void gk_array2csr(size_t n, size_t range, int *array, int *ptr, int *ind)
{
size_t i;
gk_iset(range+1, 0, ptr);
for (i=0; i<n; i++)
ptr[array[i]]++;
/* Compute the ptr, ind structure */
MAKECSR(i, range, ptr);
for (i=0; i<n; i++)
ind[ptr[array[i]]++] = i;
SHIFTCSR(i, range, ptr);
}
/*************************************************************************
* This function partitions a finite element mesh by partitioning its dual
* graph using KMETIS and then assigning nodes in a load balanced fashion.
**************************************************************************/
int METIS_PartMeshDual(idx_t *ne, idx_t *nn, idx_t *eptr, idx_t *eind,
idx_t *vwgt, idx_t *vsize, idx_t *ncommon, idx_t *nparts,
real_t *tpwgts, idx_t *options, idx_t *objval, idx_t *epart,
idx_t *npart)
{
int sigrval=0, renumber=0, ptype;
idx_t i, j;
idx_t *xadj=NULL, *adjncy=NULL, *nptr=NULL, *nind=NULL;
idx_t ncon=1, pnumflag=0;
int rstatus = METIS_OK;
/* 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;
renumber = GETOPTION(options, METIS_OPTION_NUMBERING, 0);
ptype = GETOPTION(options, METIS_OPTION_PTYPE, METIS_PTYPE_KWAY);
/* renumber the mesh */
if (renumber) {
ChangeMesh2CNumbering(*ne, eptr, eind);
options[METIS_OPTION_NUMBERING] = 0;
}
/* get the dual graph */
rstatus = METIS_MeshToDual(ne, nn, eptr, eind, ncommon, &pnumflag, &xadj, &adjncy);
if (rstatus != METIS_OK)
raise(SIGERR);
/* partition the graph */
if (ptype == METIS_PTYPE_KWAY)
rstatus = METIS_PartGraphKway(ne, &ncon, xadj, adjncy, vwgt, vsize, NULL,
nparts, tpwgts, NULL, options, objval, epart);
else
rstatus = METIS_PartGraphRecursive(ne, &ncon, xadj, adjncy, vwgt, vsize, NULL,
nparts, tpwgts, NULL, options, objval, epart);
if (rstatus != METIS_OK)
raise(SIGERR);
/* construct the node-element list */
nptr = ismalloc(*nn+1, 0, "METIS_PartMeshDual: nptr");
nind = imalloc(eptr[*ne], "METIS_PartMeshDual: 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);
/* partition the other side of the mesh */
InduceRowPartFromColumnPart(*nn, nptr, nind, npart, epart, *nparts, tpwgts);
gk_free((void **)&nptr, &nind, LTERM);
SIGTHROW:
if (renumber) {
ChangeMesh2FNumbering2(*ne, *nn, eptr, eind, epart, npart);
options[METIS_OPTION_NUMBERING] = 1;
}
METIS_Free(xadj);
METIS_Free(adjncy);
gk_siguntrap();
gk_malloc_cleanup(0);
return metis_rcode(sigrval);
}
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;
}
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);
}
/*************************************************************************
* This function converts a mesh into a dual graph
**************************************************************************/
void ParMETIS_V3_Mesh2Dual(idxtype *elmdist, idxtype *eptr, idxtype *eind,
int *numflag, int *ncommonnodes, idxtype **xadj,
idxtype **adjncy, MPI_Comm *comm)
{
int i, j, jj, k, kk, m;
int npes, mype, pe, count, mask, pass;
int nelms, lnns, my_nns, node;
int firstelm, firstnode, lnode, nrecv, nsend;
int *scounts, *rcounts, *sdispl, *rdispl;
idxtype *nodedist, *nmap, *auxarray;
idxtype *gnptr, *gnind, *nptr, *nind, *myxadj, *myadjncy = NULL;
idxtype *sbuffer, *rbuffer, *htable;
KeyValueType *nodelist, *recvbuffer;
idxtype ind[200], wgt[200];
int gmaxnode, gminnode;
CtrlType ctrl;
SetUpCtrl(&ctrl, -1, 0, *comm);
npes = ctrl.npes;
mype = ctrl.mype;
nelms = elmdist[mype+1]-elmdist[mype];
if (*numflag == 1)
ChangeNumberingMesh2(elmdist, eptr, eind, NULL, NULL, NULL, npes, mype, 1);
