void gk_gkmcoreAdd(gk_mcore_t *mcore, int type, size_t nbytes, void *ptr)
{
if (mcore->cmop == mcore->nmops) {
mcore->nmops *= 2;
mcore->mops = realloc(mcore->mops, mcore->nmops*sizeof(gk_mop_t));
if (mcore->mops == NULL)
gk_errexit(SIGMEM, "***Memory allocation for gkmcore failed.\n");
}
mcore->mops[mcore->cmop].type = type;
mcore->mops[mcore->cmop].nbytes = nbytes;
mcore->mops[mcore->cmop].ptr = ptr;
mcore->cmop++;
switch (type) {
case GK_MOPT_MARK:
break;
case GK_MOPT_HEAP:
mcore->num_hallocs++;
mcore->size_hallocs += nbytes;
mcore->cur_hallocs += nbytes;
if (mcore->max_hallocs < mcore->cur_hallocs)
mcore->max_hallocs = mcore->cur_hallocs;
break;
default:
gk_errexit(SIGMEM, "Incorrect mcore type operation.\n");
}
}
/*************************************************************************
* This function is my wrapper around realloc
**************************************************************************/
void *gk_realloc(void *oldptr, size_t nbytes, char *msg)
{
void *ptr=NULL;
if (nbytes == 0)
nbytes++; /* Force mallocs to actually allocate some memory */
/* remove this memory de-allocation */
if (gkmcore != NULL && oldptr != NULL) gk_gkmcoreDel(gkmcore, oldptr);
ptr = (void *)realloc(oldptr, nbytes);
if (ptr == NULL) {
fprintf(stderr, " Maximum memory used: %10zu bytes\n", gk_GetMaxMemoryUsed());
fprintf(stderr, " Current memory used: %10zu bytes\n", gk_GetCurMemoryUsed());
gk_errexit(SIGMEM, "***Memory realloc failed for %s. " "Requested size: %zu bytes",
msg, nbytes);
return NULL;
}
/* add this memory allocation */
if (gkmcore != NULL) gk_gkmcoreAdd(gkmcore, GK_MOPT_HEAP, nbytes, ptr);
return ptr;
}
void gk_mcorePop(gk_mcore_t *mcore)
{
while (mcore->cmop > 0) {
mcore->cmop--;
switch (mcore->mops[mcore->cmop].type) {
case GK_MOPT_MARK: /* push marker */
goto DONE;
break;
case GK_MOPT_CORE: /* core free */
mcore->corecpos -= mcore->mops[mcore->cmop].nbytes;
mcore->cur_callocs -= mcore->mops[mcore->cmop].nbytes;
if (mcore->corecpos < 0)
errexit("Internal Error: wspace's core over-freed [%zu, %zu]\n",
mcore->coresize, mcore->corecpos);
break;
case GK_MOPT_HEAP: /* heap free */
gk_free((void **)&mcore->mops[mcore->cmop].ptr, LTERM);
mcore->cur_hallocs -= mcore->mops[mcore->cmop].nbytes;
break;
default:
gk_errexit(SIGMEM, "Unknown mop type of %d\n", mcore->mops[mcore->cmop].type);
}
}
DONE:
;
/*printf("MCPPOP: %zu\n", mcore->cmop); */
}
void AllocateKWayPartitionMemory(ctrl_t *ctrl, graph_t *graph)
{
graph->pwgts = imalloc(ctrl->nparts*graph->ncon, "AllocateKWayPartitionMemory: pwgts");
graph->where = imalloc(graph->nvtxs, "AllocateKWayPartitionMemory: where");
graph->bndptr = imalloc(graph->nvtxs, "AllocateKWayPartitionMemory: bndptr");
graph->bndind = imalloc(graph->nvtxs, "AllocateKWayPartitionMemory: bndind");
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
graph->ckrinfo = (ckrinfo_t *)gk_malloc(graph->nvtxs*sizeof(ckrinfo_t),
"AllocateKWayPartitionMemory: ckrinfo");
break;
case METIS_OBJTYPE_VOL:
graph->vkrinfo = (vkrinfo_t *)gk_malloc(graph->nvtxs*sizeof(vkrinfo_t),
"AllocateKWayVolPartitionMemory: vkrinfo");
/* This is to let the cut-based -minconn and -contig large-scale graph
changes to go through */
graph->ckrinfo = (ckrinfo_t *)graph->vkrinfo;
break;
default:
gk_errexit(SIGERR, "Unknown objtype of %d\n", ctrl->objtype);
}
}
void AllocateRefinementWorkSpace(ctrl_t *ctrl, idx_t nbrpoolsize) {
ctrl->nbrpoolsize = nbrpoolsize;
ctrl->nbrpoolcpos = 0;
ctrl->nbrpoolreallocs = 0;
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
ctrl->cnbrpool = (cnbr_t *) gk_malloc(ctrl->nbrpoolsize * sizeof(cnbr_t),
"AllocateRefinementWorkSpace: cnbrpool");
break;
case METIS_OBJTYPE_VOL:
ctrl->vnbrpool = (vnbr_t *) gk_malloc(ctrl->nbrpoolsize * sizeof(vnbr_t),
"AllocateRefinementWorkSpace: vnbrpool");
break;
default:gk_errexit(SIGERR, "Unknown objtype of %d\n", ctrl->objtype);
}
/* Allocate the memory for the sparse subdomain graph */
if (ctrl->minconn) {
ctrl->pvec1 = imalloc(ctrl->nparts + 1, "AllocateRefinementWorkSpace: pvec1");
ctrl->pvec2 = imalloc(ctrl->nparts + 1, "AllocateRefinementWorkSpace: pvec2");
ctrl->maxnads = ismalloc(ctrl->nparts, INIT_MAXNAD, "AllocateRefinementWorkSpace: maxnads");
ctrl->nads = imalloc(ctrl->nparts, "AllocateRefinementWorkSpace: nads");
ctrl->adids = iAllocMatrix(ctrl->nparts, INIT_MAXNAD, 0, "AllocateRefinementWorkSpace: adids");
ctrl->adwgts = iAllocMatrix(ctrl->nparts, INIT_MAXNAD, 0, "AllocateRefinementWorkSpace: adwgts");
}
}
void *gk_malloc(size_t nbytes, char *msg)
{
void *ptr=NULL;
if (nbytes == 0)
nbytes++; /* Force mallocs to actually allocate some memory */
ptr = (void *)malloc(nbytes);
if (ptr == NULL) {
fprintf(stderr, " Current memory used: %10zu bytes\n", gk_GetCurMemoryUsed());
fprintf(stderr, " Maximum memory used: %10zu bytes\n", gk_GetMaxMemoryUsed());
gk_errexit(SIGMEM, "***Memory allocation failed for %s. Requested size: %zu bytes",
msg, nbytes);
return NULL;
}
/* add this memory allocation */
if (gkmcore != NULL) gk_gkmcoreAdd(gkmcore, GK_MOPT_HEAP, nbytes, ptr);
/* zero-out the allocated space */
#ifndef NDEBUG
memset(ptr, 0, nbytes);
#endif
return ptr;
}
void gk_graph_Write(gk_graph_t *graph, char *filename, int format)
{
ssize_t i, j;
int hasvwgts, hasvsizes, hasewgts;
FILE *fpout;
if (format != GK_GRAPH_FMT_METIS)
gk_errexit(SIGERR, "Unknown file format. %d\n", format);
if (filename)
fpout = gk_fopen(filename, "w", "gk_graph_Write: fpout");
else
fpout = stdout;
hasewgts = (graph->iadjwgt || graph->fadjwgt);
hasvwgts = (graph->ivwgts || graph->fvwgts);
hasvsizes = (graph->ivsizes || graph->fvsizes);
/* write the header line */
fprintf(fpout, "%d %zd", graph->nvtxs, graph->xadj[graph->nvtxs]/2);
if (hasvwgts || hasvsizes || hasewgts)
fprintf(fpout, " %d%d%d", hasvsizes, hasvwgts, hasewgts);
fprintf(fpout, "\n");
for (i=0; i<graph->nvtxs; i++) {
if (hasvsizes) {
if (graph->ivsizes)
fprintf(fpout, " %d", graph->ivsizes[i]);
else
fprintf(fpout, " %f", graph->fvsizes[i]);
}
if (hasvwgts) {
if (graph->ivwgts)
fprintf(fpout, " %d", graph->ivwgts[i]);
else
fprintf(fpout, " %f", graph->fvwgts[i]);
}
for (j=graph->xadj[i]; j<graph->xadj[i+1]; j++) {
fprintf(fpout, " %d", graph->adjncy[j]+1);
if (hasewgts) {
if (graph->iadjwgt)
fprintf(fpout, " %d", graph->iadjwgt[j]);
else
fprintf(fpout, " %f", graph->fadjwgt[j]);
}
}
fprintf(fpout, "\n");
}
if (filename)
gk_fclose(fpout);
}
void gk_gkmcoreDel(gk_mcore_t *mcore, void *ptr)
{
int i;
for (i=mcore->cmop-1; i>=0; i--) {
if (mcore->mops[i].type == GK_MOPT_MARK)
gk_errexit(SIGMEM, "Could not find pointer %p in mcore\n", ptr);
if (mcore->mops[i].ptr == ptr) {
if (mcore->mops[i].type != GK_MOPT_HEAP)
gk_errexit(SIGMEM, "Trying to delete a non-HEAP mop.\n");
mcore->cur_hallocs -= mcore->mops[i].nbytes;
mcore->mops[i] = mcore->mops[--mcore->cmop];
return;
}
}
gk_errexit(SIGMEM, "gkmcoreDel should never have been here!\n");
}
void ComputeKWayBoundary(ctrl_t *ctrl, graph_t *graph, idx_t bndtype)
{
idx_t i, nvtxs, nbnd;
idx_t *bndind, *bndptr;
nvtxs = graph->nvtxs;
bndind = graph->bndind;
bndptr = iset(nvtxs, -1, graph->bndptr);
nbnd = 0;
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
/* Compute the boundary */
if (bndtype == BNDTYPE_REFINE) {
for (i=0; i<nvtxs; i++) {
if (graph->ckrinfo[i].ed-graph->ckrinfo[i].id >= 0)
BNDInsert(nbnd, bndind, bndptr, i);
}
}
else { /* BNDTYPE_BALANCE */
for (i=0; i<nvtxs; i++) {
if (graph->ckrinfo[i].ed > 0)
BNDInsert(nbnd, bndind, bndptr, i);
}
}
break;
case METIS_OBJTYPE_VOL:
/* Compute the boundary */
if (bndtype == BNDTYPE_REFINE) {
for (i=0; i<nvtxs; i++) {
if (graph->vkrinfo[i].gv >= 0)
BNDInsert(nbnd, bndind, bndptr, i);
}
}
else { /* BNDTYPE_BALANCE */
for (i=0; i<nvtxs; i++) {
if (graph->vkrinfo[i].ned > 0)
BNDInsert(nbnd, bndind, bndptr, i);
}
}
break;
default:
gk_errexit(SIGERR, "Unknown objtype of %d\n", ctrl->objtype);
}
graph->nbnd = nbnd;
}
void InitKWayPartitioning(ctrl_t *ctrl, graph_t *graph)
{
switch (ctrl->iptype) {
case METIS_IPTYPE_METISRB:
InitKWayPartitioningRB(ctrl, graph);
break;
case METIS_IPTYPE_GROW:
InitKWayPartitioningGrow(ctrl, graph);
break;
default:
gk_errexit(SIGERR, "Unknown iptype: %d\n", ctrl->iptype);
}
}
void Init2WayPartition(ctrl_t *ctrl, graph_t *graph, real_t *ntpwgts,
idx_t niparts)
{
mdbglvl_et dbglvl;
