/*************************************************************************
* This function initializes the data structures of the priority queue
**************************************************************************/
void PQueueInit(CtrlType *ctrl, PQueueType *queue, int maxnodes, int maxgain)
{
int i, j, ncore;
queue->nnodes = 0;
queue->maxnodes = maxnodes;
queue->buckets = NULL;
queue->nodes = NULL;
queue->heap = NULL;
queue->locator = NULL;
if (maxgain > PLUS_GAINSPAN || maxnodes < 500)
queue->type = 2;
else
queue->type = 1;
if (queue->type == 1) {
queue->pgainspan = amin(PLUS_GAINSPAN, maxgain);
queue->ngainspan = amin(NEG_GAINSPAN, maxgain);
j = queue->ngainspan+queue->pgainspan+1;
ncore = 2 + (sizeof(ListNodeType)/sizeof(idxtype))*maxnodes + (sizeof(ListNodeType *)/sizeof(idxtype))*j;
if (WspaceAvail(ctrl) > ncore) {
queue->nodes = (ListNodeType *)idxwspacemalloc(ctrl, (sizeof(ListNodeType)/sizeof(idxtype))*maxnodes);
queue->buckets = (ListNodeType **)idxwspacemalloc(ctrl, (sizeof(ListNodeType *)/sizeof(idxtype))*j);
queue->mustfree = 0;
}
else { /* Not enough memory in the wspace, allocate it */
queue->nodes = (ListNodeType *)idxmalloc((sizeof(ListNodeType)/sizeof(idxtype))*maxnodes, "PQueueInit: queue->nodes");
queue->buckets = (ListNodeType **)idxmalloc((sizeof(ListNodeType *)/sizeof(idxtype))*j, "PQueueInit: queue->buckets");
queue->mustfree = 1;
}
for (i=0; i<maxnodes; i++)
queue->nodes[i].id = i;
for (i=0; i<j; i++)
queue->buckets[i] = NULL;
queue->buckets += queue->ngainspan; /* Advance buckets by the ngainspan proper indexing */
queue->maxgain = -queue->ngainspan;
}
else {
queue->heap = (KeyValueType *)idxwspacemalloc(ctrl, (sizeof(KeyValueType)/sizeof(idxtype))*maxnodes);
queue->locator = idxwspacemalloc(ctrl, maxnodes);
idxset(maxnodes, -1, queue->locator);
}
}
/***********************************************************************************
* This function is the entry point of the parallel ordering algorithm.
* This function assumes that the graph is already nice partitioned among the
* processors and then proceeds to perform recursive bisection.
************************************************************************************/
void ParMETIS_V3_PartGeom(idxtype *vtxdist, int *ndims, float *xyz, idxtype *part, MPI_Comm *comm)
{
int i, npes, mype, nvtxs, firstvtx, dbglvl;
idxtype *xadj, *adjncy;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
int zeroflg = 0;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
if (npes == 1) {
idxset(vtxdist[mype+1]-vtxdist[mype], 0, part);
return;
}
/* Setup a fake graph to allow the rest of the code to work unchanged */
dbglvl = 0;
nvtxs = vtxdist[mype+1]-vtxdist[mype];
firstvtx = vtxdist[mype];
xadj = idxmalloc(nvtxs+1, "ParMETIS_PartGeom: xadj");
adjncy = idxmalloc(nvtxs, "ParMETIS_PartGeom: adjncy");
for (i=0; i<nvtxs; i++) {
xadj[i] = i;
adjncy[i] = firstvtx + (i+1)%nvtxs;
}
xadj[nvtxs] = nvtxs;
/* Proceed with the rest of the code */
SetUpCtrl(&ctrl, npes, dbglvl, *comm);
ctrl.seed = mype;
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*npes);
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, NULL, adjncy, NULL, &zeroflg);
PreAllocateMemory(&ctrl, graph, &wspace);
/*=======================================================
* Compute the initial geometric partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Coordinate_Partition(&ctrl, graph, *ndims, xyz, 0, &wspace);
idxcopy(graph->nvtxs, graph->where, part);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
FreeInitialGraphAndRemap(graph, 0);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
GKfree((void **)&xadj, (void **)&adjncy, LTERM);
}
// ****************************************************************************
// Method: avtAxisRestrictionToolInterface::ResetNumberOfAxes
//
// Purpose:
// Sets the number of available axes.
//
// Arguments:
// n the number of axes
//
// Programmer: Jeremy Meredith
// Creation: February 1, 2008
//
// Modifications:
// Jeremy Meredith, Fri Feb 15 13:21:20 EST 2008
// Added axis names to the axis restriction attributes.
//
// ****************************************************************************
void
avtAxisRestrictionToolInterface::ResetNumberOfAxes(int n)
{
AxisRestrictionAttributes *a = (AxisRestrictionAttributes *)atts;
stringVector aname(n, "");
doubleVector amin(n, -1e+37);
doubleVector amax(n, +1e+37);
a->SetNames(aname);
a->SetMinima(amin);
a->SetMaxima(amax);
}
/*************************************************************************
* This function is the entry point for ONWMETIS. It requires weights on the
* vertices. It is for the case that the matrix has been pre-compressed.
**************************************************************************/
void METIS_EdgeComputeSeparator(int *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *adjwgt, int *options, int *sepsize, idxtype *part)
{
int i, j, tvwgt, tpwgts[2];
GraphType graph;
CtrlType ctrl;
SetUpGraph(&graph, OP_ONMETIS, *nvtxs, 1, xadj, adjncy, vwgt, adjwgt, 3);
tvwgt = idxsum(*nvtxs, graph.vwgt);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = ONMETIS_CTYPE;
ctrl.IType = ONMETIS_ITYPE;
ctrl.RType = ONMETIS_RTYPE;
ctrl.dbglvl = ONMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.oflags = 0;
ctrl.pfactor = 0;
ctrl.nseps = 5;
ctrl.optype = OP_OEMETIS;
ctrl.CoarsenTo = amin(100, *nvtxs-1);
ctrl.maxvwgt = 1.5*tvwgt/ctrl.CoarsenTo;
InitRandom(options[7]);
AllocateWorkSpace(&ctrl, &graph, 2);
/*============================================================
* Perform the bisection
*============================================================*/
tpwgts[0] = tvwgt/2;
tpwgts[1] = tvwgt-tpwgts[0];
MlevelEdgeBisection(&ctrl, &graph, tpwgts, 1.05);
ConstructMinCoverSeparator(&ctrl, &graph, 1.05);
*sepsize = graph.pwgts[2];
idxcopy(*nvtxs, graph.where, part);
GKfree((void**)&graph.gdata, &graph.rdata, &graph.label, LTERM);
FreeWorkSpace(&ctrl, &graph);
}
/*************************************************************************
* This function is the entry point for ONWMETIS. It requires weights on the
* vertices. It is for the case that the matrix has been pre-compressed.
**************************************************************************/
void METIS_NodeComputeSeparator(idxtype *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *adjwgt, idxtype *options, idxtype *sepsize, idxtype *part)
{
idxtype i, j, tvwgt, tpwgts[2];
GraphType graph;
CtrlType ctrl;
SetUpGraph(&graph, OP_ONMETIS, *nvtxs, 1, xadj, adjncy, vwgt, adjwgt, 3);
tvwgt = idxsum(*nvtxs, graph.vwgt, 1);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = ONMETIS_CTYPE;
ctrl.IType = ONMETIS_ITYPE;
ctrl.RType = ONMETIS_RTYPE;
ctrl.dbglvl = ONMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.oflags = 0;
ctrl.pfactor = 0;
ctrl.nseps = 3;
ctrl.optype = OP_ONMETIS;
ctrl.CoarsenTo = amin(100, *nvtxs-1);
ctrl.maxvwgt = 1.5*tvwgt/ctrl.CoarsenTo;
InitRandom(options[7]);
AllocateWorkSpace(&ctrl, &graph, 2);
/*============================================================
* Perform the bisection
*============================================================*/
tpwgts[0] = tvwgt/2;
tpwgts[1] = tvwgt-tpwgts[0];
MlevelNodeBisectionMultiple(&ctrl, &graph, tpwgts, 1.02);
*sepsize = graph.pwgts[2];
idxcopy(*nvtxs, graph.where, part);
FreeGraph(&graph, 0);
FreeWorkSpace(&ctrl, &graph);
}
/***********************************************************************************
* This function is the entry point of the parallel multilevel local diffusion
* algorithm. It uses parallel undirected diffusion followed by adaptive k-way
* refinement. This function utilizes local coarsening.
