/*************************************************************************
* This function is the entry point of refinement
**************************************************************************/
void MocRefineKWayHorizontal(CtrlType *ctrl, GraphType *orggraph, GraphType *graph, int nparts,
float *ubvec)
{
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->UncoarsenTmr));
/* Compute the parameters of the coarsest graph */
MocComputeKWayPartitionParams(ctrl, graph, nparts);
for (;;) {
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->RefTmr));
if (!MocIsHBalanced(graph->ncon, nparts, graph->npwgts, ubvec)) {
MocComputeKWayBalanceBoundary(ctrl, graph, nparts);
MCGreedy_KWayEdgeBalanceHorizontal(ctrl, graph, nparts, ubvec, 4);
ComputeKWayBoundary(ctrl, graph, nparts);
}
MCRandom_KWayEdgeRefineHorizontal(ctrl, graph, nparts, ubvec, 10);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->RefTmr));
if (graph == orggraph)
break;
graph = graph->finer;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ProjectTmr));
MocProjectKWayPartition(ctrl, graph, nparts);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ProjectTmr));
}
if (!MocIsHBalanced(graph->ncon, nparts, graph->npwgts, ubvec)) {
MocComputeKWayBalanceBoundary(ctrl, graph, nparts);
MCGreedy_KWayEdgeBalanceHorizontal(ctrl, graph, nparts, ubvec, 4);
ComputeKWayBoundary(ctrl, graph, nparts);
MCRandom_KWayEdgeRefineHorizontal(ctrl, graph, nparts, ubvec, 10);
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->UncoarsenTmr));
}
/*************************************************************************
* This function is the entry point for PWMETIS that accepts exact weights
* for the target partitions
**************************************************************************/
void METIS_WPartGraphRecursive(int *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *adjwgt, int *wgtflag, int *numflag, int *nparts,
floattype *tpwgts, int *options, int *edgecut, idxtype *part)
{
int i, j;
GraphType graph;
CtrlType ctrl;
floattype *mytpwgts;
if (*numflag == 1)
Change2CNumbering(*nvtxs, xadj, adjncy);
SetUpGraph(&graph, OP_PMETIS, *nvtxs, 1, xadj, adjncy, vwgt, adjwgt, *wgtflag);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = PMETIS_CTYPE;
ctrl.IType = PMETIS_ITYPE;
ctrl.RType = PMETIS_RTYPE;
ctrl.dbglvl = PMETIS_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 = 20;
ctrl.maxvwgt = 1.5*(idxsum(*nvtxs, graph.vwgt)/ctrl.CoarsenTo);
mytpwgts = fmalloc(*nparts, "PWMETIS: mytpwgts");
for (i=0; i<*nparts; i++)
mytpwgts[i] = tpwgts[i];
InitRandom(-1);
AllocateWorkSpace(&ctrl, &graph, *nparts);
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
*edgecut = MlevelRecursiveBisection(&ctrl, &graph, *nparts, part, mytpwgts, 1.000, 0);
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimers(&ctrl));
FreeWorkSpace(&ctrl, &graph);
free(mytpwgts);
if (*numflag == 1)
Change2FNumbering(*nvtxs, xadj, adjncy, part);
}
/*************************************************************************
* This function is the entry point for OEMETIS
**************************************************************************/
void METIS_EdgeND(int *nvtxs, idxtype *xadj, idxtype *adjncy, int *numflag, int *options,
idxtype *perm, idxtype *iperm)
{
int i, j;
GraphType graph;
CtrlType ctrl;
if (*numflag == 1)
Change2CNumbering(*nvtxs, xadj, adjncy);
SetUpGraph(&graph, OP_OEMETIS, *nvtxs, 1, xadj, adjncy, NULL, NULL, 0);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = OEMETIS_CTYPE;
ctrl.IType = OEMETIS_ITYPE;
ctrl.RType = OEMETIS_RTYPE;
ctrl.dbglvl = OEMETIS_DBGLVL;
}
else {
ctrl.CType = options[OPTION_CTYPE];
ctrl.IType = options[OPTION_ITYPE];
ctrl.RType = options[OPTION_RTYPE];
ctrl.dbglvl = options[OPTION_DBGLVL];
}
ctrl.oflags = 0;
ctrl.pfactor = -1;
ctrl.nseps = 1;
ctrl.optype = OP_OEMETIS;
ctrl.CoarsenTo = 20;
ctrl.maxvwgt = 1.5*(idxsum(*nvtxs, graph.vwgt)/ctrl.CoarsenTo);
InitRandom(-1);
AllocateWorkSpace(&ctrl, &graph, 2);
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
MlevelNestedDissection(&ctrl, &graph, iperm, ORDER_UNBALANCE_FRACTION, *nvtxs);
IFSET(ctrl.dbglvl, DBG_TIME, stoptimer(ctrl.TotalTmr));
IFSET(ctrl.dbglvl, DBG_TIME, PrintTimers(&ctrl));
for (i=0; i<*nvtxs; i++)
perm[iperm[i]] = i;
FreeWorkSpace(&ctrl, &graph);
if (*numflag == 1)
Change2FNumberingOrder(*nvtxs, xadj, adjncy, perm, iperm);
}
/* called from layer 3, when a packet arrives for layer 4
In this practical this will always be an ACK as B never sends data.
