tmp_rinfo[i].ndegrees].ewgt;
}
else {
my_edegrees[k].edge = from;
}
for (h=0; h<ncon; h++) {
lnpwgts[to*ncon+h] += nvwgt[h];
lnpwgts[from*ncon+h] -= nvwgt[h];
}
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
/* no need to bother about vertices on different pe's */
if (ladjncy[j] >= nvtxs)
continue;
me = ladjncy[j];
mydomain = tmp_where[me];
myrinfo = tmp_rinfo+me;
your_edegrees = myrinfo->degrees;
if (mydomain == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
}
else {
if (mydomain == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
}
}
/* Remove contribution from the .ed of 'from' */
if (mydomain != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == from) {
if (your_edegrees[k].ewgt == adjwgt[j]) {
myrinfo->ndegrees--;
your_edegrees[k].edge = your_edegrees[myrinfo->ndegrees].edge;
your_edegrees[k].ewgt = your_edegrees[myrinfo->ndegrees].ewgt;
}
else {
your_edegrees[k].ewgt -= adjwgt[j];
}
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (mydomain != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == to) {
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
your_edegrees[myrinfo->ndegrees].edge = to;
your_edegrees[myrinfo->ndegrees++].ewgt = adjwgt[j];
}
}
}
}
}
}
/*************************************************/
/* PASS TWO -- commit the remainder of the moves */
/*************************************************/
nlupd = nsupd = nmoves = nchanged = 0;
for (iii=0; iii<nmoved; iii++) {
i = moved[iii];
if (i == -1)
continue;
where[i] = tmp_where[i];
/* Make sure to update the vertex information */
if (htable[i] == 0) {
/* make sure you do the update */
htable[i] = 1;
update[nlupd++] = i;
}
/* Put the vertices adjacent to i into the update array */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = ladjncy[j];
if (htable[k] == 0) {
htable[k] = 1;
if (k<nvtxs)
update[nlupd++] = k;
else
supdate[nsupd++] = k;
}
}
nmoves++;
nswaps++;
/* check number of zero-gain moves */
for (k=0; k<rinfo[i].ndegrees; k++)
if (rinfo[i].degrees[k].edge == to)
break;
if (rinfo[i].id == rinfo[i].degrees[k].ewgt)
nzgswaps++;
if (graph->pexadj[i+1]-graph->pexadj[i] > 0)
changed[nchanged++] = i;
}
/* Tell interested pe's the new where[] info for the interface vertices */
CommChangedInterfaceData(ctrl, graph, nchanged, changed, where,
swchanges, rwchanges, wspace->pv4);
IFSET(ctrl->dbglvl, DBG_RMOVEINFO,
rprintf(ctrl, "\t[%d %d], [%.4f], [%d %d %d]\n",
pass, c, badmaxpwgt[0],
GlobalSESum(ctrl, nmoves),
GlobalSESum(ctrl, nsupd),
GlobalSESum(ctrl, nlupd)));
/*-------------------------------------------------------------
/ Time to communicate with processors to send the vertices
/ whose degrees need to be update.
/-------------------------------------------------------------*/
/* Issue the receives first */
for (i=0; i<nnbrs; i++) {
MPI_Irecv((void *)(rupdate+sendptr[i]), sendptr[i+1]-sendptr[i], IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. This needs some preporcessing */
for (i=0; i<nsupd; i++) {
htable[supdate[i]] = 0;
supdate[i] = graph->imap[supdate[i]];
}
iidxsort(nsupd, supdate);
for (j=i=0; i<nnbrs; i++) {
yourlastvtx = vtxdist[peind[i]+1];
for (k=j; k<nsupd && supdate[k] < yourlastvtx; k++);
MPI_Isend((void *)(supdate+j), k-j, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i);
j = k;
}
/* OK, now get into the loop waiting for the send/recv operations to finish */
MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++)
MPI_Get_count(ctrl->statuses+i, IDX_DATATYPE, nupds_pe+i);
MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/*-------------------------------------------------------------
/ Place the recieved to-be updated vertices into update[]
/-------------------------------------------------------------*/
for (i=0; i<nnbrs; i++) {
pe_updates = rupdate+sendptr[i];
for (j=0; j<nupds_pe[i]; j++) {
k = pe_updates[j];
if (htable[k-firstvtx] == 0) {
htable[k-firstvtx] = 1;
update[nlupd++] = k-firstvtx;
}
}
}
/*-------------------------------------------------------------
/ Update the rinfo of the vertices in the update[] array
/-------------------------------------------------------------*/
for (ii=0; ii<nlupd; ii++) {
i = update[ii];
ASSERT(ctrl, htable[i] == 1);
htable[i] = 0;
mydomain = where[i];
myrinfo = rinfo+i;
tmp_myrinfo = tmp_rinfo+i;
my_edegrees = myrinfo->degrees;
your_edegrees = tmp_myrinfo->degrees;
graph->lmincut -= myrinfo->ed;
myrinfo->ndegrees = 0;
myrinfo->id = 0;
myrinfo->ed = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
yourdomain = where[ladjncy[j]];
if (mydomain != yourdomain) {
myrinfo->ed += adjwgt[j];
for (k=0; k<myrinfo->ndegrees; k++) {
if (my_edegrees[k].edge == yourdomain) {
my_edegrees[k].ewgt += adjwgt[j];
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
my_edegrees[k].edge = yourdomain;
my_edegrees[k].ewgt = adjwgt[j];
your_edegrees[k].edge = yourdomain;
your_edegrees[k].ewgt = adjwgt[j];
myrinfo->ndegrees++;
}
ASSERT(ctrl, myrinfo->ndegrees <= xadj[i+1]-xadj[i]);
ASSERT(ctrl, tmp_myrinfo->ndegrees <= xadj[i+1]-xadj[i]);
}
else {
myrinfo->id += adjwgt[j];
}
}
graph->lmincut += myrinfo->ed;
tmp_myrinfo->id = myrinfo->id;
tmp_myrinfo->ed = myrinfo->ed;
tmp_myrinfo->ndegrees = myrinfo->ndegrees;
}
/* finally, sum-up the partition weights */
MPI_Allreduce((void *)lnpwgts, (void *)gnpwgts, nparts*ncon,
MPI_DOUBLE, MPI_SUM, ctrl->comm);
}
graph->mincut = GlobalSESum(ctrl, graph->lmincut)/2;
if (graph->mincut == oldcut)
break;
}
/*
gnswaps = GlobalSESum(ctrl, nswaps);
gnzgswaps = GlobalSESum(ctrl, nzgswaps);
if (mype == 0)
printf("niters: %d, nswaps: %d, nzgswaps: %d\n", pass+1, gnswaps, gnzgswaps);
*/
GKfree((void **)&badmaxpwgt, (void **)&update, (void **)&nupds_pe, (void **)&htable, LTERM);
GKfree((void **)&changed, (void **)&pperm, (void **)&perm, (void **)&moved, LTERM);
GKfree((void **)&pgnpwgts, (void **)&ognpwgts, (void **)&overfill, (void **)&movewgts, LTERM);
GKfree((void **)&tmp_where, (void **)&tmp_rinfo, (void **)&tmp_edegrees, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->KWayTmr));
}
/*************************************************************************
* This function projects a partition, and at the same time computes the
* parameters for refinement.