************************************************************************************/
void ParMETIS_RepartLDiffusion(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, realtype *adjwgt, int *wgtflag, int *numflag, int *options,
int *edgecut, idxtype *part, MPI_Comm *comm)
{
int npes, mype;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
if (npes == 1) { /* Take care the npes = 1 case */
idxset(vtxdist[1], 0, part);
*edgecut = 0;
return;
}
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
SetUpCtrl(&ctrl, npes, options, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 70*npes);
graph = SetUpGraph(&ctrl, vtxdist, xadj, vwgt, adjncy, adjwgt, *wgtflag);
graph->vsize = idxsmalloc(graph->nvtxs, 1, "Par_KMetis: vsize");
PreAllocateMemory(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TRACK, printf("%d ParMETIS_RepartLDiffusion about to call AdaptiveUndirected_Partition\n",mype));
AdaptiveUndirected_Partition(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TRACK, printf("%d ParMETIS_RepartLDiffusion about to call ReMapGraph\n",mype));
ReMapGraph(&ctrl, graph, 0, &wspace);
idxcopy(graph->nvtxs, graph->where, part);
*edgecut = graph->mincut;
IMfree((void**)&graph->vsize, LTERM);
FreeInitialGraphAndRemap(graph, *wgtflag);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
}
/***********************************************************************************
* This function creates the fused-element-graph and returns the partition
************************************************************************************/
void ParMETIS_FusedElementGraph(idxtype *vtxdist, idxtype *xadj, realtype *vvol,
realtype *vsurf, idxtype *adjncy, idxtype *vwgt, realtype *adjwgt,
int *wgtflag, int *numflag, int *nparts, int *options,
idxtype *part, MPI_Comm *comm)
{
int npes, mype, nvtxs;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
nvtxs = vtxdist[mype+1]-vtxdist[mype];
/* IFSET(options[OPTION_DBGLVL], DBG_TRACK, printf("%d ParMETIS_FEG npes=%d\n",mype, npes)); */
SetUpCtrl(&ctrl, *nparts, options, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*amax(npes, *nparts));
graph = SetUpGraph(&ctrl, vtxdist, xadj, vwgt, adjncy, adjwgt, *wgtflag);
graph->where = part;
PreAllocateMemory(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
CreateFusedElementGraph(&ctrl, graph, &wspace, numflag);
idxcopy(nvtxs, graph->where, part);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
if (((*wgtflag)&2) == 0)
IMfree((void**)&graph->vwgt, LTERM);
IMfree((void**)&graph->lperm, &graph->peind, &graph->pexadj, &graph->peadjncy,
&graph->peadjloc, &graph->recvptr, &graph->recvind, &graph->sendptr,
&graph->imap, &graph->sendind, &graph, LTERM);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
}
double minAreaFreeRect(int**P, int Psz, int*P0sz){
int N = Psz;
bool find = false;
ll ans = 1LL<<60;
for(int a=0, a_len=(N); a<a_len; ++a) for(int b=0, b_len=(N); b<b_len; ++b) for(int c=0, c_len=(N); c<c_len; ++c) for(int d=0, d_len=(N); d<d_len; ++d){
if (a == b || a == c || a == d || b == c || b == d || c == d) continue;
if (P[a][0] - P[b][0] != P[d][0] - P[c][0]) continue;
if (P[a][1] - P[b][1] != P[d][1] - P[c][1]) continue;
ll x1 = P[a][0] - P[b][0];
ll x2 = P[c][0] - P[b][0];
ll y1 = P[a][1] - P[b][1];
ll y2 = P[c][1] - P[b][1];
if (x1 * x2 + y1 * y2 == 0) {
ll area = abs(x1 * y2 - x2 * y1);
if (area > 0) {
find = true;
amin(&ans, area);
}
}
}
if (!find)
return 0;
return ans;
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void FM_2WayEdgeRefine(CtrlType *ctrl, GraphType *graph, int *tpwgts, int npasses)
{
int i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, limit, tmp;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts;
idxtype *moved, *swaps, *perm;
PQueueType parts[2];
int higain, oldgain, mincut, mindiff, origdiff, initcut, newcut, mincutorder, avgvwgt;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
pwgts = graph->pwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
limit = (int) amin(amax(0.01*nvtxs, 15), 100);
avgvwgt = amin((pwgts[0]+pwgts[1])/20, 2*(pwgts[0]+pwgts[1])/nvtxs);
tmp = graph->adjwgtsum[idxamax(nvtxs, graph->adjwgtsum)];
PQueueInit(ctrl, &parts[0], nvtxs, tmp);
PQueueInit(ctrl, &parts[1], nvtxs, tmp);
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("Partitions: [%6d %6d] T[%6d %6d], Nv-Nb[%6d %6d]. ICut: %6d\n",
pwgts[0], pwgts[1], tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
origdiff = abs(tpwgts[0]-pwgts[0]);
idxset(nvtxs, -1, moved);
for (pass=0; pass<npasses; pass++) { /* Do a number of passes */
PQueueReset(&parts[0]);
PQueueReset(&parts[1]);
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
mindiff = abs(tpwgts[0]-pwgts[0]);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = perm[ii];
ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0);
ASSERT(bndptr[bndind[i]] != -1);
PQueueInsert(&parts[where[bndind[i]]], bndind[i], ed[bndind[i]]-id[bndind[i]]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
from = (tpwgts[0]-pwgts[0] < tpwgts[1]-pwgts[1] ? 0 : 1);
to = (from+1)%2;
if ((higain = PQueueGetMax(&parts[from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
newcut -= (ed[higain]-id[higain]);
INC_DEC(pwgts[to], pwgts[from], vwgt[higain]);
if ((newcut < mincut && abs(tpwgts[0]-pwgts[0]) <= origdiff+avgvwgt) ||
(newcut == mincut && abs(tpwgts[0]-pwgts[0]) < mindiff)) {
mincut = newcut;
mindiff = abs(tpwgts[0]-pwgts[0]);
mincutorder = nswaps;
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
INC_DEC(pwgts[from], pwgts[to], vwgt[higain]);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
printf("Moved %6d from %d. [%3d %3d] %5d [%4d %4d]\n", higain, from, ed[higain]-id[higain], vwgt[higain], newcut, pwgts[0], pwgts[1]));
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
PQueueDelete(&parts[where[k]], k, oldgain);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
PQueueUpdate(&parts[where[k]], k, oldgain, ed[k]-id[k]);
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
PQueueInsert(&parts[where[k]], k, ed[k]-id[k]);
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
INC_DEC(pwgts[to], pwgts[(to+1)%2], vwgt[higain]);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
IFSET(ctrl->dbglvl, DBG_REFINE,
printf("\tMinimum cut: %6d at %5d, PWGTS: [%6d %6d], NBND: %6d\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut == initcut)
break;
}
PQueueFree(ctrl, &parts[0]);
PQueueFree(ctrl, &parts[1]);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/***********************************************************************************
* This function is the entry point of the parallel ordering algorithm.
* This function assumes that the graph is already nice partitioned among the
* processors and then proceeds to perform recursive bisection.
************************************************************************************/
void ParMETIS_V3_NodeND(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy, int *numflag,
int *options, idxtype *order, idxtype *sizes, MPI_Comm *comm)
{
int i, j;
int ltvwgts[MAXNCON];
int nparts, npes, mype, wgtflag = 0, seed = GLOBAL_SEED;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph, *mgraph;
idxtype *morder;
int minnvtxs;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
nparts = npes;
if (!ispow2(npes)) {
if (mype == 0)
printf("Error: The number of processors must be a power of 2!\n");
return;
}
if (vtxdist[npes] < (int)((float)(npes*npes)*1.2)) {
if (mype == 0)
printf("Error: Too many processors for this many vertices.\n");
return;
}
minnvtxs = vtxdist[1]-vtxdist[0];
for (i=0; i<npes; i++)
minnvtxs = (minnvtxs < vtxdist[i+1]-vtxdist[i]) ? minnvtxs : vtxdist[i+1]-vtxdist[i];
if (minnvtxs < (int)((float)npes*1.1)) {
if (mype == 0)
printf("Error: vertices are not distributed equally.\n");
return;
}
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 1);
SetUpCtrl(&ctrl, nparts, options[PMV3_OPTION_DBGLVL], *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*npes);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*amax(npes, nparts));
ctrl.seed = mype;
ctrl.sync = seed;
ctrl.partType = STATIC_PARTITION;
ctrl.ps_relation = -1;
ctrl.tpwgts = fsmalloc(nparts, 1.0/(float)(nparts), "tpwgts");
ctrl.ubvec[0] = 1.03;
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, NULL, adjncy, NULL, &wgtflag);
PreAllocateMemory(&ctrl, graph, &wspace);
/*=======================================================
* Compute the initial k-way partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Moc_Global_Partition(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
/*=======================================================
* Move the graph according to the partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.MoveTmr));
MALLOC_CHECK(NULL);
graph->ncon = 1;
mgraph = Moc_MoveGraph(&ctrl, graph, &wspace);
MALLOC_CHECK(NULL);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.MoveTmr));
/*=======================================================
* Now compute an ordering of the moved graph
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
FreeWSpace(&wspace);
PreAllocateMemory(&ctrl, mgraph, &wspace);
ctrl.ipart = ISEP_NODE;
ctrl.CoarsenTo = amin(vtxdist[npes]+1, amax(20*npes, 1000));
/* compute tvwgts */
for (j=0; j<mgraph->ncon; j++)
ltvwgts[j] = 0;
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<mgraph->ncon; j++)
ltvwgts[j] += mgraph->vwgt[i*mgraph->ncon+j];
for (j=0; j<mgraph->ncon; j++)
ctrl.tvwgts[j] = GlobalSESum(&ctrl, ltvwgts[j]);
mgraph->nvwgt = fmalloc(mgraph->nvtxs*mgraph->ncon, "mgraph->nvwgt");
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<mgraph->ncon; j++)
mgraph->nvwgt[i*mgraph->ncon+j] = (float)(mgraph->vwgt[i*mgraph->ncon+j]) / (float)(ctrl.tvwgts[j]);
morder = idxmalloc(mgraph->nvtxs, "PAROMETIS: morder");
MultilevelOrder(&ctrl, mgraph, morder, sizes, &wspace);
MALLOC_CHECK(NULL);
/* Invert the ordering back to the original graph */
ProjectInfoBack(&ctrl, graph, order, morder, &wspace);
MALLOC_CHECK(NULL);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
free(ctrl.tpwgts);
free(morder);
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph, 0);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 0);
MALLOC_CHECK(NULL);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void MocGeneral2WayBalance(CtrlType *ctrl, GraphType *graph, float *tpwgts, float lbfactor)
{
