/*************************************************************************
* This function is the entry point for KWMETIS
**************************************************************************/
void METIS_WPartGraphKway(int *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *adjwgt, int *wgtflag, int *numflag, int *nparts,
float *tpwgts, int *options, int *edgecut, idxtype *part)
{
int i, j;
GraphType graph;
CtrlType ctrl;
if (*numflag == 1)
Change2CNumbering(*nvtxs, xadj, adjncy);
SetUpGraph(&graph, OP_KMETIS, *nvtxs, 1, xadj, adjncy, vwgt, adjwgt, *wgtflag);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = KMETIS_CTYPE;
ctrl.IType = KMETIS_ITYPE;
ctrl.RType = KMETIS_RTYPE;
ctrl.dbglvl = KMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.optype = OP_KMETIS;
ctrl.CoarsenTo = amax((*nvtxs)/(40*log2i(*nparts)), 20*(*nparts));
ctrl.maxvwgt = 1.5*((graph.vwgt ? idxsum(*nvtxs, graph.vwgt) : (*nvtxs))/ctrl.CoarsenTo);
InitRandom(-1);
AllocateWorkSpace(&ctrl, &graph, *nparts);
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
*edgecut = MlevelKWayPartitioning(&ctrl, &graph, *nparts, part, tpwgts, 1.03);
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimers(&ctrl));
FreeWorkSpace(&ctrl, &graph);
if (*numflag == 1)
Change2FNumbering(*nvtxs, xadj, adjncy, part);
}
/*************************************************************************
* This function finds a matching using the HEM heuristic
**************************************************************************/
void Match_RM_NVW(CtrlType *ctrl, GraphType *graph)
{
int i, ii, j, nvtxs, cnvtxs, maxidx;
idxtype *xadj, *adjncy;
idxtype *match, *cmap, *perm;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
cmap = graph->cmap;
match = idxset(nvtxs, UNMATCHED, idxwspacemalloc(ctrl, nvtxs));
perm = idxwspacemalloc(ctrl, nvtxs);
RandomPermute(nvtxs, perm, 1);
cnvtxs = 0;
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
maxidx = i;
/* Find a random matching, subject to maxvwgt constraints */
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (match[adjncy[j]] == UNMATCHED) {
maxidx = adjncy[j];
break;
}
}
cmap[i] = cmap[maxidx] = cnvtxs++;
match[i] = maxidx;
match[maxidx] = i;
}
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
CreateCoarseGraph_NVW(ctrl, graph, cnvtxs, match, perm);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/***********************************************************************************
* 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);
}
void informdoorbell (void)
{
packet pkt;
char buzzer_pressed=BUZZER_PRESSED;
/* Make a noise */
starttimer(40000);
buzzerled(BUZZER);
while(gettimer()>10000);
buzzerled(ALLOFF);
stoptimer();
/* Send packet */
pkt.datatype=CONTROL;
pkt.data=&(buzzer_pressed);
pkt.length=1;
senddata(&pkt);
}
/*
static int broadcast_cpuhp_notify(struct notifier_block *n,
unsigned long action, void *hcpu){
int hotcpu = (unsigned long)hcpu;
struct jz_timerevent *evt = &per_cpu(jzclockevent, hotcpu);
if(hotcpu != 0) {
switch(action & 0xf) {
case CPU_DEAD:
jz_set_mode(CLOCK_EVT_MODE_SHUTDOWN,&evt->clkevt);
break;
case CPU_ONLINE:
jz_set_mode(CLOCK_EVT_MODE_RESUME,&evt->clkevt);
break;
}
}
return NOTIFY_OK;
}
*/
static void jz_clockevent_init(struct jz_timerevent *evt_dev,int cpu) {
struct clock_event_device *cd = &evt_dev->clkevt;
struct clk *ext_clk = clk_get(NULL,"ext1");
spin_lock_init(&evt_dev->lock);
evt_dev->rate = clk_get_rate(ext_clk) / CLKEVENT_DIV;
clk_put(ext_clk);
stoptimer(evt_dev);
outl((1 << evt_dev->ch),evt_dev->ctrl_addr + TCU_TMCR);
outl(CSRDIV(CLKEVENT_DIV) | CSR_EXT_EN,evt_dev->config_addr);
if(evt_dev->requestirq == 0){
evt_dev->evt_action.handler = jz_timer_interrupt;
evt_dev->evt_action.thread_fn = NULL;
evt_dev->evt_action.flags = IRQF_DISABLED | IRQF_PERCPU | IRQF_TIMER;
evt_dev->evt_action.name = "jz-timerirq";
evt_dev->evt_action.dev_id = (void*)evt_dev;
if(setup_irq(evt_dev->irq, &evt_dev->evt_action) < 0) {
pr_err("timer request irq error\n");
BUG();
}
evt_dev->requestirq = 1;
}
memset(cd,0,sizeof(struct clock_event_device));
cd->name = "jz-clockenvent";
cd->features = CLOCK_EVT_FEAT_ONESHOT | CLOCK_EVT_FEAT_PERIODIC;
cd->shift = 10;
cd->rating = 400;
cd->set_mode = jz_set_mode;
cd->set_next_event = jz_set_next_event;
cd->irq = evt_dev->irq;
cd->cpumask = cpumask_of(cpu);
clockevents_config_and_register(cd,evt_dev->rate,4,65536);
/*
evt_dev->cpu_notify.notifier_call = broadcast_cpuhp_notify;
if(cpu == 0){
register_cpu_notifier(&evt_dev->cpu_notify);
}
*/
}
/*************************************************************************
* This function computes the initial bisection of the coarsest graph
**************************************************************************/
void InitSeparator(CtrlType *ctrl, GraphType *graph, float ubfactor)
{
int dbglvl;
dbglvl = ctrl->dbglvl;
IFSET(ctrl->dbglvl, DBG_REFINE, ctrl->dbglvl -= DBG_REFINE);
IFSET(ctrl->dbglvl, DBG_MOVEINFO, ctrl->dbglvl -= DBG_MOVEINFO);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
GrowBisectionNode(ctrl, graph, ubfactor);
Compute2WayNodePartitionParams(ctrl, graph);
IFSET(ctrl->dbglvl, DBG_IPART, printf("Initial Sep: %d\n", graph->mincut));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
ctrl->dbglvl = dbglvl;
}
/***********************************************************************************
* This function is the testing routine for the adaptive multilevel partitioning code.
* It computes a partition from scratch, it then moves the graph and changes some
* of the vertex weights and then call the adaptive code.
************************************************************************************/
void TestParMGridGen(char *filename, int *options, int minsize, int maxsize, MPI_Comm comm)
{
int i, nparts, npes, mype;
MGridGraphType graph;
idxtype *part;
double tmr;
MPI_Comm_size(comm, &npes);
MPI_Comm_rank(comm, &mype);
MGridReadTestGraph(&graph, filename, comm);
part = idxmalloc(graph.nvtxs, "TestParMGridGen: part");
/*======================================================================
/ ParMETIS_AspectRatio
/=======================================================================*/
if (mype==0)
printf("------------------------ PARAMETERS --------------------------------------\n");
for (i=0; i<npes; i++)
if (mype == i)
printf("%s, Dim=%d [%2d %2d] CType=%d RType=%d Nvtxs=%d Nedges=%d\n", filename,
options[OPTION_DIM], minsize, maxsize, options[OPTION_CTYPE],
options[OPTION_RTYPE], graph.nvtxs, graph.nedges);
cleartimer(tmr);
MPI_Barrier(comm);
starttimer(tmr);
ParMGridGen(graph.vtxdist, graph.xadj, graph.vvol, graph.vsurf, graph.adjncy,
graph.adjwgt, &nparts, minsize, maxsize, options, part, &comm);
MPI_Barrier(comm);
stoptimer(tmr);
printf("Total Time = %lf\n", gettimer(tmr));
WriteParallelPartition(filename, part, graph.vtxdist, nparts, mype, npes);
IMfree(&graph.vtxdist, &graph.xadj, &graph.vvol, &graph.vsurf, &graph.vwgt,
&graph.adjncy, &graph.adjwgt, &part, LTERM);
}
static void jz_clockevent_init(struct jz_timerevent *evt_dev)
{
struct clock_event_device *cd = &evt_dev->clkevt;
struct clk *ext_clk = clk_get(NULL, "ext1");
spin_lock_init(&evt_dev->lock);
evt_dev->rate = clk_get_rate(ext_clk) / CLKEVENT_DIV;
clk_put(ext_clk);
evt_dev->clk_gate = clk_get(NULL, "tcu");
if (IS_ERR(evt_dev->clk_gate)) {
evt_dev->clk_gate = NULL;
printk("warning: tcu clk get fail!\n");
}
if (evt_dev->clk_gate)
clk_enable(evt_dev->clk_gate);
stoptimer();
tcu_writel(CH_TCSR(CLKEVENT_CH), CSRDIV(CLKEVENT_DIV) | CSR_EXT_EN);
evt_dev->evt_action.handler = jz_timer_interrupt;
evt_dev->evt_action.thread_fn = NULL;
evt_dev->evt_action.flags = IRQF_DISABLED | IRQF_TIMER;
evt_dev->evt_action.name = "jz-timerirq";
evt_dev->evt_action.dev_id = (void*)evt_dev;
if (setup_irq(IRQ_TCU1, &evt_dev->evt_action) < 0) {
pr_err("timer request irq error\n");
BUG();
}
memset(cd, 0, sizeof(struct clock_event_device));
cd->name = "jz-clockenvent";
cd->features = CLOCK_EVT_FEAT_ONESHOT | CLOCK_EVT_FEAT_PERIODIC;
cd->shift = 10;
cd->rating = 400;
cd->set_mode = jz_set_mode;
cd->set_next_event = jz_set_next_event;
cd->irq = IRQ_TCU1;
cd->cpumask = cpumask_of(0);
clockevents_config_and_register(cd, evt_dev->rate, 4, 65536);
printk("clockevents_config_and_register success.\n");
}
/*************************************************************************
* This function takes a graph and produces a bisection of it
**************************************************************************/
int MlevelVolKWayPartitioning(CtrlType *ctrl, GraphType *graph, int nparts, idxtype *part,
float *tpwgts, float ubfactor)
{
int i, j, nvtxs, tvwgt, tpwgts2[2];
GraphType *cgraph;
int wgtflag=3, numflag=0, options[10], edgecut;
cgraph = Coarsen2Way(ctrl, graph);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
AllocateVolKWayPartitionMemory(ctrl, cgraph, nparts);
options[0] = 1;
options[OPTION_CTYPE] = MATCH_SHEMKWAY;
options[OPTION_ITYPE] = IPART_GGPKL;
options[OPTION_RTYPE] = RTYPE_FM;
options[OPTION_DBGLVL] = 0;
METIS_WPartGraphRecursive(&cgraph->nvtxs, cgraph->xadj, cgraph->adjncy, cgraph->vwgt,
cgraph->adjwgt, &wgtflag, &numflag, &nparts, tpwgts, options,
&edgecut, cgraph->where);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
IFSET(ctrl->dbglvl, DBG_IPART, printf("Initial %d-way partitioning cut: %d\n", nparts, edgecut));
IFSET(ctrl->dbglvl, DBG_KWAYPINFO, ComputePartitionInfo(cgraph, nparts, cgraph->where));
RefineVolKWay(ctrl, graph, cgraph, nparts, tpwgts, ubfactor);
idxcopy(graph->nvtxs, graph->where, part);
GKfree((void**)&graph->gdata, &graph->rdata, LTERM);
return graph->minvol;
}
/*************************************************************************
* This function computes the initial id/ed
**************************************************************************/
void ComputeVolKWayPartitionParams(CtrlType *ctrl, GraphType *graph, int nparts)
{
int i, ii, j, k, kk, l, nvtxs, nbnd, mincut, minvol, me, other, pid;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *pwgts, *where;
VRInfoType *rinfo, *myrinfo, *orinfo;
VEDegreeType *myedegrees, *oedegrees;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
pwgts = idxset(nparts, 0, graph->pwgts);
rinfo = graph->vrinfo;
starttimer(ctrl->AuxTmr1);
/*------------------------------------------------------------
/ Compute now the id/ed degrees
/------------------------------------------------------------*/
ctrl->wspace.cdegree = 0;
mincut = 0;
for (i=0; i<nvtxs; i++) {
me = where[i];
pwgts[me] += vwgt[i];
myrinfo = rinfo+i;
myrinfo->id = myrinfo->ed = myrinfo->nid = myrinfo->ndegrees = 0;
myrinfo->edegrees = NULL;
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (me == where[adjncy[j]]) {
myrinfo->id += adjwgt[j];
myrinfo->nid++;
}
}
myrinfo->ed = graph->adjwgtsum[i] - myrinfo->id;
mincut += myrinfo->ed;
/* Time to compute the particular external degrees */
if (myrinfo->ed > 0) {
myedegrees = myrinfo->edegrees = ctrl->wspace.vedegrees+ctrl->wspace.cdegree;
ctrl->wspace.cdegree += xadj[i+1]-xadj[i];
for (j=xadj[i]; j<xadj[i+1]; j++) {
other = where[adjncy[j]];
if (me != other) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (myedegrees[k].pid == other) {
myedegrees[k].ed += adjwgt[j];
myedegrees[k].ned++;
break;
}
}
if (k == myrinfo->ndegrees) {
myedegrees[myrinfo->ndegrees].gv = 0;
myedegrees[myrinfo->ndegrees].pid = other;
myedegrees[myrinfo->ndegrees].ed = adjwgt[j];
myedegrees[myrinfo->ndegrees++].ned = 1;
}
}
}
ASSERT(myrinfo->ndegrees <= xadj[i+1]-xadj[i]);
}
}
graph->mincut = mincut/2;
stoptimer(ctrl->AuxTmr1);
ComputeKWayVolGains(ctrl, graph, nparts);
}
/*************************************************************************
* This function is the entry point for the node ND code for ParMETIS
**************************************************************************/
void METIS_NodeNDP(int nvtxs, idxtype *xadj, idxtype *adjncy, int npes,