**************************************************************************/
void Project2WayPartition(CtrlType *ctrl, GraphType *graph)
{
int i, j, k, nvtxs, nbnd, me;
idxtype *xadj, *adjncy, *adjwgt, *adjwgtsum;
idxtype *cmap, *where, *id, *ed, *bndptr, *bndind;
idxtype *cwhere, *cid, *ced, *cbndptr;
GraphType *cgraph;
cgraph = graph->coarser;
cwhere = cgraph->where;
cid = cgraph->id;
ced = cgraph->ed;
cbndptr = cgraph->bndptr;
nvtxs = graph->nvtxs;
cmap = graph->cmap;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
Allocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
id = idxset(nvtxs, 0, graph->id);
ed = idxset(nvtxs, 0, graph->ed);
bndptr = idxset(nvtxs, -1, graph->bndptr);
bndind = graph->bndind;
/* Go through and project partition and compute id/ed for the nodes */
for (i=0; i<nvtxs; i++) {
k = cmap[i];
where[i] = cwhere[k];
cmap[i] = cbndptr[k];
}
for (nbnd=0, i=0; i<nvtxs; i++) {
me = where[i];
id[i] = adjwgtsum[i];
if (xadj[i] == xadj[i+1]) {
bndptr[i] = nbnd;
bndind[nbnd++] = i;
}
else {
if (cmap[i] != -1) { /* If it is an interface node. Note that cmap[i] = cbndptr[cmap[i]] */
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (me != where[adjncy[j]])
ed[i] += adjwgt[j];
}
id[i] -= ed[i];
if (ed[i] > 0 || xadj[i] == xadj[i+1]) {
bndptr[i] = nbnd;
bndind[nbnd++] = i;
}
}
}
}
graph->mincut = cgraph->mincut;
graph->nbnd = nbnd;
idxcopy(2, cgraph->pwgts, graph->pwgts);
FreeGraph(graph->coarser);
graph->coarser = NULL;
}
/*************************************************************************
* This function performs the gather/scatter for the boundary vertices
**************************************************************************/
void CommChangedInterfaceData(CtrlType *ctrl, GraphType *graph,
int nchanged, idxtype *changed, idxtype *data,
KeyValueType *sendpairs, KeyValueType *recvpairs, idxtype *psendptr)
{
int i, j, k, n, penum, nnbrs, firstvtx, nrecv;
idxtype *peind, *sendptr, *recvptr, *recvind, *pexadj, *peadjncy, *peadjloc;
KeyValueType *pairs;
firstvtx = graph->vtxdist[ctrl->mype];
nnbrs = graph->nnbrs;
nrecv = graph->nrecv;
peind = graph->peind;
sendptr = graph->sendptr;
recvptr = graph->recvptr;
recvind = graph->recvind;
pexadj = graph->pexadj;
peadjncy = graph->peadjncy;
peadjloc = graph->peadjloc;
/* Issue the receives first */
for (i=0; i<nnbrs; i++) {
MPI_Irecv((void *)(recvpairs+recvptr[i]), 2*(recvptr[i+1]-recvptr[i]), IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
if (nchanged != 0) {
idxcopy(ctrl->npes, sendptr, psendptr);
/* Copy the changed values into the sendvector */
for (i=0; i<nchanged; i++) {
j = changed[i];
for (k=pexadj[j]; k<pexadj[j+1]; k++) {
penum = peadjncy[k];
sendpairs[psendptr[penum]].key = peadjloc[k];
sendpairs[psendptr[penum]].val = data[j];
psendptr[penum]++;
}
}
for (i=0; i<nnbrs; i++) {
MPI_Isend((void *)(sendpairs+sendptr[i]), 2*(psendptr[i]-sendptr[i]), IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->sreq+i);
}
}
else {
for (i=0; i<nnbrs; i++)
MPI_Isend((void *)(sendpairs), 0, IDX_DATATYPE, 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++) {
MPI_Wait(ctrl->rreq+i, &(ctrl->status));
MPI_Get_count(&ctrl->status, IDX_DATATYPE, &n);
if (n != 0) {
n = n/2;
pairs = recvpairs+graph->recvptr[i];
for (k=0; k<n; k++)
data[pairs[k].key] = pairs[k].val;
}
}
MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void MocRandomBisection(CtrlType *ctrl, GraphType *graph, float *tpwgts, float ubfactor)
{
int i, ii, j, k, nvtxs, ncon, from, bestcut, mincut, nbfs, qnum;
idxtype *bestwhere, *where, *perm;
int counts[MAXNCON];
float *nvwgt;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
nvwgt = graph->nvwgt;
MocAllocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(graph->nedges, graph->adjwgt);
perm = idxmalloc(nvtxs, "BisectGraph: perm");
for (; nbfs>0; nbfs--) {
for (i=0; i<ncon; i++)
counts[i] = 0;
RandomPermute(nvtxs, perm, 1);
/* Partition by spliting the queues randomly */
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
qnum = samax(ncon, nvwgt+i*ncon);
where[i] = counts[qnum];
counts[qnum] = (counts[qnum]+1)%2;
}
MocCompute2WayPartitionParams(ctrl, graph);
MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 6);
MocBalance2Way(ctrl, graph, tpwgts, 1.02);
MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 6);
MocBalance2Way(ctrl, graph, tpwgts, 1.02);
MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 6);
/*
printf("Edgecut: %6d, NPwgts: [", graph->mincut);
for (i=0; i<graph->ncon; i++)
printf("(%.3f %.3f) ", graph->npwgts[i], graph->npwgts[graph->ncon+i]);
printf("]\n");
*/