mask = (1<<11)-1;
/*****************************/
/* Determine number of nodes */
/*****************************/
gminnode = GlobalSEMin(&ctrl, eind[idxamin(eptr[nelms], eind)]);
for (i=0; i<eptr[nelms]; i++)
eind[i] -= gminnode;
gmaxnode = GlobalSEMax(&ctrl, eind[idxamax(eptr[nelms], eind)]);
/**************************/
/* Check for input errors */
/**************************/
ASSERTS(nelms > 0);
/* construct node distribution array */
nodedist = idxsmalloc(npes+1, 0, "nodedist");
for (nodedist[0]=0, i=0,j=gmaxnode+1; i<npes; i++) {
k = j/(npes-i);
nodedist[i+1] = nodedist[i]+k;
j -= k;
}
my_nns = nodedist[mype+1]-nodedist[mype];
firstnode = nodedist[mype];
nodelist = (KeyValueType *)GKmalloc(eptr[nelms]*sizeof(KeyValueType), "nodelist");
auxarray = idxmalloc(eptr[nelms], "auxarray");
htable = idxsmalloc(amax(my_nns, mask+1), -1, "htable");
scounts = imalloc(4*npes+2, "scounts");
rcounts = scounts+npes;
sdispl = scounts+2*npes;
rdispl = scounts+3*npes+1;
/*********************************************/
/* first find a local numbering of the nodes */
/*********************************************/
for (i=0; i<nelms; i++) {
for (j=eptr[i]; j<eptr[i+1]; j++) {
nodelist[j].key = eind[j];
nodelist[j].val = j;
auxarray[j] = i; /* remember the local element ID that uses this node */
}
}
ikeysort(eptr[nelms], nodelist);
for (count=1, i=1; i<eptr[nelms]; i++) {
if (nodelist[i].key > nodelist[i-1].key)
count++;
}
lnns = count;
nmap = idxmalloc(lnns, "nmap");
/* renumber the nodes of the elements array */
count = 1;
nmap[0] = nodelist[0].key;
eind[nodelist[0].val] = 0;
nodelist[0].val = auxarray[nodelist[0].val]; /* Store the local element ID */
for (i=1; i<eptr[nelms]; i++) {
if (nodelist[i].key > nodelist[i-1].key) {
nmap[count] = nodelist[i].key;
count++;
}
eind[nodelist[i].val] = count-1;
nodelist[i].val = auxarray[nodelist[i].val]; /* Store the local element ID */
}
MPI_Barrier(*comm);
/**********************************************************/
/* perform comms necessary to construct node-element list */
/**********************************************************/
iset(npes, 0, scounts);
for (pe=i=0; i<eptr[nelms]; i++) {
while (nodelist[i].key >= nodedist[pe+1])
pe++;
scounts[pe] += 2;
}
ASSERTS(pe < npes);
MPI_Alltoall((void *)scounts, 1, MPI_INT, (void *)rcounts, 1, MPI_INT, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
icopy(npes, rcounts, rdispl);
MAKECSR(i, npes, rdispl);
ASSERTS(sdispl[npes] == eptr[nelms]*2);
nrecv = rdispl[npes]/2;
recvbuffer = (KeyValueType *)GKmalloc(amax(1, nrecv)*sizeof(KeyValueType), "recvbuffer");
MPI_Alltoallv((void *)nodelist, scounts, sdispl, IDX_DATATYPE, (void *)recvbuffer,
rcounts, rdispl, IDX_DATATYPE, *comm);
/**************************************/
/* construct global node-element list */
/**************************************/
gnptr = idxsmalloc(my_nns+1, 0, "gnptr");
for (i=0; i<npes; i++) {
for (j=rdispl[i]/2; j<rdispl[i+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
ASSERTS(lnode >= 0 && lnode < my_nns)
gnptr[lnode]++;
}
}
MAKECSR(i, my_nns, gnptr);
gnind = idxmalloc(amax(1, gnptr[my_nns]), "gnind");
for (pe=0; pe<npes; pe++) {
firstelm = elmdist[pe];
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
gnind[gnptr[lnode]++] = recvbuffer[j].val+firstelm;
}
}
SHIFTCSR(i, my_nns, gnptr);
/*********************************************************/