ASSERT(graph->tvwgt[0] >= 0);
dbglvl = ctrl->dbglvl;
IFSET(ctrl->dbglvl, METIS_DBG_REFINE, ctrl->dbglvl -= METIS_DBG_REFINE);
IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO, ctrl->dbglvl -= METIS_DBG_MOVEINFO);
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startcputimer(ctrl->InitPartTmr));
switch (ctrl->iptype) {
case METIS_IPTYPE_RANDOM:
if (graph->ncon == 1)
RandomBisection(ctrl, graph, ntpwgts, niparts);
else
McRandomBisection(ctrl, graph, ntpwgts, niparts);
break;
case METIS_IPTYPE_GROW:
if (graph->nedges == 0)
if (graph->ncon == 1)
RandomBisection(ctrl, graph, ntpwgts, niparts);
else
McRandomBisection(ctrl, graph, ntpwgts, niparts);
else
if (graph->ncon == 1)
GrowBisection(ctrl, graph, ntpwgts, niparts);
else
McGrowBisection(ctrl, graph, ntpwgts, niparts);
break;
default:
gk_errexit(SIGERR, "Unknown initial partition type: %d\n", ctrl->iptype);
}
IFSET(ctrl->dbglvl, METIS_DBG_IPART, printf("Initial Cut: %"PRIDX"\n", graph->mincut));
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopcputimer(ctrl->InitPartTmr));
ctrl->dbglvl = dbglvl;
}
void InitKWayPartitioningGrow(ctrl_t *ctrl, graph_t *graph)
{
idx_t ntrials, curobj=0, bestobj=0;
idx_t *bestwhere=NULL;
WCOREPUSH;
bestwhere = iwspacemalloc(ctrl, graph->nvtxs);
for (ntrials=0; ntrials<ctrl->nIparts; ntrials++) {
GrowKWayPartitioning(ctrl, graph);
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
curobj = graph->mincut;
break;
case METIS_OBJTYPE_VOL:
curobj = graph->minvol;
break;
default:
gk_errexit(SIGERR, "Unknown objtype: %d\n", ctrl->objtype);
}
printf(" Ipart.%"PRIDX" curobj: %"PRIDX"\n", ntrials, curobj);
if (ntrials == 0 || bestobj > curobj) {
icopy(graph->nvtxs, graph->where, bestwhere);
bestobj = curobj;
}
}
if (bestobj != curobj)
icopy(graph->nvtxs, bestwhere, graph->where);
printf(" Ipart Select bestobj: %"PRIDX"\n", bestobj);
WCOREPOP;
}
void gk_gkmcorePop(gk_mcore_t *mcore)
{
while (mcore->cmop > 0) {
mcore->cmop--;
switch (mcore->mops[mcore->cmop].type) {
case GK_MOPT_MARK: /* push marker */
goto DONE;
break;
case GK_MOPT_HEAP: /* heap free */
free(mcore->mops[mcore->cmop].ptr);
mcore->cur_hallocs -= mcore->mops[mcore->cmop].nbytes;
break;
default:
gk_errexit(SIGMEM, "Unknown mop type of %d\n", mcore->mops[mcore->cmop].type);
}
}
DONE:
;
}
void InitSeparator(ctrl_t *ctrl, graph_t *graph, idx_t niparts)
{
real_t ntpwgts[2] = {0.5, 0.5};
mdbglvl_et dbglvl;
dbglvl = ctrl->dbglvl;
IFSET(ctrl->dbglvl, METIS_DBG_REFINE, ctrl->dbglvl -= METIS_DBG_REFINE);
IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO, ctrl->dbglvl -= METIS_DBG_MOVEINFO);
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startcputimer(ctrl->InitPartTmr));
/* this is required for the cut-based part of the refinement */
Setup2WayBalMultipliers(ctrl, graph, ntpwgts);
switch (ctrl->iptype) {
case METIS_IPTYPE_EDGE:
if (graph->nedges == 0)
RandomBisection(ctrl, graph, ntpwgts, niparts);
else
GrowBisection(ctrl, graph, ntpwgts, niparts);
Compute2WayPartitionParams(ctrl, graph);
ConstructSeparator(ctrl, graph);
break;
case METIS_IPTYPE_NODE:
GrowBisectionNode(ctrl, graph, ntpwgts, niparts);
break;
default:
gk_errexit(SIGERR, "Unkown iptype of %"PRIDX"\n", ctrl->iptype);
}
IFSET(ctrl->dbglvl, METIS_DBG_IPART, printf("Initial Sep: %"PRIDX"\n", graph->mincut));
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopcputimer(ctrl->InitPartTmr));
ctrl->dbglvl = dbglvl;
}
void Refine2WayNode(ctrl_t *ctrl, graph_t *orggraph, graph_t *graph)
{
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startcputimer(ctrl->UncoarsenTmr));
if (graph == orggraph) {
Compute2WayNodePartitionParams(ctrl, graph);
}
else {
do {
graph = graph->finer;
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startcputimer(ctrl->ProjectTmr));
Project2WayNodePartition(ctrl, graph);
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopcputimer(ctrl->ProjectTmr));
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startcputimer(ctrl->RefTmr));
FM_2WayNodeBalance(ctrl, graph);
ASSERT(CheckNodePartitionParams(graph));
switch (ctrl->rtype) {
case METIS_RTYPE_SEP2SIDED:
FM_2WayNodeRefine2Sided(ctrl, graph, ctrl->niter);
break;
case METIS_RTYPE_SEP1SIDED:
FM_2WayNodeRefine1Sided(ctrl, graph, ctrl->niter);
break;
default:
gk_errexit(SIGERR, "Unknown rtype of %d\n", ctrl->rtype);
}
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopcputimer(ctrl->RefTmr));
} while (graph != orggraph);
}
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopcputimer(ctrl->UncoarsenTmr));
}
void ComputeKWayPartitionParams(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, j, k, l, nvtxs, ncon, nparts, nbnd, mincut, me, other;
idx_t *xadj, *vwgt, *adjncy, *adjwgt, *pwgts, *where, *bndind, *bndptr;
nparts = ctrl->nparts;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
pwgts = iset(nparts*ncon, 0, graph->pwgts);
bndind = graph->bndind;
bndptr = iset(nvtxs, -1, graph->bndptr);
nbnd = mincut = 0;
/* Compute pwgts */
if (ncon == 1) {
for (i=0; i<nvtxs; i++) {
ASSERT(where[i] >= 0 && where[i] < nparts);
pwgts[where[i]] += vwgt[i];
}
}
else {
for (i=0; i<nvtxs; i++) {
me = where[i];
for (j=0; j<ncon; j++)
pwgts[me*ncon+j] += vwgt[i*ncon+j];
}
}
/* Compute the required info for refinement */
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
{
ckrinfo_t *myrinfo;
cnbr_t *mynbrs;
memset(graph->ckrinfo, 0, sizeof(ckrinfo_t)*nvtxs);
cnbrpoolReset(ctrl);
for (i=0; i<nvtxs; i++) {
me = where[i];
myrinfo = graph->ckrinfo+i;
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (me == where[adjncy[j]])
myrinfo->id += adjwgt[j];
else
myrinfo->ed += adjwgt[j];
}
/* Time to compute the particular external degrees */
if (myrinfo->ed > 0) {
mincut += myrinfo->ed;
myrinfo->inbr = cnbrpoolGetNext(ctrl, xadj[i+1]-xadj[i]+1);
mynbrs = ctrl->cnbrpool + myrinfo->inbr;
for (j=xadj[i]; j<xadj[i+1]; j++) {
other = where[adjncy[j]];
if (me != other) {
for (k=0; k<myrinfo->nnbrs; k++) {
if (mynbrs[k].pid == other) {
mynbrs[k].ed += adjwgt[j];
break;
}
}
if (k == myrinfo->nnbrs) {
mynbrs[k].pid = other;
mynbrs[k].ed = adjwgt[j];
myrinfo->nnbrs++;
}
}
}
ASSERT(myrinfo->nnbrs <= xadj[i+1]-xadj[i]);
/* Only ed-id>=0 nodes are considered to be in the boundary */
if (myrinfo->ed-myrinfo->id >= 0)
BNDInsert(nbnd, bndind, bndptr, i);
}
else {
myrinfo->inbr = -1;
}
}
graph->mincut = mincut/2;
graph->nbnd = nbnd;
}
ASSERT(CheckBnd2(graph));
break;
case METIS_OBJTYPE_VOL:
{
vkrinfo_t *myrinfo;
vnbr_t *mynbrs;
memset(graph->vkrinfo, 0, sizeof(vkrinfo_t)*nvtxs);
vnbrpoolReset(ctrl);
/* Compute now the id/ed degrees */
for (i=0; i<nvtxs; i++) {
me = where[i];
myrinfo = graph->vkrinfo+i;
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (me == where[adjncy[j]])
myrinfo->nid++;
else
myrinfo->ned++;
}
/* Time to compute the particular external degrees */
if (myrinfo->ned > 0) {
mincut += myrinfo->ned;
myrinfo->inbr = vnbrpoolGetNext(ctrl, xadj[i+1]-xadj[i]+1);
mynbrs = ctrl->vnbrpool + myrinfo->inbr;
for (j=xadj[i]; j<xadj[i+1]; j++) {
other = where[adjncy[j]];
if (me != other) {
for (k=0; k<myrinfo->nnbrs; k++) {
if (mynbrs[k].pid == other) {
mynbrs[k].ned++;
break;
}
}
if (k == myrinfo->nnbrs) {
mynbrs[k].gv = 0;
mynbrs[k].pid = other;
mynbrs[k].ned = 1;
myrinfo->nnbrs++;
}
}
}
ASSERT(myrinfo->nnbrs <= xadj[i+1]-xadj[i]);
}
else {
myrinfo->inbr = -1;
}
}
graph->mincut = mincut/2;
ComputeKWayVolGains(ctrl, graph);
}
ASSERT(graph->minvol == ComputeVolume(graph, graph->where));
break;
default:
gk_errexit(SIGERR, "Unknown objtype of %d\n", ctrl->objtype);
}
}
void gk_graph_ComputeBestFOrdering0(gk_graph_t *graph, int v, int type,
int32_t **r_perm, int32_t **r_iperm)
{
ssize_t j, jj, *xadj;
int i, k, u, nvtxs;
int32_t *adjncy, *perm, *degrees, *minIDs, *open;
gk_i32pq_t *queue;
if (graph->nvtxs <= 0)
return;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
/* the degree of the vertices in the closed list */
degrees = gk_i32smalloc(nvtxs, 0, "gk_graph_ComputeBestFOrdering: degrees");
/* the minimum vertex ID of an open vertex to the closed list */
minIDs = gk_i32smalloc(nvtxs, nvtxs+1, "gk_graph_ComputeBestFOrdering: minIDs");
/* the open list */
open = gk_i32malloc(nvtxs, "gk_graph_ComputeBestFOrdering: open");
/* if perm[i] >= 0, then perm[i] is the order of vertex i;
otherwise perm[i] == -1.