*/
void A_input(struct pkt packet)
{
int ackcount = 0;
int i;
/* if received ACK is not corrupted */
if (!IsCorrupted(packet)) {
if (TRACE > 0)
printf("----A: uncorrupted ACK %d is received\n",packet.acknum);
total_ACKs_received++;
/* check if new ACK or duplicate */
if (windowcount != 0) {
int seqfirst = buffer[windowfirst].seqnum;
int seqlast = buffer[windowlast].seqnum;
/* check case when seqnum has and hasn't wrapped */
if (((seqfirst <= seqlast) && (packet.acknum >= seqfirst && packet.acknum <= seqlast)) ||
((seqfirst > seqlast) && (packet.acknum >= seqfirst || packet.acknum <= seqlast))) {
/* packet is a new ACK */
if (TRACE > 0)
printf("----A: ACK %d is not a duplicate\n",packet.acknum);
new_ACKs++;
/* cumulative acknowledgement - determine how many packets are ACKed */
if (packet.acknum >= seqfirst)
ackcount = packet.acknum + 1 - seqfirst;
else
ackcount = SEQSPACE - seqfirst + packet.acknum;
/* slide window by the number of packets ACKed */
windowfirst = (windowfirst + ackcount) % WINDOWSIZE;
/* delete the acked packets from window buffer */
for (i=0; i<ackcount; i++)
windowcount--;
/* start timer again if there are still more unacked packets in window */
stoptimer(A);
if (windowcount > 0)
starttimer(A, RTT);
}
}
else
if (TRACE > 0)
printf ("----A: duplicate ACK received, do nothing!\n");
}
else
if (TRACE > 0)
printf ("----A: corrupted ACK is received, do nothing!\n");
}
void main(int argc, char *argv[])
{
FILE *fp;
char infile[96]="in.txt", outfile[96]="out.txt";
int propternary=0;
if(argc>1) strcpy(infile, argv[1]);
if(argc>2) strcpy(outfile,argv[2]);
fp = (FILE*)gfopen(outfile,"w");
puts("results go into out.txt");
starttimer();
DescriptiveStatistics(fp, infile, 50, 100, propternary);
}
/*Initial Sender State*/
void A_Initial_State()
{
//printf("before if loop nextseq %d, base+window %d\n",nextSequenceNum,(base+windowSize));
if(nextSequenceNum<(base+windowSize))
{
//printf("Entered if loop\n");
tolayer3(A,senderBuffer[nextSequenceNum]);
if(base==nextSequenceNum)
starttimer(A,timeout);
nextSequenceNum++;
}
}
/*************************************************************************
* This function finds a matching using the HEM heuristic
**************************************************************************/
void MCMatch_RM(CtrlType *ctrl, GraphType *graph)
{
int i, ii, j, k, nvtxs, ncon, cnvtxs, maxidx;
idxtype *xadj, *adjncy, *adjwgt;
idxtype *match, *cmap, *perm;
float *nvwgt;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
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++) {
k = adjncy[j];
if (match[k] == UNMATCHED && AreAllVwgtsBelowFast(ncon, nvwgt+i*ncon, nvwgt+k*ncon, ctrl->nmaxvwgt)) {
maxidx = k;
break;
}
}
cmap[i] = cmap[maxidx] = cnvtxs++;
match[i] = maxidx;
match[maxidx] = i;
}
}
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
CreateCoarseGraph(ctrl, graph, cnvtxs, match, perm);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/* called when A's timer goes off */
void A_timerinterrupt() //ram's comment - changed the return type to void.