/* send the node-element info to the relevant processors */
/*********************************************************/
iset(npes, 0, scounts);
/* use a hash table to ensure that each node is sent to a proc only once */
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
if (htable[lnode] == -1) {
scounts[pe] += gnptr[lnode+1]-gnptr[lnode];
htable[lnode] = 1;
}
}
/* now reset the hash table */
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
htable[lnode] = -1;
}
}
MPI_Alltoall((void *)scounts, 1, MPI_INT, (void *)rcounts, 1, MPI_INT, *comm);
icopy(npes, scounts, sdispl);
MAKECSR(i, npes, sdispl);
/* create the send buffer */
nsend = sdispl[npes];
sbuffer = (idxtype *)realloc(nodelist, sizeof(idxtype)*amax(1, nsend));
count = 0;
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
if (htable[lnode] == -1) {
for (k=gnptr[lnode]; k<gnptr[lnode+1]; k++) {
if (k == gnptr[lnode])
sbuffer[count++] = -1*(gnind[k]+1);
else
sbuffer[count++] = gnind[k];
}
htable[lnode] = 1;
}
}
ASSERTS(count == sdispl[pe+1]);
/* now reset the hash table */
for (j=rdispl[pe]/2; j<rdispl[pe+1]/2; j++) {
lnode = recvbuffer[j].key-firstnode;
htable[lnode] = -1;
}
}
icopy(npes, rcounts, rdispl);
MAKECSR(i, npes, rdispl);
nrecv = rdispl[npes];
rbuffer = (idxtype *)realloc(recvbuffer, sizeof(idxtype)*amax(1, nrecv));
MPI_Alltoallv((void *)sbuffer, scounts, sdispl, IDX_DATATYPE, (void *)rbuffer,
rcounts, rdispl, IDX_DATATYPE, *comm);
k = -1;
nptr = idxsmalloc(lnns+1, 0, "nptr");
nind = rbuffer;
for (pe=0; pe<npes; pe++) {
for (j=rdispl[pe]; j<rdispl[pe+1]; j++) {
if (nind[j] < 0) {
k++;
nind[j] = (-1*nind[j])-1;
}
nptr[k]++;
}
}
MAKECSR(i, lnns, nptr);
ASSERTS(k+1 == lnns);
ASSERTS(nptr[lnns] == nrecv)
myxadj = *xadj = idxsmalloc(nelms+1, 0, "xadj");
idxset(mask+1, -1, htable);
firstelm = elmdist[mype];
/* Two passes -- in first pass, simply find out the memory requirements */
for (pass=0; pass<2; pass++) {
for (i=0; i<nelms; i++) {
for (count=0, j=eptr[i]; j<eptr[i+1]; j++) {
node = eind[j];
for (k=nptr[node]; k<nptr[node+1]; k++) {
if ((kk=nind[k]) == firstelm+i)
continue;
m = htable[(kk&mask)];
if (m == -1) {
ind[count] = kk;
wgt[count] = 1;
htable[(kk&mask)] = count++;
}
else {
if (ind[m] == kk) {
wgt[m]++;
}
else {
for (jj=0; jj<count; jj++) {
if (ind[jj] == kk) {
wgt[jj]++;
break;
}
}
if (jj == count) {
ind[count] = kk;
wgt[count++] = 1;
}
}
}
}
}
for (j=0; j<count; j++) {
htable[(ind[j]&mask)] = -1;
if (wgt[j] >= *ncommonnodes) {
if (pass == 0)
myxadj[i]++;
else
myadjncy[myxadj[i]++] = ind[j];
}
}
}
if (pass == 0) {
MAKECSR(i, nelms, myxadj);
myadjncy = *adjncy = idxmalloc(myxadj[nelms], "adjncy");
}
else {
SHIFTCSR(i, nelms, myxadj);
}
}
/*****************************************/
/* correctly renumber the elements array */
/*****************************************/
for (i=0; i<eptr[nelms]; i++)
eind[i] = nmap[eind[i]] + gminnode;
if (*numflag == 1)
ChangeNumberingMesh2(elmdist, eptr, eind, myxadj, myadjncy, NULL, npes, mype, 0);
/* do not free nodelist, recvbuffer, rbuffer */
GKfree((void **)&scounts, (void **)&nodedist, (void **)&nmap, (void **)&sbuffer,
(void **)&htable, (void **)&nptr, (void **)&nind, (void **)&gnptr,
(void **)&gnind, (void **)&auxarray, LTERM);
FreeCtrl(&ctrl);
return;
}
/*************************************************************************
* This function is the entry point of the initial partitioning algorithm.