int *options, idxtype *perm, idxtype *iperm, idxtype *sizes)
{
int i, ii, j, l, wflag, nflag;
GraphType graph;
CtrlType ctrl;
idxtype *cptr, *cind;
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = ONMETIS_CTYPE;
ctrl.IType = ONMETIS_ITYPE;
ctrl.RType = ONMETIS_RTYPE;
ctrl.dbglvl = ONMETIS_DBGLVL;
ctrl.oflags = ONMETIS_OFLAGS;
ctrl.pfactor = ONMETIS_PFACTOR;
ctrl.nseps = ONMETIS_NSEPS;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
ctrl.oflags = options[OPTION_OFLAGS];
ctrl.pfactor = options[OPTION_PFACTOR];
ctrl.nseps = options[OPTION_NSEPS];
}
if (ctrl.nseps < 1)
ctrl.nseps = 1;
ctrl.optype = OP_ONMETIS;
ctrl.CoarsenTo = 100;
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
InitRandom(-1);
if (ctrl.oflags&OFLAG_COMPRESS) {
/*============================================================
* Compress the graph
==============================================================*/
cptr = idxmalloc(nvtxs+1, "ONMETIS: cptr");
cind = idxmalloc(nvtxs, "ONMETIS: cind");
CompressGraph(&ctrl, &graph, nvtxs, xadj, adjncy, cptr, cind);
if (graph.nvtxs >= COMPRESSION_FRACTION*(nvtxs)) {
ctrl.oflags--; /* We actually performed no compression */
GKfree((void**)&cptr, &cind, LTERM);
}
else if (2*graph.nvtxs < nvtxs && ctrl.nseps == 1)
ctrl.nseps = 2;
}
else {
SetUpGraph(&graph, OP_ONMETIS, nvtxs, 1, xadj, adjncy, NULL, NULL, 0);
}
/*=============================================================
* Do the nested dissection ordering
--=============================================================*/
ctrl.maxvwgt = 1.5*(idxsum(graph.nvtxs, graph.vwgt)/ctrl.CoarsenTo);
AllocateWorkSpace(&ctrl, &graph, 2);
idxset(2*npes-1, 0, sizes);
MlevelNestedDissectionP(&ctrl, &graph, iperm, graph.nvtxs, npes, 0, sizes);
FreeWorkSpace(&ctrl, &graph);
if (ctrl.oflags&OFLAG_COMPRESS) { /* Uncompress the ordering */
if (graph.nvtxs < COMPRESSION_FRACTION*(nvtxs)) {
/* construct perm from iperm */
for (i=0; i<graph.nvtxs; i++)
perm[iperm[i]] = i;
for (l=ii=0; ii<graph.nvtxs; ii++) {
i = perm[ii];
for (j=cptr[i]; j<cptr[i+1]; j++)
iperm[cind[j]] = l++;
}
}
GKfree((void**)&cptr, &cind, LTERM);
}
for (i=0; i<nvtxs; i++)
perm[iperm[i]] = i;
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimers(&ctrl));
}
/*************************************************************************
* This function is the entry point for PWMETIS that accepts exact weights
* for the target partitions
**************************************************************************/
void METIS_mCPartGraphRecursive2(int *nvtxs, int *ncon, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, int *wgtflag, int *numflag, int *nparts,
float *tpwgts, int *options, int *edgecut, idxtype *part)
{
int i, j;
GraphType graph;
CtrlType ctrl;
float *mytpwgts;
float avgwgt;
if (*numflag == 1)
Change2CNumbering(*nvtxs, xadj, adjncy);
SetUpGraph(&graph, OP_PMETIS, *nvtxs, *ncon, xadj, adjncy, vwgt, adjwgt, *wgtflag);
graph.npwgts = NULL;
mytpwgts = fmalloc(*nparts, "mytpwgts");
scopy(*nparts, tpwgts, mytpwgts);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = McPMETIS_CTYPE;
ctrl.IType = McPMETIS_ITYPE;
ctrl.RType = McPMETIS_RTYPE;
ctrl.dbglvl = McPMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.optype = OP_PMETIS;
ctrl.CoarsenTo = 100;
ctrl.nmaxvwgt = 1.5/(1.0*ctrl.CoarsenTo);
InitRandom(options[7]);
AllocateWorkSpace(&ctrl, &graph, *nparts);
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
ASSERT(CheckGraph(&graph));
*edgecut = MCMlevelRecursiveBisection2(&ctrl, &graph, *nparts, mytpwgts, part, 1.000, 0);
/*
{
idxtype wgt[2048], minwgt, maxwgt, sumwgt;
printf("nvtxs: %d, nparts: %d, ncon: %d\n", graph.nvtxs, *nparts, *ncon);
for (i=0; i<(*nparts)*(*ncon); i++)
wgt[i] = 0;
for (i=0; i<graph.nvtxs; i++)
for (j=0; j<*ncon; j++)
wgt[part[i]*(*ncon)+j] += vwgt[i*(*ncon)+j];
for (j=0; j<*ncon; j++) {
minwgt = maxwgt = sumwgt = 0;
for (i=0; i<(*nparts); i++) {
minwgt = (wgt[i*(*ncon)+j] < wgt[minwgt*(*ncon)+j]) ? i : minwgt;
maxwgt = (wgt[i*(*ncon)+j] > wgt[maxwgt*(*ncon)+j]) ? i : maxwgt;
sumwgt += wgt[i*(*ncon)+j];
}
avgwgt = (float)sumwgt / (float)*nparts;
printf("min: %5d, max: %5d, avg: %5.2f, balance: %6.3f\n", wgt[minwgt*(*ncon)+j], wgt[maxwgt*(*ncon)+j], avgwgt, (float)wgt[maxwgt*(*ncon)+j] / avgwgt);
}
printf("\n");
}
*/
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimers(&ctrl));
FreeWorkSpace(&ctrl, &graph);
GKfree((void**)(void *)&mytpwgts, LTERM);
if (*numflag == 1)
Change2FNumbering(*nvtxs, xadj, adjncy, part);
}
/*************************************************************************
* This function is the entry point of refinement
**************************************************************************/
void RefineKWay(CtrlType *ctrl, GraphType *orggraph, GraphType *graph, int nparts, float *tpwgts, float ubfactor)
{
int i, nlevels, mustfree=0;
GraphType *ptr;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->UncoarsenTmr));
/* Compute the parameters of the coarsest graph */
ComputeKWayPartitionParams(ctrl, graph, nparts);
/* Take care any non-contiguity */
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->AuxTmr1));
if (ctrl->RType == RTYPE_KWAYRANDOM_MCONN) {
EliminateComponents(ctrl, graph, nparts, tpwgts, 1.25);
EliminateSubDomainEdges(ctrl, graph, nparts, tpwgts);
EliminateComponents(ctrl, graph, nparts, tpwgts, 1.25);
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->AuxTmr1));
/* Determine how many levels are there */
for (ptr=graph, nlevels=0; ptr!=orggraph; ptr=ptr->finer, nlevels++);
for (i=0; ;i++) {
/* PrintSubDomainGraph(graph, nparts, graph->where); */
if (ctrl->RType == RTYPE_KWAYRANDOM_MCONN && (i == nlevels/2 || i == nlevels/2+1))
EliminateSubDomainEdges(ctrl, graph, nparts, tpwgts);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->RefTmr));
if (2*i >= nlevels && !IsBalanced(graph->pwgts, nparts, tpwgts, 1.04*ubfactor)) {
ComputeKWayBalanceBoundary(ctrl, graph, nparts);
if (ctrl->RType == RTYPE_KWAYRANDOM_MCONN)
Greedy_KWayEdgeBalanceMConn(ctrl, graph, nparts, tpwgts, ubfactor, 1);
else
Greedy_KWayEdgeBalance(ctrl, graph, nparts, tpwgts, ubfactor, 1);
ComputeKWayBoundary(ctrl, graph, nparts);
}
switch (ctrl->RType) {
case RTYPE_KWAYRANDOM:
Random_KWayEdgeRefine(ctrl, graph, nparts, tpwgts, ubfactor, 10, 1);
break;
case RTYPE_KWAYGREEDY:
Greedy_KWayEdgeRefine(ctrl, graph, nparts, tpwgts, ubfactor, 10);
break;
case RTYPE_KWAYRANDOM_MCONN:
Random_KWayEdgeRefineMConn(ctrl, graph, nparts, tpwgts, ubfactor, 10, 1);
break;
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->RefTmr));
if (graph == orggraph)
break;
GKfree(&graph->gdata, LTERM); /* Deallocate the graph related arrays */
graph = graph->finer;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ProjectTmr));
if (graph->vwgt == NULL) {
graph->vwgt = idxsmalloc(graph->nvtxs, 1, "RefineKWay: graph->vwgt");
graph->adjwgt = idxsmalloc(graph->nedges, 1, "RefineKWay: graph->adjwgt");
mustfree = 1;
}
ProjectKWayPartition(ctrl, graph, nparts);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ProjectTmr));
}
if (!IsBalanced(graph->pwgts, nparts, tpwgts, ubfactor)) {
ComputeKWayBalanceBoundary(ctrl, graph, nparts);
if (ctrl->RType == RTYPE_KWAYRANDOM_MCONN) {
Greedy_KWayEdgeBalanceMConn(ctrl, graph, nparts, tpwgts, ubfactor, 8);
Random_KWayEdgeRefineMConn(ctrl, graph, nparts, tpwgts, ubfactor, 10, 0);
}
else {
Greedy_KWayEdgeBalance(ctrl, graph, nparts, tpwgts, ubfactor, 8);
Random_KWayEdgeRefine(ctrl, graph, nparts, tpwgts, ubfactor, 10, 0);
}
}
/* Take care any trivial non-contiguity */
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->AuxTmr2));
EliminateComponents(ctrl, graph, nparts, tpwgts, ubfactor);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->AuxTmr2));
if (mustfree)
GKfree(&graph->vwgt, &graph->adjwgt, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->UncoarsenTmr));
}
/***********************************************************************************
* This function is the entry point of the parallel ordering algorithm.
* This function assumes that the graph is already nice partitioned among the
* processors and then proceeds to perform recursive bisection.
************************************************************************************/
void ParMETIS_V3_NodeND(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy, int *numflag,
int *options, idxtype *order, idxtype *sizes, MPI_Comm *comm)
{
int i, j;
int ltvwgts[MAXNCON];
int nparts, npes, mype, wgtflag = 0, seed = GLOBAL_SEED;
CtrlType ctrl;
WorkSpaceType wspace;
GraphType *graph, *mgraph;
idxtype *morder;
int minnvtxs;
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
nparts = npes;
if (!ispow2(npes)) {
if (mype == 0)
printf("Error: The number of processors must be a power of 2!\n");
return;
}
if (vtxdist[npes] < (int)((float)(npes*npes)*1.2)) {
if (mype == 0)
printf("Error: Too many processors for this many vertices.\n");
return;
}
minnvtxs = vtxdist[1]-vtxdist[0];
for (i=0; i<npes; i++)
minnvtxs = (minnvtxs < vtxdist[i+1]-vtxdist[i]) ? minnvtxs : vtxdist[i+1]-vtxdist[i];
if (minnvtxs < (int)((float)npes*1.1)) {
if (mype == 0)
printf("Error: vertices are not distributed equally.\n");
return;
}
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 1);
SetUpCtrl(&ctrl, nparts, options[PMV3_OPTION_DBGLVL], *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*npes);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*amax(npes, nparts));
ctrl.seed = mype;
ctrl.sync = seed;
ctrl.partType = STATIC_PARTITION;
ctrl.ps_relation = -1;
ctrl.tpwgts = fsmalloc(nparts, 1.0/(float)(nparts), "tpwgts");
ctrl.ubvec[0] = 1.03;
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, NULL, adjncy, NULL, &wgtflag);
PreAllocateMemory(&ctrl, graph, &wspace);
/*=======================================================
* Compute the initial k-way partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Moc_Global_Partition(&ctrl, graph, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
/*=======================================================
* Move the graph according to the partitioning
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.MoveTmr));
MALLOC_CHECK(NULL);
graph->ncon = 1;
mgraph = Moc_MoveGraph(&ctrl, graph, &wspace);
MALLOC_CHECK(NULL);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.MoveTmr));
/*=======================================================
* Now compute an ordering of the moved graph
=======================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
FreeWSpace(&wspace);
PreAllocateMemory(&ctrl, mgraph, &wspace);
ctrl.ipart = ISEP_NODE;
ctrl.CoarsenTo = amin(vtxdist[npes]+1, amax(20*npes, 1000));
/* compute tvwgts */
for (j=0; j<mgraph->ncon; j++)
ltvwgts[j] = 0;
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<mgraph->ncon; j++)
ltvwgts[j] += mgraph->vwgt[i*mgraph->ncon+j];
for (j=0; j<mgraph->ncon; j++)
ctrl.tvwgts[j] = GlobalSESum(&ctrl, ltvwgts[j]);
mgraph->nvwgt = fmalloc(mgraph->nvtxs*mgraph->ncon, "mgraph->nvwgt");
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<mgraph->ncon; j++)
mgraph->nvwgt[i*mgraph->ncon+j] = (float)(mgraph->vwgt[i*mgraph->ncon+j]) / (float)(ctrl.tvwgts[j]);
morder = idxmalloc(mgraph->nvtxs, "PAROMETIS: morder");
MultilevelOrder(&ctrl, mgraph, morder, sizes, &wspace);
MALLOC_CHECK(NULL);
/* Invert the ordering back to the original graph */
ProjectInfoBack(&ctrl, graph, order, morder, &wspace);
MALLOC_CHECK(NULL);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
free(ctrl.tpwgts);
free(morder);
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph, 0);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (*numflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, order, npes, mype, 0);
MALLOC_CHECK(NULL);
}
/***********************************************************************************
* This function is the entry point of the parallel kmetis algorithm that uses
* coordinates to compute an initial graph distribution.