{
//printf("\n in A_timerinterrupt.. Calling startimer");
starttimer(0,TIMEOUT);
timer_A_started=1;
int i;
for(i=base;i<nextseqnum;i++)// send packets in window
{
tolayer3(0,my_packet[i]);
A_transport++;
}
}
/*************************************************************************
* This function finds a matching using the HEM heuristic
**************************************************************************/
void Match_HEM(CtrlType *ctrl, GraphType *graph)
{
int i, ii, j, k, nvtxs, cnvtxs, maxidx, maxwgt;
idxtype *xadj, *vwgt, *adjncy, *adjwgt;
idxtype *match, *cmap, *perm;
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);
RandomPermute(nvtxs, perm, 1);
cnvtxs = 0;
for (ii=0; 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++) {
k = adjncy[j];
if (match[k] == UNMATCHED && maxwgt < adjwgt[j] && vwgt[i]+vwgt[k] <= 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));
CreateCoarseGraph(ctrl, graph, cnvtxs, match, perm);
idxwspacefree(ctrl, nvtxs);
idxwspacefree(ctrl, nvtxs);
}
/*************************************************************************
* This function is the entry point for KWMETIS
**************************************************************************/
void METIS_mCPartGraphKway(int *nvtxs, int *ncon, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt, int *wgtflag, int *numflag,
int *nparts, floattype *rubvec, 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, *ncon, xadj, adjncy, vwgt, adjwgt, *wgtflag);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = McKMETIS_CTYPE;
ctrl.IType = McKMETIS_ITYPE;
ctrl.RType = McKMETIS_RTYPE;
ctrl.dbglvl = McKMETIS_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)/(20*log2Int(*nparts)), 30*(*nparts));
ctrl.nmaxvwgt = 1.5/(1.0*ctrl.CoarsenTo);
InitRandom(-1);
AllocateWorkSpace(&ctrl, &graph, *nparts);
IFSET(ctrl.dbglvl, DBG_TIME, InitTimers(&ctrl));
IFSET(ctrl.dbglvl, DBG_TIME, starttimer(ctrl.TotalTmr));
ASSERT(CheckGraph(&graph));
*edgecut = MCMlevelKWayPartitioning(&ctrl, &graph, *nparts, part, rubvec);
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);
}
/* Write a 1 */
void ringwriteone (void)
{
/* Set to output, expecting high */
spit->dra = 0xE0;
spit->ddra = 0xE0;
starttimer (10000000);
/* Pull low for 10 us - this works... */
spit->dra = 0xC0;
spit->dra = 0xC0;
spit->dra = 0xC0;
/* Set to input */
spit->ddra = 0xC0;
/* Wait for a total of 120 us */
while (gettimer () > 10000000 - 40);
}
/*Function sends packet with "Id" nextSeqNumber, Also add event to timer List*/
void A_sendPacket(int sequence) {
insertLinkedList(sequence, get_sim_time());
tolayer3(A, senderBuffer[sequence]);
//printf("Packet sent from A to B Sequence Number : %d \n",sequence);
if (nextSequenceNum == base) //whenever it starts fresh after receiving all Acks, start the timer
{
//printf("TIMER HAS BEEN STARTED at %f\n",get_sim_time());
//stoptimer(A);
starttimer(A, timeout);
}
}
/*************************************************************************
* This function takes a graph and produces a bisection of it
**************************************************************************/
int MCMlevelKWayPartitioning(CtrlType *ctrl, GraphType *graph, int nparts, idxtype *part,
float *rubvec)
{
int i, j, nvtxs;
GraphType *cgraph;
int options[10], edgecut;
cgraph = MCCoarsen2Way(ctrl, graph);
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
MocAllocateKWayPartitionMemory(ctrl, cgraph, nparts);