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
GKfree((void**)&bestwhere, &perm, LTERM);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void RandomBisection(CtrlType *ctrl, GraphType *graph, int *tpwgts, float ubfactor)
{
int i, ii, j, k, nvtxs, pwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where;
idxtype *perm, *bestwhere;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
Allocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
perm = idxmalloc(nvtxs, "BisectGraph: queue");
ASSERTP(tpwgts[0]+tpwgts[1] == idxsum(nvtxs, vwgt), ("%d %d\n", tpwgts[0]+tpwgts[1], idxsum(nvtxs, vwgt)));
maxpwgt[0] = ubfactor*tpwgts[0];
maxpwgt[1] = ubfactor*tpwgts[1];
minpwgt[0] = (1.0/ubfactor)*tpwgts[0];
minpwgt[1] = (1.0/ubfactor)*tpwgts[1];
nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(nvtxs, graph->adjwgtsum)+1; /* The +1 is for the 0 edges case */
for (; nbfs>0; nbfs--) {
RandomPermute(nvtxs, perm, 1);
idxset(nvtxs, 1, where);
pwgts[1] = tpwgts[0]+tpwgts[1];
pwgts[0] = 0;
if (nbfs != 1) {
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (pwgts[0]+vwgt[i] < maxpwgt[0]) {
where[i] = 0;
pwgts[0] += vwgt[i];
pwgts[1] -= vwgt[i];
if (pwgts[0] > minpwgt[0])
break;
}
}
}
/*************************************************************
* Do some partition refinement
**************************************************************/
Compute2WayPartitionParams(ctrl, graph);
/* printf("IPART: %3d [%5d %5d] [%5d %5d] %5d\n", graph->nvtxs, pwgts[0], pwgts[1], graph->pwgts[0], graph->pwgts[1], graph->mincut); */
Balance2Way(ctrl, graph, tpwgts, ubfactor);
/* printf("BPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut); */
FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
/* printf("RPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut); */
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
GKfree(&bestwhere, &perm, LTERM);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void GrowBisection(CtrlType *ctrl, GraphType *graph, int *tpwgts, float ubfactor)
{
int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where;
idxtype *queue, *touched, *gain, *bestwhere;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
Allocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
queue = idxmalloc(nvtxs, "BisectGraph: queue");
touched = idxmalloc(nvtxs, "BisectGraph: touched");
ASSERTP(tpwgts[0]+tpwgts[1] == idxsum(nvtxs, vwgt), ("%d %d\n", tpwgts[0]+tpwgts[1], idxsum(nvtxs, vwgt)));
maxpwgt[0] = ubfactor*tpwgts[0];
maxpwgt[1] = ubfactor*tpwgts[1];
minpwgt[0] = (1.0/ubfactor)*tpwgts[0];
minpwgt[1] = (1.0/ubfactor)*tpwgts[1];
nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(nvtxs, graph->adjwgtsum)+1; /* The +1 is for the 0 edges case */
for (; nbfs>0; nbfs--) {
idxset(nvtxs, 0, touched);
pwgts[1] = tpwgts[0]+tpwgts[1];
pwgts[0] = 0;
idxset(nvtxs, 1, where);
queue[0] = RandomInRange(nvtxs);
touched[queue[0]] = 1;
first = 0; last = 1;
nleft = nvtxs-1;
drain = 0;
/* Start the BFS from queue to get a partition */
for (;;) {
if (first == last) { /* Empty. Disconnected graph! */
if (nleft == 0 || drain)
break;
k = RandomInRange(nleft);
for (i=0; i<nvtxs; i++) {
if (touched[i] == 0) {
if (k == 0)
break;
else
k--;
}
}
queue[0] = i;
touched[i] = 1;
first = 0; last = 1;;
nleft--;
}
i = queue[first++];
if (pwgts[0] > 0 && pwgts[1]-vwgt[i] < minpwgt[1]) {
drain = 1;
continue;
}
where[i] = 0;
INC_DEC(pwgts[0], pwgts[1], vwgt[i]);
if (pwgts[1] <= maxpwgt[1])
break;
drain = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (touched[k] == 0) {
queue[last++] = k;
touched[k] = 1;
nleft--;
}
}
}
/* Check to see if we hit any bad limiting cases */
if (pwgts[1] == 0) {
i = RandomInRange(nvtxs);
where[i] = 1;
INC_DEC(pwgts[1], pwgts[0], vwgt[i]);
}
/*************************************************************
* Do some partition refinement
**************************************************************/
Compute2WayPartitionParams(ctrl, graph);
/*printf("IPART: %3d [%5d %5d] [%5d %5d] %5d\n", graph->nvtxs, pwgts[0], pwgts[1], graph->pwgts[0], graph->pwgts[1], graph->mincut); */
Balance2Way(ctrl, graph, tpwgts, ubfactor);
/*printf("BPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/
FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
/*printf("RPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
GKfree(&bestwhere, &queue, &touched, LTERM);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void GrowBisectionNode(CtrlType *ctrl, GraphType *graph, float ubfactor)
{
int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], tpwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *bndind;