************************************************************************************/
void ParMETIS_V3_PartGeomKway(idxtype *vtxdist, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, int *wgtflag, int *numflag, int *ndims,
float *xyz, int *ncon, int *nparts, float *tpwgts, float *ubvec,
int *options, int *edgecut, idxtype *part, MPI_Comm *comm)
{
int h, i, j;
int nvtxs = -1, npes, mype;
int uwgtflag, cut, gcut, maxnvtxs;
int ltvwgts[MAXNCON];
int moptions[10];
CtrlType ctrl;
idxtype *uvwgt;
WorkSpaceType wspace;
GraphType *graph, *mgraph;
float avg, maximb, balance, *mytpwgts;
int seed, dbglvl = 0;
int iwgtflag, inumflag, incon, inparts, ioptions[10];
float *itpwgts, iubvec[MAXNCON];
MPI_Comm_size(*comm, &npes);
MPI_Comm_rank(*comm, &mype);
/********************************/
/* Try and take care bad inputs */
/********************************/
if (options != NULL && options[0] == 1)
dbglvl = options[PMV3_OPTION_DBGLVL];
CheckInputs(STATIC_PARTITION, npes, dbglvl, wgtflag, &iwgtflag, numflag, &inumflag,
ncon, &incon, nparts, &inparts, tpwgts, &itpwgts, ubvec, iubvec,
NULL, NULL, options, ioptions, part, comm);
/*********************************/
/* Take care the nparts = 1 case */
/*********************************/
if (inparts <= 1) {
idxset(vtxdist[mype+1]-vtxdist[mype], 0, part);
*edgecut = 0;
return;
}
/******************************/
/* Take care of npes = 1 case */
/******************************/
if (npes == 1 && inparts > 1) {
moptions[0] = 0;
nvtxs = vtxdist[1];
if (incon == 1) {
METIS_WPartGraphKway(&nvtxs, xadj, adjncy, vwgt, adjwgt, &iwgtflag, &inumflag,
&inparts, itpwgts, moptions, edgecut, part);
}
else {
/* ADD: this is because METIS does not support tpwgts for all constraints */
mytpwgts = fmalloc(inparts, "mytpwgts");
for (i=0; i<inparts; i++)
mytpwgts[i] = itpwgts[i*incon];
moptions[7] = -1;
METIS_mCPartGraphRecursive2(&nvtxs, &incon, xadj, adjncy, vwgt, adjwgt, &iwgtflag,
&inumflag, &inparts, mytpwgts, moptions, edgecut, part);
free(mytpwgts);
}
return;
}
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
/*****************************/
/* Set up control structures */
/*****************************/
if (ioptions[0] == 1) {
dbglvl = ioptions[PMV3_OPTION_DBGLVL];
seed = ioptions[PMV3_OPTION_SEED];
}
else {
dbglvl = GLOBAL_DBGLVL;
seed = GLOBAL_SEED;
}
SetUpCtrl(&ctrl, npes, dbglvl, *comm);
ctrl.CoarsenTo = amin(vtxdist[npes]+1, 25*incon*amax(npes, inparts));
ctrl.seed = (seed == 0) ? mype : seed*mype;
ctrl.sync = GlobalSEMax(&ctrl, seed);
ctrl.partType = STATIC_PARTITION;
ctrl.ps_relation = -1;
ctrl.tpwgts = itpwgts;
scopy(incon, iubvec, ctrl.ubvec);
uwgtflag = iwgtflag|2;
uvwgt = idxsmalloc(vtxdist[mype+1]-vtxdist[mype], 1, "uvwgt");
graph = Moc_SetUpGraph(&ctrl, 1, vtxdist, xadj, uvwgt, adjncy, adjwgt, &uwgtflag);
free(graph->nvwgt); graph->nvwgt = NULL;
PreAllocateMemory(&ctrl, graph, &wspace);
/*=================================================================
* Compute the initial npes-way partitioning geometric partitioning
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
Coordinate_Partition(&ctrl, graph, *ndims, xyz, 1, &wspace);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
/*=================================================================
* Move the graph according to the partitioning
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.MoveTmr));
free(uvwgt);
graph->vwgt = ((iwgtflag&2) != 0) ? vwgt : idxsmalloc(graph->nvtxs*incon, 1, "vwgt");
graph->ncon = incon;
j = ctrl.nparts;
ctrl.nparts = ctrl.npes;
mgraph = Moc_MoveGraph(&ctrl, graph, &wspace);
ctrl.nparts = j;
/**********************************************************/
/* Do the same functionality as Moc_SetUpGraph for mgraph */
/**********************************************************/
/* compute tvwgts */
for (j=0; j<incon; j++)
ltvwgts[j] = 0;
for (i=0; i<graph->nvtxs; i++)
for (j=0; j<incon; j++)
ltvwgts[j] += mgraph->vwgt[i*incon+j];
for (j=0; j<incon; j++)
ctrl.tvwgts[j] = GlobalSESum(&ctrl, ltvwgts[j]);
/* check for zero wgt constraints */
for (i=0; i<incon; i++) {
/* ADD: take care of the case in which tvwgts is zero */
if (ctrl.tvwgts[i] == 0) {
if (ctrl.mype == 0) printf("ERROR: sum weight for constraint %d is zero\n", i);
MPI_Finalize();
exit(-1);
}
}
/* compute nvwgt */
mgraph->nvwgt = fmalloc(mgraph->nvtxs*incon, "mgraph->nvwgt");
for (i=0; i<mgraph->nvtxs; i++)
for (j=0; j<incon; j++)
mgraph->nvwgt[i*incon+j] = (float)(mgraph->vwgt[i*incon+j]) / (float)(ctrl.tvwgts[j]);
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.MoveTmr));
if (ctrl.dbglvl&DBG_INFO) {
cut = 0;
for (i=0; i<graph->nvtxs; i++)
for (j=graph->xadj[i]; j<graph->xadj[i+1]; j++)
if (graph->where[i] != graph->where[graph->adjncy[j]])
cut += graph->adjwgt[j];
gcut = GlobalSESum(&ctrl, cut)/2;
maxnvtxs = GlobalSEMax(&ctrl, mgraph->nvtxs);
balance = (float)(maxnvtxs)/((float)(graph->gnvtxs)/(float)(npes));
rprintf(&ctrl, "XYZ Cut: %6d \tBalance: %6.3f [%d %d %d]\n",
gcut, balance, maxnvtxs, graph->gnvtxs, npes);
}
/*=================================================================
* Set up the newly moved graph
=================================================================*/
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
ctrl.nparts = inparts;
FreeWSpace(&wspace);
PreAllocateMemory(&ctrl, mgraph, &wspace);
/*=======================================================
* Now compute the partition of the moved graph
=======================================================*/
if (vtxdist[npes] < SMALLGRAPH || vtxdist[npes] < npes*20 || GlobalSESum(&ctrl, mgraph->nedges) == 0) {
IFSET(ctrl.dbglvl, DBG_INFO, rprintf(&ctrl, "Partitioning a graph of size %d serially\n", vtxdist[npes]));
PartitionSmallGraph(&ctrl, mgraph, &wspace);
}
else {
Moc_Global_Partition(&ctrl, mgraph, &wspace);
}
ParallelReMapGraph(&ctrl, mgraph, &wspace);
/* Invert the ordering back to the original graph */
ctrl.nparts = npes;
ProjectInfoBack(&ctrl, graph, part, mgraph->where, &wspace);
*edgecut = mgraph->mincut;
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
/*******************/
/* Print out stats */
/*******************/
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimingInfo(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, MPI_Barrier(ctrl.gcomm));
if (ctrl.dbglvl&DBG_INFO) {
rprintf(&ctrl, "Final %d-way CUT: %6d \tBalance: ", inparts, mgraph->mincut);
avg = 0.0;
for (h=0; h<incon; h++) {
maximb = 0.0;
for (i=0; i<inparts; i++)
maximb = amax(maximb, mgraph->gnpwgts[i*incon+h]/itpwgts[i*incon+h]);
avg += maximb;
rprintf(&ctrl, "%.3f ", maximb);
}
rprintf(&ctrl, " avg: %.3f\n", avg/(float)incon);
}
GKfree((void **)&itpwgts, LTERM);
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph, iwgtflag);
FreeWSpace(&wspace);
FreeCtrl(&ctrl);
if (inumflag == 1)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
}
/*************************************************************************
* This function creates the coarser graph
**************************************************************************/
void CreateCoarseGraphNoMask(CtrlType *ctrl, GraphType *graph, int cnvtxs, idxtype *match, idxtype *perm)
{
int i, j, k, m, istart, iend, nvtxs, nedges, ncon, cnedges, v, u, dovsize;
idxtype *xadj, *vwgt, *vsize, *adjncy, *adjwgt, *adjwgtsum, *auxadj;
idxtype *cmap, *htable;
idxtype *cxadj, *cvwgt, *cvsize, *cadjncy, *cadjwgt, *cadjwgtsum;
float *nvwgt, *cnvwgt;
GraphType *cgraph;
dovsize = (ctrl->optype == OP_KVMETIS ? 1 : 0);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
vwgt = graph->vwgt;
vsize = graph->vsize;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
adjwgtsum = graph->adjwgtsum;
cmap = graph->cmap;
/* Initialize the coarser graph */
cgraph = SetUpCoarseGraph(graph, cnvtxs, dovsize);
cxadj = cgraph->xadj;
cvwgt = cgraph->vwgt;
cvsize = cgraph->vsize;
cnvwgt = cgraph->nvwgt;
cadjwgtsum = cgraph->adjwgtsum;
cadjncy = cgraph->adjncy;
cadjwgt = cgraph->adjwgt;
htable = idxset(cnvtxs, -1, idxwspacemalloc(ctrl, cnvtxs));
iend = xadj[nvtxs];
auxadj = ctrl->wspace.auxcore;
memcpy(auxadj, adjncy, iend*sizeof(idxtype));
for (i=0; i<iend; i++)
auxadj[i] = cmap[auxadj[i]];
cxadj[0] = cnvtxs = cnedges = 0;
for (i=0; i<nvtxs; i++) {
v = perm[i];
if (cmap[v] != cnvtxs)
continue;
u = match[v];
if (ncon == 1)
cvwgt[cnvtxs] = vwgt[v];
else
scopy(ncon, nvwgt+v*ncon, cnvwgt+cnvtxs*ncon);
if (dovsize)
cvsize[cnvtxs] = vsize[v];
cadjwgtsum[cnvtxs] = adjwgtsum[v];
nedges = 0;
istart = xadj[v];
iend = xadj[v+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
if ((m = htable[k]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = adjwgt[j];
htable[k] = nedges++;
}
else {
cadjwgt[m] += adjwgt[j];
}
}
if (v != u) {
if (ncon == 1)
cvwgt[cnvtxs] += vwgt[u];
else
saxpy(ncon, 1.0, nvwgt+u*ncon, 1, cnvwgt+cnvtxs*ncon, 1);
if (dovsize)
cvsize[cnvtxs] += vsize[u];
cadjwgtsum[cnvtxs] += adjwgtsum[u];
istart = xadj[u];
iend = xadj[u+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
if ((m = htable[k]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = adjwgt[j];
htable[k] = nedges++;
}
else {
cadjwgt[m] += adjwgt[j];
}
}
/* Remove the contracted adjacency weight */
if ((j = htable[cnvtxs]) != -1) {
ASSERT(cadjncy[j] == cnvtxs);
cadjwgtsum[cnvtxs] -= cadjwgt[j];
cadjncy[j] = cadjncy[--nedges];
cadjwgt[j] = cadjwgt[nedges];
htable[cnvtxs] = -1;
}
}
ASSERTP(cadjwgtsum[cnvtxs] == idxsum(nedges, cadjwgt), ("%d %d\n", cadjwgtsum[cnvtxs], idxsum(nedges, cadjwgt)));
for (j=0; j<nedges; j++)
htable[cadjncy[j]] = -1; /* Zero out the htable */
cnedges += nedges;
cxadj[++cnvtxs] = cnedges;
cadjncy += nedges;
cadjwgt += nedges;
}
cgraph->nedges = cnedges;
ReAdjustMemory(graph, cgraph, dovsize);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
idxwspacefree(ctrl, cnvtxs);
}
/*************************************************************************
* This function performs k-way refinement
**************************************************************************/
void Moc_KWayFM(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace, int npasses)
{
int h, i, ii, iii, j, k, c;
int pass, nvtxs, nedges, ncon;
int nmoves, nmoved, nswaps, nzgswaps;
/* int gnswaps, gnzgswaps; */
int me, firstvtx, lastvtx, yourlastvtx;
int from, to = -1, oldto, oldcut, mydomain, yourdomain, imbalanced, overweight;
int npes = ctrl->npes, mype = ctrl->mype, nparts = ctrl->nparts;
int nlupd, nsupd, nnbrs, nchanged;
idxtype *xadj, *ladjncy, *adjwgt, *vtxdist;
idxtype *where, *tmp_where, *moved;
floattype *lnpwgts, *gnpwgts, *ognpwgts, *pgnpwgts, *movewgts, *overfill;
idxtype *update, *supdate, *rupdate, *pe_updates;
idxtype *changed, *perm, *pperm, *htable;
idxtype *peind, *recvptr, *sendptr;
KeyValueType *swchanges, *rwchanges;
RInfoType *rinfo, *myrinfo, *tmp_myrinfo, *tmp_rinfo;
EdgeType *tmp_edegrees, *my_edegrees, *your_edegrees;
floattype lbvec[MAXNCON], *nvwgt, *badmaxpwgt, *ubvec, *tpwgts, lbavg, ubavg;