options[0] = 1;
options[OPTION_CTYPE] = MATCH_SBHEM_INFNORM;
options[OPTION_ITYPE] = IPART_RANDOM;
options[OPTION_RTYPE] = RTYPE_FM;
options[OPTION_DBGLVL] = 0;
/* Determine what you will use as the initial partitioner, based on tolerances */
for (i=0; i<graph->ncon; i++) {
if (rubvec[i] > 1.2)
break;
}
if (i == graph->ncon)
METIS_mCPartGraphRecursiveInternal(&cgraph->nvtxs, &cgraph->ncon,
cgraph->xadj, cgraph->adjncy, cgraph->nvwgt, cgraph->adjwgt, &nparts,
options, &edgecut, cgraph->where);
else
METIS_mCHPartGraphRecursiveInternal(&cgraph->nvtxs, &cgraph->ncon,
cgraph->xadj, cgraph->adjncy, cgraph->nvwgt, cgraph->adjwgt, &nparts,
rubvec, options, &edgecut, cgraph->where);
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
IFSET(ctrl->dbglvl, DBG_IPART, printf("Initial %d-way partitioning cut: %d\n", nparts, edgecut));
IFSET(ctrl->dbglvl, DBG_KWAYPINFO, ComputePartitionInfo(cgraph, nparts, cgraph->where));
MocRefineKWayHorizontal(ctrl, graph, cgraph, nparts, rubvec);
idxcopy(graph->nvtxs, graph->where, part);
GKfree(&graph->nvwgt, &graph->npwgts, &graph->gdata, &graph->rdata, LTERM);
return graph->mincut;
}
void plot() {
printf("Plotting %i ...\n",bs.len);
RGBmp out=RGBmp(rm.width,rm.height,myRGB::black);
starttimer();
for (int i=1;i<=bs.len;i++) {
Blob *b=bs.p2num(i);
if (!b->blobbed) {
b->plot(&out);
}
}
printf("%.1f seconds\n",gettimer());
String fname=getnextfilename("out","bmp");
out.writefile(fname);
out.freedom();
if (usexv)
system(Sconc("xv ",fname," &"));
}
/* called when A's timer goes off */
void A_timerinterrupt(void)
{
int i;
if (TRACE > 0)
printf("----A: time out,resend packets!\n");
for(i=0; i<windowcount; i++) {
if (TRACE > 0)
printf ("---A: resending packet %d\n", (buffer[(windowfirst+i) % WINDOWSIZE]).seqnum);
tolayer3(A,buffer[(windowfirst+i) % WINDOWSIZE]);
packets_resent++;
if (i==0) starttimer(A,RTT);
}
}
/*************************************************************************
* This function is the entry point for KWMETIS
**************************************************************************/
void METIS_WPartGraphVKway(int *nvtxs, idxtype *xadj, idxtype *adjncy, idxtype *vwgt,
idxtype *vsize, int *wgtflag, int *numflag, int *nparts,
float *tpwgts, int *options, int *volume, idxtype *part)
{
int i, j;
GraphType graph;
CtrlType ctrl;
if (*numflag == 1)
Change2CNumbering(*nvtxs, xadj, adjncy);
VolSetUpGraph(&graph, OP_KVMETIS, *nvtxs, 1, xadj, adjncy, vwgt, vsize, *wgtflag);
if (options[0] == 0) { /* Use the default parameters */
ctrl.CType = KVMETIS_CTYPE;
ctrl.IType = KVMETIS_ITYPE;
ctrl.RType = KVMETIS_RTYPE;
ctrl.dbglvl = KVMETIS_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_KVMETIS;
ctrl.CoarsenTo = amax((*nvtxs)/(40*log2Int(*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));
*volume = MlevelVolKWayPartitioning(&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);
}
/*************************************************************************
* 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;
}
/* called from layer 5, passed the data to be sent to other side */
void A_output(struct msg message) //ram's comment - students can change the return type of the function from struct to pointers if necessary
{
//Buffer 1000 packets, used it to send when window moves ahead
if(SND_BUFSIZE<1000)
{
A_application++;
strncpy(buffer[SND_BUFSIZE].data, message.data,20);
SND_BUFSIZE++;
}