*/
perm = gk_i32smalloc(nvtxs, -1, "gk_graph_ComputeBestFOrdering: perm");
/* create the queue and put everything in it */
queue = gk_i32pqCreate(nvtxs);
for (i=0; i<nvtxs; i++)
gk_i32pqInsert(queue, i, 0);
gk_i32pqUpdate(queue, v, 1);
open[0] = v;
/* start processing the nodes */
for (i=0; i<nvtxs; i++) {
if ((v = gk_i32pqGetTop(queue)) == -1)
gk_errexit(SIGERR, "The priority queue got empty ahead of time [i=%d].\n", i);
if (perm[v] != -1)
gk_errexit(SIGERR, "The perm[%d] has already been set.\n", v);
perm[v] = i;
for (j=xadj[v]; j<xadj[v+1]; j++) {
u = adjncy[j];
if (perm[u] == -1) {
degrees[u]++;
minIDs[u] = (i < minIDs[u] ? i : minIDs[u]);
switch (type) {
case 1: /* DFS */
gk_i32pqUpdate(queue, u, 1);
break;
case 2: /* Max in closed degree */
gk_i32pqUpdate(queue, u, degrees[u]);
break;
case 3: /* Sum of orders in closed list */
for (k=0, jj=xadj[u]; jj<xadj[u+1]; jj++) {
if (perm[adjncy[jj]] != -1)
k += perm[adjncy[jj]];
}
gk_i32pqUpdate(queue, u, k);
break;
case 4: /* Sum of order-differences (w.r.t. current number) in closed
list (updated once in a while) */
for (k=0, jj=xadj[u]; jj<xadj[u+1]; jj++) {
if (perm[adjncy[jj]] != -1)
k += (i-perm[adjncy[jj]]);
}
gk_i32pqUpdate(queue, u, k);
break;
default:
;
}
}
}
}
/* time to decide what to return */
if (r_perm != NULL) {
*r_perm = perm;
perm = NULL;
}
if (r_iperm != NULL) {
/* use the 'degrees' array to build the iperm array */
for (i=0; i<nvtxs; i++)
degrees[perm[i]] = i;
*r_iperm = degrees;
degrees = NULL;
}
/* cleanup memory */
gk_i32pqDestroy(queue);
gk_free((void **)&perm, °rees, &minIDs, &open, LTERM);
}
void CreateGraphNodal(idx_t ne, idx_t nn, idx_t *eptr, idx_t *eind,
idx_t **r_xadj, idx_t **r_adjncy)
{
idx_t i, j, nnbrs;
idx_t *nptr, *nind;
idx_t *xadj, *adjncy;
idx_t *marker, *nbrs;
/* construct the node-element list first */
nptr = ismalloc(nn+1, 0, "CreateGraphNodal: nptr");
nind = imalloc(eptr[ne], "CreateGraphNodal: nind");
for (i=0; i<ne; i++) {
for (j=eptr[i]; j<eptr[i+1]; j++)
nptr[eind[j]]++;
}
MAKECSR(i, nn, nptr);
for (i=0; i<ne; i++) {
for (j=eptr[i]; j<eptr[i+1]; j++)
nind[nptr[eind[j]]++] = i;
}
SHIFTCSR(i, nn, nptr);
/* Allocate memory for xadj, since you know its size.
These are done using standard malloc as they are returned
to the calling function */
if ((xadj = (idx_t *)malloc((nn+1)*sizeof(idx_t))) == NULL)
gk_errexit(SIGMEM, "***Failed to allocate memory for xadj.\n");
*r_xadj = xadj;
iset(nn+1, 0, xadj);
/* allocate memory for working arrays used by FindCommonElements */
marker = ismalloc(nn, 0, "CreateGraphNodal: marker");
nbrs = imalloc(nn, "CreateGraphNodal: nbrs");
for (i=0; i<nn; i++) {
xadj[i] = FindCommonNodes(i, nptr[i+1]-nptr[i], nind+nptr[i], eptr,
eind, marker, nbrs);
}
MAKECSR(i, nn, xadj);
/* Allocate memory for adjncy, since you now know its size.
These are done using standard malloc as they are returned
to the calling function */
if ((adjncy = (idx_t *)malloc(xadj[nn]*sizeof(idx_t))) == NULL) {
free(xadj);
*r_xadj = NULL;
gk_errexit(SIGMEM, "***Failed to allocate memory for adjncy.\n");
}
*r_adjncy = adjncy;
for (i=0; i<nn; i++) {
nnbrs = FindCommonNodes(i, nptr[i+1]-nptr[i], nind+nptr[i], eptr,
eind, marker, nbrs);
for (j=0; j<nnbrs; j++)
adjncy[xadj[i]++] = nbrs[j];
}
SHIFTCSR(i, nn, xadj);
gk_free((void **)&nptr, &nind, &marker, &nbrs, LTERM);
}
void InitKWayPartitioningRB(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, options[METIS_NOPTIONS], curobj=0;
idx_t *bestwhere=NULL;
real_t *ubvec=NULL;
int status;
METIS_SetDefaultOptions(options);
options[METIS_OPTION_NITER] = 10;
options[METIS_OPTION_OBJTYPE] = METIS_OBJTYPE_CUT;
options[METIS_OPTION_NO2HOP] = ctrl->no2hop;
options[METIS_OPTION_ONDISK] = ctrl->ondisk;
ubvec = rmalloc(graph->ncon, "InitKWayPartitioning: ubvec");
for (i=0; i<graph->ncon; i++)
ubvec[i] = (real_t)pow(ctrl->ubfactors[i], 1.0/log(ctrl->nparts));
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
case METIS_OBJTYPE_VOL:
options[METIS_OPTION_NCUTS] = ctrl->nIparts;
status = METIS_PartGraphRecursive(&graph->nvtxs, &graph->ncon,
graph->xadj, graph->adjncy, graph->vwgt, graph->vsize,
graph->adjwgt, &ctrl->nparts, ctrl->tpwgts, ubvec,
options, &curobj, graph->where);
if (status != METIS_OK)
gk_errexit(SIGERR, "Failed during initial partitioning\n");
break;
#ifdef XXX /* This does not seem to help */
case METIS_OBJTYPE_VOL:
bestwhere = imalloc(graph->nvtxs, "InitKWayPartitioning: bestwhere");
options[METIS_OPTION_NCUTS] = 2;
ntrials = (ctrl->nIparts+1)/2;
for (i=0; i<ntrials; i++) {
status = METIS_PartGraphRecursive(&graph->nvtxs, &graph->ncon,
graph->xadj, graph->adjncy, graph->vwgt, graph->vsize,
graph->adjwgt, &ctrl->nparts, ctrl->tpwgts, ubvec,
options, &curobj, graph->where);
if (status != METIS_OK)
gk_errexit(SIGERR, "Failed during initial partitioning\n");
curobj = ComputeVolume(graph, graph->where);
if (i == 0 || bestobj > curobj) {
bestobj = curobj;
if (i < ntrials-1)
icopy(graph->nvtxs, graph->where, bestwhere);
}
if (bestobj == 0)
break;
}
if (bestobj != curobj)
icopy(graph->nvtxs, bestwhere, graph->where);
break;
#endif
default:
gk_errexit(SIGERR, "Unknown objtype: %d\n", ctrl->objtype);
}
gk_free((void **)&ubvec, &bestwhere, LTERM);
}
int METIS_PartGraphKway(idx_t *nvtxs, idx_t *ncon, idx_t *xadj, idx_t *adjncy,
idx_t *vwgt, idx_t *vsize, idx_t *adjwgt, idx_t *nparts,
real_t *tpwgts, real_t *ubvec, idx_t *options, idx_t *objval,
idx_t *part)
{
int sigrval=0, renumber=0;
graph_t *graph;
ctrl_t *ctrl;
/* set up malloc cleaning code and signal catchers */
if (!gk_malloc_init())
return METIS_ERROR_MEMORY;
gk_sigtrap();
if ((sigrval = gk_sigcatch()) != 0)
goto SIGTHROW;
/* set up the run parameters */
ctrl = SetupCtrl(METIS_OP_KMETIS, options, *ncon, *nparts, tpwgts, ubvec);
if (!ctrl) {
gk_siguntrap();
return METIS_ERROR_INPUT;
}
/* if required, change the numbering to 0 */
if (ctrl->numflag == 1) {
Change2CNumbering(*nvtxs, xadj, adjncy);
renumber = 1;
}
/* set up the graph */
graph = SetupGraph(ctrl, *nvtxs, *ncon, xadj, adjncy, vwgt, vsize, adjwgt);
/* set up multipliers for making balance computations easier */
SetupKWayBalMultipliers(ctrl, graph);
/* set various run parameters that depend on the graph */
if (ctrl->iptype == METIS_IPTYPE_METISRB) {
ctrl->CoarsenTo = gk_max((*nvtxs)/(40*gk_log2(*nparts)), 30*(*nparts));
ctrl->CoarsenTo = 10*(*nparts);
ctrl->nIparts = (ctrl->CoarsenTo == 30*(*nparts) ? 4 : 5);
}
else {
ctrl->CoarsenTo = 10*(*nparts);
ctrl->nIparts = 10;
}
/* take care contiguity requests for disconnected graphs */
if (ctrl->contig && !IsConnected(graph, 0))
gk_errexit(SIGERR, "METIS Error: A contiguous partition is requested for a non-contiguous input graph.\n");
/* allocate workspace memory */
AllocateWorkSpace(ctrl, graph);
/* start the partitioning */
IFSET(ctrl->dbglvl, METIS_DBG_TIME, InitTimers(ctrl));
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->TotalTmr));
*objval = MlevelKWayPartitioning(ctrl, graph, part);
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->TotalTmr));
IFSET(ctrl->dbglvl, METIS_DBG_TIME, PrintTimers(ctrl));
/* clean up */
FreeCtrl(&ctrl);
SIGTHROW:
/* if required, change the numbering back to 1 */
if (renumber)
Change2FNumbering(*nvtxs, xadj, adjncy, part);
gk_siguntrap();
gk_malloc_cleanup(0);
return metis_rcode(sigrval);
}
idx_t MlevelKWayPartitioning(ctrl_t *ctrl, graph_t *graph, idx_t *part)
{
idx_t i, objval=0, curobj=0, bestobj=0;
real_t curbal=0.0, bestbal=0.0;
graph_t *cgraph;
for (i=0; i<ctrl->ncuts; i++) {
cgraph = CoarsenGraph(ctrl, graph);
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->InitPartTmr));
AllocateKWayPartitionMemory(ctrl, cgraph);
/* Compute the initial partitioning */
switch (ctrl->iptype) {
case METIS_IPTYPE_METISRB:
FreeWorkSpace(ctrl); /* Release the work space, for the recursive metis call */
InitKWayPartitioningRB(ctrl, cgraph);
AllocateWorkSpace(ctrl, graph); /* Re-allocate the work space */
break;
case METIS_IPTYPE_GROW:
AllocateRefinementWorkSpace(ctrl, 2*cgraph->nedges);
InitKWayPartitioningGrow(ctrl, cgraph);
break;