int *nupds_pe;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->KWayTmr));
/*************************/
/* set up common aliases */
/*************************/
nvtxs = graph->nvtxs;
nedges = graph->nedges;
ncon = graph->ncon;
vtxdist = graph->vtxdist;
xadj = graph->xadj;
ladjncy = graph->adjncy;
adjwgt = graph->adjwgt;
firstvtx = vtxdist[mype];
lastvtx = vtxdist[mype+1];
where = graph->where;
rinfo = graph->rinfo;
lnpwgts = graph->lnpwgts;
gnpwgts = graph->gnpwgts;
ubvec = ctrl->ubvec;
tpwgts = ctrl->tpwgts;
nnbrs = graph->nnbrs;
peind = graph->peind;
recvptr = graph->recvptr;
sendptr = graph->sendptr;
changed = idxmalloc(nvtxs, "KWR: changed");
rwchanges = wspace->pairs;
swchanges = rwchanges + recvptr[nnbrs];
/************************************/
/* set up important data structures */
/************************************/
perm = idxmalloc(nvtxs, "KWR: perm");
pperm = idxmalloc(nparts, "KWR: pperm");
update = idxmalloc(nvtxs, "KWR: update");
supdate = wspace->indices;
rupdate = supdate + recvptr[nnbrs];
nupds_pe = imalloc(npes, "KWR: nupds_pe");
htable = idxsmalloc(nvtxs+graph->nrecv, 0, "KWR: lhtable");
badmaxpwgt = fmalloc(nparts*ncon, "badmaxpwgt");
for (i=0; i<nparts; i++) {
for (h=0; h<ncon; h++) {
badmaxpwgt[i*ncon+h] = ubvec[h]*tpwgts[i*ncon+h];
}
}
movewgts = fmalloc(nparts*ncon, "KWR: movewgts");
ognpwgts = fmalloc(nparts*ncon, "KWR: ognpwgts");
pgnpwgts = fmalloc(nparts*ncon, "KWR: pgnpwgts");
overfill = fmalloc(nparts*ncon, "KWR: overfill");
moved = idxmalloc(nvtxs, "KWR: moved");
tmp_where = idxmalloc(nvtxs+graph->nrecv, "KWR: tmp_where");
tmp_rinfo = (RInfoType *)GKmalloc(sizeof(RInfoType)*nvtxs, "KWR: tmp_rinfo");
tmp_edegrees = (EdgeType *)GKmalloc(sizeof(EdgeType)*nedges, "KWR: tmp_edegrees");
idxcopy(nvtxs+graph->nrecv, where, tmp_where);
for (i=0; i<nvtxs; i++) {
tmp_rinfo[i].id = rinfo[i].id;
tmp_rinfo[i].ed = rinfo[i].ed;
tmp_rinfo[i].ndegrees = rinfo[i].ndegrees;
tmp_rinfo[i].degrees = tmp_edegrees+xadj[i];
for (j=0; j<rinfo[i].ndegrees; j++) {
tmp_rinfo[i].degrees[j].edge = rinfo[i].degrees[j].edge;
tmp_rinfo[i].degrees[j].ewgt = rinfo[i].degrees[j].ewgt;
}
}
nswaps = nzgswaps = 0;
/*********************************************************/
/* perform a small number of passes through the vertices */
/*********************************************************/
for (pass=0; pass<npasses; pass++) {
if (mype == 0)
RandomPermute(nparts, pperm, 1);
MPI_Bcast((void *)pperm, nparts, IDX_DATATYPE, 0, ctrl->comm);
FastRandomPermute(nvtxs, perm, 1);
oldcut = graph->mincut;
/* check to see if the partitioning is imbalanced */
Moc_ComputeParallelBalance(ctrl, graph, graph->where, lbvec);
ubavg = savg(ncon, ubvec);
lbavg = savg(ncon, lbvec);
imbalanced = (lbavg > ubavg) ? 1 : 0;
for (c=0; c<2; c++) {
scopy(ncon*nparts, gnpwgts, ognpwgts);
sset(ncon*nparts, 0.0, movewgts);
nmoved = 0;
/**********************************************/
/* PASS ONE -- record stats for desired moves */
/**********************************************/
for (iii=0; iii<nvtxs; iii++) {
i = perm[iii];
from = tmp_where[i];
nvwgt = graph->nvwgt+i*ncon;
for (h=0; h<ncon; h++)
if (fabs(nvwgt[h]-gnpwgts[from*ncon+h]) < SMALLFLOAT)
break;
if (h < ncon) {
continue;
}
/* check for a potential improvement */
if (tmp_rinfo[i].ed >= tmp_rinfo[i].id) {
my_edegrees = tmp_rinfo[i].degrees;
for (k=0; k<tmp_rinfo[i].ndegrees; k++) {
to = my_edegrees[k].edge;
if (ProperSide(c, pperm[from], pperm[to])) {
for (h=0; h<ncon; h++)
if (gnpwgts[to*ncon+h]+nvwgt[h] > badmaxpwgt[to*ncon+h] && nvwgt[h] > 0.0)
break;
if (h == ncon)
break;
}
}
oldto = to;
/* check if a subdomain was found that fits */
if (k < tmp_rinfo[i].ndegrees) {
for (j=k+1; j<tmp_rinfo[i].ndegrees; j++) {
to = my_edegrees[j].edge;
if (ProperSide(c, pperm[from], pperm[to])) {
for (h=0; h<ncon; h++)
if (gnpwgts[to*ncon+h]+nvwgt[h] > badmaxpwgt[to*ncon+h] && nvwgt[h] > 0.0)
break;
if (h == ncon) {
if (my_edegrees[j].ewgt > my_edegrees[k].ewgt ||
(my_edegrees[j].ewgt == my_edegrees[k].ewgt &&
IsHBalanceBetterTT(ncon,gnpwgts+oldto*ncon,gnpwgts+to*ncon,nvwgt,ubvec))){
k = j;
oldto = my_edegrees[k].edge;
}
}
}
}
to = oldto;
if (my_edegrees[k].ewgt > tmp_rinfo[i].id ||
(my_edegrees[k].ewgt == tmp_rinfo[i].id &&
(imbalanced || graph->level > 3 || iii % 8 == 0) &&
IsHBalanceBetterFT(ncon,gnpwgts+from*ncon,gnpwgts+to*ncon,nvwgt,ubvec))){
/****************************************/
/* Update tmp arrays of the moved vertex */
/****************************************/
tmp_where[i] = to;
moved[nmoved++] = i;
for (h=0; h<ncon; h++) {
lnpwgts[to*ncon+h] += nvwgt[h];
lnpwgts[from*ncon+h] -= nvwgt[h];
gnpwgts[to*ncon+h] += nvwgt[h];
gnpwgts[from*ncon+h] -= nvwgt[h];
movewgts[to*ncon+h] += nvwgt[h];
movewgts[from*ncon+h] -= nvwgt[h];
}
tmp_rinfo[i].ed += tmp_rinfo[i].id-my_edegrees[k].ewgt;
SWAP(tmp_rinfo[i].id, my_edegrees[k].ewgt, j);
if (my_edegrees[k].ewgt == 0) {
tmp_rinfo[i].ndegrees--;
my_edegrees[k].edge = my_edegrees[tmp_rinfo[i].ndegrees].edge;
my_edegrees[k].ewgt = my_edegrees[tmp_rinfo[i].ndegrees].ewgt;
}
else {
my_edegrees[k].edge = from;
}
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
/* no need to bother about vertices on different pe's */
if (ladjncy[j] >= nvtxs)
continue;
me = ladjncy[j];
mydomain = tmp_where[me];
myrinfo = tmp_rinfo+me;
your_edegrees = myrinfo->degrees;
if (mydomain == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
}
else {
if (mydomain == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
}
}
/* Remove contribution from the .ed of 'from' */
if (mydomain != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == from) {
if (your_edegrees[k].ewgt == adjwgt[j]) {
myrinfo->ndegrees--;
your_edegrees[k].edge = your_edegrees[myrinfo->ndegrees].edge;
your_edegrees[k].ewgt = your_edegrees[myrinfo->ndegrees].ewgt;
}
else {
your_edegrees[k].ewgt -= adjwgt[j];
}
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (mydomain != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == to) {
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
your_edegrees[myrinfo->ndegrees].edge = to;
your_edegrees[myrinfo->ndegrees++].ewgt = adjwgt[j];
}
}
}
}
}
}
}
/******************************************/
/* Let processors know the subdomain wgts */
/* if all proposed moves commit. */
/******************************************/
MPI_Allreduce((void *)lnpwgts, (void *)pgnpwgts, nparts*ncon,
MPI_DOUBLE, MPI_SUM, ctrl->comm);
/**************************/
/* compute overfill array */
/**************************/
overweight = 0;
for (j=0; j<nparts; j++) {
for (h=0; h<ncon; h++) {
if (pgnpwgts[j*ncon+h] > ognpwgts[j*ncon+h]) {
overfill[j*ncon+h] =
(pgnpwgts[j*ncon+h]-badmaxpwgt[j*ncon+h]) /
(pgnpwgts[j*ncon+h]-ognpwgts[j*ncon+h]);
}
else {
overfill[j*ncon+h] = 0.0;
}
overfill[j*ncon+h] = amax(overfill[j*ncon+h], 0.0);
overfill[j*ncon+h] *= movewgts[j*ncon+h];
if (overfill[j*ncon+h] > 0.0)
overweight = 1;
ASSERTP(ctrl, ognpwgts[j*ncon+h] <= badmaxpwgt[j*ncon+h] ||
pgnpwgts[j*ncon+h] <= ognpwgts[j*ncon+h],
(ctrl, "%.4f %.4f %.4f\n", ognpwgts[j*ncon+h],
badmaxpwgt[j*ncon+h], pgnpwgts[j*ncon+h]));
}
}
/****************************************************/
/* select moves to undo according to overfill array */
/****************************************************/
if (overweight == 1) {
for (iii=0; iii<nmoved; iii++) {
i = moved[iii];
oldto = tmp_where[i];
nvwgt = graph->nvwgt+i*ncon;
my_edegrees = tmp_rinfo[i].degrees;
for (k=0; k<tmp_rinfo[i].ndegrees; k++)
if (my_edegrees[k].edge == where[i])
break;
for (h=0; h<ncon; h++)
if (nvwgt[h] > 0.0 && overfill[oldto*ncon+h] > nvwgt[h]/4.0)
break;
/**********************************/
/* nullify this move if necessary */
/**********************************/
if (k != tmp_rinfo[i].ndegrees && h != ncon) {
moved[iii] = -1;
from = oldto;
to = where[i];
for (h=0; h<ncon; h++) {
overfill[oldto*ncon+h] = amax(overfill[oldto*ncon+h]-nvwgt[h], 0.0);
}
tmp_where[i] = to;
tmp_rinfo[i].ed += tmp_rinfo[i].id-my_edegrees[k].ewgt;
SWAP(tmp_rinfo[i].id, my_edegrees[k].ewgt, j);
if (my_edegrees[k].ewgt == 0) {
tmp_rinfo[i].ndegrees--;
my_edegrees[k].edge = my_edegrees[tmp_rinfo[i].ndegrees].edge;
my_edegrees[k].ewgt = my_edegrees[tmp_rinfo[i].ndegrees].ewgt;
}
else {
my_edegrees[k].edge = from;
}
for (h=0; h<ncon; h++) {
lnpwgts[to*ncon+h] += nvwgt[h];
lnpwgts[from*ncon+h] -= nvwgt[h];
}
/* Update the degrees of adjacent vertices */
for (j=xadj[i]; j<xadj[i+1]; j++) {
/* no need to bother about vertices on different pe's */
if (ladjncy[j] >= nvtxs)
continue;
me = ladjncy[j];
mydomain = tmp_where[me];
myrinfo = tmp_rinfo+me;
your_edegrees = myrinfo->degrees;
if (mydomain == from) {
INC_DEC(myrinfo->ed, myrinfo->id, adjwgt[j]);
}
else {
if (mydomain == to) {
INC_DEC(myrinfo->id, myrinfo->ed, adjwgt[j]);
}
}
/* Remove contribution from the .ed of 'from' */
if (mydomain != from) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == from) {
if (your_edegrees[k].ewgt == adjwgt[j]) {
myrinfo->ndegrees--;
your_edegrees[k].edge = your_edegrees[myrinfo->ndegrees].edge;
your_edegrees[k].ewgt = your_edegrees[myrinfo->ndegrees].ewgt;
}
else {
your_edegrees[k].ewgt -= adjwgt[j];
}
break;
}
}
}
/* Add contribution to the .ed of 'to' */
if (mydomain != to) {
for (k=0; k<myrinfo->ndegrees; k++) {
if (your_edegrees[k].edge == to) {
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
your_edegrees[myrinfo->ndegrees].edge = to;
your_edegrees[myrinfo->ndegrees++].ewgt = adjwgt[j];
}
}
}
}
}
}
/*************************************************/
/* PASS TWO -- commit the remainder of the moves */
/*************************************************/
nlupd = nsupd = nmoves = nchanged = 0;
for (iii=0; iii<nmoved; iii++) {
i = moved[iii];
if (i == -1)
continue;
where[i] = tmp_where[i];
/* Make sure to update the vertex information */
if (htable[i] == 0) {
/* make sure you do the update */
htable[i] = 1;
update[nlupd++] = i;
}
/* Put the vertices adjacent to i into the update array */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = ladjncy[j];
if (htable[k] == 0) {
htable[k] = 1;
if (k<nvtxs)
update[nlupd++] = k;
else
supdate[nsupd++] = k;
}
}
nmoves++;
nswaps++;
/* check number of zero-gain moves */
for (k=0; k<rinfo[i].ndegrees; k++)
if (rinfo[i].degrees[k].edge == to)
break;
if (rinfo[i].id == rinfo[i].degrees[k].ewgt)
nzgswaps++;
if (graph->pexadj[i+1]-graph->pexadj[i] > 0)
changed[nchanged++] = i;
}
/* Tell interested pe's the new where[] info for the interface vertices */
CommChangedInterfaceData(ctrl, graph, nchanged, changed, where,
swchanges, rwchanges, wspace->pv4);
IFSET(ctrl->dbglvl, DBG_RMOVEINFO,
rprintf(ctrl, "\t[%d %d], [%.4f], [%d %d %d]\n",
pass, c, badmaxpwgt[0],
GlobalSESum(ctrl, nmoves),
GlobalSESum(ctrl, nsupd),
GlobalSESum(ctrl, nlupd)));
/*-------------------------------------------------------------
/ Time to communicate with processors to send the vertices
/ whose degrees need to be update.