int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, me, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *perm, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, newcut, mincutorder;
int qsizes[MAXNCON][2];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
npwgts = graph->npwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
qnum = idxwspacemalloc(ctrl, nvtxs);
limit = amin(amax(0.01*nvtxs, 15), 100);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1);
PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1);
qsizes[i][0] = qsizes[i][1] = 0;
}
for (i=0; i<nvtxs; i++) {
qnum[i] = samax(ncon, nvwgt+i*ncon);
qsizes[qnum[i]][where[i]]++;
}
/*
printf("Weight Distribution: \t");
for (i=0; i<ncon; i++)
printf(" [%d %d]", qsizes[i][0], qsizes[i][1]);
printf("\n");
*/
for (from=0; from<2; from++) {
for (j=0; j<ncon; j++) {
if (qsizes[j][from] == 0) {
for (i=0; i<nvtxs; i++) {
if (where[i] != from)
continue;
k = samax2(ncon, nvwgt+i*ncon);
if (k == j && qsizes[qnum[i]][from] > qsizes[j][from] && nvwgt[i*ncon+qnum[i]] < 1.3*nvwgt[i*ncon+j]) {
qsizes[qnum[i]][from]--;
qsizes[j][from]++;
qnum[i] = j;
}
}
}
}
}
/*
printf("Weight Distribution (after):\t ");
for (i=0; i<ncon; i++)
printf(" [%d %d]", qsizes[i][0], qsizes[i][1]);
printf("\n");
*/
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
minbal = origbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
newcut = mincut = graph->mincut;
mincutorder = -1;
if (ctrl->dbglvl&DBG_REFINE) {
printf("Parts: [");
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: %.3f [B]\n", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut, origbal);
}
idxset(nvtxs, -1, moved);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert all nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nvtxs, perm, 1);
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (minbal < lbfactor)
break;
SelectQueue(ncon, npwgts, tpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if (newbal < minbal || (newbal == minbal &&
(newcut < mincut || (newcut == mincut && BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
if (ctrl->dbglvl&DBG_MOVEINFO) {
printf("Moved %6d from %d(%d). Gain: %5d, Cut: %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf(", %.3f LB: %.3f\n", minbal, newbal);
}
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (moved[k] == -1)
PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]);
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
if (ctrl->dbglvl&DBG_REFINE) {
printf("\tMincut: %6d at %5d, NBND: %6d, NPwgts: [", mincut, mincutorder, nbnd);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("], LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts));
}
graph->mincut = mincut;
graph->nbnd = nbnd;
for (i=0; i<ncon; i++) {
PQueueFree(ctrl, &parts[i][0]);
PQueueFree(ctrl, &parts[i][1]);
}
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void Mc_Serial_Balance2Way(GraphType *graph, float *tpwgts, float lbfactor)
{
int i, ii, j, k, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
FPQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, newcut, mincutorder;
int qsizes[MAXNCON][2];
KeyValueType *cand;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->sendind;
ed = graph->recvind;
npwgts = graph->gnpwgts;
bndptr = graph->sendptr;
bndind = graph->recvptr;
moved = idxmalloc(nvtxs, "moved");
swaps = idxmalloc(nvtxs, "swaps");
qnum = idxmalloc(nvtxs, "qnum");
cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand");
limit = amin(amax(0.01*nvtxs, 15), 100);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
FPQueueInit(&parts[i][0], nvtxs);
FPQueueInit(&parts[i][1], nvtxs);
qsizes[i][0] = qsizes[i][1] = 0;
}
for (i=0; i<nvtxs; i++) {
qnum[i] = samax(ncon, nvwgt+i*ncon);
qsizes[qnum[i]][where[i]]++;
}
for (from=0; from<2; from++) {
for (j=0; j<ncon; j++) {
if (qsizes[j][from] == 0) {
for (i=0; i<nvtxs; i++) {
if (where[i] != from)
continue;
k = samax2(ncon, nvwgt+i*ncon);
if (k == j &&
qsizes[qnum[i]][from] > qsizes[j][from] &&
nvwgt[i*ncon+qnum[i]] < 1.3*nvwgt[i*ncon+j]) {
qsizes[qnum[i]][from]--;
qsizes[j][from]++;
qnum[i] = j;
}
}
}
}
}
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
minbal = origbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
newcut = mincut = graph->mincut;
mincutorder = -1;
idxset(nvtxs, -1, moved);
/* Insert all nodes in the priority queues */
nbnd = graph->gnvtxs;
for (i=0; i<nvtxs; i++) {
cand[i].key = id[i]-ed[i];
cand[i].val = i;
}
ikeysort(nvtxs, cand);
for (ii=0; ii<nvtxs; ii++) {
i = cand[ii].val;
FPQueueInsert(&parts[qnum[i]][where[i]], i, (float)(ed[i]-id[i]));
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (minbal < lbfactor)
break;
Serial_SelectQueue(ncon, npwgts, tpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = FPQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if (newbal < minbal || (newbal == minbal &&
(newcut < mincut || (newcut == mincut &&
Serial_BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (moved[k] == -1)
FPQueueUpdate(&parts[qnum[k]][where[k]], k, (float)(oldgain), (float)(ed[k]-id[k]));
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->gnvtxs = nbnd;
for (i=0; i<ncon; i++) {
FPQueueFree(&parts[i][0]);
FPQueueFree(&parts[i][1]);
}
GKfree((void **)&cand, (void **)&qnum, (void **)&moved, (void **)&swaps, LTERM);
return;
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void Mc_Serial_FM_2WayRefine(GraphType *graph, float *tpwgts, int npasses)
{
int i, ii, j, k;
int kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
FPQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, initcut, newcut, mincutorder;
float rtpwgts[MAXNCON*2];
KeyValueType *cand;
int mype;
MPI_Comm_rank(MPI_COMM_WORLD, &mype);
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->sendind;
ed = graph->recvind;
npwgts = graph->gnpwgts;
bndptr = graph->sendptr;
bndind = graph->recvptr;
moved = idxmalloc(nvtxs, "moved");
swaps = idxmalloc(nvtxs, "swaps");
qnum = idxmalloc(nvtxs, "qnum");
cand = (KeyValueType *)GKmalloc(nvtxs*sizeof(KeyValueType), "cand");
limit = amin(amax(0.01*nvtxs, 25), 150);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
FPQueueInit(&parts[i][0], nvtxs);
FPQueueInit(&parts[i][1], nvtxs);
}
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
origbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
for (i=0; i<ncon; i++) {
rtpwgts[i] = origbal*tpwgts[i];
rtpwgts[ncon+i] = origbal*tpwgts[ncon+i];
}
idxset(nvtxs, -1, moved);
for (pass=0; pass<npasses; pass++) { /* Do a number of passes */
for (i=0; i<ncon; i++) {
FPQueueReset(&parts[i][0]);
FPQueueReset(&parts[i][1]);
}
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
minbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
/* Insert boundary nodes in the priority queues */
nbnd = graph->gnvtxs;
for (i=0; i<nbnd; i++) {
cand[i].key = id[i]-ed[i];
cand[i].val = i;
}
ikeysort(nbnd, cand);
for (ii=0; ii<nbnd; ii++) {
i = bndind[cand[ii].val];
FPQueueInsert(&parts[qnum[i]][where[i]], i, (float)(ed[i]-id[i]));
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
Serial_SelectQueue(ncon, npwgts, rtpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = FPQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Serial_Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if ((newcut < mincut && newbal-origbal <= .00001) ||
(newcut == mincut && (newbal < minbal ||
(newbal == minbal && Serial_BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[i]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
FPQueueDelete(&parts[qnum[k]][where[k]], k);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
FPQueueUpdate(&parts[qnum[k]][where[k]], k, (float)oldgain, (float)(ed[k]-id[k]));
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
FPQueueInsert(&parts[qnum[k]][where[k]], k, (float)(ed[k]-id[k]));
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy2(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy2(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->gnvtxs = nbnd;
if (mincutorder == -1 || mincut == initcut)
break;
}
for (i=0; i<ncon; i++) {
FPQueueFree(&parts[i][0]);
FPQueueFree(&parts[i][1]);
}
GKfree((void **)&cand, (void **)&qnum, (void **)&moved, (void **)&swaps, LTERM);
return;
}
/**************************************************************
* This subroutine remaps a partitioning on a single processor
**************************************************************/
void SerialRemap(GraphType *graph, int nparts, idxtype *base, idxtype *scratch,
idxtype *remap, float *tpwgts)
{
int i, ii, j, k;
int nvtxs, nmapped, max_mult;
int from, to, current_from, smallcount, bigcount;
KeyValueType *flowto, *bestflow;
KeyKeyValueType *sortvtx;
idxtype *vsize, *htable, *map, *rowmap;
nvtxs = graph->nvtxs;
vsize = graph->vsize;
max_mult = amin(MAX_NPARTS_MULTIPLIER, nparts);
sortvtx = (KeyKeyValueType *)GKmalloc(nvtxs*sizeof(KeyKeyValueType), "sortvtx");
flowto = (KeyValueType *)GKmalloc((nparts*max_mult+nparts)*sizeof(KeyValueType), "flowto");
bestflow = flowto+nparts;
map = htable = idxsmalloc(nparts*2, -1, "htable");
rowmap = map+nparts;
for (i=0; i<nvtxs; i++) {
sortvtx[i].key1 = base[i];
sortvtx[i].key2 = vsize[i];
sortvtx[i].val = i;
}
qsort((void *)sortvtx, (size_t)nvtxs, (size_t)sizeof(KeyKeyValueType), SSMIncKeyCmp);
for (j=0; j<nparts; j++) {
flowto[j].key = 0;
flowto[j].val = j;
}
/* this step has nparts*nparts*log(nparts) computational complexity */
bigcount = smallcount = current_from = 0;
for (ii=0; ii<nvtxs; ii++) {
i = sortvtx[ii].val;
from = base[i];
to = scratch[i];
if (from > current_from) {
/* reset the hash table */
for (j=0; j<smallcount; j++)
htable[flowto[j].val] = -1;
ASSERTS(idxsum(nparts, htable) == -nparts);
ikeysort(smallcount, flowto);
for (j=0; j<amin(smallcount, max_mult); j++, bigcount++) {
bestflow[bigcount].key = flowto[j].key;
bestflow[bigcount].val = current_from*nparts+flowto[j].val;
}
smallcount = 0;
current_from = from;
}
if (htable[to] == -1) {
htable[to] = smallcount;
flowto[smallcount].key = -vsize[i];
flowto[smallcount].val = to;
smallcount++;
}
else {
flowto[htable[to]].key += -vsize[i];
}
}
/* reset the hash table */
for (j=0; j<smallcount; j++)
htable[flowto[j].val] = -1;
ASSERTS(idxsum(nparts, htable) == -nparts);
ikeysort(smallcount, flowto);
for (j=0; j<amin(smallcount, max_mult); j++, bigcount++) {
bestflow[bigcount].key = flowto[j].key;
bestflow[bigcount].val = current_from*nparts+flowto[j].val;
}
ikeysort(bigcount, bestflow);
ASSERTS(idxsum(nparts, map) == -nparts);