default:
gk_errexit(SIGERR, "Unknown iptype: %d\n", ctrl->iptype);
}
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->InitPartTmr));
IFSET(ctrl->dbglvl, METIS_DBG_IPART, printf("Initial %"PRIDX \
"-way partitioning cut: %"PRIDX"\n", ctrl->nparts, objval));
RefineKWay(ctrl, graph, cgraph);
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
curobj = graph->mincut;
break;
case METIS_OBJTYPE_VOL:
curobj = graph->minvol;
break;
default:
gk_errexit(SIGERR, "Unknown objtype: %d\n", ctrl->objtype);
}
curbal = ComputeLoadImbalanceDiff(graph, ctrl->nparts, ctrl->pijbm, ctrl->ubfactors);
if (i == 0
|| (curbal <= 0.0005 && bestobj > curobj)
|| (bestbal > 0.0005 && curbal < bestbal)) {
icopy(graph->nvtxs, graph->where, part);
bestobj = curobj;
bestbal = curbal;
}
FreeRData(graph);
if (bestobj == 0)
break;
}
FreeGraph(&graph);
return bestobj;
}
gk_graph_t *gk_graph_Read(char *filename, int format, int isfewgts,
int isfvwgts, int isfvsizes)
{
ssize_t i, k, l;
size_t nfields, nvtxs, nedges, fmt, ncon, lnlen;
int32_t ival;
float fval;
int readsizes=0, readwgts=0, readvals=0, numbering=0;
char *line=NULL, *head, *tail, fmtstr[256];
FILE *fpin=NULL;
gk_graph_t *graph=NULL;
if (!gk_fexists(filename))
gk_errexit(SIGERR, "File %s does not exist!\n", filename);
if (format == GK_GRAPH_FMT_METIS) {
fpin = gk_fopen(filename, "r", "gk_graph_Read: fpin");
do {
if (gk_getline(&line, &lnlen, fpin) <= 0)
gk_errexit(SIGERR, "Premature end of input file: file:%s\n", filename);
} while (line[0] == '%');
fmt = ncon = 0;
nfields = sscanf(line, "%zu %zu %zu %zu", &nvtxs, &nedges, &fmt, &ncon);
if (nfields < 2)
gk_errexit(SIGERR, "Header line must contain at least 2 integers (#vtxs and #edges).\n");
nedges *= 2;
if (fmt > 111)
gk_errexit(SIGERR, "Cannot read this type of file format [fmt=%zu]!\n", fmt);
sprintf(fmtstr, "%03zu", fmt%1000);
readsizes = (fmtstr[0] == '1');
readwgts = (fmtstr[1] == '1');
readvals = (fmtstr[2] == '1');
numbering = 1;
ncon = (ncon == 0 ? 1 : ncon);
}
else {
gk_errexit(SIGERR, "Unrecognized format: %d\n", format);
}
graph = gk_graph_Create();
graph->nvtxs = nvtxs;
graph->xadj = gk_zmalloc(nvtxs+1, "gk_graph_Read: xadj");
graph->adjncy = gk_i32malloc(nedges, "gk_graph_Read: adjncy");
if (readvals) {
if (isfewgts)
graph->fadjwgt = gk_fmalloc(nedges, "gk_graph_Read: fadjwgt");
else
graph->iadjwgt = gk_i32malloc(nedges, "gk_graph_Read: iadjwgt");
}
if (readsizes) {
if (isfvsizes)
graph->fvsizes = gk_fmalloc(nvtxs, "gk_graph_Read: fvsizes");
else
graph->ivsizes = gk_i32malloc(nvtxs, "gk_graph_Read: ivsizes");
}
if (readwgts) {
if (isfvwgts)
graph->fvwgts = gk_fmalloc(nvtxs*ncon, "gk_graph_Read: fvwgts");
else
graph->ivwgts = gk_i32malloc(nvtxs*ncon, "gk_graph_Read: ivwgts");
}
/*----------------------------------------------------------------------
* Read the sparse graph file
*---------------------------------------------------------------------*/
numbering = (numbering ? - 1 : 0);
for (graph->xadj[0]=0, k=0, i=0; i<nvtxs; i++) {
do {
if (gk_getline(&line, &lnlen, fpin) == -1)
gk_errexit(SIGERR, "Pregraphure end of input file: file while reading row %d\n", i);
} while (line[0] == '%');
head = line;
tail = NULL;
/* Read vertex sizes */
if (readsizes) {
if (isfvsizes) {
#ifdef __MSC__
graph->fvsizes[i] = (float)strtod(head, &tail);
#else
graph->fvsizes[i] = strtof(head, &tail);
#endif
if (tail == head)
gk_errexit(SIGERR, "The line for vertex %zd does not have size information\n", i+1);
if (graph->fvsizes[i] < 0)
gk_errexit(SIGERR, "The size for vertex %zd must be >= 0\n", i+1);
}
else {
graph->ivsizes[i] = strtol(head, &tail, 0);
if (tail == head)
gk_errexit(SIGERR, "The line for vertex %zd does not have size information\n", i+1);
if (graph->ivsizes[i] < 0)
gk_errexit(SIGERR, "The size for vertex %zd must be >= 0\n", i+1);
}
head = tail;
}
/* Read vertex weights */
if (readwgts) {
for (l=0; l<ncon; l++) {
if (isfvwgts) {
#ifdef __MSC__
graph->fvwgts[i*ncon+l] = (float)strtod(head, &tail);
#else
graph->fvwgts[i*ncon+l] = strtof(head, &tail);
#endif
if (tail == head)
gk_errexit(SIGERR, "The line for vertex %zd does not have enough weights "
"for the %d constraints.\n", i+1, ncon);
if (graph->fvwgts[i*ncon+l] < 0)
gk_errexit(SIGERR, "The weight vertex %zd and constraint %zd must be >= 0\n", i+1, l);
}
else {
graph->ivwgts[i*ncon+l] = strtol(head, &tail, 0);
if (tail == head)
gk_errexit(SIGERR, "The line for vertex %zd does not have enough weights "
"for the %d constraints.\n", i+1, ncon);
if (graph->ivwgts[i*ncon+l] < 0)
gk_errexit(SIGERR, "The weight vertex %zd and constraint %zd must be >= 0\n", i+1, l);
}
head = tail;
}
}
/* Read the rest of the row */
while (1) {
ival = (int)strtol(head, &tail, 0);
if (tail == head)
break;
head = tail;
if ((graph->adjncy[k] = ival + numbering) < 0)
gk_errexit(SIGERR, "Error: Invalid column number %d at row %zd.\n", ival, i);
if (readvals) {
if (isfewgts) {
#ifdef __MSC__
fval = (float)strtod(head, &tail);
#else
fval = strtof(head, &tail);
#endif
if (tail == head)
gk_errexit(SIGERR, "Value could not be found for edge! Vertex:%zd, NNZ:%zd\n", i, k);
graph->fadjwgt[k] = fval;
}
else {
ival = strtol(head, &tail, 0);
if (tail == head)
gk_errexit(SIGERR, "Value could not be found for edge! Vertex:%zd, NNZ:%zd\n", i, k);
graph->iadjwgt[k] = ival;
}
head = tail;
}
k++;
}
graph->xadj[i+1] = k;
}
if (k != nedges)
gk_errexit(SIGERR, "gk_graph_Read: Something wrong with the number of edges in "
"the input file. nedges=%zd, Actualnedges=%zd.\n", nedges, k);
gk_fclose(fpin);
gk_free((void **)&line, LTERM);
return graph;
}
void ProjectKWayPartition(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, j, k, nvtxs, nbnd, nparts, me, other, istart, iend, tid, ted;
idx_t *xadj, *adjncy, *adjwgt;
idx_t *cmap, *where, *bndptr, *bndind, *cwhere, *htable;
graph_t *cgraph;
WCOREPUSH;
nparts = ctrl->nparts;
cgraph = graph->coarser;
cwhere = cgraph->where;
nvtxs = graph->nvtxs;
cmap = graph->cmap;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
AllocateKWayPartitionMemory(ctrl, graph);
where = graph->where;
bndind = graph->bndind;
bndptr = iset(nvtxs, -1, graph->bndptr);
htable = iset(nparts, -1, iwspacemalloc(ctrl, nparts));
/* Compute the required info for refinement */
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
ASSERT(CheckBnd2(cgraph));
{
ckrinfo_t *myrinfo;
cnbr_t *mynbrs;
/* 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] = cgraph->ckrinfo[k].ed; /* For optimization */
}
memset(graph->ckrinfo, 0, sizeof(ckrinfo_t)*nvtxs);
cnbrpoolReset(ctrl);
for (nbnd=0, i=0; i<nvtxs; i++) {
istart = xadj[i];
iend = xadj[i+1];
myrinfo = graph->ckrinfo+i;
if (cmap[i] == 0) { /* Interior node. Note that cmap[i] = crinfo[cmap[i]].ed */
for (tid=0, j=istart; j<iend; j++)
tid += adjwgt[j];
myrinfo->id = tid;
myrinfo->inbr = -1;
}
else { /* Potentially an interface node */
myrinfo->inbr = cnbrpoolGetNext(ctrl, iend-istart+1);
mynbrs = ctrl->cnbrpool + myrinfo->inbr;
me = where[i];
for (tid=0, ted=0, j=istart; j<iend; j++) {
other = where[adjncy[j]];
if (me == other) {
tid += adjwgt[j];
}
else {
ted += adjwgt[j];
if ((k = htable[other]) == -1) {
htable[other] = myrinfo->nnbrs;
mynbrs[myrinfo->nnbrs].pid = other;
mynbrs[myrinfo->nnbrs++].ed = adjwgt[j];
}
else {
mynbrs[k].ed += adjwgt[j];
}
}
}
myrinfo->id = tid;
myrinfo->ed = ted;
/* Remove space for edegrees if it was interior */
if (ted == 0) {
ctrl->nbrpoolcpos -= iend-istart+1;
myrinfo->inbr = -1;
}
else {
if (ted-tid >= 0)
BNDInsert(nbnd, bndind, bndptr, i);
for (j=0; j<myrinfo->nnbrs; j++)
htable[mynbrs[j].pid] = -1;
}
}
}
graph->nbnd = nbnd;
}
ASSERT(CheckBnd2(graph));
break;
case METIS_OBJTYPE_VOL:
{
vkrinfo_t *myrinfo;
vnbr_t *mynbrs;
ASSERT(cgraph->minvol == ComputeVolume(cgraph, cgraph->where));
/* 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] = cgraph->vkrinfo[k].ned; /* For optimization */
}
memset(graph->vkrinfo, 0, sizeof(vkrinfo_t)*nvtxs);
vnbrpoolReset(ctrl);
for (i=0; i<nvtxs; i++) {
istart = xadj[i];
iend = xadj[i+1];
myrinfo = graph->vkrinfo+i;