/-------------------------------------------------------------*/
/* Issue the receives first */
for (i=0; i<nnbrs; i++) {
MPI_Irecv((void *)(rupdate+sendptr[i]), sendptr[i+1]-sendptr[i], IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. This needs some preporcessing */
for (i=0; i<nsupd; i++) {
htable[supdate[i]] = 0;
supdate[i] = graph->imap[supdate[i]];
}
iidxsort(nsupd, supdate);
for (j=i=0; i<nnbrs; i++) {
yourlastvtx = vtxdist[peind[i]+1];
for (k=j; k<nsupd && supdate[k] < yourlastvtx; k++);
MPI_Isend((void *)(supdate+j), k-j, IDX_DATATYPE, peind[i], 1, ctrl->comm, ctrl->sreq+i);
j = k;
}
/* OK, now get into the loop waiting for the send/recv operations to finish */
MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++)
MPI_Get_count(ctrl->statuses+i, IDX_DATATYPE, nupds_pe+i);
MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/*-------------------------------------------------------------
/ Place the recieved to-be updated vertices into update[]
/-------------------------------------------------------------*/
for (i=0; i<nnbrs; i++) {
pe_updates = rupdate+sendptr[i];
for (j=0; j<nupds_pe[i]; j++) {
k = pe_updates[j];
if (htable[k-firstvtx] == 0) {
htable[k-firstvtx] = 1;
update[nlupd++] = k-firstvtx;
}
}
}
/*-------------------------------------------------------------
/ Update the rinfo of the vertices in the update[] array
/-------------------------------------------------------------*/
for (ii=0; ii<nlupd; ii++) {
i = update[ii];
ASSERT(ctrl, htable[i] == 1);
htable[i] = 0;
mydomain = where[i];
myrinfo = rinfo+i;
tmp_myrinfo = tmp_rinfo+i;
my_edegrees = myrinfo->degrees;
your_edegrees = tmp_myrinfo->degrees;
graph->lmincut -= myrinfo->ed;
myrinfo->ndegrees = 0;
myrinfo->id = 0;
myrinfo->ed = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
yourdomain = where[ladjncy[j]];
if (mydomain != yourdomain) {
myrinfo->ed += adjwgt[j];
for (k=0; k<myrinfo->ndegrees; k++) {
if (my_edegrees[k].edge == yourdomain) {
my_edegrees[k].ewgt += adjwgt[j];
your_edegrees[k].ewgt += adjwgt[j];
break;
}
}
if (k == myrinfo->ndegrees) {
my_edegrees[k].edge = yourdomain;
my_edegrees[k].ewgt = adjwgt[j];
your_edegrees[k].edge = yourdomain;
your_edegrees[k].ewgt = adjwgt[j];
myrinfo->ndegrees++;
}
ASSERT(ctrl, myrinfo->ndegrees <= xadj[i+1]-xadj[i]);
ASSERT(ctrl, tmp_myrinfo->ndegrees <= xadj[i+1]-xadj[i]);
}
else {
myrinfo->id += adjwgt[j];
}
}
graph->lmincut += myrinfo->ed;
tmp_myrinfo->id = myrinfo->id;
tmp_myrinfo->ed = myrinfo->ed;
tmp_myrinfo->ndegrees = myrinfo->ndegrees;
}
/* finally, sum-up the partition weights */
MPI_Allreduce((void *)lnpwgts, (void *)gnpwgts, nparts*ncon,
MPI_DOUBLE, MPI_SUM, ctrl->comm);
}
graph->mincut = GlobalSESum(ctrl, graph->lmincut)/2;
if (graph->mincut == oldcut)
break;
}
/*
gnswaps = GlobalSESum(ctrl, nswaps);
gnzgswaps = GlobalSESum(ctrl, nzgswaps);
if (mype == 0)
printf("niters: %d, nswaps: %d, nzgswaps: %d\n", pass+1, gnswaps, gnzgswaps);
*/
GKfree((void **)&badmaxpwgt, (void **)&update, (void **)&nupds_pe, (void **)&htable, LTERM);
GKfree((void **)&changed, (void **)&pperm, (void **)&perm, (void **)&moved, LTERM);
GKfree((void **)&pgnpwgts, (void **)&ognpwgts, (void **)&overfill, (void **)&movewgts, LTERM);
GKfree((void **)&tmp_where, (void **)&tmp_rinfo, (void **)&tmp_edegrees, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->KWayTmr));
}
int main(int argc, char *argv[])
{
int i, npart;
idxtype *part;
float ncut=0;
GraphType graph;
char filename[256],outputFile[256];
int wgtflag = 0, addSelfLoop=1, outputFileGiven=0, txtFormat=0 ;
int randomInit = 0;
idxtype minEdgeWeight = 0;
Options opt;
timer TOTALTmr, METISTmr, IOTmr;
initOptions(&opt);
if (argc < 2) {
print_help(argv[0]);
exit(0);
}
for (argv++; *argv != NULL; argv++){
if ((*argv)[0] == '-')
{
int temp;
switch ((*argv)[1])
{
case 'b':
case 'B':
opt.penalty_power=atof(*(++argv));
break;
case 'i':
case 'I':
opt.gamma=atof(*(++argv));
break;
case 'o':
case 'O':
strcpy(outputFile,*(++argv));
outputFileGiven=1;
break;
case 'D'://quality threshold. This is a post-processing step proposed in SR-MCL. If you dont want post-processing (this is what original MLR-MCL, R-MCL, MCL do, please set "-d 0"
case 'd':
opt.quality_threshold = atof(*(++argv));
break;
case 'w':
case 'W':
opt.weighted_density = true;
break;
case 'c':
case 'C':
opt.coarsenTo= atoi(*(++argv));
break;
default:
printf("Invalid option %s\n", *argv);
print_help(argv[0]);
exit(0);
}
}
else
{
strcpy(filename, *argv);
}
}
if ( randomInit > 0 )
InitRandom(time(NULL));
else
InitRandom(-1);
cleartimer(TOTALTmr);
cleartimer(METISTmr);
cleartimer(IOTmr);
starttimer(TOTALTmr);
starttimer(IOTmr);
ReadGraph(&graph, filename, &wgtflag, addSelfLoop, txtFormat);
if ( opt.matchType == MATCH_UNSPECIFIED )
{
// opt.matchType = (graph.nvtxs>50000) ? MATCH_POWERLAW_FC :
// MATCH_SHEMN;
opt.matchType = MATCH_SHEMN;
}
stoptimer(IOTmr);
if (graph.nvtxs <= 0) {
printf("Empty graph. Nothing to do.\n");
exit(0);
}
int noOfSingletons = 0;
GraphType *noSingletonGraph ;
idxtype* nodeMap = lookForSingletons(&graph, &noOfSingletons);
if ( noOfSingletons > 0 )
{
getSubgraph(&graph, nodeMap, graph.nvtxs-noOfSingletons,
wgtflag, &noSingletonGraph);
GKfree((void**)&(graph.xadj), (void**)&(graph.adjncy), LTERM);
if ( wgtflag&1 > 0 )
GKfree( (void**)&(graph.adjwgt), LTERM);
// free(graph.gdata);
printf("Found %d singleton nodes in the", noOfSingletons);
printf(" input graph. Removing them.\n");
}
if ( !outputFileGiven )
{
strcpy(outputFile, filename);
sprintf(outputFile,"%s.c%d.i%1.1f.b%1.1f",outputFile,opt.coarsenTo,opt.gamma,opt.penalty_power);
}
printf("Input graph information ---------------------------------------------------\n");
printf(" Name: %s, #Vertices: %d, #Edges: %d\n", filename, graph.nvtxs, graph.nedges/2);
printf("Output shall be placed in the file %s\n",
outputFile);
fflush(stdout);
part = idxmalloc(graph.nvtxs, "main: part");
printf("------------------------------------------------\n");
printf("Clustering....\n");
fflush(stdout);
starttimer(METISTmr); //YK: main algorithm starts here!
if ( noOfSingletons > 0 )
{
mlmcl(&(noSingletonGraph->nvtxs), noSingletonGraph->xadj, noSingletonGraph->adjncy,
noSingletonGraph->vwgt,noSingletonGraph->adjwgt, &wgtflag, part, opt );
}
else
{
mlmcl(&graph.nvtxs, graph.xadj, graph.adjncy,graph.vwgt,
graph.adjwgt, &wgtflag, part, opt );
}
stoptimer(METISTmr);
printf("------------------------------------------------\n");
if ( noOfSingletons > 0 )
{
npart=mapPartition(part,noSingletonGraph->nvtxs);
ncut=ComputeNCut(noSingletonGraph, part,npart);
// printf("In graph that does not include singletons,");
// printf("No. of Clusters: %d, N-Cut value: %.2f\n",npart,ncut);
idxtype *clusterSizes = histogram(part,
graph.nvtxs-noOfSingletons, npart);
int maxSize = clusterSizes[idxamax(npart, clusterSizes)];
float avgClusterSize =
(graph.nvtxs-noOfSingletons)*1.0/(npart);
float balance = (maxSize*1.0) /
((graph.nvtxs-noOfSingletons)*1.0/npart);
float stdDevn = stdDeviation(clusterSizes, npart);
float avgNcut = ncut * 1.0/npart;
float normStdDevn = stdDevn/avgClusterSize;
// Warning: This computation only works if the singletons
// have been placed in their own clusters. This works for
// MLR-MCL, in other words, because it is guaranteed to
// place singletons in their own clusters.