* This algorithm assembles the graph to all the processors and preceed
* serially.
**************************************************************************/
idx_t Mc_Diffusion(ctrl_t *ctrl, graph_t *graph, idx_t *vtxdist, idx_t *where,
idx_t *home, idx_t npasses)
{
idx_t h, i, j;
idx_t nvtxs, nedges, ncon, pass, iter, domain, processor;
idx_t nparts, mype, npes, nlinks, me, you, wsize;
idx_t nvisited, nswaps = -1, tnswaps, done, alldone = -1;
idx_t *rowptr, *colind, *diff_where, *sr_where, *ehome, *map, *rmap;
idx_t *pack, *unpack, *match, *proc2sub, *sub2proc;
idx_t *visited, *gvisited;
real_t *transfer, *npwgts, maxdiff, minflow, maxflow;
real_t lbavg, oldlbavg, ubavg, *lbvec;
real_t *diff_flows, *sr_flows;
real_t diff_lbavg, sr_lbavg, diff_cost, sr_cost;
idx_t *rbuffer, *sbuffer;
idx_t *rcount, *rdispl;
real_t *solution, *load, *workspace;
matrix_t matrix;
graph_t *egraph;
if (graph->ncon > 3)
return 0;
WCOREPUSH;
nvtxs = graph->nvtxs;
nedges = graph->nedges;
ncon = graph->ncon;
nparts = ctrl->nparts;
mype = ctrl->mype;
npes = ctrl->npes;
ubavg = ravg(ncon, ctrl->ubvec);
/* initialize variables and allocate memory */
lbvec = rwspacemalloc(ctrl, ncon);
diff_flows = rwspacemalloc(ctrl, ncon);
sr_flows = rwspacemalloc(ctrl, ncon);
load = rwspacemalloc(ctrl, nparts);
solution = rwspacemalloc(ctrl, nparts);
npwgts = graph->gnpwgts = rwspacemalloc(ctrl, ncon*nparts);
matrix.values = rwspacemalloc(ctrl, nedges);
transfer = matrix.transfer = rwspacemalloc(ctrl, ncon*nedges);
proc2sub = iwspacemalloc(ctrl, gk_max(nparts, npes*2));
sub2proc = iwspacemalloc(ctrl, nparts);
match = iwspacemalloc(ctrl, nparts);
rowptr = matrix.rowptr = iwspacemalloc(ctrl, nparts+1);
colind = matrix.colind = iwspacemalloc(ctrl, nedges);
rcount = iwspacemalloc(ctrl, npes);
rdispl = iwspacemalloc(ctrl, npes+1);
pack = iwspacemalloc(ctrl, nvtxs);
unpack = iwspacemalloc(ctrl, nvtxs);
rbuffer = iwspacemalloc(ctrl, nvtxs);
sbuffer = iwspacemalloc(ctrl, nvtxs);
map = iwspacemalloc(ctrl, nvtxs);
rmap = iwspacemalloc(ctrl, nvtxs);
diff_where = iwspacemalloc(ctrl, nvtxs);
ehome = iwspacemalloc(ctrl, nvtxs);
wsize = gk_max(sizeof(real_t)*nparts*6, sizeof(idx_t)*(nvtxs+nparts*2+1));
workspace = (real_t *)gk_malloc(wsize, "Mc_Diffusion: workspace");
graph->ckrinfo = (ckrinfo_t *)gk_malloc(nvtxs*sizeof(ckrinfo_t), "Mc_Diffusion: rinfo");
/* construct subdomain connectivity graph */
matrix.nrows = nparts;
SetUpConnectGraph(graph, &matrix, (idx_t *)workspace);
nlinks = (matrix.nnzs-nparts) / 2;
visited = iwspacemalloc(ctrl, matrix.nnzs);
gvisited = iwspacemalloc(ctrl, matrix.nnzs);
for (pass=0; pass<npasses; pass++) {
rset(matrix.nnzs*ncon, 0.0, transfer);
iset(matrix.nnzs, 0, gvisited);
iset(matrix.nnzs, 0, visited);
iter = nvisited = 0;
/* compute ncon flow solutions */
for (h=0; h<ncon; h++) {
rset(nparts, 0.0, solution);
ComputeLoad(graph, nparts, load, ctrl->tpwgts, h);
