std::vector< std::set<idxtype> > srmclWithGraph(GraphType *graph,std::vector<std::set<idxtype> >& indices,Options opt )
{
GraphType *cgraph,*t,*cgraphFinest, *cgraphCoarest;
int nvtxs=graph->nvtxs;
CtrlType ctrl;
int levels=0;
Matrix *M=NULL;
int ct = ctrl.CoarsenTo = opt.coarsenTo, runMcl=0;
ctrl.CType=opt.matchType;
ctrl.optype=OP_KMETIS;
ctrl.dbglvl=1;
if ( ct < 5 )
ctrl.maxvwgt = (int) ceil( (1.5*nvtxs)/ct );
else if ( ct <= 100 )
ctrl.maxvwgt = (int) ceil( (2.0*nvtxs)/ct );
else if ( ct > 100 )
{
// we can allow some imbalance here.
ctrl.maxvwgt = (int) ceil( 10.0 * nvtxs/ct );
}
my_AllocateWorkSpace(&ctrl,graph);
cgraphCoarest = cgraph = Coarsen2Way(&ctrl,graph);
FreeWorkSpace(&ctrl, graph);
for(levels=0,t=cgraph ; t->finer!=NULL ; t=t->finer,levels++);
int num_level = levels;
printf("Coarsened %d levels\n", levels);
printf("Number of vertices in coarsest graph:%d\n", cgraph->nvtxs);
fflush(stdout);
int times_rmcl= 1;
while ( cgraph != NULL )
{
Matrix *M0, *Mnew;
int iters;
M0=setupCanonicalMatrix(cgraph->nvtxs, cgraph->nedges, cgraph->xadj, cgraph->adjncy, cgraph->adjwgt, opt.ncutify); //do weight normalization here
#ifdef TEST_OUTPUT
//printf("level: %d, M0:\n",levels);
//printMatrix(M0);
#endif
if ( cgraph->coarser == NULL )
{
// if this is the coarsest graph, flow matrix is same as
// the transition matrix.
M=M0;
// dumpMatrix(M);
}
if (cgraph->finer == NULL){
iters=opt.num_last_iter;
cgraphFinest = cgraph;
}
else{
iters=opt.iter_per_level;
}
//printMatrix(M);
Mnew=dprmcl(M,M0,cgraph, opt, iters, levels); //YK: main process!
M=Mnew;
if ( cgraph->finer != NULL ) //YK: map nodes to the finer graph
{
int nnzRefinedGraph = 2*M->nnz;
if ( ctrl.CType == MATCH_POWERLAW_FC )
{
int ii;
nnzRefinedGraph = 0;
for ( ii=0; ii<cgraph->finer->nvtxs; ii++ )
{
int tx=cgraph->finer->cmap[ii];
nnzRefinedGraph +=
M->xadj[tx+1]-M->xadj[tx];
}
}
else if ( ctrl.CType == MATCH_HASH && levels <= 3 )
{
int ii;
nnzRefinedGraph=0;
for ( ii=0; ii<cgraph->nvtxs; ii++ )
{
nnzRefinedGraph += cgraph->vwgt[ii] *
(M->xadj[ii+1] - M->xadj[ii]);
}
}
// dumpMatrix(M);
Mnew=propagateFlow(M,cgraph,cgraph->finer,nnzRefinedGraph);
//printf("times:%d, levels:%d Mnew\n",times_rmcl, levels);
//printMatrix(Mnew);
if ( M != NULL )
{
freeMatrix(M);
}
M=Mnew;
}
cgraph=cgraph->finer; //change to the finer graph
levels--;
printf("Done level %d (%d times MLR-MCL)\n", levels+1,times_rmcl);
fflush(stdout);
}
fflush(stdout);
idxtype* firstIndices = (idxtype*) malloc(sizeof(idxtype) * M->nvtxs);
idxcopy(nvtxs, M->attractors, firstIndices);
bool* isAttractor = (bool*) malloc(sizeof(bool) * M->nvtxs);
getAttractorsForAll(M, isAttractor);
int** countAttractor = (int**) malloc(sizeof(int*) * (num_level+1)); //count the number of times the node being an attractor node
for(int i=0; i <= num_level ; i++){
countAttractor[i] = (int*) malloc (sizeof(int) * nvtxs);
for(int j=0; j<M->nvtxs; j++){
countAttractor[i][j] = 0;
}
}
#ifdef TEST_OUTPUT