printf("Output statistics for graph without singletons\n");
printf("Clusters: %d N-Cut: %.3f",
npart, ncut);
printf(" AvgN-Cut: %.3f Balance in cluster sizes: %.2f ",avgNcut,
balance);
printf("Std_Deviation in cluster sizes: %.2f ", stdDevn);
printf("Coefficient_of_Variation: %.2f\n", normStdDevn);
free( clusterSizes );
npart += noOfSingletons;
// ncut += noOfSingletons;
printf("Output statistics for original graph\n");
mapIndices(part, nodeMap, graph.nvtxs, npart-noOfSingletons);
}
else
{
npart=mapPartition(part,graph.nvtxs);
ncut=ComputeNCut(&graph, part,npart);
}
idxtype* clusterSizes = histogram(part, graph.nvtxs, npart);
int maxSize = clusterSizes[idxamax(npart, clusterSizes)];
float avgClusterSize = (graph.nvtxs)*1.0/(npart);
float balance = (maxSize*1.0)/(graph.nvtxs*1.0/npart);
float stdDevn = stdDeviation(clusterSizes, npart);
float avgNcut = ncut * 1.0/npart;
float normStdDevn = stdDevn/avgClusterSize;
printf("Clusters: %d N-Cut: %.3f AvgN-Cut: %.3f", npart,
ncut, avgNcut );
printf(" Balance in cluster sizes: %.2f Std.Deviation in cluster sizes: %.2f ",
balance, stdDevn);
printf("Coefficient_of_Variation: %.2f\n", normStdDevn);
starttimer(IOTmr);
my_WritePartition(outputFile, part, graph.nvtxs, opt.gamma);
if ( noOfSingletons > 0 )
{
free(nodeMap);
nodeMap = NULL;
}
printf("\nOutput is written to file: %s\n", outputFile);
stoptimer(IOTmr);
stoptimer(TOTALTmr);
printf("\nTiming Information --------------------------------------------------\n");
printf(" I/O: \t\t %7.3f\n", gettimer(IOTmr));
printf(" Partitioning: \t\t %7.3f (MLR-MCL time)\n", gettimer(METISTmr));
printf(" Total: \t\t %7.3f\n", gettimer(TOTALTmr));
printf("**********************************************************************\n");
GKfree((void**)&graph.xadj, (void**)&graph.adjncy, (void**)&graph.vwgt,
(void**)&graph.adjwgt, (void**)&part, LTERM);
}
/*************************************************************************
* Let the game begin
**************************************************************************/
main(int argc, char *argv[])
{
int i, options[10];
idxtype *perm, *iperm;
GraphType graph;
char filename[256];
int numflag = 0, wgtflag;
timer TOTALTmr, METISTmr, IOTmr, SMBTmr;
if (argc != 2) {
printf("Usage: %s <GraphFile>\n",argv[0]);
exit(0);
}
strcpy(filename, argv[1]);
cleartimer(TOTALTmr);
cleartimer(METISTmr);
cleartimer(IOTmr);
cleartimer(SMBTmr);
starttimer(TOTALTmr);
starttimer(IOTmr);
ReadGraph(&graph, filename, &wgtflag);
if (graph.nvtxs <= 0) {
printf("Empty graph. Nothing to do.\n");
exit(0);
}
if (graph.ncon != 1) {
printf("Ordering can only be applied to graphs with one constraint.\n");
exit(0);
}
stoptimer(IOTmr);
/* Ordering does not use weights! */
GKfree(&graph.vwgt, &graph.adjwgt, LTERM);
printf("**********************************************************************\n");
printf("%s", METISTITLE);
printf("Graph Information ---------------------------------------------------\n");
printf(" Name: %s, #Vertices: %d, #Edges: %d\n\n", filename, graph.nvtxs, graph.nedges/2);
printf("Node-Based Ordering... ----------------------------------------------\n");
perm = idxmalloc(graph.nvtxs, "main: perm");
iperm = idxmalloc(graph.nvtxs, "main: iperm");
options[0] = 0;
starttimer(METISTmr);
METIS_NodeND(&graph.nvtxs, graph.xadj, graph.adjncy, &numflag, options, perm, iperm);
stoptimer(METISTmr);
starttimer(IOTmr);
WritePermutation(filename, iperm, graph.nvtxs);
stoptimer(IOTmr);
starttimer(SMBTmr);
ComputeFillIn(&graph, iperm);
stoptimer(SMBTmr);
stoptimer(TOTALTmr);
printf("\nTiming Information --------------------------------------------------\n");
printf(" I/O: \t %7.3f\n", gettimer(IOTmr));
printf(" Ordering: \t %7.3f (ONMETIS time)\n", gettimer(METISTmr));
printf(" Symbolic Factorization: \t %7.3f\n", gettimer(SMBTmr));
printf(" Total: \t %7.3f\n", gettimer(TOTALTmr));
printf("**********************************************************************\n");
GKfree(&graph.xadj, &graph.adjncy, &perm, &iperm, LTERM);
}
/*************************************************************************
* 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));
}
/*************************************************************************
* Let the game begin
**************************************************************************/
main(int argc, char *argv[])
{
int i, j, ne, nn, etype, numflag=0, nparts, edgecut;
idxtype *elmnts, *epart, *npart;
timer IOTmr, DUALTmr;
char etypestr[4][5] = {"TRI", "TET", "HEX", "QUAD"};
GraphType graph;
if (argc != 3) {
printf("Usage: %s <meshfile> <nparts>\n",argv[0]);
exit(0);
}
nparts = atoi(argv[2]);
if (nparts < 2) {
printf("nparts must be greater than one.\n");
exit(0);
}
cleartimer(IOTmr);
cleartimer(DUALTmr);
starttimer(IOTmr);
elmnts = ReadMesh(argv[1], &ne, &nn, &etype);
stoptimer(IOTmr);
epart = idxmalloc(ne, "main: epart");
npart = idxmalloc(nn, "main: npart");
printf("**********************************************************************\n");
printf("%s", METISTITLE);
printf("Mesh Information ----------------------------------------------------\n");
printf(" Name: %s, #Elements: %d, #Nodes: %d, Etype: %s\n\n", argv[1], ne, nn, etypestr[etype-1]);
printf("Partitioning Nodal Graph... -----------------------------------------\n");
starttimer(DUALTmr);
METIS_PartMeshNodal(&ne, &nn, elmnts, &etype, &numflag, &nparts, &edgecut, epart, npart);
stoptimer(DUALTmr);
printf(" %d-way Edge-Cut: %7d, Balance: %5.2f\n", nparts, edgecut, ComputeElementBalance(ne, nparts, epart));
starttimer(IOTmr);
WriteMeshPartition(argv[1], nparts, ne, epart, nn, npart);
stoptimer(IOTmr);
printf("\nTiming Information --------------------------------------------------\n");
printf(" I/O: \t\t %7.3f\n", gettimer(IOTmr));
printf(" Partitioning: \t\t %7.3f\n", gettimer(DUALTmr));
printf("**********************************************************************\n");
/*
graph.nvtxs = ne;
graph.xadj = idxmalloc(ne+1, "xadj");
graph.vwgt = idxsmalloc(ne, 1, "vwgt");
graph.adjncy = idxmalloc(10*ne, "adjncy");
graph.adjwgt = idxsmalloc(10*ne, 1, "adjncy");
METIS_MeshToDual(&ne, &nn, elmnts, &etype, &numflag, graph.xadj, graph.adjncy);
ComputePartitionInfo(&graph, nparts, epart);
GKfree(&graph.xadj, &graph.adjncy, &graph.vwgt, &graph.adjwgt, LTERM);
*/
GKfree(&elmnts, &epart, &npart, LTERM);
}
/*************************************************************************
* This function computes the initial id/ed
**************************************************************************/
void ComputeKWayVolGains(CtrlType *ctrl, GraphType *graph, int nparts)
{
int i, ii, j, k, kk, l, nvtxs, me, other, pid, myndegrees;
idxtype *xadj, *vsize, *adjncy, *adjwgt, *where, *bndind, *bndptr, *ophtable;
VRInfoType *rinfo, *myrinfo, *orinfo;
VEDegreeType *myedegrees, *oedegrees;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vsize = graph->vsize;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
bndind = graph->bndind;
bndptr = idxset(nvtxs, -1, graph->bndptr);
rinfo = graph->vrinfo;
starttimer(ctrl->AuxTmr2);
ophtable = idxset(nparts, -1, idxwspacemalloc(ctrl, nparts));
/*------------------------------------------------------------
/ Compute now the iv/ev degrees
/------------------------------------------------------------*/
graph->minvol = graph->nbnd = 0;
for (i=0; i<nvtxs; i++) {
myrinfo = rinfo+i;
myrinfo->gv = -MAXIDX;
if (myrinfo->ndegrees > 0) {
me = where[i];
myedegrees = myrinfo->edegrees;
myndegrees = myrinfo->ndegrees;
graph->minvol += myndegrees*vsize[i];
for (j=xadj[i]; j<xadj[i+1]; j++) {
ii = adjncy[j];
other = where[ii];
orinfo = rinfo+ii;
oedegrees = orinfo->edegrees;
for (k=0; k<orinfo->ndegrees; k++)
ophtable[oedegrees[k].pid] = k;
ophtable[other] = 1; /* this is to simplify coding */
if (me == other) {
/* Find which domains 'i' is connected and 'ii' is not and update their gain */
for (k=0; k<myndegrees; k++) {
if (ophtable[myedegrees[k].pid] == -1)
myedegrees[k].gv -= vsize[ii];
}
}
else {
ASSERT(ophtable[me] != -1);
if (oedegrees[ophtable[me]].ned == 1) { /* I'm the only connection of 'ii' in 'me' */
/* Increase the gains for all the common domains between 'i' and 'ii' */
for (k=0; k<myndegrees; k++) {
if (ophtable[myedegrees[k].pid] != -1)
myedegrees[k].gv += vsize[ii];
}
}
else {
/* Find which domains 'i' is connected and 'ii' is not and update their gain */
for (k=0; k<myndegrees; k++) {
if (ophtable[myedegrees[k].pid] == -1)
myedegrees[k].gv -= vsize[ii];
}
}
}
for (kk=0; kk<orinfo->ndegrees; kk++)
ophtable[oedegrees[kk].pid] = -1;
ophtable[other] = -1;
}
/* Compute the max vgain */
for (k=0; k<myndegrees; k++) {
if (myedegrees[k].gv > myrinfo->gv)
myrinfo->gv = myedegrees[k].gv;
}
}
if (myrinfo->ed > 0 && myrinfo->id == 0)
myrinfo->gv += vsize[i];
if (myrinfo->gv >= 0 || myrinfo->ed-myrinfo->id >= 0)
BNDInsert(graph->nbnd, bndind, bndptr, i);
}
stoptimer(ctrl->AuxTmr2);
idxwspacefree(ctrl, nparts);
}
/*************************************************************************
* This function takes a graph and creates a sequence of coarser graphs
**************************************************************************/
GraphType *Coarsen2Way(CtrlType *ctrl, GraphType *graph)
{
int clevel;
GraphType *cgraph;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->CoarsenTmr));
cgraph = graph;
/* The following is ahack to allow the multiple bisections to go through with correct
coarsening */
if (ctrl->CType > 20) {
clevel = 1;
ctrl->CType -= 20;
}
else
clevel = 0;
do {
IFSET(ctrl->dbglvl, DBG_COARSEN, printf("%6d %7d [%d] [%d %d]\n",
cgraph->nvtxs, cgraph->nedges, ctrl->CoarsenTo, ctrl->maxvwgt,
(cgraph->vwgt ? idxsum(cgraph->nvtxs, cgraph->vwgt) : cgraph->nvtxs)));
if (cgraph->adjwgt) {
switch (ctrl->CType) {
case MATCH_RM:
Match_RM(ctrl, cgraph);
break;
case MATCH_HEM:
if (clevel < 1)
Match_RM(ctrl, cgraph);
else
Match_HEM(ctrl, cgraph);
break;
case MATCH_SHEM:
if (clevel < 1)
Match_RM(ctrl, cgraph);
else
Match_SHEM(ctrl, cgraph);
break;
case MATCH_SHEMKWAY:
Match_SHEM(ctrl, cgraph);
break;
default:
errexit("Unknown CType: %d\n", ctrl->CType);
}
}
else {
Match_RM_NVW(ctrl, cgraph);
}
cgraph = cgraph->coarser;
clevel++;
} while (cgraph->nvtxs > ctrl->CoarsenTo && cgraph->nvtxs < COARSEN_FRACTION2*cgraph->finer->nvtxs && cgraph->nedges > cgraph->nvtxs/2);
IFSET(ctrl->dbglvl, DBG_COARSEN, printf("%6d %7d [%d] [%d %d]\n",
cgraph->nvtxs, cgraph->nedges, ctrl->CoarsenTo, ctrl->maxvwgt,
(cgraph->vwgt ? idxsum(cgraph->nvtxs, cgraph->vwgt) : cgraph->nvtxs)));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->CoarsenTmr));
return cgraph;
}
/*************************************************************************
* This function is the entry point of refinement
**************************************************************************/
void RefineVolKWay(CtrlType *ctrl, GraphType *orggraph, GraphType *graph, int nparts,
float *tpwgts, float ubfactor)
{
int i, nlevels;
GraphType *ptr;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->UncoarsenTmr));
/* Take care any non-contiguity */
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->AuxTmr1));
if (ctrl->RType == RTYPE_KWAYRANDOM_MCONN) {
ComputeVolKWayPartitionParams(ctrl, graph, nparts);
EliminateVolComponents(ctrl, graph, nparts, tpwgts, 1.25);
EliminateVolSubDomainEdges(ctrl, graph, nparts, tpwgts);
EliminateVolComponents(ctrl, graph, nparts, tpwgts, 1.25);
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->AuxTmr1));
/* Determine how many levels are there */
for (ptr=graph, nlevels=0; ptr!=orggraph; ptr=ptr->finer, nlevels++);
/* Compute the parameters of the coarsest graph */
ComputeVolKWayPartitionParams(ctrl, graph, nparts);
for (i=0; ;i++) {
/*PrintSubDomainGraph(graph, nparts, graph->where);*/
MALLOC_CHECK(NULL);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->RefTmr));
if (2*i >= nlevels && !IsBalanced(graph->pwgts, nparts, tpwgts, 1.04*ubfactor)) {
ComputeVolKWayBalanceBoundary(ctrl, graph, nparts);
switch (ctrl->RType) {
case RTYPE_KWAYRANDOM:
Greedy_KWayVolBalance(ctrl, graph, nparts, tpwgts, ubfactor, 1);
break;
case RTYPE_KWAYRANDOM_MCONN:
Greedy_KWayVolBalanceMConn(ctrl, graph, nparts, tpwgts, ubfactor, 1);
break;
}
ComputeVolKWayBoundary(ctrl, graph, nparts);
}
switch (ctrl->RType) {
case RTYPE_KWAYRANDOM:
Random_KWayVolRefine(ctrl, graph, nparts, tpwgts, ubfactor, 10, 0);
break;
case RTYPE_KWAYRANDOM_MCONN:
Random_KWayVolRefineMConn(ctrl, graph, nparts, tpwgts, ubfactor, 10, 0);
break;
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->RefTmr));
if (graph == orggraph)
break;
GKfree(&graph->gdata, LTERM); /* Deallocate the graph related arrays */
graph = graph->finer;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ProjectTmr));
ProjectVolKWayPartition(ctrl, graph, nparts);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ProjectTmr));
}
if (!IsBalanced(graph->pwgts, nparts, tpwgts, ubfactor)) {
ComputeVolKWayBalanceBoundary(ctrl, graph, nparts);
switch (ctrl->RType) {
case RTYPE_KWAYRANDOM:
Greedy_KWayVolBalance(ctrl, graph, nparts, tpwgts, ubfactor, 8);
Random_KWayVolRefine(ctrl, graph, nparts, tpwgts, ubfactor, 10, 0);
break;
case RTYPE_KWAYRANDOM_MCONN:
Greedy_KWayVolBalanceMConn(ctrl, graph, nparts, tpwgts, ubfactor, 8);
Random_KWayVolRefineMConn(ctrl, graph, nparts, tpwgts, ubfactor, 10, 0);
break;
}
}
EliminateVolComponents(ctrl, graph, nparts, tpwgts, ubfactor);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->UncoarsenTmr));
}
/*************************************************************************
* This function finds a matching using the HEM heuristic
**************************************************************************/
void Match_SHEM(CtrlType *ctrl, GraphType *graph)
{
int i, ii, j, k, nvtxs, cnvtxs, maxidx, maxwgt, avgdegree;
idxtype *xadj, *vwgt, *adjncy, *adjwgt;
idxtype *match, *cmap, *degrees, *perm, *tperm;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
cmap = graph->cmap;
match = idxset(nvtxs, UNMATCHED, idxwspacemalloc(ctrl, nvtxs));
perm = idxwspacemalloc(ctrl, nvtxs);
tperm = idxwspacemalloc(ctrl, nvtxs);
degrees = idxwspacemalloc(ctrl, nvtxs);
RandomPermute(nvtxs, tperm, 1);
avgdegree = 0.7*(xadj[nvtxs]/nvtxs);
for (i=0; i<nvtxs; i++)
degrees[i] = (xadj[i+1]-xadj[i] > avgdegree ? avgdegree : xadj[i+1]-xadj[i]);
BucketSortKeysInc(nvtxs, avgdegree, degrees, tperm, perm);
cnvtxs = 0;
/* Take care any islands. Islands are matched with non-islands due to coarsening */
for (ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
if (xadj[i] < xadj[i+1])
break;
maxidx = i;
for (j=nvtxs-1; j>ii; j--) {
k = perm[j];
if (match[k] == UNMATCHED && xadj[k] < xadj[k+1]) {
maxidx = k;
break;
}
}
cmap[i] = cmap[maxidx] = cnvtxs++;
match[i] = maxidx;
match[maxidx] = i;
}
}
/* Continue with normal matching */
for (; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
maxidx = i;
maxwgt = 0;
/* Find a heavy-edge matching, subject to maxvwgt constraints */
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (match[adjncy[j]] == UNMATCHED && maxwgt < adjwgt[j] && vwgt[i]+vwgt[adjncy[j]] <= ctrl->maxvwgt) {
maxwgt = adjwgt[j];
maxidx = adjncy[j];
}
}
cmap[i] = cmap[maxidx] = cnvtxs++;
match[i] = maxidx;
match[maxidx] = i;
}
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
idxwspacefree(ctrl, nvtxs); /* degrees */
idxwspacefree(ctrl, nvtxs); /* tperm */
CreateCoarseGraph(ctrl, graph, cnvtxs, match, perm);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
void MLKKMRefine(CtrlType *ctrl, GraphType *orggraph, GraphType *graph, int nparts, int chain_length, float *tpwgts, float ubfactor)
{
int i, nlevels, mustfree=0, temp_cl;
GraphType *ptr;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->UncoarsenTmr));
/* Compute the parameters of the coarsest graph */
ComputeKWayPartitionParams(ctrl, graph, nparts);
temp_cl = chain_length;
/* Determine how many levels are there */
for (ptr=graph, nlevels=0; ptr!=orggraph; ptr=ptr->finer, nlevels++);
//printf("Number of levels is %d\n", nlevels);
for (i=0; ;i++) {
timer tmr;
float result;
cleartimer(tmr);
starttimer(tmr);
//pingpong(ctrl, graph, nparts, chain_length, tpwgts, ubfactor);
//chain_length /= 1.5;
//printf("Level: %d\n", i+1);
if (graph == orggraph){
//chain_length = chain_length>0 ? chain_length : 1;
pingpong(ctrl, graph, nparts, chain_length, tpwgts, ubfactor, 1);
break;
}
else{
//pingpong(ctrl, graph, nparts, 0, tpwgts, ubfactor, 0);
pingpong(ctrl, graph, nparts, chain_length, tpwgts, ubfactor, 0);
//chain_length /= 2;
}
//pingpong(ctrl, graph, nparts, chain_length, tpwgts, ubfactor);
// /* for time and quality each level
stoptimer(tmr);
//printf("Level %d: %7.3f", i+1, tmr);
if (cutType == NCUT){
result = ComputeNCut(graph, graph->where, nparts);
//printf(" %7f", result);
}
else{
result = ComputeRAsso(graph, graph->where, nparts);
//printf(" %7f", result);
}
//printf(" (%d)\n\n", graph->nvtxs);
//ends here*/
if (graph == orggraph)
break;
/*
if(i>1)
chain_length /= 10;
*/
GKfree((void **) &graph->gdata, LTERM); /* Deallocate the graph related arrays */
graph = graph->finer;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ProjectTmr));
if (graph->vwgt == NULL) {
graph->vwgt = idxsmalloc(graph->nvtxs, 1, "RefineKWay: graph->vwgt");
graph->adjwgt = idxsmalloc(graph->nedges, 1, "RefineKWay: graph->adjwgt");
mustfree = 1;
}
ProjectKWayPartition(ctrl, graph, nparts);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ProjectTmr));
}
if (mustfree)
GKfree((void **) &graph->vwgt, (void **) &graph->adjwgt, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->UncoarsenTmr));
}
/*************************************************************************
* 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 InitPartition(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, j, ncon, mype, npes, gnvtxs, ngroups;
idx_t *xadj, *adjncy, *adjwgt, *vwgt;
idx_t *part, *gwhere0, *gwhere1;
idx_t *tmpwhere, *tmpvwgt, *tmpxadj, *tmpadjncy, *tmpadjwgt;
graph_t *agraph;
idx_t lnparts, fpart, fpe, lnpes;
idx_t twoparts=2, moptions[METIS_NOPTIONS], edgecut, max_cut;
real_t *tpwgts, *tpwgts2, *lbvec, lbsum, min_lbsum, wsum;
MPI_Comm ipcomm;
struct {
double sum;
int rank;
} lpesum, gpesum;
WCOREPUSH;
ncon = graph->ncon;
ngroups = gk_max(gk_min(RIP_SPLIT_FACTOR, ctrl->npes), 1);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
lbvec = rwspacemalloc(ctrl, ncon);
/* assemble the graph to all the processors */
agraph = AssembleAdaptiveGraph(ctrl, graph);
gnvtxs = agraph->nvtxs;
/* make a copy of the graph's structure for later */
xadj = icopy(gnvtxs+1, agraph->xadj, iwspacemalloc(ctrl, gnvtxs+1));
vwgt = icopy(gnvtxs*ncon, agraph->vwgt, iwspacemalloc(ctrl, gnvtxs*ncon));
adjncy = icopy(agraph->nedges, agraph->adjncy, iwspacemalloc(ctrl, agraph->nedges));
adjwgt = icopy(agraph->nedges, agraph->adjwgt, iwspacemalloc(ctrl, agraph->nedges));
part = iwspacemalloc(ctrl, gnvtxs);
/* create different processor groups */
gkMPI_Comm_split(ctrl->gcomm, ctrl->mype % ngroups, 0, &ipcomm);
gkMPI_Comm_rank(ipcomm, &mype);
gkMPI_Comm_size(ipcomm, &npes);
/* Go into the recursive bisection */
METIS_SetDefaultOptions(moptions);
moptions[METIS_OPTION_SEED] = ctrl->sync + (ctrl->mype % ngroups) + 1;
tpwgts = ctrl->tpwgts;
tpwgts2 = rwspacemalloc(ctrl, 2*ncon);
lnparts = ctrl->nparts;
fpart = fpe = 0;
lnpes = npes;
while (lnpes > 1 && lnparts > 1) {
/* determine the weights of the two partitions as a function of the
weight of the target partition weights */
for (j=(lnparts>>1), i=0; i<ncon; i++) {
tpwgts2[i] = rsum(j, tpwgts+fpart*ncon+i, ncon);
tpwgts2[ncon+i] = rsum(lnparts-j, tpwgts+(fpart+j)*ncon+i, ncon);
wsum = 1.0/(tpwgts2[i] + tpwgts2[ncon+i]);
tpwgts2[i] *= wsum;
tpwgts2[ncon+i] *= wsum;
}
METIS_PartGraphRecursive(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, NULL, agraph->adjwgt, &twoparts, tpwgts2, NULL, moptions,
&edgecut, part);
/* pick one of the branches */
if (mype < fpe+lnpes/2) {
KeepPart(ctrl, agraph, part, 0);
lnpes = lnpes/2;
lnparts = lnparts/2;
}
else {
KeepPart(ctrl, agraph, part, 1);
fpart = fpart + lnparts/2;
fpe = fpe + lnpes/2;
lnpes = lnpes - lnpes/2;
lnparts = lnparts - lnparts/2;
}
}
gwhere0 = iset(gnvtxs, 0, iwspacemalloc(ctrl, gnvtxs));
gwhere1 = iwspacemalloc(ctrl, gnvtxs);
if (lnparts == 1) { /* Case npes is greater than or equal to nparts */
/* 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;
}
}
else { /* Case in which npes is smaller than nparts */
/* create the normalized tpwgts for the lnparts from ctrl->tpwgts */
tpwgts = rwspacemalloc(ctrl, lnparts*ncon);
for (j=0; j<ncon; j++) {
for (wsum=0.0, i=0; i<lnparts; i++) {
tpwgts[i*ncon+j] = ctrl->tpwgts[(fpart+i)*ncon+j];
wsum += tpwgts[i*ncon+j];
}
for (wsum=1.0/wsum, i=0; i<lnparts; i++)
tpwgts[i*ncon+j] *= wsum;
}
METIS_PartGraphKway(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, NULL, agraph->adjwgt, &lnparts, tpwgts, NULL, moptions,
&edgecut, part);
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart + part[i];
}
gkMPI_Allreduce((void *)gwhere0, (void *)gwhere1, gnvtxs, IDX_T, 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;
edgecut = ComputeSerialEdgeCut(agraph);
ComputeSerialBalance(ctrl, agraph, gwhere1, lbvec);
lbsum = rsum(ncon, lbvec, 1);
gkMPI_Allreduce((void *)&edgecut, (void *)&max_cut, 1, IDX_T, MPI_MAX, ctrl->gcomm);
gkMPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, REAL_T, MPI_MIN, ctrl->gcomm);
lpesum.sum = lbsum;
if (min_lbsum < UNBALANCE_FRACTION*ncon) {
if (lbsum < UNBALANCE_FRACTION*ncon)
lpesum.sum = edgecut;
else
lpesum.sum = max_cut;
}
lpesum.rank = ctrl->mype;
gkMPI_Allreduce((void *)&lpesum, (void *)&gpesum, 1, MPI_DOUBLE_INT,
MPI_MINLOC, ctrl->gcomm);
gkMPI_Bcast((void *)gwhere1, gnvtxs, IDX_T, gpesum.rank, ctrl->gcomm);
agraph->xadj = tmpxadj;
agraph->adjncy = tmpadjncy;
agraph->adjwgt = tmpadjwgt;
agraph->vwgt = tmpvwgt;
agraph->where = tmpwhere;
}
icopy(graph->nvtxs, gwhere1+graph->vtxdist[ctrl->mype], graph->where);
FreeGraph(agraph);
gkMPI_Comm_free(&ipcomm);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
WCOREPOP;
}
void KWayNodeRefine(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace,
int npasses, float ubfraction)
{
int i, ii, iii, j, k, pass, nvtxs, firstvtx, lastvtx, otherlastvtx, c, nmoves,
nlupd, nsupd, nnbrs, nchanged, nsep;
int npes = ctrl->npes, mype = ctrl->mype, nparts = ctrl->nparts;
idxtype *xadj, *adjncy, *adjwgt, *vtxdist, *vwgt;
idxtype *where, *lpwgts, *gpwgts, *sepind;
idxtype *peind, *recvptr, *sendptr;
idxtype *update, *supdate, *rupdate, *pe_updates, *htable, *changed;
idxtype *badminpwgt, *badmaxpwgt;
KeyValueType *swchanges, *rwchanges;
int *nupds_pe;
NRInfoType *rinfo, *myrinfo;
int from, to, me, other, oldcut;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->KWayTmr));
nvtxs = graph->nvtxs;
vtxdist = graph->vtxdist;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
vwgt = graph->vwgt;
firstvtx = vtxdist[mype];
lastvtx = vtxdist[mype+1];
where = graph->where;
rinfo = graph->nrinfo;
lpwgts = graph->lpwgts;
gpwgts = graph->gpwgts;
nsep = graph->nsep;
sepind = graph->sepind;
nnbrs = graph->nnbrs;
peind = graph->peind;
recvptr = graph->recvptr;
sendptr = graph->sendptr;
changed = idxmalloc(nvtxs, "KWayRefine: changed");
rwchanges = wspace->pairs;
swchanges = rwchanges + recvptr[nnbrs];
update = idxmalloc(nvtxs, "KWayRefine: update");
supdate = wspace->indices;
rupdate = supdate + recvptr[nnbrs];
nupds_pe = imalloc(npes, "KWayRefine: nupds_pe");
htable = idxsmalloc(nvtxs+graph->nrecv, 0, "KWayRefine: lhtable");
badminpwgt = wspace->pv1;
badmaxpwgt = wspace->pv2;
for (i=0; i<nparts; i+=2) {
badminpwgt[i] = badminpwgt[i+1] = (1.0/ubfraction)*(gpwgts[i]+gpwgts[i+1])/2;
badmaxpwgt[i] = badmaxpwgt[i+1] = ubfraction*(gpwgts[i]+gpwgts[i+1])/2;
}
//myprintf(ctrl, "%6d %6d %6d %6d %6d %6d %6d\n", lpwgts[0], lpwgts[1], lpwgts[2], gpwgts[0], gpwgts[1], gpwgts[2], badmaxpwgt[0]);
IFSET(ctrl->dbglvl, DBG_REFINEINFO,
PrintNodeBalanceInfo(ctrl, nparts, gpwgts, badminpwgt, badmaxpwgt, 1));
for (pass=0; pass<npasses; pass++) {
oldcut = graph->mincut;
for (c=0; c<2; c++) {
for (i=0; i<nparts; i+=2) {
badminpwgt[i] = badminpwgt[i+1] = (1.0/ubfraction)*(gpwgts[i]+gpwgts[i+1])/2;
badmaxpwgt[i] = badmaxpwgt[i+1] = ubfraction*(gpwgts[i]+gpwgts[i+1])/2;
}
nlupd = nsupd = nmoves = nchanged = 0;
for (ii=0; ii<nsep; ii++) {
i = sepind[ii];
from = where[i];
ASSERT(ctrl, from >= nparts);
/* Go through the loop to see if gain is possible for the separator vertex */
if (rinfo[i].edegrees[(c+1)%2] <= vwgt[i]) {
/* It is a one-sded move so it will go to the other partition.