if (cmap[i] == 0) { /* Note that cmap[i] = crinfo[cmap[i]].ed */
myrinfo->nid = iend-istart;
myrinfo->inbr = -1;
}
else { /* Potentially an interface node */
myrinfo->inbr = vnbrpoolGetNext(ctrl, iend-istart+1);
mynbrs = ctrl->vnbrpool + myrinfo->inbr;
me = where[i];
for (tid=0, ted=0, j=istart; j<iend; j++) {
other = where[adjncy[j]];
if (me == other) {
tid++;
}
else {
ted++;
if ((k = htable[other]) == -1) {
htable[other] = myrinfo->nnbrs;
mynbrs[myrinfo->nnbrs].gv = 0;
mynbrs[myrinfo->nnbrs].pid = other;
mynbrs[myrinfo->nnbrs++].ned = 1;
}
else {
mynbrs[k].ned++;
}
}
}
myrinfo->nid = tid;
myrinfo->ned = ted;
/* Remove space for edegrees if it was interior */
if (ted == 0) {
ctrl->nbrpoolcpos -= iend-istart+1;
myrinfo->inbr = -1;
}
else {
for (j=0; j<myrinfo->nnbrs; j++)
htable[mynbrs[j].pid] = -1;
}
}
}
ComputeKWayVolGains(ctrl, graph);
ASSERT(graph->minvol == ComputeVolume(graph, graph->where));
}
break;
default:
gk_errexit(SIGERR, "Unknown objtype of %d\n", ctrl->objtype);
}
graph->mincut = cgraph->mincut;
icopy(nparts*graph->ncon, cgraph->pwgts, graph->pwgts);
FreeGraph(&graph->coarser);
graph->coarser = NULL;
WCOREPOP;
}
/*************************************************************************
* This function converts a mesh into a dual graph
**************************************************************************/
int ParMETIS_V3_Mesh2Dual(idx_t *elmdist, idx_t *eptr, idx_t *eind,
idx_t *numflag, idx_t *ncommon, idx_t **r_xadj,
idx_t **r_adjncy, MPI_Comm *comm)
{
idx_t i, j, jj, k, kk, m;
idx_t npes, mype, pe, count, mask, pass;
idx_t nelms, lnns, my_nns, node;
idx_t firstelm, firstnode, lnode, nrecv, nsend;
idx_t *scounts, *rcounts, *sdispl, *rdispl;
idx_t *nodedist, *nmap, *auxarray;
idx_t *gnptr, *gnind, *nptr, *nind, *myxadj=NULL, *myadjncy = NULL;
idx_t *sbuffer, *rbuffer, *htable;
ikv_t *nodelist, *recvbuffer;
idx_t maxcount, *ind, *wgt;
idx_t gmaxnode, gminnode;
size_t curmem;
gk_malloc_init();
curmem = gk_GetCurMemoryUsed();
/* Get basic comm info */
gkMPI_Comm_size(*comm, &npes);
gkMPI_Comm_rank(*comm, &mype);
nelms = elmdist[mype+1]-elmdist[mype];
if (*numflag > 0)
ChangeNumberingMesh(elmdist, eptr, eind, NULL, NULL, NULL, npes, mype, 1);
mask = (1<<11)-1;
/*****************************/
/* Determine number of nodes */
/*****************************/
gminnode = GlobalSEMinComm(*comm, imin(eptr[nelms], eind));
for (i=0; i<eptr[nelms]; i++)
eind[i] -= gminnode;
gmaxnode = GlobalSEMaxComm(*comm, imax(eptr[nelms], eind));
/**************************/
/* Check for input errors */
/**************************/
ASSERT(nelms > 0);
/* construct node distribution array */
nodedist = ismalloc(npes+1, 0, "nodedist");
for (nodedist[0]=0, i=0,j=gmaxnode+1; i<npes; i++) {
k = j/(npes-i);
nodedist[i+1] = nodedist[i]+k;
j -= k;
}
my_nns = nodedist[mype+1]-nodedist[mype];
firstnode = nodedist[mype];
nodelist = ikvmalloc(eptr[nelms], "nodelist");
auxarray = imalloc(eptr[nelms], "auxarray");
htable = ismalloc(gk_max(my_nns, mask+1), -1, "htable");
scounts = imalloc(npes, "scounts");
rcounts = imalloc(npes, "rcounts");
sdispl = imalloc(npes+1, "sdispl");
rdispl = imalloc(npes+1, "rdispl");
/*********************************************/
/* first find a local numbering of the nodes */
/*********************************************/
for (i=0; i<nelms; i++) {
for (j=eptr[i]; j<eptr[i+1]; j++) {
nodelist[j].key = eind[j];
nodelist[j].val = j;
auxarray[j] = i; /* remember the local element ID that uses this node */
}
}
ikvsorti(eptr[nelms], nodelist);
for (count=1, i=1; i<eptr[nelms]; i++) {
if (nodelist[i].key > nodelist[i-1].key)
count++;
}
lnns = count;
nmap = imalloc(lnns, "nmap");
/* renumber the nodes of the elements array */
count = 1;
nmap[0] = nodelist[0].key;
eind[nodelist[0].val] = 0;
nodelist[0].val = auxarray[nodelist[0].val]; /* Store the local element ID */
for (i=1; i<eptr[nelms]; i++) {
if (nodelist[i].key > nodelist[i-1].key) {
nmap[count] = nodelist[i].key;
count++;
}
eind[nodelist[i].val] = count-1;
nodelist[i].val = auxarray[nodelist[i].val]; /* Store the local element ID */
}
gkMPI_Barrier(*comm);
/**********************************************************/
/* perform comms necessary to construct node-element list */
/**********************************************************/
iset(npes, 0, scounts);
for (pe=i=0; i<eptr[nelms]; i++) {
while (nodelist[i].key >= nodedist[pe+1])
pe++;
scounts[pe] += 2;
}
ASSERT(pe < npes);
gkMPI_Alltoall((void *)scounts, 1, IDX_T, (void *)rcounts, 1, IDX_T, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
icopy(npes, rcounts, rdispl);
MAKECSR(i, npes, rdispl);
ASSERT(sdispl[npes] == eptr[nelms]*2);
nrecv = rdispl[npes]/2;
recvbuffer = ikvmalloc(gk_max(1, nrecv), "recvbuffer");
gkMPI_Alltoallv((void *)nodelist, scounts, sdispl, IDX_T, (void *)recvbuffer,
rcounts, rdispl, IDX_T, *comm);
/**************************************/
/* construct global node-element list */
/**************************************/
gnptr = ismalloc(my_nns+1, 0, "gnptr");
for (i=0; i<npes; i++) {
for (j=rdispl[i]/2; j<rdispl[i+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
ASSERT(lnode >= 0 && lnode < my_nns)
gnptr[lnode]++;
}
}
MAKECSR(i, my_nns, gnptr);
gnind = imalloc(gk_max(1, gnptr[my_nns]), "gnind");
for (pe=0; pe<npes; pe++) {
firstelm = elmdist[pe];
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
gnind[gnptr[lnode]++] = recvbuffer[j].val+firstelm;
}
}
SHIFTCSR(i, my_nns, gnptr);
/*********************************************************/
/* send the node-element info to the relevant processors */
/*********************************************************/
iset(npes, 0, scounts);
/* use a hash table to ensure that each node is sent to a proc only once */
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
if (htable[lnode] == -1) {
scounts[pe] += gnptr[lnode+1]-gnptr[lnode];
htable[lnode] = 1;
}
}
/* now reset the hash table */
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
htable[lnode] = -1;
}
}
gkMPI_Alltoall((void *)scounts, 1, IDX_T, (void *)rcounts, 1, IDX_T, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
/* create the send buffer */
nsend = sdispl[npes];
sbuffer = imalloc(gk_max(1, nsend), "sbuffer");
count = 0;
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
if (htable[lnode] == -1) {
for (k=gnptr[lnode]; k<gnptr[lnode+1]; k++) {
if (k == gnptr[lnode])
sbuffer[count++] = -1*(gnind[k]+1);
else
sbuffer[count++] = gnind[k];
}
htable[lnode] = 1;
}
}
ASSERT(count == sdispl[pe+1]);
/* now reset the hash table */
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
htable[lnode] = -1;
}
}
icopy(npes, rcounts, rdispl);
MAKECSR(i, npes, rdispl);
nrecv = rdispl[npes];
rbuffer = imalloc(gk_max(1, nrecv), "rbuffer");
gkMPI_Alltoallv((void *)sbuffer, scounts, sdispl, IDX_T, (void *)rbuffer,
rcounts, rdispl, IDX_T, *comm);
k = -1;
nptr = ismalloc(lnns+1, 0, "nptr");
nind = rbuffer;
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]; j<rdispl[pe+1]; j++) {
if (nind[j] < 0) {
k++;
nind[j] = (-1*nind[j])-1;
}
nptr[k]++;
}
}
MAKECSR(i, lnns, nptr);
ASSERT(k+1 == lnns);
ASSERT(nptr[lnns] == nrecv)
myxadj = *r_xadj = (idx_t *)malloc(sizeof(idx_t)*(nelms+1));
if (myxadj == NULL)
gk_errexit(SIGMEM, "Failed to allocate memory for the dual graph's xadj array.\n");
iset(nelms+1, 0, myxadj);
iset(mask+1, -1, htable);
firstelm = elmdist[mype];
/* Two passes -- in first pass, simply find out the memory requirements */
maxcount = 200;
ind = imalloc(maxcount, "ParMETIS_V3_Mesh2Dual: ind");
wgt = imalloc(maxcount, "ParMETIS_V3_Mesh2Dual: wgt");
for (pass=0; pass<2; pass++) {
for (i=0; i<nelms; i++) {
for (count=0, j=eptr[i]; j<eptr[i+1]; j++) {
node = eind[j];
for (k=nptr[node]; k<nptr[node+1]; k++) {
if ((kk=nind[k]) == firstelm+i)
continue;
m = htable[(kk&mask)];
if (m == -1) {
ind[count] = kk;
wgt[count] = 1;
htable[(kk&mask)] = count++;
}
else {
if (ind[m] == kk) {
wgt[m]++;
}
else {
for (jj=0; jj<count; jj++) {
if (ind[jj] == kk) {
wgt[jj]++;
break;
}
}
if (jj == count) {
ind[count] = kk;
wgt[count++] = 1;
}
}
}
/* Adjust the memory.