idxtype *queue, *touched, *gain, *bestwhere;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
queue = idxmalloc(nvtxs, "BisectGraph: queue");
touched = idxmalloc(nvtxs, "BisectGraph: touched");
tpwgts[0] = idxsum(nvtxs, vwgt);
tpwgts[1] = tpwgts[0]/2;
tpwgts[0] -= tpwgts[1];
maxpwgt[0] = ubfactor*tpwgts[0];
maxpwgt[1] = ubfactor*tpwgts[1];
minpwgt[0] = (1.0/ubfactor)*tpwgts[0];
minpwgt[1] = (1.0/ubfactor)*tpwgts[1];
/* Allocate memory for graph->rdata. Allocate sufficient memory for both edge and node */
graph->rdata = idxmalloc(5*nvtxs+3, "GrowBisectionNode: graph->rdata");
graph->pwgts = graph->rdata;
graph->where = graph->rdata + 3;
graph->bndptr = graph->rdata + nvtxs + 3;
graph->bndind = graph->rdata + 2*nvtxs + 3;
graph->nrinfo = (NRInfoType *)(graph->rdata + 3*nvtxs + 3);
graph->id = graph->rdata + 3*nvtxs + 3;
graph->ed = graph->rdata + 4*nvtxs + 3;
where = graph->where;
bndind = graph->bndind;
nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = tpwgts[0]+tpwgts[1];
for (nbfs++; nbfs>0; nbfs--) {
idxset(nvtxs, 0, touched);
pwgts[1] = tpwgts[0]+tpwgts[1];
pwgts[0] = 0;
idxset(nvtxs, 1, where);
queue[0] = RandomInRange(nvtxs);
touched[queue[0]] = 1;
first = 0; last = 1;
nleft = nvtxs-1;
drain = 0;
/* Start the BFS from queue to get a partition */
if (nbfs >= 1) {
for (;;) {
if (first == last) { /* Empty. Disconnected graph! */
if (nleft == 0 || drain)
break;
k = RandomInRange(nleft);
for (i=0; i<nvtxs; i++) {
if (touched[i] == 0) {
if (k == 0)
break;
else
k--;
}
}
queue[0] = i;
touched[i] = 1;
first = 0; last = 1;;
nleft--;
}
i = queue[first++];
if (pwgts[1]-vwgt[i] < minpwgt[1]) {
drain = 1;
continue;
}
where[i] = 0;
INC_DEC(pwgts[0], pwgts[1], vwgt[i]);
if (pwgts[1] <= maxpwgt[1])
break;
drain = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (touched[k] == 0) {
queue[last++] = k;
touched[k] = 1;
nleft--;
}
}
}
}
/*************************************************************
* Do some partition refinement
**************************************************************/
Compute2WayPartitionParams(ctrl, graph);
Balance2Way(ctrl, graph, tpwgts, ubfactor);
FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
/* Construct and refine the vertex separator */
for (i=0; i<graph->nbnd; i++)
where[bndind[i]] = 2;
Compute2WayNodePartitionParams(ctrl, graph);
FM_2WayNodeRefine(ctrl, graph, ubfactor, 6);
/* printf("ISep: [%d %d %d] %d\n", graph->pwgts[0], graph->pwgts[1], graph->pwgts[2], bestcut); */
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
Compute2WayNodePartitionParams(ctrl, graph);
GKfree(&bestwhere, &queue, &touched, LTERM);
}
/*************************************************************************
* This function is the entry point for ONCMETIS
**************************************************************************/
void METIS_NodeND(int *nvtxs, idxtype *xadj, idxtype *adjncy, int *numflag, int *options,
idxtype *perm, idxtype *iperm)
{
int i, ii, j, l, wflag, nflag;
GraphType graph;
CtrlType ctrl;
idxtype *cptr, *cind, *piperm;
if (*numflag == 1)
Change2CNumbering(*nvtxs, xadj, adjncy);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = ONMETIS_CTYPE;
ctrl.IType = ONMETIS_ITYPE;
ctrl.RType = ONMETIS_RTYPE;
ctrl.dbglvl = ONMETIS_DBGLVL;
ctrl.oflags = ONMETIS_OFLAGS;
ctrl.pfactor = ONMETIS_PFACTOR;
ctrl.nseps = ONMETIS_NSEPS;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
ctrl.oflags = options[OPTION_OFLAGS];
ctrl.pfactor = options[OPTION_PFACTOR];
ctrl.nseps = options[OPTION_NSEPS];
}
if (ctrl.nseps < 1)
ctrl.nseps = 1;
ctrl.optype = OP_ONMETIS;
ctrl.CoarsenTo = 100;
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
InitRandom(-1);
if (ctrl.pfactor > 0) {
/*============================================================
* Prune the dense columns
==============================================================*/
piperm = idxmalloc(*nvtxs, "ONMETIS: piperm");
PruneGraph(&ctrl, &graph, *nvtxs, xadj, adjncy, piperm, (float)(0.1*ctrl.pfactor));
}
else if (ctrl.oflags&OFLAG_COMPRESS) {
/*============================================================
* Compress the graph
==============================================================*/
cptr = idxmalloc(*nvtxs+1, "ONMETIS: cptr");
cind = idxmalloc(*nvtxs, "ONMETIS: cind");
CompressGraph(&ctrl, &graph, *nvtxs, xadj, adjncy, cptr, cind);
if (graph.nvtxs >= COMPRESSION_FRACTION*(*nvtxs)) {
ctrl.oflags--; /* We actually performed no compression */
GKfree(&cptr, &cind, LTERM);
}
else if (2*graph.nvtxs < *nvtxs && ctrl.nseps == 1)
ctrl.nseps = 2;
}
else {
SetUpGraph(&graph, OP_ONMETIS, *nvtxs, 1, xadj, adjncy, NULL, NULL, 0);
}