ASSERTS(idxsum(nparts, rowmap) == -nparts);
nmapped = 0;
/* now make as many assignments as possible */
for (ii=0; ii<bigcount; ii++) {
i = bestflow[ii].val;
j = i % nparts; /* to */
k = i / nparts; /* from */
if (map[j] == -1 && rowmap[k] == -1 && SimilarTpwgts(tpwgts, graph->ncon, j, k)) {
map[j] = k;
rowmap[k] = j;
nmapped++;
}
if (nmapped == nparts)
break;
}
/* remap the rest */
/* it may help try remapping to the same label first */
if (nmapped < nparts) {
for (j=0; j<nparts && nmapped<nparts; j++) {
if (map[j] == -1) {
for (ii=0; ii<nparts; ii++) {
i = (j+ii) % nparts;
if (rowmap[i] == -1 && SimilarTpwgts(tpwgts, graph->ncon, i, j)) {
map[j] = i;
rowmap[i] = j;
nmapped++;
break;
}
}
}
}
}
/* check to see if remapping fails (due to dis-similar tpwgts) */
/* if remapping fails, revert to original mapping */
if (nmapped < nparts)
for (i=0; i<nparts; i++)
map[i] = i;
for (i=0; i<nvtxs; i++)
remap[i] = map[remap[i]];
GKfree((void **)&sortvtx, (void **)&flowto, (void **)&htable, LTERM);
}
/*************************************************************************
* This is the top level ordering routine
**************************************************************************/
void MultilevelOrder(CtrlType *ctrl, GraphType *graph, idxtype *order, idxtype *sizes, WorkSpaceType *wspace)
{
int i, nparts, nvtxs, npes;
idxtype *perm, *lastnode, *morder, *porder;
GraphType *mgraph;
npes = ctrl->npes;
nvtxs = graph->nvtxs;
perm = idxmalloc(nvtxs, "MultilevelOrder: perm");
lastnode = idxsmalloc(4*npes, -1, "MultilevelOrder: lastnode");
for (i=0; i<nvtxs; i++)
perm[i] = i;
lastnode[2] = graph->gnvtxs;
idxset(nvtxs, -1, order);
sizes[0] = 2*npes-1;
graph->where = idxsmalloc(nvtxs, 0, "MultilevelOrder: graph->where");
for (nparts=2; nparts<=ctrl->npes; nparts*=2) {
ctrl->nparts = nparts;
Order_Partition(ctrl, graph, wspace);
LabelSeparators(ctrl, graph, lastnode, perm, order, sizes, wspace);
CompactGraph(ctrl, graph, perm, wspace);
if (ctrl->CoarsenTo < 100*nparts) {
ctrl->CoarsenTo = 1.5*ctrl->CoarsenTo;
}
ctrl->CoarsenTo = amin(ctrl->CoarsenTo, graph->gnvtxs-1);
}
/*-----------------------------------------------------------------
/ Move the graph so that each processor gets its partition
-----------------------------------------------------------------*/
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MoveTmr));
SetUp(ctrl, graph, wspace);
graph->ncon = 1; /*needed for Moc_MoveGraph */
mgraph = Moc_MoveGraph(ctrl, graph, wspace);
/* Fill in the sizes[] array for the local part. Just the vtxdist of the mgraph */
for (i=0; i<npes; i++)
sizes[i] = mgraph->vtxdist[i+1]-mgraph->vtxdist[i];
porder = idxmalloc(graph->nvtxs, "MultilevelOrder: porder");
morder = idxmalloc(mgraph->nvtxs, "MultilevelOrder: morder");
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MoveTmr));
/* Find the local ordering */
LocalNDOrder(ctrl, mgraph, morder, lastnode[2*(ctrl->npes+ctrl->mype)]-mgraph->nvtxs, wspace);
/* Project the ordering back to the before-move graph */
ProjectInfoBack(ctrl, graph, porder, morder, wspace);
/* Copy the ordering from porder to order using perm */
for (i=0; i<graph->nvtxs; i++) {
ASSERT(ctrl, order[perm[i]] == -1);
order[perm[i]] = porder[i];
}
FreeGraph(mgraph);
GKfree((void **)&perm, (void **)&lastnode, (void **)&porder, (void **)&morder, LTERM);
/* PrintVector(ctrl, 2*npes-1, 0, sizes, "SIZES"); */
}
/*************************************************************************
* This function performs a node-based FM refinement
**************************************************************************/
void FM_2WayNodeRefineEqWgt(CtrlType *ctrl, GraphType *graph, idxtype npasses)
{
idxtype i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind;
idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idxtype *mptr, *mind, *moved, *swaps, *perm;
PQueueType parts[2];
NRInfoType *rinfo;
idxtype higain, oldgain, mincut, initcut, mincutorder;
idxtype pass, to, other, limit;
idxtype mindiff, newdiff;
idxtype u[2], g[2];
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
i = ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt);
PQueueInit(ctrl, &parts[0], nvtxs, i);
PQueueInit(ctrl, &parts[1], nvtxs, i);
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
mptr = idxwspacemalloc(ctrl, nvtxs+1);
mind = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("Partitions: [%6D %6D] Nv-Nb[%6D %6D]. ISep: %6D\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
for (pass=0; pass<npasses; pass++) {
idxset(nvtxs, -1, moved);
PQueueReset(&parts[0]);
PQueueReset(&parts[1]);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(where[i] == 2);
PQueueInsert(&parts[0], i, vwgt[i]-rinfo[i].edegrees[1]);
PQueueInsert(&parts[1], i, vwgt[i]-rinfo[i].edegrees[0]);
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = (ctrl->oflags&OFLAG_COMPRESS ? amin(5*nbnd, 400) : amin(2*nbnd, 300));
/******************************************************
* Get into the FM loop
*******************************************************/
mptr[0] = nmind = 0;
mindiff = idxtype_abs(pwgts[0]-pwgts[1]);
to = (pwgts[0] < pwgts[1] ? 0 : 1);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
to = (pwgts[0] < pwgts[1] ? 0 : 1);
if (pwgts[0] == pwgts[1]) {
u[0] = PQueueSeeMax(&parts[0]);
u[1] = PQueueSeeMax(&parts[1]);
if (u[0] != -1 && u[1] != -1) {
g[0] = vwgt[u[0]]-rinfo[u[0]].edegrees[1];
g[1] = vwgt[u[1]]-rinfo[u[1]].edegrees[0];
to = (g[0] > g[1] ? 0 : (g[0] < g[1] ? 1 : pass%2));
}
}
other = (to+1)%2;
if ((higain = PQueueGetMax(&parts[to])) == -1)
break;
if (moved[higain] == -1) /* Delete if it was in the separator originally */
PQueueDelete(&parts[other], higain, vwgt[higain]-rinfo[higain].edegrees[to]);
ASSERT(bndptr[higain] != -1);
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = idxtype_abs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > limit) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
oldgain = vwgt[k]-rinfo[k].edegrees[to];
rinfo[k].edegrees[to] += vwgt[higain];
if (moved[k] == -1 || moved[k] == -(2+other))
PQueueUpdate(&parts[other], k, oldgain, oldgain-vwgt[higain]);
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
oldgain = vwgt[kk]-rinfo[kk].edegrees[other];
rinfo[kk].edegrees[other] -= vwgt[k];
if (moved[kk] == -1 || moved[kk] == -(2+to))
PQueueUpdate(&parts[to], kk, oldgain, oldgain+vwgt[k]);
}
}
/* Insert the new vertex into the priority queue. Only one side! */
if (moved[k] == -1) {
PQueueInsert(&parts[to], k, vwgt[k]-edegrees[other]);
moved[k] = -(2+to);
}
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
mprintf("Moved %6D to %3D, Gain: %5D [%5D] [%4D %4D] \t[%5D %5D %5D]\n", higain, to, g[to], g[other], vwgt[u[to]], vwgt[u[other]], pwgts[0], pwgts[1], pwgts[2]));
}
/****************************************************************
* Roll back computation
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
to = where[higain];
other = (to+1)%2;
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("\tMinimum sep: %6D at %5D, PWGTS: [%6D %6D], NBND: %6D\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut >= initcut)
break;
}
PQueueFree(ctrl, &parts[0]);
PQueueFree(ctrl, &parts[1]);
idxwspacefree(ctrl, nvtxs+1);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void MocFM_2WayEdgeRefine(CtrlType *ctrl, GraphType *graph, float *tpwgts, int npasses)
{
int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass, me, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *perm, *qnum;
float *nvwgt, *npwgts, mindiff[MAXNCON], origbal, minbal, newbal;
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, initcut, newcut, mincutorder;
float rtpwgts[2];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
npwgts = graph->npwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
qnum = idxwspacemalloc(ctrl, nvtxs);
limit = amin(amax(0.01*nvtxs, 25), 150);
/* Initialize the queues */
for (i=0; i<ncon; i++) {
PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1);
PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1);
}
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
origbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
rtpwgts[0] = origbal*tpwgts[0];
rtpwgts[1] = origbal*tpwgts[1];
/*
if (ctrl->dbglvl&DBG_REFINE) {
printf("Parts: [");
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: %.3f\n", tpwgts[0], tpwgts[1], graph->nvtxs, graph->nbnd, graph->mincut, origbal);
}
*/
idxset(nvtxs, -1, moved);
for (pass=0; pass<npasses; pass++) { /* Do a number of passes */
for (i=0; i<ncon; i++) {
PQueueReset(&parts[i][0]);
PQueueReset(&parts[i][1]);
}
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
minbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(ed[i] > 0 || id[i] == 0);
ASSERT(bndptr[i] != -1);
PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
SelectQueue(ncon, npwgts, rtpwgts, &from, &cnum, parts);
to = (from+1)%2;
if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
newbal = Compute2WayHLoadImbalance(ncon, npwgts, tpwgts);
if ((newcut < mincut && newbal-origbal <= .00001) ||
(newcut == mincut && (newbal < minbal ||
(newbal == minbal && BetterBalance(ncon, npwgts, tpwgts, mindiff))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
for (i=0; i<ncon; i++)
mindiff[i] = fabs(tpwgts[0]-npwgts[i]);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/*
if (ctrl->dbglvl&DBG_MOVEINFO) {
printf("Moved %6d from %d(%d). Gain: %5d, Cut: %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf(", %.3f LB: %.3f\n", minbal, newbal);
}
*/
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
PQueueDelete(&parts[qnum[k]][where[k]], k, oldgain);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]);
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
PQueueInsert(&parts[qnum[k]][where[k]], k, ed[k]-id[k]);
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/*
if (ctrl->dbglvl&DBG_REFINE) {
printf("\tMincut: %6d at %5d, NBND: %6d, NPwgts: [", mincut, mincutorder, nbnd);
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("], LB: %.3f\n", Compute2WayHLoadImbalance(ncon, npwgts, tpwgts));
}
*/
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut == initcut)
break;
}
for (i=0; i<ncon; i++) {
PQueueFree(ctrl, &parts[i][0]);
PQueueFree(ctrl, &parts[i][1]);
}
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function is the entry point of the initial balancing algorithm.
* This algorithm assembles the graph to all the processors and preceeds
* with the balancing step.