lbvec[h] = (rmax(nparts, load)+1.0/nparts) * (real_t)nparts;
ConjGrad2(&matrix, load, solution, 0.001, workspace);
ComputeTransferVector(ncon, &matrix, solution, transfer, h);
}
oldlbavg = ravg(ncon, lbvec);
tnswaps = 0;
maxdiff = 0.0;
for (i=0; i<nparts; i++) {
for (j=rowptr[i]; j<rowptr[i+1]; j++) {
maxflow = rmax(ncon, transfer+j*ncon);
minflow = rmin(ncon, transfer+j*ncon);
maxdiff = (maxflow - minflow > maxdiff) ? maxflow - minflow : maxdiff;
}
}
while (nvisited < nlinks) {
/* compute independent sets of subdomains */
iset(gk_max(nparts, npes*2), UNMATCHED, proc2sub);
CSR_Match_SHEM(&matrix, match, proc2sub, gvisited, ncon);
/* set up the packing arrays */
iset(nparts, UNMATCHED, sub2proc);
for (i=0; i<npes*2; i++) {
if (proc2sub[i] == UNMATCHED)
break;
sub2proc[proc2sub[i]] = i/2;
}
iset(npes, 0, rcount);
for (i=0; i<nvtxs; i++) {
domain = where[i];
processor = sub2proc[domain];
if (processor != UNMATCHED)
rcount[processor]++;
}
rdispl[0] = 0;
for (i=1; i<npes+1; i++)
rdispl[i] = rdispl[i-1] + rcount[i-1];
iset(nvtxs, UNMATCHED, unpack);
for (i=0; i<nvtxs; i++) {
domain = where[i];
processor = sub2proc[domain];
if (processor != UNMATCHED)
unpack[rdispl[processor]++] = i;
}
SHIFTCSR(i, npes, rdispl);
iset(nvtxs, UNMATCHED, pack);
for (i=0; i<rdispl[npes]; i++) {
ASSERT(unpack[i] != UNMATCHED);
domain = where[unpack[i]];
processor = sub2proc[domain];
if (processor != UNMATCHED)
pack[unpack[i]] = i;
}
/* Compute the flows */
if (proc2sub[mype*2] != UNMATCHED) {
me = proc2sub[2*mype];
you = proc2sub[2*mype+1];
ASSERT(me != you);
for (j=rowptr[me]; j<rowptr[me+1]; j++) {
if (colind[j] == you) {
visited[j] = 1;
rcopy(ncon, transfer+j*ncon, diff_flows);
break;
}
}
for (j=rowptr[you]; j<rowptr[you+1]; j++) {
if (colind[j] == me) {
visited[j] = 1;
for (h=0; h<ncon; h++) {
if (transfer[j*ncon+h] > 0.0)
diff_flows[h] = -1.0 * transfer[j*ncon+h];
}
break;
}
}
nswaps = 1;
rcopy(ncon, diff_flows, sr_flows);
iset(nvtxs, 0, sbuffer);
for (i=0; i<nvtxs; i++) {
if (where[i] == me || where[i] == you)
sbuffer[i] = 1;
}
egraph = ExtractGraph(ctrl, graph, sbuffer, map, rmap);
if (egraph != NULL) {
icopy(egraph->nvtxs, egraph->where, diff_where);
for (j=0; j<egraph->nvtxs; j++)
ehome[j] = home[map[j]];
RedoMyLink(ctrl, egraph, ehome, me, you, sr_flows, &sr_cost, &sr_lbavg);
if (ncon <= 4) {
sr_where = egraph->where;
egraph->where = diff_where;
nswaps = BalanceMyLink(ctrl, egraph, ehome, me, you, diff_flows, maxdiff,
&diff_cost, &diff_lbavg, 1.0/(real_t)nvtxs);
if ((sr_lbavg < diff_lbavg &&
(diff_lbavg >= ubavg-1.0 || sr_cost == diff_cost)) ||
(sr_lbavg < ubavg-1.0 && sr_cost < diff_cost)) {
for (i=0; i<egraph->nvtxs; i++)
where[map[i]] = sr_where[i];
}
else {
for (i=0; i<egraph->nvtxs; i++)
where[map[i]] = diff_where[i];
}
}
else {
for (i=0; i<egraph->nvtxs; i++)
where[map[i]] = egraph->where[i];
}
gk_free((void **)&egraph->xadj, &egraph->nvwgt, &egraph->adjncy, &egraph, LTERM);