/*=============================================================
* Do the nested dissection ordering
--=============================================================*/
ctrl.maxvwgt = 1.5*(idxsum(graph.nvtxs, graph.vwgt)/ctrl.CoarsenTo);
AllocateWorkSpace(&ctrl, &graph, 2);
if (ctrl.oflags&OFLAG_CCMP)
MlevelNestedDissectionCC(&ctrl, &graph, iperm, ORDER_UNBALANCE_FRACTION, graph.nvtxs);
else
MlevelNestedDissection(&ctrl, &graph, iperm, ORDER_UNBALANCE_FRACTION, graph.nvtxs);
FreeWorkSpace(&ctrl, &graph);
if (ctrl.pfactor > 0) { /* Order any prunned vertices */
if (graph.nvtxs < *nvtxs) {
idxcopy(graph.nvtxs, iperm, perm); /* Use perm as an auxiliary array */
for (i=0; i<graph.nvtxs; i++)
iperm[piperm[i]] = perm[i];
for (i=graph.nvtxs; i<*nvtxs; i++)
iperm[piperm[i]] = i;
}
GKfree(&piperm, LTERM);
}
else if (ctrl.oflags&OFLAG_COMPRESS) { /* Uncompress the ordering */
if (graph.nvtxs < COMPRESSION_FRACTION*(*nvtxs)) {
/* construct perm from iperm */
for (i=0; i<graph.nvtxs; i++)
perm[iperm[i]] = i;
for (l=ii=0; ii<graph.nvtxs; ii++) {
i = perm[ii];
for (j=cptr[i]; j<cptr[i+1]; j++)
iperm[cind[j]] = l++;
}
}
GKfree(&cptr, &cind, LTERM);
}
for (i=0; i<*nvtxs; i++)
perm[iperm[i]] = i;
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimers(&ctrl));
if (*numflag == 1)
Change2FNumberingOrder(*nvtxs, xadj, adjncy, perm, iperm);
}
/*************************************************************************
* Let the game begin
**************************************************************************/
int main(int argc, char *argv[])
{
idxtype i, j, istep, options[10], nn, ne, fstep, lstep, nparts, nboxes, u[3], dim, nchanges, ncomm;
char filename[256];
idxtype *mien, *mrng, *part, *oldpart, *sflag, *bestdims, *fepart;
double *mxyz, *bxyz;
idxtype *xadj, *adjncy, *cntptr, *cntind;
idxtype numflag = 0, wgtflag = 0, edgecut, etype=2;
void *cinfo;
FILE *fpin;
long long int *ltmp;
if (argc != 6) {
mfprintf(stderr, "Usage: %s <nn> <ne> <fstep> <lstep> <nparts>\n", argv[0]);
exit(0);
}
nn = atoi(argv[1]);
ne = atoi(argv[2]);
fstep = atoi(argv[3]);
lstep = atoi(argv[4]);
nparts = atoi(argv[5]);
mprintf("Reading %s, nn: %D, ne: %D, fstep: %D, lstep: %D, nparts: %D\n", filename, nn, ne, fstep, lstep, nparts);
mien = idxmalloc(4*ne, "main: mien");
mxyz = gk_dmalloc(3*nn, "main: mxyz");
mrng = idxmalloc(4*ne, "main: mrng");
bxyz = gk_dmalloc(6*ne*4, "main: bxyz");
fepart = idxmalloc(nn, "main: fepart");
part = idxmalloc(nn, "main: part");
oldpart = idxmalloc(nn, "main: oldpart");
sflag = idxmalloc(nn, "main: sflag");
bestdims = idxsmalloc(2*nparts, -1, "main: bestdims");
xadj = idxmalloc(nn+1, "main: xadj");
adjncy = idxmalloc(50*nn, "main: adjncy");
/*========================================================================
* Read the initial mesh and setup the graph and contact information
*========================================================================*/
msprintf(filename, "mien.%04D", fstep);
fpin = GKfopen(filename, "rb", "main: mien");
fread(mien, sizeof(int), 4*ne, fpin);
for (i=0; i<4*ne; i++)
mien[i] = Flip_int32(mien[i]);
GKfclose(fpin);
msprintf(filename, "mxyz.%04D", fstep);
fpin = GKfopen(filename, "rb", "main: mxyz");
fread(mxyz, sizeof(double), 3*nn, fpin);
for (i=0; i<3*nn; i++) {
ltmp = (long long int *)(mxyz+i);
*ltmp = Flip_int64(*ltmp);
}
GKfclose(fpin);
mprintf("%e %e %e\n", mxyz[3*0+0], mxyz[3*0+1], mxyz[3*0+2]);
msprintf(filename, "mrng.%04D", fstep);
fpin = GKfopen(filename, "rb", "main: mrng");
fread(mrng, sizeof(int), 4*ne, fpin);
for (i=0; i<4*ne; i++)
mrng[i] = Flip_int32(mrng[i]);
GKfclose(fpin);
/*========================================================================
* Determine which nodes are in the surface
*========================================================================*/
iset(nn, 0, sflag);
for (i=0; i<ne; i++) {
if (mrng[4*i+0] > 0) { /* 1, 2, 3 */
sflag[mien[4*i+0]-1] = 1;
sflag[mien[4*i+1]-1] = 1;
sflag[mien[4*i+2]-1] = 1;
}
if (mrng[4*i+1] > 0) { /* 1, 2, 4 */
sflag[mien[4*i+0]-1] = 1;
sflag[mien[4*i+1]-1] = 1;
sflag[mien[4*i+3]-1] = 1;
}
if (mrng[4*i+2] > 0) { /* 2, 3, 4 */
sflag[mien[4*i+1]-1] = 1;
sflag[mien[4*i+2]-1] = 1;
sflag[mien[4*i+3]-1] = 1;
}
if (mrng[4*i+3] > 0) { /* 1, 3, 4 */
sflag[mien[4*i+0]-1] = 1;
sflag[mien[4*i+2]-1] = 1;
sflag[mien[4*i+3]-1] = 1;
}
}
mprintf("Contact Nodes: %D of %D\n", isum(nn, sflag), nn);
/*========================================================================
* Compute the FE partition
*========================================================================*/
numflag = mien[idxargmin(4*ne, mien)];
METIS_MeshToNodal(&ne, &nn, mien, &etype, &numflag, xadj, adjncy);