**************************************************************************/
void Balance_Partition(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int i, j, mype, npes, nvtxs, nedges, ncon;
idxtype *vtxdist, *xadj, *adjncy, *adjwgt, *vwgt, *vsize;
idxtype *part, *lwhere, *home;
GraphType *agraph, cgraph;
CtrlType myctrl;
int lnparts, fpart, fpe, lnpes, ngroups, srnpes, srmype;
int twoparts=2, numflag = 0, wgtflag = 3, moptions[10], edgecut, max_cut;
int sr_pe, gd_pe, sr, gd, who_wins, *rcounts, *rdispls;
float my_cut, my_totalv, my_cost = -1.0, my_balance = -1.0, wsum;
float rating, max_rating, your_cost = -1.0, your_balance = -1.0;
float lbvec[MAXNCON], lbsum, min_lbsum, *mytpwgts, mytpwgts2[2], buffer[2];
MPI_Status status;
MPI_Comm ipcomm, srcomm;
struct {
float cost;
int rank;
} lpecost, gpecost;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
vtxdist = graph->vtxdist;
agraph = Mc_AssembleAdaptiveGraph(ctrl, graph, wspace);
nvtxs = cgraph.nvtxs = agraph->nvtxs;
nedges = cgraph.nedges = agraph->nedges;
ncon = cgraph.ncon = agraph->ncon;
xadj = cgraph.xadj = idxmalloc(nvtxs*(5+ncon)+1+nedges*2, "U_IP: xadj");
vwgt = cgraph.vwgt = xadj + nvtxs+1;
vsize = cgraph.vsize = xadj + nvtxs*(1+ncon)+1;
cgraph.where = agraph->where = part = xadj + nvtxs*(2+ncon)+1;
lwhere = xadj + nvtxs*(3+ncon)+1;
home = xadj + nvtxs*(4+ncon)+1;
adjncy = cgraph.adjncy = xadj + nvtxs*(5+ncon)+1;
adjwgt = cgraph.adjwgt = xadj + nvtxs*(5+ncon)+1 + nedges;
/* ADD: this assumes that tpwgts for all constraints is the same */
/* ADD: this is necessary because serial metis does not support the general case */
mytpwgts = fsmalloc(ctrl->nparts, 0.0, "mytpwgts");
for (i=0; i<ctrl->nparts; i++)
for (j=0; j<ncon; j++)
mytpwgts[i] += ctrl->tpwgts[i*ncon+j];
for (i=0; i<ctrl->nparts; i++)
mytpwgts[i] /= (float)ncon;
idxcopy(nvtxs+1, agraph->xadj, xadj);
idxcopy(nvtxs*ncon, agraph->vwgt, vwgt);
idxcopy(nvtxs, agraph->vsize, vsize);
idxcopy(nedges, agraph->adjncy, adjncy);
idxcopy(nedges, agraph->adjwgt, adjwgt);
/****************************************/
/****************************************/
if (ctrl->ps_relation == DISCOUPLED) {
rcounts = imalloc(ctrl->npes, "rcounts");
rdispls = imalloc(ctrl->npes+1, "rdispls");
for (i=0; i<ctrl->npes; i++) {
rdispls[i] = rcounts[i] = vtxdist[i+1]-vtxdist[i];
}
MAKECSR(i, ctrl->npes, rdispls);
MPI_Allgatherv((void *)graph->home, graph->nvtxs, IDX_DATATYPE,
(void *)part, rcounts, rdispls, IDX_DATATYPE, ctrl->comm);
for (i=0; i<agraph->nvtxs; i++)
home[i] = part[i];
GKfree((void **)&rcounts, (void **)&rdispls, LTERM);
}
else {
for (i=0; i<ctrl->npes; i++)
for (j=vtxdist[i]; j<vtxdist[i+1]; j++)
part[j] = home[j] = i;
}
/* Ensure that the initial partitioning is legal */
for (i=0; i<agraph->nvtxs; i++) {
if (part[i] >= ctrl->nparts)
part[i] = home[i] = part[i] % ctrl->nparts;
if (part[i] < 0)
part[i] = home[i] = (-1*part[i]) % ctrl->nparts;
}
/****************************************/
/****************************************/
IFSET(ctrl->dbglvl, DBG_REFINEINFO, Mc_ComputeSerialBalance(ctrl, agraph, agraph->where, lbvec));
IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "input cut: %d, balance: ", ComputeSerialEdgeCut(agraph)));
for (i=0; i<agraph->ncon; i++)
IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "%.3f ", lbvec[i]));
IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "\n"));
/****************************************/
/* Split the processors into two groups */
/****************************************/
sr = (ctrl->mype % 2 == 0) ? 1 : 0;
gd = (ctrl->mype % 2 == 1) ? 1 : 0;
if (graph->ncon > MAX_NCON_FOR_DIFFUSION || ctrl->npes == 1) {
sr = 1;
gd = 0;
}
sr_pe = 0;
gd_pe = 1;
MPI_Comm_split(ctrl->gcomm, sr, 0, &ipcomm);
MPI_Comm_rank(ipcomm, &mype);
MPI_Comm_size(ipcomm, &npes);
myctrl.dbglvl = 0;
myctrl.mype = mype;
myctrl.npes = npes;
myctrl.comm = ipcomm;
myctrl.sync = ctrl->sync;
myctrl.seed = ctrl->seed;
myctrl.nparts = ctrl->nparts;
myctrl.ipc_factor = ctrl->ipc_factor;
myctrl.redist_factor = ctrl->redist_base;
myctrl.partType = ADAPTIVE_PARTITION;
myctrl.ps_relation = DISCOUPLED;
myctrl.tpwgts = ctrl->tpwgts;
icopy(ncon, ctrl->tvwgts, myctrl.tvwgts);
icopy(ncon, ctrl->ubvec, myctrl.ubvec);
if (sr == 1) {
/*******************************************/
/* Half of the processors do scratch-remap */
/*******************************************/
ngroups = amax(amin(RIP_SPLIT_FACTOR, npes), 1);
MPI_Comm_split(ipcomm, mype % ngroups, 0, &srcomm);
MPI_Comm_rank(srcomm, &srmype);
MPI_Comm_size(srcomm, &srnpes);
moptions[0] = 0;
moptions[7] = ctrl->sync + (mype % ngroups) + 1;
idxset(nvtxs, 0, lwhere);
lnparts = ctrl->nparts;
fpart = fpe = 0;
lnpes = srnpes;
while (lnpes > 1 && lnparts > 1) {
ASSERT(ctrl, agraph->nvtxs > 1);
/* Determine the weights of the partitions */
mytpwgts2[0] = ssum(lnparts/2, mytpwgts+fpart);
mytpwgts2[1] = 1.0-mytpwgts2[0];
if (agraph->ncon == 1) {
METIS_WPartGraphKway2(&agraph->nvtxs, agraph->xadj, agraph->adjncy, agraph->vwgt,
agraph->adjwgt, &wgtflag, &numflag, &twoparts, mytpwgts2, moptions, &edgecut,
part);
}
else {
METIS_mCPartGraphRecursive2(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag, &twoparts, mytpwgts2,
moptions, &edgecut, part);
}
wsum = ssum(lnparts/2, mytpwgts+fpart);
sscale(lnparts/2, 1.0/wsum, mytpwgts+fpart);
sscale(lnparts-lnparts/2, 1.0/(1.0-wsum), mytpwgts+fpart+lnparts/2);
/* I'm picking the left branch */
if (srmype < fpe+lnpes/2) {
Mc_KeepPart(agraph, wspace, part, 0);
lnpes = lnpes/2;
lnparts = lnparts/2;
}
else {
Mc_KeepPart(agraph, wspace, part, 1);
fpart = fpart + lnparts/2;
fpe = fpe + lnpes/2;
lnpes = lnpes - lnpes/2;
lnparts = lnparts - lnparts/2;
}
}
/* In case srnpes is greater than or equal to nparts */
if (lnparts == 1) {
/* Only the first process will assign labels (for the reduction to work) */
if (srmype == fpe) {
for (i=0; i<agraph->nvtxs; i++)
lwhere[agraph->label[i]] = fpart;
}
}
/* In case srnpes is smaller than nparts */
else {
if (ncon == 1)
METIS_WPartGraphKway2(&agraph->nvtxs, agraph->xadj, agraph->adjncy, agraph->vwgt,
agraph->adjwgt, &wgtflag, &numflag, &lnparts, mytpwgts+fpart, moptions,
&edgecut, part);
else
METIS_mCPartGraphRecursive2(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag, &lnparts, mytpwgts+fpart,
moptions, &edgecut, part);
for (i=0; i<agraph->nvtxs; i++)
lwhere[agraph->label[i]] = fpart + part[i];
}
MPI_Allreduce((void *)lwhere, (void *)part, nvtxs, IDX_DATATYPE, MPI_SUM, srcomm);
edgecut = ComputeSerialEdgeCut(&cgraph);
Mc_ComputeSerialBalance(ctrl, &cgraph, part, lbvec);
lbsum = ssum(ncon, lbvec);
MPI_Allreduce((void *)&edgecut, (void *)&max_cut, 1, MPI_INT, MPI_MAX, ipcomm);
MPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, MPI_FLOAT, MPI_MIN, ipcomm);
lpecost.rank = ctrl->mype;
lpecost.cost = lbsum;
if (min_lbsum < UNBALANCE_FRACTION * (float)(ncon)) {
if (lbsum < UNBALANCE_FRACTION * (float)(ncon))
lpecost.cost = (float)edgecut;
else
lpecost.cost = (float)max_cut + lbsum;
}
MPI_Allreduce((void *)&lpecost, (void *)&gpecost, 1, MPI_FLOAT_INT, MPI_MINLOC, ipcomm);
if (ctrl->mype == gpecost.rank && ctrl->mype != sr_pe) {
MPI_Send((void *)part, nvtxs, IDX_DATATYPE, sr_pe, 1, ctrl->comm);
}
if (ctrl->mype != gpecost.rank && ctrl->mype == sr_pe) {
MPI_Recv((void *)part, nvtxs, IDX_DATATYPE, gpecost.rank, 1, ctrl->comm, &status);
}
if (ctrl->mype == sr_pe) {
idxcopy(nvtxs, part, lwhere);
SerialRemap(&cgraph, ctrl->nparts, home, lwhere, part, ctrl->tpwgts);
}
MPI_Comm_free(&srcomm);
}
/**************************************/
/* The other half do global diffusion */
/**************************************/
else {
/******************************************************************/
/* The next stmt is required to balance out the sr MPI_Comm_split */
/******************************************************************/
MPI_Comm_split(ipcomm, MPI_UNDEFINED, 0, &srcomm);
if (ncon == 1) {
rating = WavefrontDiffusion(&myctrl, agraph, home);
Mc_ComputeSerialBalance(ctrl, &cgraph, part, lbvec);
lbsum = ssum(ncon, lbvec);
/* Determine which PE computed the best partitioning */
MPI_Allreduce((void *)&rating, (void *)&max_rating, 1, MPI_FLOAT, MPI_MAX, ipcomm);
MPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, MPI_FLOAT, MPI_MIN, ipcomm);
lpecost.rank = ctrl->mype;
lpecost.cost = lbsum;
if (min_lbsum < UNBALANCE_FRACTION * (float)(ncon)) {
if (lbsum < UNBALANCE_FRACTION * (float)(ncon))
lpecost.cost = rating;
else
lpecost.cost = max_rating + lbsum;
}
MPI_Allreduce((void *)&lpecost, (void *)&gpecost, 1, MPI_FLOAT_INT, MPI_MINLOC, ipcomm);
/* Now send this to the coordinating processor */
if (ctrl->mype == gpecost.rank && ctrl->mype != gd_pe)
MPI_Send((void *)part, nvtxs, IDX_DATATYPE, gd_pe, 1, ctrl->comm);
if (ctrl->mype != gpecost.rank && ctrl->mype == gd_pe)
MPI_Recv((void *)part, nvtxs, IDX_DATATYPE, gpecost.rank, 1, ctrl->comm, &status);
if (ctrl->mype == gd_pe) {
idxcopy(nvtxs, part, lwhere);
SerialRemap(&cgraph, ctrl->nparts, home, lwhere, part, ctrl->tpwgts);
}
}
else {
Mc_Diffusion(&myctrl, agraph, graph->vtxdist, agraph->where, home, wspace, N_MOC_GD_PASSES);
}
}
if (graph->ncon <= MAX_NCON_FOR_DIFFUSION) {
if (ctrl->mype == sr_pe || ctrl->mype == gd_pe) {
/********************************************************************/
/* The coordinators from each group decide on the best partitioning */
/********************************************************************/
my_cut = (float) ComputeSerialEdgeCut(&cgraph);
my_totalv = (float) Mc_ComputeSerialTotalV(&cgraph, home);
Mc_ComputeSerialBalance(ctrl, &cgraph, part, lbvec);
my_balance = ssum(cgraph.ncon, lbvec);
my_balance /= (float) cgraph.ncon;
my_cost = ctrl->ipc_factor * my_cut + REDIST_WGT * ctrl->redist_base * my_totalv;
IFSET(ctrl->dbglvl, DBG_REFINEINFO, printf("%s initial cut: %.1f, totalv: %.1f, balance: %.3f\n",
(ctrl->mype == sr_pe ? "scratch-remap" : "diffusion"), my_cut, my_totalv, my_balance));
if (ctrl->mype == gd_pe) {
buffer[0] = my_cost;
buffer[1] = my_balance;
MPI_Send((void *)buffer, 2, MPI_FLOAT, sr_pe, 1, ctrl->comm);
}
else {
MPI_Recv((void *)buffer, 2, MPI_FLOAT, gd_pe, 1, ctrl->comm, &status);
your_cost = buffer[0];
your_balance = buffer[1];
}
}
if (ctrl->mype == sr_pe) {
who_wins = gd_pe;
if ((my_balance < 1.1 && your_balance > 1.1) ||
(my_balance < 1.1 && your_balance < 1.1 && my_cost < your_cost) ||
(my_balance > 1.1 && your_balance > 1.1 && my_balance < your_balance)) {
who_wins = sr_pe;
}
}
MPI_Bcast((void *)&who_wins, 1, MPI_INT, sr_pe, ctrl->comm);
}
else {
who_wins = sr_pe;
}
MPI_Bcast((void *)part, nvtxs, IDX_DATATYPE, who_wins, ctrl->comm);
idxcopy(graph->nvtxs, part+vtxdist[ctrl->mype], graph->where);
MPI_Comm_free(&ipcomm);
GKfree((void **)&xadj, (void **)&mytpwgts, LTERM);
/* For whatever reason, FreeGraph crashes here...so explicitly free the memory.