{
for(int i=0; i<M->nvtxs; i++){
int level = 1, nodeIdx = i;
for(cgraph = cgraphFinest; cgraph->coarser!=NULL; cgraph = cgraph->coarser, level++){
idxtype oldNodeIdx = nodeIdx;
nodeIdx = cgraph->cmap[oldNodeIdx];
printf(" cmap: level:%d vid: %d to %d \n", level, oldNodeIdx+1, nodeIdx+1);
}
}
}
#endif
for(int i=0; i<M->nvtxs; i++){
if(isAttractor[i]){
countAttractor[0][i]++;
#ifdef TEST_OUTPUT
printf("attractor vid:%d(%d times), %d-th times MLRMCL\n",i+1,countAttractor[0][i],times_rmcl);fflush(stdout);
#endif
int level = 1, nodeIdx = i;
for(cgraph = cgraphFinest; cgraph->coarser!=NULL; cgraph = cgraph->coarser, level++){
idxtype oldNodeIdx = nodeIdx;
nodeIdx = cgraph->cmap[nodeIdx];
countAttractor[level][nodeIdx]++;
#ifdef TEST_OUTPUT
//printf("countAttractor[%d][%d] = %d;\n", level, nodeIdx,countAttractor[level][nodeIdx]);
#endif
}
}
}
//construct hash table
idxtype* hashtable = idxmalloc(nvtxs,"Hashtable");
for(int i=0;i<nvtxs;i++)
hashtable[i]=-1; // clear hashtable.
int num_clusters = 0; //include singleton
for(int i=0;i<nvtxs;i++){
if ( hashtable[firstIndices[i]] == -1){
hashtable[firstIndices[i]] = num_clusters;
num_clusters++;
}
}
/*idxtype* invHashTable = idxmalloc(num_clusters,"inverse hashtable"); //map cluster id back to the attractor id
for(int i=0;i<nvtxs;i++){
if ( hashtable[firstIndices[i]] != -1){
invHashTable[hashtable[firstIndices[i]]] = firstIndices[i];
}
}*/
//construct clusters
std::vector< std::set<idxtype> > clusters(0); //clusters[i] contains nodes' numbers (in original graph) in cluster i
for(int i=0; i<num_clusters ; i++){
std::set<idxtype> newCluster;
clusters.push_back(newCluster);
}
for(int i=0; i<nvtxs ; i++){
if( firstIndices[i] != -1){
idxtype cID = hashtable[firstIndices[i]];
indices[i].insert(cID);
if(cID == -1)
continue;
clusters[ cID ].insert( i );
}
}
idxtype* lastIndices = firstIndices;
bool* lastAttractor = isAttractor;
idxtype* oldHashtable = hashtable;
printf("number of clusters in 1st time MLR-MCL: %d\n", num_clusters);fflush(stdout);
idxtype* thisIndices = (idxtype*) malloc(sizeof(idxtype) * M->nvtxs);
idxtype* newHashtable = idxmalloc(nvtxs,"newHashtable");
bool* thisAttractor = (bool*) malloc(sizeof(bool) * M->nvtxs);
while(times_rmcl < opt.time_rmcl){ //do more than one times R-MCL but penalize attractor nodes (by assigning them higher inflation rate)
times_rmcl++;
cgraph = cgraphCoarest;
for(levels=0,t=cgraph ; t->finer!=NULL ; t=t->finer,levels++);
while ( cgraph != NULL )
{
Matrix *M0, *Mnew;
int iters;
M0=setupCanonicalMatrix(cgraph->nvtxs, cgraph->nedges, cgraph->xadj, cgraph->adjncy, cgraph->adjwgt, opt.ncutify); //do weight normalization here
if ( cgraph->coarser == NULL )
{
M=M0;
}
if (cgraph->finer == NULL)
iters=opt.num_last_iter;
else
iters=opt.iter_per_level;
//Mnew=dprmcl(M,M0,cgraph, opt, iters,levels); //original
Mnew=dprmcl_penalizeAttractors(M,M0,cgraph, opt, iters, levels, countAttractor); //YK: main process!