This will be replaced by a idxrealloc() when GKlib will be incorporated */
if (count == maxcount-1) {
maxcount *= 2;
ind = irealloc(ind, maxcount, "ParMETIS_V3_Mesh2Dual: ind");
wgt = irealloc(wgt, maxcount, "ParMETIS_V3_Mesh2Dual: wgt");
}
}
}
for (j=0; j<count; j++) {
htable[(ind[j]&mask)] = -1;
if (wgt[j] >= *ncommon) {
if (pass == 0)
myxadj[i]++;
else
myadjncy[myxadj[i]++] = ind[j];
}
}
}
if (pass == 0) {
MAKECSR(i, nelms, myxadj);
myadjncy = *r_adjncy = (idx_t *)malloc(sizeof(idx_t)*myxadj[nelms]);
if (myadjncy == NULL)
gk_errexit(SIGMEM, "Failed to allocate memory for dual graph's adjncy array.\n");
}
else {
SHIFTCSR(i, nelms, myxadj);
}
}
/*****************************************/
/* correctly renumber the elements array */
/*****************************************/
for (i=0; i<eptr[nelms]; i++)
eind[i] = nmap[eind[i]] + gminnode;
if (*numflag == 1)
ChangeNumberingMesh(elmdist, eptr, eind, myxadj, myadjncy, NULL, npes, mype, 0);
/* do not free nodelist, recvbuffer, rbuffer */
gk_free((void **)&nodedist, &nodelist, &auxarray, &htable, &scounts, &rcounts,
&sdispl, &rdispl, &nmap, &recvbuffer, &gnptr, &gnind, &sbuffer, &rbuffer,
&nptr, &ind, &wgt, LTERM);
if (gk_GetCurMemoryUsed() - curmem > 0) {
printf("ParMETIS appears to have a memory leak of %zdbytes. Report this.\n",
(ssize_t)(gk_GetCurMemoryUsed() - curmem));
}
gk_malloc_cleanup(0);
return METIS_OK;
}
void ComputeSubDomainGraph(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, ii, j, pid, other, nparts, nvtxs, nnbrs;
idx_t *xadj, *adjncy, *adjwgt, *where;
idx_t *pptr, *pind;
idx_t nads=0, *vadids, *vadwgts;
WCOREPUSH;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
nparts = ctrl->nparts;
vadids = ctrl->pvec1;
vadwgts = iset(nparts, 0, ctrl->pvec2);
pptr = iwspacemalloc(ctrl, nparts+1);
pind = iwspacemalloc(ctrl, nvtxs);
iarray2csr(nvtxs, nparts, where, pptr, pind);
for (pid=0; pid<nparts; pid++) {
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
{
ckrinfo_t *rinfo;
cnbr_t *nbrs;
rinfo = graph->ckrinfo;
for (nads=0, ii=pptr[pid]; ii<pptr[pid+1]; ii++) {
i = pind[ii];
ASSERT(pid == where[i]);
if (rinfo[i].ed > 0) {
nnbrs = rinfo[i].nnbrs;
nbrs = ctrl->cnbrpool + rinfo[i].inbr;
for (j=0; j<nnbrs; j++) {
other = nbrs[j].pid;
if (vadwgts[other] == 0)
vadids[nads++] = other;
vadwgts[other] += nbrs[j].ed;
}
}
}
}
break;
case METIS_OBJTYPE_VOL:
{
vkrinfo_t *rinfo;
vnbr_t *nbrs;
rinfo = graph->vkrinfo;
for (nads=0, ii=pptr[pid]; ii<pptr[pid+1]; ii++) {
i = pind[ii];
ASSERT(pid == where[i]);
if (rinfo[i].ned > 0) {
nnbrs = rinfo[i].nnbrs;
nbrs = ctrl->vnbrpool + rinfo[i].inbr;
for (j=0; j<nnbrs; j++) {
other = nbrs[j].pid;
if (vadwgts[other] == 0)
vadids[nads++] = other;
vadwgts[other] += nbrs[j].ned;
}
}
}
}
break;
default:
gk_errexit(SIGERR, "Unknown objtype: %d\n", ctrl->objtype);
}
/* See if you have enough memory to store the adjacent info for that subdomain */
if (ctrl->maxnads[pid] < nads) {
ctrl->maxnads[pid] = 2*nads;
ctrl->adids[pid] = irealloc(ctrl->adids[pid], ctrl->maxnads[pid],
"ComputeSubDomainGraph: adids[pid]");
ctrl->adwgts[pid] = irealloc(ctrl->adwgts[pid], ctrl->maxnads[pid],
"ComputeSubDomainGraph: adids[pid]");
}
ctrl->nads[pid] = nads;
for (j=0; j<nads; j++) {
ctrl->adids[pid][j] = vadids[j];
ctrl->adwgts[pid][j] = vadwgts[vadids[j]];
vadwgts[vadids[j]] = 0;
}
}
WCOREPOP;
}
void EliminateSubDomainEdges(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, ii, j, k, ncon, nparts, scheme, pid_from, pid_to, me, other, nvtxs,
total, max, avg, totalout, nind=0, ncand=0, ncand2, target, target2,
nadd, bestnadd=0;
idx_t min, move, *cpwgt;
idx_t *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *maxpwgt,
*mypmat, *otherpmat, *kpmat, *ind;
idx_t *nads, **adids, **adwgts;
ikv_t *cand, *cand2;
ipq_t queue;
real_t *tpwgts, badfactor=1.4;
idx_t *pptr, *pind;
idx_t *vmarker=NULL, *pmarker=NULL, *modind=NULL; /* volume specific work arrays */
WCOREPUSH;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
adjwgt = (ctrl->objtype == METIS_OBJTYPE_VOL ? NULL : graph->adjwgt);
where = graph->where;
pwgts = graph->pwgts; /* We assume that this is properly initialized */
nparts = ctrl->nparts;
tpwgts = ctrl->tpwgts;
cpwgt = iwspacemalloc(ctrl, ncon);
maxpwgt = iwspacemalloc(ctrl, nparts*ncon);
ind = iwspacemalloc(ctrl, nvtxs);
otherpmat = iset(nparts, 0, iwspacemalloc(ctrl, nparts));
cand = ikvwspacemalloc(ctrl, nparts);
cand2 = ikvwspacemalloc(ctrl, nparts);
pptr = iwspacemalloc(ctrl, nparts+1);
pind = iwspacemalloc(ctrl, nvtxs);
iarray2csr(nvtxs, nparts, where, pptr, pind);
if (ctrl->objtype == METIS_OBJTYPE_VOL) {
/* Vol-refinement specific working arrays */
modind = iwspacemalloc(ctrl, nvtxs);
vmarker = iset(nvtxs, 0, iwspacemalloc(ctrl, nvtxs));
pmarker = iset(nparts, -1, iwspacemalloc(ctrl, nparts));
}
/* Compute the pmat matrix and ndoms */
ComputeSubDomainGraph(ctrl, graph);
nads = ctrl->nads;
adids = ctrl->adids;
adwgts = ctrl->adwgts;
mypmat = iset(nparts, 0, ctrl->pvec1);
kpmat = iset(nparts, 0, ctrl->pvec2);
/* Compute the maximum allowed weight for each domain */
for (i=0; i<nparts; i++) {
for (j=0; j<ncon; j++)
maxpwgt[i*ncon+j] =
(ncon == 1 ? 1.25 : 1.025)*tpwgts[i]*graph->tvwgt[j]*ctrl->ubfactors[j];
}
ipqInit(&queue, nparts);
/* Get into the loop eliminating subdomain connections */
while (1) {
total = isum(nparts, nads, 1);
avg = total/nparts;
max = nads[iargmax(nparts, nads)];
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("Adjacent Subdomain Stats: Total: %3"PRIDX", "
"Max: %3"PRIDX"[%zu], Avg: %3"PRIDX"\n",
total, max, iargmax(nparts, nads), avg));
if (max < badfactor*avg)
break;
/* Add the subdomains that you will try to reduce their connectivity */
ipqReset(&queue);
for (i=0; i<nparts; i++) {
if (nads[i] >= avg + (max-avg)/2)
ipqInsert(&queue, i, nads[i]);
}
move = 0;
while ((me = ipqGetTop(&queue)) != -1) {
totalout = isum(nads[me], adwgts[me], 1);
for (ncand2=0, i=0; i<nads[me]; i++) {
mypmat[adids[me][i]] = adwgts[me][i];
/* keep track of the weakly connected adjacent subdomains */
if (2*nads[me]*adwgts[me][i] < totalout) {
cand2[ncand2].val = adids[me][i];
cand2[ncand2++].key = adwgts[me][i];
}
}
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("Me: %"PRIDX", Degree: %4"PRIDX", TotalOut: %"PRIDX",\n",
me, nads[me], totalout));
/* Sort the connections according to their cut */
ikvsorti(ncand2, cand2);
/* Two schemes are used for eliminating subdomain edges.
The first, tries to eliminate subdomain edges by moving remote groups
of vertices to subdomains that 'me' is already connected to.
The second, tries to eliminate subdomain edges by moving entire sets of
my vertices that connect to the 'other' subdomain to a subdomain that
I'm already connected to.
These two schemes are applied in sequence. */
target = target2 = -1;
for (scheme=0; scheme<2; scheme++) {
for (min=0; min<ncand2; min++) {
other = cand2[min].val;
/* pid_from is the subdomain from where the vertices will be removed.
pid_to is the adjacent subdomain to pid_from that defines the
(me, other) subdomain edge that needs to be removed */
if (scheme == 0) {
pid_from = other;
pid_to = me;
}
else {
pid_from = me;
pid_to = other;
}
/* Go and find the vertices in 'other' that are connected in 'me' */
for (nind=0, ii=pptr[pid_from]; ii<pptr[pid_from+1]; ii++) {
i = pind[ii];
ASSERT(where[i] == pid_from);
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (where[adjncy[j]] == pid_to) {
ind[nind++] = i;
break;
}
}
}
/* Go and construct the otherpmat to see where these nind vertices are
connected to */
iset(ncon, 0, cpwgt);
for (ncand=0, ii=0; ii<nind; ii++) {
i = ind[ii];
iaxpy(ncon, 1, vwgt+i*ncon, 1, cpwgt, 1);
for (j=xadj[i]; j<xadj[i+1]; j++) {
if ((k = where[adjncy[j]]) == pid_from)
continue;
if (otherpmat[k] == 0)
cand[ncand++].val = k;
otherpmat[k] += (adjwgt ? adjwgt[j] : 1);
}
}
for (i=0; i<ncand; i++) {
cand[i].key = otherpmat[cand[i].val];
ASSERT(cand[i].key > 0);
}
ikvsortd(ncand, cand);
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("\tMinOut: %4"PRIDX", to: %3"PRIDX", TtlWgt: %5"PRIDX"[#:%"PRIDX"]\n",
mypmat[other], other, isum(ncon, cpwgt, 1), nind));
/* Go through and select the first domain that is common with 'me', and does
not increase the nads[target] higher than nads[me], subject to the maxpwgt
constraint. Traversal is done from the mostly connected to the least. */
for (i=0; i<ncand; i++) {
k = cand[i].val;
if (mypmat[k] > 0) {
/* Check if balance will go off */
if (!ivecaxpylez(ncon, 1, cpwgt, pwgts+k*ncon, maxpwgt+k*ncon))
continue;
/* get a dense vector out of k's connectivity */
for (j=0; j<nads[k]; j++)
kpmat[adids[k][j]] = adwgts[k][j];
/* Check if the move to domain k will increase the nads of another
subdomain j that the set of vertices being moved are connected
to but domain k is not connected to. */
for (j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && kpmat[j] == 0 && nads[j]+1 >= nads[me])
break;
}
/* There were no bad second level effects. See if you can find a
subdomain to move to. */
if (j == nparts) {
for (nadd=0, j=0; j<nparts; j++) {
if (otherpmat[j] > 0 && kpmat[j] == 0)
nadd++;
}
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("\t\tto=%"PRIDX", nadd=%"PRIDX", %"PRIDX"\n", k, nadd, nads[k]));