FreeGraph(agraph);
*/
GKfree((void **)&agraph->xadj, (void **)&agraph->adjncy, (void **)&agraph->vwgt, (void **)&agraph->nvwgt, LTERM);
GKfree((void **)&agraph->vsize, (void **)&agraph->adjwgt, (void **)&agraph->label, LTERM);
GKfree((void **)&agraph, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
}
/***********************************************************************************
* This function is the entry point of the parallel multilevel local diffusion
* algorithm. It uses parallel undirected diffusion followed by adaptive k-way
* refinement. This function utilizes local coarsening.
************************************************************************************/
void ParMETIS_V3_RefineKway(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, int *wgtflag, int *numflag, int *ncon,
int *nparts, float *tpwgts, float *ubvec, int *options, int *edgecut,
idxtype *part, MPI_Comm *comm)
{
int h, i;
int npes, mype;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph;
int tewgt, tvsize, nmoved, maxin, maxout;
float gtewgt, gtvsize, avg, maximb;
int ps_relation, seed, dbglvl = 0;
int iwgtflag, inumflag, incon, inparts, ioptions[10];
float *itpwgts, iubvec[MAXNCON];
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
/********************************/
/* Try and take care bad inputs */
/********************************/
if (options != NULL && options[0] == 1)
dbglvl = options[PMV3_OPTION_DBGLVL];
CheckInputs(REFINE_PARTITION, npes, dbglvl, wgtflag, &iwgtflag, numflag, &inumflag,
ncon, &incon, nparts, &inparts, tpwgts, &itpwgts, ubvec, iubvec,
NULL, NULL, options, ioptions, part, comm);
/* ADD: take care of disconnected graph */
/* ADD: take care of highly unbalanced vtxdist */
/*********************************/
/* Take care the nparts = 1 case */
/*********************************/
if (inparts <= 1) {
idxset(vtxdist[mype+1]-vtxdist[mype], 0, part);
*edgecut = 0;
return;
}
/**************************/
/* Set up data structures */
/**************************/
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
/*****************************/
/* Set up control structures */
/*****************************/
if (ioptions[0] == 1) {
dbglvl = ioptions[PMV3_OPTION_DBGLVL];
seed = ioptions[PMV3_OPTION_SEED];
ps_relation = (npes == inparts) ? ioptions[PMV3_OPTION_PSR] : DISCOUPLED;
}
else {
dbglvl = GLOBAL_DBGLVL;
seed = GLOBAL_SEED;
ps_relation = (npes == inparts) ? COUPLED : DISCOUPLED;
}
SetUpCtrl(&ctrl, inparts, dbglvl, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 50*incon*amax(npes, inparts));
ctrl.ipc_factor = 1000.0;
ctrl.redist_factor = 1.0;
ctrl.redist_base = 1.0;
ctrl.seed = (seed == 0) ? mype : seed*mype;
ctrl.sync = GlobalSEMax(&ctrl, seed);
ctrl.partType = REFINE_PARTITION;
ctrl.ps_relation = ps_relation;
ctrl.tpwgts = itpwgts;
graph = Moc_SetUpGraph(&ctrl, incon, vtxdist, xadj, vwgt, adjncy, adjwgt, &iwgtflag);
graph->vsize = idxsmalloc(graph->nvtxs, 1, "vsize");
graph->home = idxmalloc(graph->nvtxs, "home");
if (ctrl.ps_relation == COUPLED)
idxset(graph->nvtxs, mype, graph->home);
else
idxcopy(graph->nvtxs, part, graph->home);
tewgt = idxsum(graph->nedges, graph->adjwgt);
tvsize = idxsum(graph->nvtxs, graph->vsize);
gtewgt = (float) GlobalSESum(&ctrl, tewgt) + 1.0/graph->gnvtxs;
gtvsize = (float) GlobalSESum(&ctrl, tvsize) + 1.0/graph->gnvtxs;
ctrl.edge_size_ratio = gtewgt/gtvsize;
scopy(incon, iubvec, ctrl.ubvec);
PreAllocateMemory(&ctrl, graph, &wspace);
/***********************/
/* Partition and Remap */
/***********************/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Adaptive_Partition(&ctrl, graph, &wspace);
ParallelReMapGraph(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
idxcopy(graph->nvtxs, graph->where, part);
if (edgecut != NULL)
*edgecut = graph->mincut;
/***********************/
/* Take care of output */
/***********************/
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
if (ctrl.dbglvl&DBG_INFO) {
Mc_ComputeMoveStatistics(&ctrl, graph, &nmoved, &maxin, &maxout);
rprintf(&ctrl, "Final %3d-way Cut: %6d \tBalance: ", inparts, graph->mincut);
avg = 0.0;
for (h=0; h<incon; h++) {
maximb = 0.0;
for (i=0; i<inparts; i++)
maximb = amax(maximb, graph->gnpwgts[i*incon+h]/itpwgts[i*incon+h]);
avg += maximb;
rprintf(&ctrl, "%.3f ", maximb);
}
rprintf(&ctrl, "\nNMoved: %d %d %d %d\n", nmoved, maxin, maxout, maxin+maxout);
}
/*************************************/
/* Free memory, renumber, and return */
/*************************************/
GKfree((void **)&graph->lnpwgts, (void **)&graph->gnpwgts, (void **)&graph->nvwgt, (void **)(&graph->home), (void **)(&graph->vsize), LTERM);
GKfree((void **)&itpwgts, LTERM);
FreeInitialGraphAndRemap(graph, iwgtflag);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
return;
}
/*************************************************************************
* This function performs a node-based FM refinement. This is the
* one-way version
**************************************************************************/
void FM_2WayNodeRefine_OneSided(CtrlType *ctrl, GraphType *graph, float ubfactor, idxtype npasses)
{
idxtype i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind;
idxtype *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idxtype *mptr, *mind, *swaps, *perm;
PQueueType parts;
NRInfoType *rinfo;
idxtype higain, oldgain, mincut, initcut, mincutorder;
idxtype pass, to, other, limit;
idxtype badmaxpwgt, mindiff, newdiff;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
PQueueInit(ctrl, &parts, nvtxs, ComputeMaxNodeGain(nvtxs, xadj, adjncy, vwgt));
perm = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
mptr = idxwspacemalloc(ctrl, nvtxs+1);
mind = idxwspacemalloc(ctrl, nvtxs);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("Partitions-N1: [%6D %6D] Nv-Nb[%6D %6D]. ISep: %6D\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
badmaxpwgt = (int)(ubfactor*(pwgts[0]+pwgts[1]+pwgts[2])/2);
to = (pwgts[0] < pwgts[1] ? 1 : 0);
for (pass=0; pass<npasses; pass++) {
other = to;
to = (to+1)%2;
PQueueReset(&parts);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
RandomPermute(nbnd, perm, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(where[i] == 2);
PQueueInsert(&parts, i, vwgt[i]-rinfo[i].edegrees[other]);
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = (ctrl->oflags&OFLAG_COMPRESS ? amin(5*nbnd, 400) : amin(2*nbnd, 300));
/******************************************************
* Get into the FM loop
*******************************************************/
mptr[0] = nmind = 0;
mindiff = idxtype_abs(pwgts[0]-pwgts[1]);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if ((higain = PQueueGetMax(&parts)) == -1)
break;
ASSERT(bndptr[higain] != -1);
if (pwgts[to]+vwgt[higain] > badmaxpwgt)
break; /* No point going any further. Balance will be bad */
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = idxtype_abs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > limit) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
rinfo[k].edegrees[to] += vwgt[higain];
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%d %d %d\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
oldgain = vwgt[kk]-rinfo[kk].edegrees[other];
rinfo[kk].edegrees[other] -= vwgt[k];
/* Since the moves are one-sided this vertex has not been moved yet */
PQueueUpdateUp(&parts, kk, oldgain, oldgain+vwgt[k]);
}
}
/* Insert the new vertex into the priority queue. Safe due to one-sided moves */
PQueueInsert(&parts, k, vwgt[k]-edegrees[other]);
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, DBG_MOVEINFO,
mprintf("Moved %6D to %3D, Gain: %5D [%5D] \t[%5D %5D %5D] [%3D %2D]\n",
higain, to, (vwgt[higain]-rinfo[higain].edegrees[other]), vwgt[higain], pwgts[0], pwgts[1], pwgts[2], nswaps, limit));
}
/****************************************************************
* Roll back computation
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
ASSERT(where[higain] == to);
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, DBG_REFINE,
mprintf("\tMinimum sep: %6D at %5D, PWGTS: [%6D %6D], NBND: %6D\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut))
break;
}
PQueueFree(ctrl, &parts);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs+1);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function is the entry point of the initial partition algorithm
* that does recursive bissection.