options[0] = 0;
METIS_PartGraphVKway(&nn, xadj, adjncy, NULL, NULL, &wgtflag, &numflag, &nparts,
options, &edgecut, fepart);
mprintf("K-way partitioning Volume: %D\n", edgecut);
/*========================================================================
* Get into the loop in which you go over the different configurations
*========================================================================*/
for (istep=fstep; istep<=lstep; istep++) {
msprintf(filename, "mxyz.%04D", istep);
mprintf("Reading %s...............................................................\n", filename);
fpin = GKfopen(filename, "rb", "main: mxyz");
fread(mxyz, sizeof(double), 3*nn, fpin);
for (i=0; i<3*nn; i++) {
ltmp = (long long int *)(mxyz+i);
*ltmp = Flip_int64(*ltmp);
}
GKfclose(fpin);
msprintf(filename, "mrng.%04D", istep);
fpin = GKfopen(filename, "rb", "main: mrng");
fread(mrng, sizeof(int), 4*ne, fpin);
for (i=0; i<4*ne; i++)
mrng[i] = Flip_int32(mrng[i]);
GKfclose(fpin);
/* Determine which nodes are in the surface */
iset(nn, 0, sflag);
for (i=0; i<ne; i++) {
if (mrng[4*i+0] > 0) { /* 1, 2, 3 */
sflag[mien[4*i+0]-1] = 1;
sflag[mien[4*i+1]-1] = 1;
sflag[mien[4*i+2]-1] = 1;
}
if (mrng[4*i+1] > 0) { /* 1, 2, 4 */
sflag[mien[4*i+0]-1] = 1;
sflag[mien[4*i+1]-1] = 1;
sflag[mien[4*i+3]-1] = 1;
}
if (mrng[4*i+2] > 0) { /* 2, 3, 4 */
sflag[mien[4*i+1]-1] = 1;
sflag[mien[4*i+2]-1] = 1;
sflag[mien[4*i+3]-1] = 1;
}
if (mrng[4*i+3] > 0) { /* 1, 3, 4 */
sflag[mien[4*i+0]-1] = 1;
sflag[mien[4*i+2]-1] = 1;
sflag[mien[4*i+3]-1] = 1;
}
}
mprintf("Contact Nodes: %D of %D\n", isum(nn, sflag), nn);
/* Determine the bounding boxes of the surface elements */
for (nboxes=0, i=0; i<ne; i++) {
if (mrng[4*i+0] > 0) { /* 1, 2, 3 */
u[0] = mien[4*i+0]-1;
u[1] = mien[4*i+1]-1;
u[2] = mien[4*i+2]-1;
bxyz[6*nboxes+0] = bxyz[6*nboxes+3] = mxyz[3*u[0]+0];
bxyz[6*nboxes+1] = bxyz[6*nboxes+4] = mxyz[3*u[0]+1];
bxyz[6*nboxes+2] = bxyz[6*nboxes+5] = mxyz[3*u[0]+2];
for (j=1; j<3; j++) {
for (dim=0; dim<3; dim++) {
bxyz[6*nboxes+dim] = (bxyz[6*nboxes+dim] > mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+dim]);
bxyz[6*nboxes+3+dim] = (bxyz[6*nboxes+3+dim] < mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+3+dim]);
}
}
nboxes++;
}
if (mrng[4*i+1] > 0) { /* 1, 2, 4 */
u[0] = mien[4*i+0]-1;
u[1] = mien[4*i+1]-1;
u[2] = mien[4*i+3]-1;
bxyz[6*nboxes+0] = bxyz[6*nboxes+3] = mxyz[3*u[0]+0];
bxyz[6*nboxes+1] = bxyz[6*nboxes+4] = mxyz[3*u[0]+1];
bxyz[6*nboxes+2] = bxyz[6*nboxes+5] = mxyz[3*u[0]+2];
for (j=1; j<3; j++) {
for (dim=0; dim<3; dim++) {
bxyz[6*nboxes+dim] = (bxyz[6*nboxes+dim] > mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+dim]);
bxyz[6*nboxes+3+dim] = (bxyz[6*nboxes+3+dim] < mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+3+dim]);
}
}
nboxes++;
}
if (mrng[4*i+2] > 0) { /* 2, 3, 4 */
u[0] = mien[4*i+1]-1;
u[1] = mien[4*i+2]-1;
u[2] = mien[4*i+3]-1;
bxyz[6*nboxes+0] = bxyz[6*nboxes+3] = mxyz[3*u[0]+0];
bxyz[6*nboxes+1] = bxyz[6*nboxes+4] = mxyz[3*u[0]+1];
bxyz[6*nboxes+2] = bxyz[6*nboxes+5] = mxyz[3*u[0]+2];
for (j=1; j<3; j++) {
for (dim=0; dim<3; dim++) {
bxyz[6*nboxes+dim] = (bxyz[6*nboxes+dim] > mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+dim]);
bxyz[6*nboxes+3+dim] = (bxyz[6*nboxes+3+dim] < mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+3+dim]);
}
}
nboxes++;
}
if (mrng[4*i+3] > 0) { /* 1, 3, 4 */
u[0] = mien[4*i+0]-1;
u[1] = mien[4*i+2]-1;
u[2] = mien[4*i+3]-1;
bxyz[6*nboxes+0] = bxyz[6*nboxes+3] = mxyz[3*u[0]+0];
bxyz[6*nboxes+1] = bxyz[6*nboxes+4] = mxyz[3*u[0]+1];
bxyz[6*nboxes+2] = bxyz[6*nboxes+5] = mxyz[3*u[0]+2];
for (j=1; j<3; j++) {
for (dim=0; dim<3; dim++) {
bxyz[6*nboxes+dim] = (bxyz[6*nboxes+dim] > mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+dim]);
bxyz[6*nboxes+3+dim] = (bxyz[6*nboxes+3+dim] < mxyz[3*u[j]+dim] ? mxyz[3*u[j]+dim] : bxyz[6*nboxes+3+dim]);
}
}
nboxes++;
}
}
cinfo = METIS_PartSurfForContactRCB(&nn, mxyz, sflag, &nparts, part, bestdims);
METIS_FindContacts(cinfo, &nboxes, bxyz, &nparts, &cntptr, &cntind);
METIS_FreeContactInfo(cinfo);
nchanges = 0;
if (istep > fstep) {
for (i=0; i<nn; i++)
nchanges += (part[i] != oldpart[i] ? 1 : 0);
}
idxcopy(nn, part, oldpart);
ncomm = ComputeMapCost(nn, nparts, fepart, part);
mprintf("Contacting Elements: %D Indices: %D Nchanges: %D MapCost: %D\n", nboxes, cntptr[nboxes]-nboxes, nchanges, ncomm);
gk_free((void **)&cntptr, &cntind, LTERM);
}
}
/*************************************************************************
* This function projects a partition, and at the same time computes the
* parameters for refinement.