if (nads[k]+nadd < nads[me]) {
if (target2 == -1 || nads[target2]+bestnadd > nads[k]+nadd ||
(nads[target2]+bestnadd == nads[k]+nadd && bestnadd > nadd)) {
target2 = k;
bestnadd = nadd;
}
}
if (nadd == 0)
target = k;
}
/* reset kpmat for the next iteration */
for (j=0; j<nads[k]; j++)
kpmat[adids[k][j]] = 0;
}
if (target != -1)
break;
}
/* reset the otherpmat for the next iteration */
for (i=0; i<ncand; i++)
otherpmat[cand[i].val] = 0;
if (target == -1 && target2 != -1)
target = target2;
if (target != -1) {
IFSET(ctrl->dbglvl, METIS_DBG_CONNINFO,
printf("\t\tScheme: %"PRIDX". Moving to %"PRIDX"\n", scheme, target));
move = 1;
break;
}
}
if (target != -1)
break; /* A move was found. No need to try the other scheme */
}
/* reset the mypmat for next iteration */
for (i=0; i<nads[me]; i++)
mypmat[adids[me][i]] = 0;
/* Note that once a target is found the above loops exit right away. So the
following variables are valid */
if (target != -1) {
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:
MoveGroupMinConnForCut(ctrl, graph, target, nind, ind);
break;
case METIS_OBJTYPE_VOL:
MoveGroupMinConnForVol(ctrl, graph, target, nind, ind, vmarker,
pmarker, modind);
break;
default:
gk_errexit(SIGERR, "Unknown objtype of %d\n", ctrl->objtype);
}
/* Update the csr representation of the partitioning vector */
iarray2csr(nvtxs, nparts, where, pptr, pind);
}
}
if (move == 0)
break;
}
ipqFree(&queue);
WCOREPOP;
}
ctrl_t *SetupCtrl(moptype_et optype, idx_t *options, idx_t ncon, idx_t nparts,
real_t *tpwgts, real_t *ubvec)
{
idx_t i, j;
ctrl_t *ctrl;
ctrl = (ctrl_t *)gk_malloc(sizeof(ctrl_t), "SetupCtrl: ctrl");
memset((void *)ctrl, 0, sizeof(ctrl_t));
switch (optype) {
case METIS_OP_PMETIS:
ctrl->objtype = GETOPTION(options, METIS_OPTION_OBJTYPE, METIS_OBJTYPE_CUT);
ctrl->ctype = GETOPTION(options, METIS_OPTION_CTYPE, METIS_CTYPE_SHEM);
ctrl->rtype = METIS_RTYPE_FM;
ctrl->ncuts = GETOPTION(options, METIS_OPTION_NCUTS, 1);
ctrl->niter = GETOPTION(options, METIS_OPTION_NITER, 10);
ctrl->seed = GETOPTION(options, METIS_OPTION_SEED, -1);
ctrl->dbglvl = GETOPTION(options, METIS_OPTION_DBGLVL, 0);
if (ncon == 1) {
ctrl->iptype = GETOPTION(options, METIS_OPTION_IPTYPE, METIS_IPTYPE_GROW);
ctrl->ufactor = GETOPTION(options, METIS_OPTION_UFACTOR, PMETIS_DEFAULT_UFACTOR);
ctrl->CoarsenTo = 20;
}
else {
ctrl->iptype = GETOPTION(options, METIS_OPTION_IPTYPE, METIS_IPTYPE_RANDOM);
ctrl->ufactor = GETOPTION(options, METIS_OPTION_UFACTOR, MCPMETIS_DEFAULT_UFACTOR);
ctrl->CoarsenTo = 100;
}
break;
case METIS_OP_KMETIS:
ctrl->objtype = GETOPTION(options, METIS_OPTION_OBJTYPE, METIS_OBJTYPE_CUT);
ctrl->ctype = GETOPTION(options, METIS_OPTION_CTYPE, METIS_CTYPE_SHEM);
ctrl->iptype = METIS_IPTYPE_METISRB;
ctrl->rtype = METIS_RTYPE_GREEDY;
ctrl->ncuts = GETOPTION(options, METIS_OPTION_NCUTS, 1);
ctrl->niter = GETOPTION(options, METIS_OPTION_NITER, 10);
ctrl->ufactor = GETOPTION(options, METIS_OPTION_UFACTOR, KMETIS_DEFAULT_UFACTOR);
ctrl->minconn = GETOPTION(options, METIS_OPTION_MINCONN, 0);
ctrl->contig = GETOPTION(options, METIS_OPTION_CONTIG, 0);
ctrl->seed = GETOPTION(options, METIS_OPTION_SEED, -1);
ctrl->dbglvl = GETOPTION(options, METIS_OPTION_DBGLVL, 0);
break;
case METIS_OP_OMETIS:
ctrl->objtype = GETOPTION(options, METIS_OPTION_OBJTYPE, METIS_OBJTYPE_NODE);
ctrl->ctype = GETOPTION(options, METIS_OPTION_CTYPE, METIS_CTYPE_SHEM);
ctrl->rtype = GETOPTION(options, METIS_OPTION_RTYPE, METIS_RTYPE_SEP1SIDED);
ctrl->iptype = GETOPTION(options, METIS_OPTION_IPTYPE, METIS_IPTYPE_EDGE);
ctrl->nseps = GETOPTION(options, METIS_OPTION_NSEPS, 1);
ctrl->niter = GETOPTION(options, METIS_OPTION_NITER, 10);
ctrl->ufactor = GETOPTION(options, METIS_OPTION_UFACTOR, OMETIS_DEFAULT_UFACTOR);
ctrl->compress = GETOPTION(options, METIS_OPTION_COMPRESS, 1);
ctrl->ccorder = GETOPTION(options, METIS_OPTION_CCORDER, 0);
ctrl->seed = GETOPTION(options, METIS_OPTION_SEED, -1);
ctrl->dbglvl = GETOPTION(options, METIS_OPTION_DBGLVL, 0);
ctrl->pfactor = 0.1*GETOPTION(options, METIS_OPTION_PFACTOR, 0);
ctrl->CoarsenTo = 100;
break;
default:
gk_errexit(SIGERR, "Unknown optype of %d\n", optype);
}
ctrl->numflag = GETOPTION(options, METIS_OPTION_NUMBERING, 0);
ctrl->optype = optype;
ctrl->ncon = ncon;
ctrl->nparts = nparts;
ctrl->maxvwgt = ismalloc(ncon, 0, "SetupCtrl: maxvwgt");
/* setup the target partition weights */
if (ctrl->optype != METIS_OP_OMETIS) {
ctrl->tpwgts = rmalloc(nparts*ncon, "SetupCtrl: ctrl->tpwgts");
if (tpwgts) {
rcopy(nparts*ncon, tpwgts, ctrl->tpwgts);
}
else {
for (i=0; i<nparts; i++) {
for (j=0; j<ncon; j++)
ctrl->tpwgts[i*ncon+j] = 1.0/nparts;
}
}
}
else { /* METIS_OP_OMETIS */
/* this is required to allow the pijbm to be defined properly for
the edge-based refinement during initial partitioning */
ctrl->tpwgts = rsmalloc(2, .5, "SetupCtrl: ctrl->tpwgts");
}
/* setup the ubfactors */
ctrl->ubfactors = rsmalloc(ctrl->ncon, I2RUBFACTOR(ctrl->ufactor), "SetupCtrl: ubfactors");
if (ubvec)
rcopy(ctrl->ncon, ubvec, ctrl->ubfactors);
for (i=0; i<ctrl->ncon; i++)
ctrl->ubfactors[i] += 0.0000499;
/* Allocate memory for balance multipliers.
Note that for PMETIS/OMETIS routines the memory allocated is more
than required as balance multipliers for 2 parts is sufficient. */
ctrl->pijbm = rmalloc(nparts*ncon, "SetupCtrl: ctrl->pijbm");
InitRandom(ctrl->seed);
IFSET(ctrl->dbglvl, METIS_DBG_INFO, PrintCtrl(ctrl));
if (!CheckParams(ctrl)) {
FreeCtrl(&ctrl);
return NULL;
}
else {
return ctrl;
}
}
void EliminateComponents(ctrl_t *ctrl, graph_t *graph) {
idx_t i, ii, j, jj, k, me, nparts, nvtxs, ncon, ncmps, other,
ncand, target;
idx_t *xadj, *adjncy, *vwgt, *adjwgt, *where, *pwgts;
idx_t *cptr, *cind, *cpvec, *pcptr, *pcind, *cwhere;
idx_t cid, bestcid, *cwgt, *bestcwgt;
idx_t ntodo, oldntodo, *todo;
rkv_t *cand;
real_t *tpwgts;
idx_t *vmarker = NULL, *pmarker = NULL, *modind = NULL; /* volume specific work arrays */
WCOREPUSH;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
adjwgt = (ctrl->objtype == METIS_OBJTYPE_VOL ? NULL : graph->adjwgt);
where = graph->where;
pwgts = graph->pwgts;
nparts = ctrl->nparts;
tpwgts = ctrl->tpwgts;
cptr = iwspacemalloc(ctrl, nvtxs + 1);
cind = iwspacemalloc(ctrl, nvtxs);
ncmps = FindPartitionInducedComponents(graph, where, cptr, cind);
IFSET(ctrl->dbglvl, METIS_DBG_CONTIGINFO,
printf("I found %"PRIDX" components, for this %"PRIDX"-way partition\n",
ncmps, nparts));
/* There are more components than partitions */
if (ncmps > nparts) {
cwgt = iwspacemalloc(ctrl, ncon);
bestcwgt = iwspacemalloc(ctrl, ncon);
cpvec = iwspacemalloc(ctrl, nparts);
pcptr = iset(nparts + 1, 0, iwspacemalloc(ctrl, nparts + 1));
pcind = iwspacemalloc(ctrl, ncmps);
cwhere = iset(nvtxs, -1, iwspacemalloc(ctrl, nvtxs));
todo = iwspacemalloc(ctrl, ncmps);
cand = (rkv_t *) wspacemalloc(ctrl, nparts * sizeof(rkv_t));
if (ctrl->objtype == METIS_OBJTYPE_VOL) {
/* Vol-refinement specific working arrays */
modind = iwspacemalloc(ctrl, nvtxs);
vmarker = iset(nvtxs, 0, iwspacemalloc(ctrl, nvtxs));
pmarker = iset(nparts, -1, iwspacemalloc(ctrl, nparts));
}
/* Get a CSR representation of the components-2-partitions mapping */
for (i = 0; i < ncmps; i++) {
pcptr[where[cind[cptr[i]]]]++;
}
MAKECSR(i, nparts, pcptr);
for (i = 0; i < ncmps; i++) {
pcind[pcptr[where[cind[cptr[i]]]]++] = i;
}
SHIFTCSR(i, nparts, pcptr);
/* Assign the heaviest component of each partition to its original partition */
for (ntodo = 0, i = 0; i < nparts; i++) {
if (pcptr[i + 1] - pcptr[i] == 1) {
bestcid = pcind[pcptr[i]];
} else {
for (bestcid = -1, j = pcptr[i]; j < pcptr[i + 1]; j++) {
cid = pcind[j];
iset(ncon, 0, cwgt);
for (ii = cptr[cid]; ii < cptr[cid + 1]; ii++) {
iaxpy(ncon, 1, vwgt + cind[ii] * ncon, 1, cwgt, 1);
}
if (bestcid == -1 || isum(ncon, bestcwgt, 1) < isum(ncon, cwgt, 1)) {
bestcid = cid;
icopy(ncon, cwgt, bestcwgt);
}
}
/* Keep track of those that need to be dealt with */
for (j = pcptr[i]; j < pcptr[i + 1]; j++) {
if (pcind[j] != bestcid) {
todo[ntodo++] = pcind[j];
}
}
}
for (j = cptr[bestcid]; j < cptr[bestcid + 1]; j++) {
ASSERT(where[cind[j]] == i);
cwhere[cind[j]] = i;
}
}
while (ntodo > 0) {
oldntodo = ntodo;
for (i = 0; i < ntodo; i++) {
cid = todo[i];
me = where[cind[cptr[cid]]]; /* Get the domain of this component */
/* Determine the weight of the block to be moved */
iset(ncon, 0, cwgt);
for (j = cptr[cid]; j < cptr[cid + 1]; j++) {
iaxpy(ncon, 1, vwgt + cind[j] * ncon, 1, cwgt, 1);
}
IFSET(ctrl->dbglvl, METIS_DBG_CONTIGINFO,
printf("Trying to move %"PRIDX" [%"PRIDX"] from %"PRIDX"\n",
cid, isum(ncon, cwgt, 1), me));
/* Determine the connectivity */
iset(nparts, 0, cpvec);
for (j = cptr[cid]; j < cptr[cid + 1]; j++) {
ii = cind[j];
for (jj = xadj[ii]; jj < xadj[ii + 1]; jj++) {
if (cwhere[adjncy[jj]] != -1) {
cpvec[cwhere[adjncy[jj]]] += (adjwgt ? adjwgt[jj] : 1);
}
}
}
/* Put the neighbors into a cand[] array for sorting */
for (ncand = 0, j = 0; j < nparts; j++) {
if (cpvec[j] > 0) {
cand[ncand].key = cpvec[j];
cand[ncand++].val = j;
}
}
if (ncand == 0) {
continue;
}
rkvsortd(ncand, cand);
/* Limit the moves to only the top candidates, which are defined as
those with connectivity at least 50% of the best.