}
/* Pack the flow data */
iset(nvtxs, UNMATCHED, sbuffer);
for (i=0; i<nvtxs; i++) {
domain = where[i];
if (domain == you || domain == me)
sbuffer[pack[i]] = where[i];
}
}
/* Broadcast the flow data */
gkMPI_Allgatherv((void *)&sbuffer[rdispl[mype]], rcount[mype], IDX_T,
(void *)rbuffer, rcount, rdispl, IDX_T, ctrl->comm);
/* Unpack the flow data */
for (i=0; i<rdispl[npes]; i++) {
if (rbuffer[i] != UNMATCHED)
where[unpack[i]] = rbuffer[i];
}
/* Do other stuff */
gkMPI_Allreduce((void *)visited, (void *)gvisited, matrix.nnzs,
IDX_T, MPI_MAX, ctrl->comm);
nvisited = isum(matrix.nnzs, gvisited, 1)/2;
tnswaps += GlobalSESum(ctrl, nswaps);
if (iter++ == NGD_PASSES)
break;
}
/* perform serial refinement */
Mc_ComputeSerialPartitionParams(ctrl, graph, nparts);
Mc_SerialKWayAdaptRefine(ctrl, graph, nparts, home, ctrl->ubvec, 10);
/* check for early breakout */
for (h=0; h<ncon; h++) {
lbvec[h] = (real_t)(nparts) *
npwgts[rargmax_strd(nparts,npwgts+h,ncon)*ncon+h];
}
lbavg = ravg(ncon, lbvec);
done = 0;
if (tnswaps == 0 || lbavg >= oldlbavg || lbavg <= ubavg + 0.035)
done = 1;
alldone = GlobalSEMax(ctrl, done);
if (alldone == 1)
break;
}
/* ensure that all subdomains have at least one vertex */
/*
iset(nparts, 0, match);
for (i=0; i<nvtxs; i++)
match[where[i]]++;
done = 0;
while (done == 0) {
done = 1;
me = iargmin(nparts, match);
if (match[me] == 0) {
if (ctrl->mype == PE) printf("WARNING: empty subdomain %"PRIDX" in Mc_Diffusion\n", me);
you = iargmax(nparts, match);
for (i=0; i<nvtxs; i++) {
if (where[i] == you) {
where[i] = me;
match[you]--;
match[me]++;
done = 0;
break;
}
}
}
}
*/
/* now free memory and return */
gk_free((void **)&workspace, (void **)&graph->ckrinfo, LTERM);
graph->gnpwgts = NULL;
graph->ckrinfo = NULL;
WCOREPOP;
return 0;
}
/******************************************************************************
* This function creates a coarse graph corresponding to the partitioning vector
*******************************************************************************/
GraphType *CreatePartitionGraphForContact(idxtype nvtxs, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, idxtype cnvtxs, idxtype *part)
{
idxtype i, ii, j, jj, k, cnedges;
idxtype *cxadj, *cadjncy, *cvwgt, *cadjwgt;
idxtype *ptr, *ind, *marker;
GraphType *cgraph;
ptr = idxsmalloc(cnvtxs+1, 0, "CreatePartitionGraph: ptr");
ind = idxmalloc(nvtxs, "CreatePartitionGraph: ind");
marker = idxsmalloc(cnvtxs, -1, "CreatePartitionGraph: marker");
cgraph = CreateGraph();
cgraph->ncon = 2;
cgraph->nvtxs = cnvtxs;
cxadj = cgraph->xadj = idxsmalloc(cnvtxs+1, 0, "CreatePartitionGraph: cxadj");
cadjncy = cgraph->adjncy = idxmalloc(xadj[nvtxs], "CreatePartitionGraph: cadjncy");
cvwgt = cgraph->vwgt = idxmalloc(2*cnvtxs, "CreatePartitionGraph: cvwgt");
cadjwgt = cgraph->adjwgt = idxmalloc(xadj[nvtxs], "CreatePartitionGraph: cadjwgt");
for (i=0; i<nvtxs; i++)
ptr[part[i]]++;