**************************************************************************/
void ProjectKWayPartition(CtrlType *ctrl, GraphType *graph, int nparts)
{
int i, j, k, nvtxs, nbnd, me, other, istart, iend, ndegrees;
idxtype *xadj, *adjncy, *adjwgt, *adjwgtsum;
idxtype *cmap, *where, *bndptr, *bndind;
idxtype *cwhere;
GraphType *cgraph;
RInfoType *crinfo, *rinfo, *myrinfo;
EDegreeType *myedegrees;
idxtype *htable;
cgraph = graph->coarser;
cwhere = cgraph->where;
crinfo = cgraph->rinfo;
nvtxs = graph->nvtxs;
cmap = graph->cmap;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
AllocateKWayPartitionMemory(ctrl, graph, nparts);
where = graph->where;
rinfo = graph->rinfo;
bndind = graph->bndind;
bndptr = idxset(nvtxs, -1, graph->bndptr);
/* Go through and project partition and compute id/ed for the nodes */
for (i=0; i<nvtxs; i++) {
k = cmap[i];
where[i] = cwhere[k];
cmap[i] = crinfo[k].ed; /* For optimization */
}
htable = idxset(nparts, -1, idxwspacemalloc(ctrl, nparts));
ctrl->wspace.cdegree = 0;
for (nbnd=0, i=0; i<nvtxs; i++) {
me = where[i];
myrinfo = rinfo+i;
myrinfo->id = myrinfo->ed = myrinfo->ndegrees = 0;
myrinfo->edegrees = NULL;
myrinfo->id = adjwgtsum[i];
if (cmap[i] > 0) { /* If it is an interface node. Note cmap[i] = crinfo[cmap[i]].ed */
istart = xadj[i];
iend = xadj[i+1];
myedegrees = myrinfo->edegrees = ctrl->wspace.edegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += iend-istart;
ndegrees = 0;
for (j=istart; j<iend; j++) {
other = where[adjncy[j]];
if (me != other) {
myrinfo->ed += adjwgt[j];
if ((k = htable[other]) == -1) {
htable[other] = ndegrees;
myedegrees[ndegrees].pid = other;
myedegrees[ndegrees++].ed = adjwgt[j];
}
else {
myedegrees[k].ed += adjwgt[j];
}
}
}
myrinfo->id -= myrinfo->ed;
/* Remove space for edegrees if it was interior */
if (myrinfo->ed == 0) {
myrinfo->edegrees = NULL;
ctrl->wspace.cdegree -= iend-istart;
}
else {
if (myrinfo->ed-myrinfo->id >= 0)
BNDInsert(nbnd, bndind, bndptr, i);
myrinfo->ndegrees = ndegrees;
for (j=0; j<ndegrees; j++)
htable[myedegrees[j].pid] = -1;
}
}
}
idxcopy(nparts, cgraph->pwgts, graph->pwgts);
graph->mincut = cgraph->mincut;
graph->nbnd = nbnd;
FreeGraph(graph->coarser);
graph->coarser = NULL;
idxwspacefree(ctrl, nparts);
ASSERT(CheckBnd2(graph));
}
/*************************************************************************
* 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 Mc_InitPartition_RB(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int i, j;
int ncon, mype, npes, gnvtxs, ngroups;
idxtype *xadj, *adjncy, *adjwgt, *vwgt;
idxtype *part, *gwhere0, *gwhere1;
idxtype *tmpwhere, *tmpvwgt, *tmpxadj, *tmpadjncy, *tmpadjwgt;
GraphType *agraph;
int lnparts, fpart, fpe, lnpes;
int twoparts=2, numflag = 0, wgtflag = 3, moptions[10], edgecut, max_cut;
float *mytpwgts, mytpwgts2[2], lbvec[MAXNCON], lbsum, min_lbsum, wsum;
MPI_Comm ipcomm;
struct {
float sum;
int rank;
} lpesum, gpesum;
ncon = graph->ncon;
ngroups = amax(amin(RIP_SPLIT_FACTOR, ctrl->npes), 1);
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
agraph = Mc_AssembleAdaptiveGraph(ctrl, graph, wspace);
part = idxmalloc(agraph->nvtxs, "Mc_IP_RB: part");
xadj = idxmalloc(agraph->nvtxs+1, "Mc_IP_RB: xadj");
adjncy = idxmalloc(agraph->nedges, "Mc_IP_RB: adjncy");
adjwgt = idxmalloc(agraph->nedges, "Mc_IP_RB: adjwgt");
vwgt = idxmalloc(agraph->nvtxs*ncon, "Mc_IP_RB: vwgt");
idxcopy(agraph->nvtxs*ncon, agraph->vwgt, vwgt);
idxcopy(agraph->nvtxs+1, agraph->xadj, xadj);
idxcopy(agraph->nedges, agraph->adjncy, adjncy);
idxcopy(agraph->nedges, agraph->adjwgt, adjwgt);
MPI_Comm_split(ctrl->gcomm, ctrl->mype % ngroups, 0, &ipcomm);
MPI_Comm_rank(ipcomm, &mype);