This applies only when ncon=1, as for multi-constraint, balancing
will be hard. */
if (ncon == 1) {
for (j = 1; j < ncand; j++) {
if (cand[j].key < .5 * cand[0].key) {
break;
}
}
ncand = j;
}
/* Now among those, select the one with the best balance */
target = cand[0].val;
for (j = 1; j < ncand; j++) {
if (BetterBalanceKWay(ncon, cwgt, ctrl->ubfactors,
1, pwgts + target * ncon, ctrl->pijbm + target * ncon,
1, pwgts + cand[j].val * ncon, ctrl->pijbm + cand[j].val * ncon)) {
target = cand[j].val;
}
}
IFSET(ctrl->dbglvl, METIS_DBG_CONTIGINFO,
printf("\tMoving it to %"PRIDX" [%"PRIDX"] [%"PRIDX"]\n", target, cpvec[target], ncand));
/* Note that as a result of a previous movement, a connected component may
now will like to stay to its original partition */
if (target != me) {
switch (ctrl->objtype) {
case METIS_OBJTYPE_CUT:MoveGroupContigForCut(ctrl, graph, target, cid, cptr, cind);
break;
case METIS_OBJTYPE_VOL:
MoveGroupContigForVol(ctrl, graph, target, cid, cptr, cind,
vmarker, pmarker, modind);
break;
default:gk_errexit(SIGERR, "Unknown objtype %d\n", ctrl->objtype);
}
}
/* Update the cwhere vector */
for (j = cptr[cid]; j < cptr[cid + 1]; j++) {
cwhere[cind[j]] = target;
}
todo[i] = todo[--ntodo];
}
if (oldntodo == ntodo) {
IFSET(ctrl->dbglvl, METIS_DBG_CONTIGINFO, printf("Stopped at ntodo: %"PRIDX"\n", ntodo));
break;
}
}
for (i = 0; i < nvtxs; i++) {
ASSERT(where[i] == cwhere[i]);
}
}
WCOREPOP;
}
gk_graph_t *gk_graph_Prune(gk_graph_t *graph, int what, int minf, int maxf)
{
ssize_t i, j, nnz;
int nrows, ncols;
ssize_t *rowptr, *nrowptr;
int *rowind, *nrowind, *collen;
float *rowval, *nrowval;
gk_graph_t *ngraph;
ngraph = gk_graph_Create();
nrows = ngraph->nrows = graph->nrows;
ncols = ngraph->ncols = graph->ncols;
rowptr = graph->rowptr;
rowind = graph->rowind;
rowval = graph->rowval;
nrowptr = ngraph->rowptr = gk_zmalloc(nrows+1, "gk_graph_Prune: nrowptr");
nrowind = ngraph->rowind = gk_imalloc(rowptr[nrows], "gk_graph_Prune: nrowind");
nrowval = ngraph->rowval = gk_fmalloc(rowptr[nrows], "gk_graph_Prune: nrowval");
switch (what) {
case GK_CSR_COL:
collen = gk_ismalloc(ncols, 0, "gk_graph_Prune: collen");
for (i=0; i<nrows; i++) {
for (j=rowptr[i]; j<rowptr[i+1]; j++) {
ASSERT(rowind[j] < ncols);
collen[rowind[j]]++;
}
}
for (i=0; i<ncols; i++)
collen[i] = (collen[i] >= minf && collen[i] <= maxf ? 1 : 0);
nrowptr[0] = 0;
for (nnz=0, i=0; i<nrows; i++) {
for (j=rowptr[i]; j<rowptr[i+1]; j++) {
if (collen[rowind[j]]) {
nrowind[nnz] = rowind[j];
nrowval[nnz] = rowval[j];
nnz++;
}
}
nrowptr[i+1] = nnz;
}
gk_free((void **)&collen, LTERM);
break;
case GK_CSR_ROW:
nrowptr[0] = 0;
for (nnz=0, i=0; i<nrows; i++) {
if (rowptr[i+1]-rowptr[i] >= minf && rowptr[i+1]-rowptr[i] <= maxf) {
for (j=rowptr[i]; j<rowptr[i+1]; j++, nnz++) {
nrowind[nnz] = rowind[j];
nrowval[nnz] = rowval[j];
}
}
nrowptr[i+1] = nnz;
}
break;
default:
gk_graph_Free(&ngraph);
gk_errexit(SIGERR, "Unknown prunning type of %d\n", what);
return NULL;
}
return ngraph;
}
void gk_graph_ComputeBestFOrdering(gk_graph_t *graph, int v, int type,
int32_t **r_perm, int32_t **r_iperm)
{
ssize_t j, jj, *xadj;
int i, k, u, nvtxs, nopen, ntodo;
int32_t *adjncy, *perm, *degrees, *wdegrees, *sod, *level, *ot, *pos;
gk_i32pq_t *queue;
if (graph->nvtxs <= 0)
return;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
/* the degree of the vertices in the closed list */
degrees = gk_i32smalloc(nvtxs, 0, "gk_graph_ComputeBestFOrdering: degrees");
/* the weighted degree of the vertices in the closed list for type==3 */
wdegrees = gk_i32smalloc(nvtxs, 0, "gk_graph_ComputeBestFOrdering: wdegrees");
/* the sum of differences for type==4 */
sod = gk_i32smalloc(nvtxs, 0, "gk_graph_ComputeBestFOrdering: sod");
/* the encountering level of a vertex type==5 */
level = gk_i32smalloc(nvtxs, 0, "gk_graph_ComputeBestFOrdering: level");
/* The open+todo list of vertices.
The vertices from [0..nopen] are the open vertices.
The vertices from [nopen..ntodo) are the todo vertices.
*/
ot = gk_i32incset(nvtxs, 0, gk_i32malloc(nvtxs, "gk_graph_FindComponents: ot"));
/* For a vertex that has not been explored, pos[i] is the position in the ot list. */
pos = gk_i32incset(nvtxs, 0, gk_i32malloc(nvtxs, "gk_graph_FindComponents: pos"));
/* if perm[i] >= 0, then perm[i] is the order of vertex i; otherwise perm[i] == -1. */
perm = gk_i32smalloc(nvtxs, -1, "gk_graph_ComputeBestFOrdering: perm");
/* create the queue and put the starting vertex in it */
queue = gk_i32pqCreate(nvtxs);
gk_i32pqInsert(queue, v, 1);
/* put v at the front of the open list */
pos[0] = ot[0] = v;
pos[v] = ot[v] = 0;
nopen = 1;
ntodo = nvtxs;
/* start processing the nodes */
for (i=0; i<nvtxs; i++) {
if (nopen == 0) { /* deal with non-connected graphs */
gk_i32pqInsert(queue, ot[0], 1);
nopen++;
}
if ((v = gk_i32pqGetTop(queue)) == -1)
gk_errexit(SIGERR, "The priority queue got empty ahead of time [i=%d].\n", i);
if (perm[v] != -1)
gk_errexit(SIGERR, "The perm[%d] has already been set.\n", v);
perm[v] = i;
if (ot[pos[v]] != v)
gk_errexit(SIGERR, "Something went wrong [ot[pos[%d]]!=%d.\n", v, v);
if (pos[v] >= nopen)
gk_errexit(SIGERR, "The position of v is not in open list. pos[%d]=%d is >=%d.\n", v, pos[v], nopen);
/* remove v from the open list and re-arrange the todo part of the list */
ot[pos[v]] = ot[nopen-1];
pos[ot[nopen-1]] = pos[v];
if (ntodo > nopen) {
ot[nopen-1] = ot[ntodo-1];
pos[ot[ntodo-1]] = nopen-1;
}
nopen--;
ntodo--;
for (j=xadj[v]; j<xadj[v+1]; j++) {
u = adjncy[j];
if (perm[u] == -1) {
/* update ot list, if u is not in the open list by putting it at the end
of the open list. */
if (degrees[u] == 0) {
ot[pos[u]] = ot[nopen];
pos[ot[nopen]] = pos[u];
ot[nopen] = u;
pos[u] = nopen;
nopen++;
level[u] = level[v]+1;
gk_i32pqInsert(queue, u, 0);
}
/* update the in-closed degree */
degrees[u]++;
/* update the queues based on the type */
switch (type) {
case 1: /* DFS */
gk_i32pqUpdate(queue, u, 1000*(i+1)+degrees[u]);
break;
case 2: /* Max in closed degree */
gk_i32pqUpdate(queue, u, degrees[u]);
break;
case 3: /* Sum of orders in closed list */
wdegrees[u] += i;
gk_i32pqUpdate(queue, u, wdegrees[u]);
break;
case 4: /* Sum of order-differences */
/* this is handled at the end of the loop */
;
break;
case 5: /* BFS with in degree priority */
gk_i32pqUpdate(queue, u, -(1000*level[u] - degrees[u]));
break;
case 6: /* Hybrid of 1+2 */
gk_i32pqUpdate(queue, u, (i+1)*degrees[u]);
break;
default:
;
}
}
}
if (type == 4) { /* update all the vertices in the open list */
for (j=0; j<nopen; j++) {
u = ot[j];
if (perm[u] != -1)
gk_errexit(SIGERR, "For i=%d, the open list contains a closed vertex: ot[%zd]=%d, perm[%d]=%d.\n", i, j, u, u, perm[u]);
sod[u] += degrees[u];
if (i<1000 || i%25==0)
gk_i32pqUpdate(queue, u, sod[u]);
}
}
/*
for (j=0; j<ntodo; j++) {
if (pos[ot[j]] != j)
gk_errexit(SIGERR, "pos[ot[%zd]] != %zd.\n", j, j);
}
*/
}
/* time to decide what to return */
if (r_perm != NULL) {
*r_perm = perm;
perm = NULL;
}
if (r_iperm != NULL) {
/* use the 'degrees' array to build the iperm array */
for (i=0; i<nvtxs; i++)
degrees[perm[i]] = i;
*r_iperm = degrees;
degrees = NULL;
}
/* cleanup memory */
gk_i32pqDestroy(queue);
gk_free((void **)&perm, °rees, &wdegrees, &sod, &ot, &pos, &level, LTERM);
}