MAKECSR(i, cnvtxs, ptr);
for (i=0; i<nvtxs; i++)
ind[ptr[part[i]]++] = i;
SHIFTCSR(i, cnvtxs, ptr);
/* Create the adjacency list of the coarse/partition graph */
for (cxadj[0]=cnedges=0, i=0; i<cnvtxs; i++) {
cvwgt[2*i+0] = cvwgt[2*i+1] = 0;
for (j=ptr[i]; j<ptr[i+1]; j++) {
ii = ind[j];
cvwgt[2*i+0] += vwgt[2*ii+0];
cvwgt[2*i+1] += vwgt[2*ii+1];
for (jj=xadj[ii]; jj<xadj[ii+1]; jj++) {
if ((k = part[adjncy[jj]]) == i)
continue;
if (marker[k] == -1) {
cadjncy[cnedges] = k;
cadjwgt[cnedges] = adjwgt[jj];
marker[k] = cnedges++;
}
else
cadjwgt[marker[k]] += adjwgt[jj];
}
}
cxadj[i+1] = cnedges;
/* reset the marker array */
for (j=cxadj[i]; j<cnedges; j++)
marker[cadjncy[j]] = -1;
}
gk_free((void **)&ptr, &ind, &marker, LTERM);
return cgraph;
}
/*************************************************************************
* 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;
}
graph_t *FixGraph(graph_t *graph)
{
idx_t i, j, k, l, nvtxs, nedges;
idx_t *xadj, *adjncy, *adjwgt;
idx_t *nxadj, *nadjncy, *nadjwgt;
graph_t *ngraph;
uvw_t *edges;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
ASSERT(adjwgt != NULL);
ngraph = CreateGraph();
ngraph->nvtxs = nvtxs;
/* deal with vertex weights/sizes */
ngraph->ncon = graph->ncon;
ngraph->vwgt = icopy(nvtxs*graph->ncon, graph->vwgt,
imalloc(nvtxs*graph->ncon, "FixGraph: vwgt"));
ngraph->vsize = ismalloc(nvtxs, 1, "FixGraph: vsize");
if (graph->vsize)
icopy(nvtxs, graph->vsize, ngraph->vsize);
/* fix graph by sorting the "superset" of edges */
edges = (uvw_t *)gk_malloc(sizeof(uvw_t)*2*xadj[nvtxs], "FixGraph: edges");
for (nedges=0, i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
/* keep only the upper-trianglular part of the adjacency matrix */
if (i < adjncy[j]) {
edges[nedges].u = i;
edges[nedges].v = adjncy[j];
edges[nedges].w = adjwgt[j];
nedges++;
}
else if (i > adjncy[j]) {
edges[nedges].u = adjncy[j];
edges[nedges].v = i;
edges[nedges].w = adjwgt[j];
nedges++;
}
}
}
uvwsorti(nedges, edges);
/* keep the unique subset */
for (k=0, i=1; i<nedges; i++) {
if (edges[k].v != edges[i].v || edges[k].u != edges[i].u) {
edges[++k] = edges[i];
}
}
nedges = k+1;
/* allocate memory for the fixed graph */
nxadj = ngraph->xadj = ismalloc(nvtxs+1, 0, "FixGraph: nxadj");
nadjncy = ngraph->adjncy = imalloc(2*nedges, "FixGraph: nadjncy");
nadjwgt = ngraph->adjwgt = imalloc(2*nedges, "FixGraph: nadjwgt");
/* create the adjacency list of the fixed graph from the upper-triangular
part of the adjacency matrix */
for (k=0; k<nedges; k++) {
nxadj[edges[k].u]++;
nxadj[edges[k].v]++;
}
MAKECSR(i, nvtxs, nxadj);
for (k=0; k<nedges; k++) {
nadjncy[nxadj[edges[k].u]] = edges[k].v;
nadjncy[nxadj[edges[k].v]] = edges[k].u;
nadjwgt[nxadj[edges[k].u]] = edges[k].w;
nadjwgt[nxadj[edges[k].v]] = edges[k].w;
nxadj[edges[k].u]++;
nxadj[edges[k].v]++;
}
SHIFTCSR(i, nvtxs, nxadj);
gk_free((void **)&edges, LTERM);
return ngraph;
}