if(nextseqnum<(base+WINSIZE)){
struct pkt packet_a;
strncpy(packet_a.payload,buffer[nextseqnum].data,20);
packet_a.seqnum=nextseqnum;
packet_a.acknum=-1;//it is not acknowledgment
packet_a.checksum=packet_a.seqnum+packet_a.acknum;
for(int i=0;i<20;i++)
packet_a.checksum+=packet_a.payload[i];
my_packet[my_packet_no++]=packet_a;
tolayer3(0,packet_a);
A_transport++;
if(base==nextseqnum)
{
starttimer(0,TIMEOUT);
timer_A_started=1;
}
nextseqnum++;
}
}
/* called from layer 5 (application layer), passed the message to be sent to other side */
void A_output(struct msg message)
{
struct pkt sendpkt;
int i;
/* if not blocked waiting on ACK */
if ( windowcount < WINDOWSIZE) {
if (TRACE > 1)
printf("----A: New message arrives, send window is not full, send new messge to layer3!\n");
/* create packet */
sendpkt.seqnum = A_nextseqnum;
sendpkt.acknum = NOTINUSE;
for ( i=0; i<20 ; i++ )
sendpkt.payload[i] = message.data[i];
sendpkt.checksum = ComputeChecksum(sendpkt);
/* put packet in window buffer */
/* windowlast will always be 0 for alternating bit; but not for GoBackN */
windowlast = (windowlast + 1) % WINDOWSIZE;
buffer[windowlast] = sendpkt;
windowcount++;
/* send out packet */
if (TRACE > 0)
printf("Sending packet %d to layer 3\n", sendpkt.seqnum);
tolayer3 (A, sendpkt);
/* start timer if first packet in window */
if (windowcount == 1)
starttimer(A,RTT);
/* get next sequence number, wrap back to 0 */
A_nextseqnum = (A_nextseqnum + 1) % SEQSPACE;
}
/* if blocked, window is full */
else {
if (TRACE > 0)
printf("----A: New message arrives, send window is full\n");
window_full++;
}
}
/***********************************************************************************
* 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);
}
void real_main() {
init_candy();
bool should_quit = false;
starttimer();
while (!should_quit) {
iterate_candy();
#ifdef X11GFX
bool ev_occ=XCheckMaskEvent(d,ExposureMask | KeyPressMask | ButtonPressMask | StructureNotifyMask,
&ev);
if (ev_occ) {
switch (ev.type) {
case Expose: break;
case ButtonPress: should_quit=true; break;
case KeyPress: break;
default: break;
}
}
#endif
#ifdef ALLEGRO
if (key[KEY_ESC] || key[KEY_SPACE])
should_quit=true;
#endif
}
savetimer();
#ifdef ALLEGRO
allegro_exit();
#endif
displayframespersecond();
}
int state10(void)
{
unsigned char RecBuf[2];
char v=1;
int otime,ctime;
unsigned char Data_next;
int Control=0;
serialinit();
initAudio();
initCompr();
/* Synch with PC... */
serialtransmit('A');
otime=10000000;
starttimer(SAMPLE_DELAY);
ctime=gettimer();
while ( otime > ctime ) {
otime=ctime;
ctime=gettimer();
}
RecBuf[0] = (getsample());
otime=1000;
while (1) {
/*starttimer(SAMPLE_DELAY);*/
while ( otime > (otime= gettimer() > 0) );
RecBuf[v] = (getsample());
if (v) {
Control = compress(RecBuf[0],RecBuf[1]);
} else {
Control = serialtransmit(Control);
if (Control) {
serialclose();
return Control;
}
}
v=1-v;
}
}
/*************************************************************************
* 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 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);
}
/*************************************************************************
* 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 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;
}
/*************************************************************************
* 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));
}