* This algorithm assembles the graph to all the processors and preceeds
* by parallelizing the recursive bisection step.
**************************************************************************/
void Mc_InitPartition_RB(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int i, j;
int ncon, mype, npes, gnvtxs, ngroups;
idxtype *xadj, *adjncy, *adjwgt, *vwgt;
idxtype *part, *gwhere0, *gwhere1;
idxtype *tmpwhere, *tmpvwgt, *tmpxadj, *tmpadjncy, *tmpadjwgt;
GraphType *agraph;
int lnparts, fpart, fpe, lnpes;
int twoparts=2, numflag = 0, wgtflag = 3, moptions[10], edgecut, max_cut;
float *mytpwgts, mytpwgts2[2], lbvec[MAXNCON], lbsum, min_lbsum, wsum;
MPI_Comm ipcomm;
struct {
float sum;
int rank;
} lpesum, gpesum;
ncon = graph->ncon;
ngroups = amax(amin(RIP_SPLIT_FACTOR, ctrl->npes), 1);
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
agraph = Mc_AssembleAdaptiveGraph(ctrl, graph, wspace);
part = idxmalloc(agraph->nvtxs, "Mc_IP_RB: part");
xadj = idxmalloc(agraph->nvtxs+1, "Mc_IP_RB: xadj");
adjncy = idxmalloc(agraph->nedges, "Mc_IP_RB: adjncy");
adjwgt = idxmalloc(agraph->nedges, "Mc_IP_RB: adjwgt");
vwgt = idxmalloc(agraph->nvtxs*ncon, "Mc_IP_RB: vwgt");
idxcopy(agraph->nvtxs*ncon, agraph->vwgt, vwgt);
idxcopy(agraph->nvtxs+1, agraph->xadj, xadj);
idxcopy(agraph->nedges, agraph->adjncy, adjncy);
idxcopy(agraph->nedges, agraph->adjwgt, adjwgt);
MPI_Comm_split(ctrl->gcomm, ctrl->mype % ngroups, 0, &ipcomm);
MPI_Comm_rank(ipcomm, &mype);
MPI_Comm_size(ipcomm, &npes);
gnvtxs = agraph->nvtxs;
gwhere0 = idxsmalloc(gnvtxs, 0, "Mc_IP_RB: gwhere0");
gwhere1 = idxmalloc(gnvtxs, "Mc_IP_RB: gwhere1");
/* ADD: this assumes that tpwgts for all constraints is the same */
/* ADD: this is necessary because serial metis does not support the general case */
mytpwgts = fsmalloc(ctrl->nparts, 0.0, "mytpwgts");
for (i=0; i<ctrl->nparts; i++)
for (j=0; j<ncon; j++)
mytpwgts[i] += ctrl->tpwgts[i*ncon+j];
for (i=0; i<ctrl->nparts; i++)
mytpwgts[i] /= (float)ncon;
/* Go into the recursive bisection */
/* ADD: consider changing this to breadth-first type bisection */
moptions[0] = 0;
moptions[7] = ctrl->sync + (ctrl->mype % ngroups) + 1;
lnparts = ctrl->nparts;
fpart = fpe = 0;
lnpes = npes;
while (lnpes > 1 && lnparts > 1) {
/* Determine the weights of the partitions */
mytpwgts2[0] = ssum(lnparts/2, mytpwgts+fpart);
mytpwgts2[1] = 1.0-mytpwgts2[0];
if (ncon == 1)
METIS_WPartGraphKway2(&agraph->nvtxs, agraph->xadj, agraph->adjncy,
agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag, &twoparts, mytpwgts2,
moptions, &edgecut, part);
else {
METIS_mCPartGraphRecursive2(&agraph->nvtxs, &ncon, agraph->xadj,
agraph->adjncy, agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag,
&twoparts, mytpwgts2, moptions, &edgecut, part);
}
wsum = ssum(lnparts/2, mytpwgts+fpart);
sscale(lnparts/2, 1.0/wsum, mytpwgts+fpart);
sscale(lnparts-lnparts/2, 1.0/(1.0-wsum), mytpwgts+fpart+lnparts/2);
/* I'm picking the left branch */
if (mype < fpe+lnpes/2) {
Mc_KeepPart(agraph, wspace, part, 0);
lnpes = lnpes/2;
lnparts = lnparts/2;
}
else {
Mc_KeepPart(agraph, wspace, part, 1);
fpart = fpart + lnparts/2;
fpe = fpe + lnpes/2;
lnpes = lnpes - lnpes/2;
lnparts = lnparts - lnparts/2;
}
}
/* In case npes is greater than or equal to nparts */
if (lnparts == 1) {
/* Only the first process will assign labels (for the reduction to work) */
if (mype == fpe) {
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart;
}
}
/* In case npes is smaller than nparts */
else {
if (ncon == 1)
METIS_WPartGraphKway2(&agraph->nvtxs, agraph->xadj, agraph->adjncy,
agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag, &lnparts, mytpwgts+fpart,
moptions, &edgecut, part);
else
METIS_mCPartGraphRecursive2(&agraph->nvtxs, &ncon, agraph->xadj,
agraph->adjncy, agraph->vwgt, agraph->adjwgt, &wgtflag, &numflag,
&lnparts, mytpwgts+fpart, moptions, &edgecut, part);
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart + part[i];
}
MPI_Allreduce((void *)gwhere0, (void *)gwhere1, gnvtxs, IDX_DATATYPE, MPI_SUM, ipcomm);
if (ngroups > 1) {
tmpxadj = agraph->xadj;
tmpadjncy = agraph->adjncy;
tmpadjwgt = agraph->adjwgt;
tmpvwgt = agraph->vwgt;
tmpwhere = agraph->where;
agraph->xadj = xadj;
agraph->adjncy = adjncy;
agraph->adjwgt = adjwgt;
agraph->vwgt = vwgt;
agraph->where = gwhere1;
agraph->vwgt = vwgt;
agraph->nvtxs = gnvtxs;
Mc_ComputeSerialBalance(ctrl, agraph, gwhere1, lbvec);
lbsum = ssum(ncon, lbvec);
edgecut = ComputeSerialEdgeCut(agraph);
MPI_Allreduce((void *)&edgecut, (void *)&max_cut, 1, MPI_INT, MPI_MAX, ctrl->gcomm);
MPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, MPI_FLOAT, MPI_MIN, ctrl->gcomm);
lpesum.sum = lbsum;
if (min_lbsum < UNBALANCE_FRACTION * (float)(ncon)) {
if (lbsum < UNBALANCE_FRACTION * (float)(ncon))
lpesum.sum = (float) (edgecut);
else
lpesum.sum = (float) (max_cut);
}
MPI_Comm_rank(ctrl->gcomm, &(lpesum.rank));
MPI_Allreduce((void *)&lpesum, (void *)&gpesum, 1, MPI_FLOAT_INT, MPI_MINLOC, ctrl->gcomm);
MPI_Bcast((void *)gwhere1, gnvtxs, IDX_DATATYPE, gpesum.rank, ctrl->gcomm);
agraph->xadj = tmpxadj;
agraph->adjncy = tmpadjncy;
agraph->adjwgt = tmpadjwgt;
agraph->vwgt = tmpvwgt;
agraph->where = tmpwhere;
}
idxcopy(graph->nvtxs, gwhere1+graph->vtxdist[ctrl->mype], graph->where);
FreeGraph(agraph);
MPI_Comm_free(&ipcomm);
GKfree((void **)&gwhere0, (void **)&gwhere1, (void **)&mytpwgts, (void **)&part, (void **)&xadj, (void **)&adjncy, (void **)&adjwgt, (void **)&vwgt, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, MPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
}
/*************************************************************************
* This function performs an edge-based FM refinement
**************************************************************************/
void MocGeneral2WayBalance2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
int i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, limit, tmp, cnum;
idxtype *xadj, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind;
idxtype *moved, *swaps, *perm, *qnum;
float *nvwgt, *npwgts, origbal[MAXNCON], minbal[MAXNCON], newbal[MAXNCON];
PQueueType parts[MAXNCON][2];
int higain, oldgain, mincut, newcut, mincutorder;
float *maxwgt, *minwgt, tvec[MAXNCON];
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
npwgts = graph->npwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = idxwspacemalloc(ctrl, nvtxs);
swaps = idxwspacemalloc(ctrl, nvtxs);
perm = idxwspacemalloc(ctrl, nvtxs);
qnum = idxwspacemalloc(ctrl, nvtxs);
limit = amin(amax(0.01*nvtxs, 15), 100);
/* Setup the weight intervals of the two subdomains */
minwgt = fwspacemalloc(ctrl, 2*ncon);
maxwgt = fwspacemalloc(ctrl, 2*ncon);
for (i=0; i<2; i++) {
for (j=0; j<ncon; j++) {
maxwgt[i*ncon+j] = tpwgts[i]*ubvec[j];
minwgt[i*ncon+j] = tpwgts[i]*(1.0/ubvec[j]);
}
}
/* Initialize the queues */
for (i=0; i<ncon; i++) {
PQueueInit(ctrl, &parts[i][0], nvtxs, PLUS_GAINSPAN+1);
PQueueInit(ctrl, &parts[i][1], nvtxs, PLUS_GAINSPAN+1);
}
for (i=0; i<nvtxs; i++)
qnum[i] = samax(ncon, nvwgt+i*ncon);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, origbal);
for (i=0; i<ncon; i++)
minbal[i] = origbal[i];
newcut = mincut = graph->mincut;
mincutorder = -1;
if (ctrl->dbglvl&DBG_REFINE) {
printf("Parts: [");
for (l=0; l<ncon; l++)
printf("(%.3f, %.3f) ", npwgts[l], npwgts[ncon+l]);
printf("] T[%.3f %.3f], Nv-Nb[%5d, %5d]. ICut: %6d, LB: ", tpwgts[0], tpwgts[1],
graph->nvtxs, graph->nbnd, graph->mincut);
for (i=0; i<ncon; i++)
printf("%.3f ", origbal[i]);
printf("[B]\n");
}
idxset(nvtxs, -1, moved);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert all nodes in the priority queues */
nbnd = graph->nbnd;
RandomPermute(nvtxs, perm, 1);
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
PQueueInsert(&parts[qnum[i]][where[i]], i, ed[i]-id[i]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if (AreAllBelow(ncon, minbal, ubvec))
break;
SelectQueue3(ncon, npwgts, tpwgts, &from, &cnum, parts, maxwgt);
to = (from+1)%2;
if (from == -1 || (higain = PQueueGetMax(&parts[cnum][from])) == -1)
break;
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
newcut -= (ed[higain]-id[higain]);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, newbal);
if (IsBetter2wayBalance(ncon, newbal, minbal, ubvec) ||
(IsBetter2wayBalance(ncon, newbal, origbal, ubvec) && newcut < mincut)) {
mincut = newcut;
for (i=0; i<ncon; i++)
minbal[i] = newbal[i];
mincutorder = nswaps;
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+from*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
if (ctrl->dbglvl&DBG_MOVEINFO) {
printf("Moved %6d from %d(%d). Gain: %5d, Cut: %5d, NPwgts: ", higain, from, cnum, ed[higain]-id[higain], newcut);
for (i=0; i<ncon; i++)
printf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]);
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
printf(", LB: ");
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
if (mincutorder == nswaps)
printf(" *\n");
else
printf("\n");
}
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
oldgain = ed[k]-id[k];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update the queue position */
if (moved[k] == -1)
PQueueUpdate(&parts[qnum[k]][where[k]], k, oldgain, ed[k]-id[k]);
/* Update its boundary information */
if (ed[k] == 0 && bndptr[k] != -1)
BNDDelete(nbnd, bndind, bndptr, k);
else if (ed[k] > 0 && bndptr[k] == -1)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
saxpy(ncon, 1.0, nvwgt+higain*ncon, 1, npwgts+to*ncon, 1);
saxpy(ncon, -1.0, nvwgt+higain*ncon, 1, npwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
if (ctrl->dbglvl&DBG_REFINE) {
printf("\tMincut: %6d at %5d, NBND: %6d, NPwgts: [", mincut, mincutorder, nbnd);
for (i=0; i<ncon; i++)
printf("(%.3f, %.3f) ", npwgts[i], npwgts[ncon+i]);
printf("], LB: ");
Compute2WayHLoadImbalanceVec(ncon, npwgts, tpwgts, tvec);
for (i=0; i<ncon; i++)
printf("%.3f ", tvec[i]);
printf("\n");
}
graph->mincut = mincut;
graph->nbnd = nbnd;
for (i=0; i<ncon; i++) {
PQueueFree(ctrl, &parts[i][0]);
PQueueFree(ctrl, &parts[i][1]);
}
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
fwspacefree(ctrl, 2*ncon);
fwspacefree(ctrl, 2*ncon);
}
/***********************************************************************************
* This function is the entry point of the parallel kmetis algorithm that uses
* coordinates to compute an initial graph distribution.