MPI_Comm_size(ipcomm, &npes);
gnvtxs = agraph->nvtxs;
gwhere0 = idxsmalloc(gnvtxs, 0, "Mc_IP_RB: gwhere0");
gwhere1 = idxmalloc(gnvtxs, "Mc_IP_RB: gwhere1");
/* ADD: this assumes that tpwgts for all constraints is the same */
/* ADD: this is necessary because serial metis does not support the general case */
mytpwgts = fsmalloc(ctrl->nparts, 0.0, "mytpwgts");
for (i=0; i<ctrl->nparts; i++)
for (j=0; j<ncon; j++)
mytpwgts[i] += ctrl->tpwgts[i*ncon+j];
for (i=0; i<ctrl->nparts; i++)
mytpwgts[i] /= (float)ncon;
/* Go into the recursive bisection */
/* ADD: consider changing this to breadth-first type bisection */
moptions[0] = 0;
moptions[7] = ctrl->sync + (ctrl->mype % ngroups) + 1;
lnparts = ctrl->nparts;
fpart = fpe = 0;
lnpes = npes;
while (lnpes > 1 && lnparts > 1) {
/* Determine the weights of the partitions */
mytpwgts2[0] = ssum(lnparts/2, mytpwgts+fpart);
mytpwgts2[1] = 1.0-mytpwgts2[0];
if (ncon == 1)
METIS_WPartGraphKway2(&agraph->nvtxs, agraph->xadj, agraph->adjncy,
agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag, &twoparts, mytpwgts2,
moptions, &edgecut, part);
else {
METIS_mCPartGraphRecursive2(&agraph->nvtxs, &ncon, agraph->xadj,
agraph->adjncy, agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag,
&twoparts, mytpwgts2, moptions, &edgecut, part);
}
wsum = ssum(lnparts/2, mytpwgts+fpart);
sscale(lnparts/2, 1.0/wsum, mytpwgts+fpart);
sscale(lnparts-lnparts/2, 1.0/(1.0-wsum), mytpwgts+fpart+lnparts/2);
/* I'm picking the left branch */
if (mype < fpe+lnpes/2) {
Mc_KeepPart(agraph, wspace, part, 0);
lnpes = lnpes/2;
lnparts = lnparts/2;
}
else {
Mc_KeepPart(agraph, wspace, part, 1);
fpart = fpart + lnparts/2;
fpe = fpe + lnpes/2;
lnpes = lnpes - lnpes/2;
lnparts = lnparts - lnparts/2;
}
}
/* In case npes is greater than or equal to nparts */
if (lnparts == 1) {
/* 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;
}
}
/* In case npes is smaller than nparts */
else {
if (ncon == 1)
METIS_WPartGraphKway2(&agraph->nvtxs, agraph->xadj, agraph->adjncy,
agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag, &lnparts, mytpwgts+fpart,
moptions, &edgecut, part);
else
METIS_mCPartGraphRecursive2(&agraph->nvtxs, &ncon, agraph->xadj,
agraph->adjncy, agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag,
&lnparts, mytpwgts+fpart, moptions, &edgecut, part);
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart + part[i];
}
MPI_Allreduce((void *)gwhere0, (void *)gwhere1, gnvtxs, IDX_DATATYPE, 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;
Mc_ComputeSerialBalance(ctrl, agraph, gwhere1, lbvec);
lbsum = ssum(ncon, lbvec);
edgecut = ComputeSerialEdgeCut(agraph);
MPI_Allreduce((void *)&edgecut, (void *)&max_cut, 1, MPI_INT, MPI_MAX, ctrl->gcomm);
MPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, MPI_FLOAT, MPI_MIN, ctrl->gcomm);
lpesum.sum = lbsum;
if (min_lbsum < UNBALANCE_FRACTION * (float)(ncon)) {
if (lbsum < UNBALANCE_FRACTION * (float)(ncon))
lpesum.sum = (float) (edgecut);
else
lpesum.sum = (float) (max_cut);
}
MPI_Comm_rank(ctrl->gcomm, &(lpesum.rank));
MPI_Allreduce((void *)&lpesum, (void *)&gpesum, 1, MPI_FLOAT_INT, MPI_MINLOC, ctrl->gcomm);
MPI_Bcast((void *)gwhere1, gnvtxs, IDX_DATATYPE, gpesum.rank, ctrl->gcomm);
agraph->xadj = tmpxadj;
agraph->adjncy = tmpadjncy;
agraph->adjwgt = tmpadjwgt;
agraph->vwgt = tmpvwgt;
agraph->where = tmpwhere;
}
idxcopy(graph->nvtxs, gwhere1+graph->vtxdist[ctrl->mype], graph->where);
FreeGraph(agraph);
MPI_Comm_free(&ipcomm);
GKfree((void **)&gwhere0, (void **)&gwhere1, (void **)&mytpwgts, (void **)&part, (void **)&xadj, (void **)&adjncy, (void **)&adjwgt, (void **)&vwgt, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
}