Look at the comments in InitMultisection to understand the meaning
of from%nparts */
to = from%nparts+c;
if (gpwgts[to]+vwgt[i] > badmaxpwgt[to]) {
/* printf("Skip because of weight! %d\n", vwgt[i]-rinfo[i].edegrees[(c+1)%2]); */
continue; /* We cannot move it there because it gets too heavy */
}
/* Update the where information of the vertex you moved */
where[i] = to;
/* Remove this vertex from the sepind. Note the trick for looking at
the sepind[ii] again */
sepind[ii--] = sepind[--nsep];
/* myprintf(ctrl, "Vertex %d [%d %d] is moving to %d from %d [%d]\n",
i+firstvtx, vwgt[i], rinfo[i].edegrees[(c+1)%2], to, from, where[i]); */
lpwgts[from] -= vwgt[i];
lpwgts[2*nparts-1] -= vwgt[i];
lpwgts[to] += vwgt[i];
gpwgts[to] += vwgt[i];
/* Put the vertices adjacent to i that belong to either the separator or
the (c+1)%2 partition into the update array */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (htable[k] == 0 && where[k] != to) {
htable[k] = 1;
if (k<nvtxs)
update[nlupd++] = k;
else
supdate[nsupd++] = k;
}
}
nmoves++;
if (graph->pexadj[i+1]-graph->pexadj[i] > 0)
changed[nchanged++] = i;
}
}
/* myprintf(ctrl, "nmoves: %d, nlupd: %d, nsupd: %d\n", nmoves, nlupd, nsupd); */
/* Tell everybody interested what the new where[] info is for the interface vertices */
CommChangedInterfaceData(ctrl, graph, nchanged, changed, where, swchanges,
rwchanges, wspace->pv4);
IFSET(ctrl->dbglvl, DBG_RMOVEINFO, rprintf(ctrl, "\t[%d %d], [%d %d %d]\n",
pass, c, GlobalSESum(ctrl, nmoves), GlobalSESum(ctrl, nsupd),
GlobalSESum(ctrl, nlupd)));
/*-----------------------------------------------------------------------
/ Time to communicate with processors to send the vertices whose degrees
/ need to be updated.
/-----------------------------------------------------------------------*/
/* Issue the receives first */
for (i=0; i<nnbrs; i++) {
MPI_Irecv((void *)(rupdate+sendptr[i]), sendptr[i+1]-sendptr[i], IDX_DATATYPE,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. This needs some preporcessing */
for (i=0; i<nsupd; i++) {
htable[supdate[i]] = 0;
supdate[i] = graph->imap[supdate[i]];
}
iidxsort(nsupd, supdate);
for (j=i=0; i<nnbrs; i++) {
otherlastvtx = vtxdist[peind[i]+1];
for (k=j; k<nsupd && supdate[k] < otherlastvtx; k++);
MPI_Isend((void *)(supdate+j), k-j, IDX_DATATYPE, peind[i], 1, ctrl->comm,
ctrl->sreq+i);
j = k;
}
/* OK, now get into the loop waiting for the send/recv operations to finish */
MPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++)
MPI_Get_count(ctrl->statuses+i, IDX_DATATYPE, nupds_pe+i);
MPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/*-------------------------------------------------------------
/ Place the received to-be updated vertices into update[]
/-------------------------------------------------------------*/
for (i=0; i<nnbrs; i++) {
pe_updates = rupdate+sendptr[i];
for (j=0; j<nupds_pe[i]; j++) {
k = pe_updates[j];
if (htable[k-firstvtx] == 0) {
htable[k-firstvtx] = 1;
update[nlupd++] = k-firstvtx;
}
}
}
/*-------------------------------------------------------------
/ Update the where information of the vertices that are pulled
/ into the separator.
/-------------------------------------------------------------*/
nchanged = 0;
for (ii=0; ii<nlupd; ii++) {
i = update[ii];
me = where[i];
if (me < nparts && me%2 == (c+1)%2) { /* This vertex is pulled into the separator */
lpwgts[me] -= vwgt[i];
where[i] = nparts+me-(me%2);
sepind[nsep++] = i; /* Put the vertex into the sepind array */
if (graph->pexadj[i+1]-graph->pexadj[i] > 0)
changed[nchanged++] = i;
lpwgts[where[i]] += vwgt[i];
lpwgts[2*nparts-1] += vwgt[i];
/* myprintf(ctrl, "Vertex %d moves into the separator from %d to %d\n",
i+firstvtx, me, where[i]); */
}
}
/* Tell everybody interested what the new where[] info is for the interface vertices */
CommChangedInterfaceData(ctrl, graph, nchanged, changed, where, swchanges,
rwchanges, wspace->pv4);
/*-------------------------------------------------------------
/ Update the rinfo of the vertices in the update[] array
/-------------------------------------------------------------*/
for (ii=0; ii<nlupd; ii++) {
i = update[ii];
ASSERT(ctrl, htable[i] == 1);
htable[i] = 0;
me = where[i];
if (me >= nparts) { /* If it is a separator vertex */
/* myprintf(ctrl, "Updating %d %d\n", i+firstvtx, me); */
myrinfo = rinfo+i;
myrinfo->edegrees[0] = myrinfo->edegrees[1] = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
other = where[adjncy[j]];
if (me != other)
myrinfo->edegrees[other%2] += vwgt[adjncy[j]];
}
}
}
/* Finally, sum-up the partition weights */
MPI_Allreduce((void *)lpwgts, (void *)gpwgts, 2*nparts, IDX_DATATYPE, MPI_SUM,
ctrl->comm);
graph->mincut = gpwgts[2*nparts-1];
IFSET(ctrl->dbglvl, DBG_REFINEINFO, PrintNodeBalanceInfo(ctrl, nparts, gpwgts,
badminpwgt, badmaxpwgt, 0));
}
if (graph->mincut == oldcut)
break;
}
GKfree((void **)&update, &nupds_pe, &htable, &changed, LTERM);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->KWayTmr));
}
/*************************************************************************
* This function creates the coarser graph
**************************************************************************/
void CreateCoarseGraph_NVW(CtrlType *ctrl, GraphType *graph, int cnvtxs, idxtype *match, idxtype *perm)
{
int i, j, jj, k, kk, l, m, istart, iend, nvtxs, nedges, ncon, cnedges, v, u, mask;
idxtype *xadj, *adjncy, *adjwgtsum, *auxadj;
idxtype *cmap, *htable;
idxtype *cxadj, *cvwgt, *cadjncy, *cadjwgt, *cadjwgtsum;
float *nvwgt, *cnvwgt;
GraphType *cgraph;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgtsum = graph->adjwgtsum;
cmap = graph->cmap;
/* Initialize the coarser graph */
cgraph = SetUpCoarseGraph(graph, cnvtxs, 0);
cxadj = cgraph->xadj;
cvwgt = cgraph->vwgt;
cnvwgt = cgraph->nvwgt;
cadjwgtsum = cgraph->adjwgtsum;
cadjncy = cgraph->adjncy;
cadjwgt = cgraph->adjwgt;
iend = xadj[nvtxs];
auxadj = ctrl->wspace.auxcore;
memcpy(auxadj, adjncy, iend*sizeof(idxtype));
for (i=0; i<iend; i++)
auxadj[i] = cmap[auxadj[i]];
mask = HTLENGTH;
htable = idxset(mask+1, -1, idxwspacemalloc(ctrl, mask+1));
cxadj[0] = cnvtxs = cnedges = 0;
for (i=0; i<nvtxs; i++) {
v = perm[i];
if (cmap[v] != cnvtxs)
continue;
u = match[v];
cvwgt[cnvtxs] = 1;
cadjwgtsum[cnvtxs] = adjwgtsum[v];
nedges = 0;
istart = xadj[v];
iend = xadj[v+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
kk = k&mask;
if ((m = htable[kk]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = 1;
htable[kk] = nedges++;
}
else if (cadjncy[m] == k) {
cadjwgt[m]++;
}
else {
for (jj=0; jj<nedges; jj++) {
if (cadjncy[jj] == k) {
cadjwgt[jj]++;
break;
}
}
if (jj == nedges) {
cadjncy[nedges] = k;
cadjwgt[nedges++] = 1;
}
}
}
if (v != u) {
cvwgt[cnvtxs]++;
cadjwgtsum[cnvtxs] += adjwgtsum[u];
istart = xadj[u];
iend = xadj[u+1];
for (j=istart; j<iend; j++) {
k = auxadj[j];
kk = k&mask;
if ((m = htable[kk]) == -1) {
cadjncy[nedges] = k;
cadjwgt[nedges] = 1;
htable[kk] = nedges++;
}
else if (cadjncy[m] == k) {
cadjwgt[m]++;
}
else {
for (jj=0; jj<nedges; jj++) {
if (cadjncy[jj] == k) {
cadjwgt[jj]++;
break;
}
}
if (jj == nedges) {
cadjncy[nedges] = k;
cadjwgt[nedges++] = 1;
}
}
}
/* Remove the contracted adjacency weight */
jj = htable[cnvtxs&mask];
if (jj >= 0 && cadjncy[jj] != cnvtxs) {
for (jj=0; jj<nedges; jj++) {
if (cadjncy[jj] == cnvtxs)
break;
}
}
if (jj >= 0 && cadjncy[jj] == cnvtxs) { /* This 2nd check is needed for non-adjacent matchings */
cadjwgtsum[cnvtxs] -= cadjwgt[jj];
cadjncy[jj] = cadjncy[--nedges];
cadjwgt[jj] = cadjwgt[nedges];
}
}
ASSERTP(cadjwgtsum[cnvtxs] == idxsum(nedges, cadjwgt), ("%d %d %d %d %d\n", cnvtxs, cadjwgtsum[cnvtxs], idxsum(nedges, cadjwgt), adjwgtsum[u], adjwgtsum[v]));
for (j=0; j<nedges; j++)
htable[cadjncy[j]&mask] = -1; /* Zero out the htable */
htable[cnvtxs&mask] = -1;
cnedges += nedges;
cxadj[++cnvtxs] = cnedges;
cadjncy += nedges;
cadjwgt += nedges;
}
cgraph->nedges = cnedges;
ReAdjustMemory(graph, cgraph, 0);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
idxwspacefree(ctrl, mask+1);
}
void ComputeNodePartitionParams(CtrlType *ctrl, GraphType *graph, WorkSpaceType *wspace)
{
int i, j, nparts, nvtxs, nsep, firstvtx, lastvtx;
idxtype *xadj, *adjncy, *adjwgt, *vtxdist, *vwgt, *lpwgts, *gpwgts, *sepind, *hmarker;
idxtype *where;
NRInfoType *rinfo, *myrinfo;
int me, other, otherwgt;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->KWayInitTmr));
nvtxs = graph->nvtxs;
nparts = ctrl->nparts;
vtxdist = graph->vtxdist;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
vwgt = graph->vwgt;
where = graph->where;
rinfo = graph->nrinfo;
lpwgts = graph->lpwgts;
gpwgts = graph->gpwgts;
sepind = graph->sepind;
hmarker = graph->hmarker;
firstvtx = vtxdist[ctrl->mype];
lastvtx = vtxdist[ctrl->mype+1];
/* Reset refinement data structures */
idxset(2*nparts, 0, lpwgts);
idxset(nvtxs, 0, hmarker);
/* Send/Receive the where and vwgt information of interface vertices. */
CommInterfaceData(ctrl, graph, where, wspace->indices, where+nvtxs);
CommInterfaceData(ctrl, graph, vwgt, wspace->indices, vwgt+nvtxs);
/*------------------------------------------------------------
/ Compute now the degrees
/------------------------------------------------------------*/
for (nsep=i=0; i<nvtxs; i++) {
me = where[i];
ASSERT(ctrl, me >= 0 && me < 2*nparts);
lpwgts[me] += vwgt[i];
if (me >= nparts) { /* If it is a separator vertex */
sepind[nsep++] = i;
lpwgts[2*nparts-1] += vwgt[i]; /* Keep track of total separator weight */
myrinfo = rinfo+i;
myrinfo->edegrees[0] = myrinfo->edegrees[1] = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
other = where[adjncy[j]];
if (me != other)
myrinfo->edegrees[other%2] += vwgt[adjncy[j]];
}
}
}
graph->nsep = nsep;
/* Finally, sum-up the partition weights */
MPI_Allreduce((void *)lpwgts, (void *)gpwgts, 2*nparts, IDX_DATATYPE, MPI_SUM, ctrl->comm);
graph->mincut = gpwgts[2*nparts-1];
/* Mark the halo vertices by determining how many non-local vertices they are
* connected to. */
idxset(nvtxs, 0, hmarker);
for (i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
if (adjncy[j] >= nvtxs) {
hmarker[i]++;
}
}
}
#ifdef XX
/* Print Weight information */
if (ctrl->mype == 0) {
for (i=0; i<nparts; i+=2)
printf("[%5d %5d %5d] ", gpwgts[i], gpwgts[i+1], gpwgts[nparts+i]);
printf("\n");
}
#endif
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->KWayInitTmr));
}