************************************************************************************/
void ParMETIS_V3_PartGeomKway(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, int *wgtflag, int *numflag, int *ndims,
float *xyz, int *ncon, int *nparts, float *tpwgts, float *ubvec,
int *options, int *edgecut, idxtype *part, MPI_Comm *comm)
{
int h, i, j;
int nvtxs = -1, npes, mype;
int uwgtflag, cut, gcut, maxnvtxs;
int ltvwgts[MAXNCON];
int moptions[10];
CtrlType ctrl;
idxtype *uvwgt;
WorkSpaceType wspace;
GraphType *graph, *mgraph;
float avg, maximb, balance, *mytpwgts;
int seed, dbglvl = 0;
int iwgtflag, inumflag, incon, inparts, ioptions[10];
float *itpwgts, iubvec[MAXNCON];
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
/********************************/
/* Try and take care bad inputs */
/********************************/
if (options != NULL && options[0] == 1)
dbglvl = options[PMV3_OPTION_DBGLVL];
CheckInputs(STATIC_PARTITION, npes, dbglvl, wgtflag, &iwgtflag, numflag, &inumflag,
ncon, &incon, nparts, &inparts, tpwgts, &itpwgts, ubvec, iubvec,
NULL, NULL, options, ioptions, part, comm);
/*********************************/
/* Take care the nparts = 1 case */
/*********************************/
if (inparts <= 1) {
idxset(vtxdist[mype+1]-vtxdist[mype], 0, part);
*edgecut = 0;
return;
}
/******************************/
/* Take care of npes = 1 case */
/******************************/
if (npes == 1 && inparts > 1) {
moptions[0] = 0;
nvtxs = vtxdist[1];
if (incon == 1) {
METIS_WPartGraphKway(&nvtxs, xadj, adjncy, vwgt, adjwgt, &iwgtflag, &inumflag,
&inparts, itpwgts, moptions, edgecut, part);
}
else {
/* ADD: this is because METIS does not support tpwgts for all constraints */
mytpwgts = fmalloc(inparts, "mytpwgts");
for (i=0; i<inparts; i++)
mytpwgts[i] = itpwgts[i*incon];
moptions[7] = -1;
METIS_mCPartGraphRecursive2(&nvtxs, &incon, xadj, adjncy, vwgt, adjwgt, &iwgtflag,
&inumflag, &inparts, mytpwgts, moptions, edgecut, part);
free(mytpwgts);
}
return;
}
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
/*****************************/
/* Set up control structures */
/*****************************/
if (ioptions[0] == 1) {
dbglvl = ioptions[PMV3_OPTION_DBGLVL];
seed = ioptions[PMV3_OPTION_SEED];
}
else {
dbglvl = GLOBAL_DBGLVL;
seed = GLOBAL_SEED;
}
SetUpCtrl(&ctrl, npes, dbglvl, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*incon*amax(npes, inparts));
ctrl.seed = (seed == 0) ? mype : seed*mype;
ctrl.sync = GlobalSEMax(&ctrl, seed);
ctrl.partType = STATIC_PARTITION;
ctrl.ps_relation = -1;
ctrl.tpwgts = itpwgts;
scopy(incon, iubvec, ctrl.ubvec);
uwgtflag = iwgtflag|2;
uvwgt = idxsmalloc(vtxdist[mype+1]-vtxdist[mype], 1, "uvwgt");
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, uvwgt, adjncy, adjwgt, &uwgtflag);
free(graph->nvwgt); graph->nvwgt = NULL;
PreAllocateMemory(&ctrl, graph, &wspace);
/*=================================================================
* Compute the initial npes-way partitioning geometric partitioning
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Coordinate_Partition(&ctrl, graph, *ndims, xyz, 1, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
/*=================================================================
* Move the graph according to the partitioning
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.MoveTmr));
free(uvwgt);
graph->vwgt = ((iwgtflag&2) != 0) ? vwgt : idxsmalloc(graph->nvtxs*incon, 1, "vwgt");
graph->ncon = incon;
j = ctrl.nparts;
ctrl.nparts = ctrl.npes;
mgraph = Moc_MoveGraph(&ctrl, graph, &wspace);
ctrl.nparts = j;
/**********************************************************/
/* Do the same functionality as Moc_SetUpGraph for mgraph */
/**********************************************************/
/* compute tvwgts */
for (j=0; j<incon; j++)
ltvwgts[j] = 0;
for (i=0; i<graph->nvtxs; i++)
for (j=0; j<incon; j++)
ltvwgts[j] += mgraph->vwgt[i*incon+j];
for (j=0; j<incon; j++)
ctrl.tvwgts[j] = GlobalSESum(&ctrl, ltvwgts[j]);
/* check for zero wgt constraints */
for (i=0; i<incon; i++) {
/* ADD: take care of the case in which tvwgts is zero */
if (ctrl.tvwgts[i] == 0) {
if (ctrl.mype == 0) printf("ERROR: sum weight for constraint %d is zero\n", i);
MPI_Finalize();
exit(-1);
}
}
/* compute nvwgt */
mgraph->nvwgt = fmalloc(mgraph->nvtxs*incon, "mgraph->nvwgt");
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<incon; j++)
mgraph->nvwgt[i*incon+j] = (float)(mgraph->vwgt[i*incon+j]) / (float)(ctrl.tvwgts[j]);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.MoveTmr));
if (ctrl.dbglvl&DBG_INFO) {
cut = 0;
for (i=0; i<graph->nvtxs; i++)
for (j=graph->xadj[i]; j<graph->xadj[i+1]; j++)
if (graph->where[i] != graph->where[graph->adjncy[j]])
cut += graph->adjwgt[j];
gcut = GlobalSESum(&ctrl, cut)/2;
maxnvtxs = GlobalSEMax(&ctrl, mgraph->nvtxs);
balance = (float)(maxnvtxs)/((float)(graph->gnvtxs)/(float)(npes));
rprintf(&ctrl, "XYZ Cut: %6d \tBalance: %6.3f [%d %d %d]\n",
gcut, balance, maxnvtxs, graph->gnvtxs, npes);
}
/*=================================================================
* Set up the newly moved graph
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
ctrl.nparts = inparts;
FreeWSpace(&wspace);
PreAllocateMemory(&ctrl, mgraph, &wspace);
/*=======================================================
* Now compute the partition of the moved graph
=======================================================*/
if (vtxdist[npes] < SMALLGRAPH || vtxdist[npes] < npes*20 || GlobalSESum(&ctrl, mgraph->nedges) == 0) {
IFSET(ctrl.dbglvl, DBG_INFO, rprintf(&ctrl, "Partitioning a graph of size %d serially\n", vtxdist[npes]));
PartitionSmallGraph(&ctrl, mgraph, &wspace);
}
else {
Moc_Global_Partition(&ctrl, mgraph, &wspace);
}
ParallelReMapGraph(&ctrl, mgraph, &wspace);
/* Invert the ordering back to the original graph */
ctrl.nparts = npes;
ProjectInfoBack(&ctrl, graph, part, mgraph->where, &wspace);
*edgecut = mgraph->mincut;
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
/*******************/
/* Print out stats */
/*******************/
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
if (ctrl.dbglvl&DBG_INFO) {
rprintf(&ctrl, "Final %d-way CUT: %6d \tBalance: ", inparts, mgraph->mincut);
avg = 0.0;
for (h=0; h<incon; h++) {
maximb = 0.0;
for (i=0; i<inparts; i++)
maximb = amax(maximb, mgraph->gnpwgts[i*incon+h]/itpwgts[i*incon+h]);
avg += maximb;
rprintf(&ctrl, "%.3f ", maximb);
}
rprintf(&ctrl, " avg: %.3f\n", avg/(float)incon);
}
GKfree((void **)&itpwgts, LTERM);
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph, iwgtflag);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
}