M=Mnew;
if ( cgraph->finer != NULL ) //YK: map nodes to the finer graph
{
int nnzRefinedGraph = 2*M->nnz;
if ( ctrl.CType == MATCH_POWERLAW_FC )
{
int ii;
nnzRefinedGraph = 0;
for ( ii=0; ii<cgraph->finer->nvtxs; ii++ )
{
int tx=cgraph->finer->cmap[ii];
nnzRefinedGraph +=
M->xadj[tx+1] - M->xadj[tx];
}
}
else if ( ctrl.CType == MATCH_HASH && levels <= 3 )
{
int ii;
nnzRefinedGraph=0;
for ( ii=0; ii<cgraph->nvtxs; ii++ )
{
nnzRefinedGraph += cgraph->vwgt[ii] *
(M->xadj[ii+1] - M->xadj[ii]);
}
}
Mnew=propagateFlow(M,cgraph,cgraph->finer,nnzRefinedGraph);
//printf("times:%d, levels:%d Mnew\n",times_rmcl, levels);
//printMatrix(Mnew);
if ( M != NULL )
{
freeMatrix(M);
}
M=Mnew;
}
cgraph=cgraph->finer; //change to the finer graph
// These two didn't get freed earlier, when we freed
// gdata.
levels--;
printf("Done level %d (%d-th times MLR-MCL)\n", levels+1,times_rmcl);fflush(stdout);
}
//update clusters
idxcopy(nvtxs, M->attractors, thisIndices);
getAttractorsForAll(M, thisAttractor);
for(int i=0; i<M->nvtxs; i++){
if(thisAttractor[i]){
countAttractor[0][i]++;
#ifdef TEST_OUTPUT
printf("attractor vid:%d(%d times), %d-th times MLRMCL\n",i+1,countAttractor[0][i],times_rmcl);
#endif
int level = 1, nodeIdx = i;
for(cgraph = cgraphFinest; cgraph->coarser!=NULL; cgraph = cgraph->coarser, level++){
nodeIdx = cgraph->cmap[nodeIdx];
countAttractor[level][nodeIdx]++;
}
//printf("attractor:%d, times:%d\n",i+1,times_rmcl);
}
}
//construct hash table
for(int i=0;i<nvtxs;i++)
newHashtable[i]=-1; // clear newHashtable.
int new_num_clusters = 0 ;
for(int i=0;i<nvtxs;i++){
if ( newHashtable[thisIndices[i]] == -1){
newHashtable[thisIndices[i]] = num_clusters+new_num_clusters; //cluster ID
if(num_clusters+new_num_clusters == 22704 || num_clusters+new_num_clusters == 29925)
printf("cluster #%d's attractor node %d\n",num_clusters+new_num_clusters,thisIndices[i]);
new_num_clusters++;
}
}
//construct clusters
for(int i=0; i<new_num_clusters; i++){
std::set<idxtype>* newCluster = new std::set<idxtype>();
clusters.push_back(*newCluster);
}
for(int i=0; i<nvtxs ; i++){
if( thisIndices[i] != -1){
idxtype cID = newHashtable[thisIndices[i]];
indices[i].insert(cID);
if(cID == -1)
continue;
clusters[ cID ].insert( i );
}
}
printf("number of clusters in %dth times MLR-MCL: %d\ttotal # clusters:%zu\n", times_rmcl, new_num_clusters,clusters.size());
num_clusters += new_num_clusters;
//detect last attractor is split to multiple clusters
/*
double* countToCluster = (double*) malloc(sizeof(double) * num_clusters) ;
for(idxtype vID=0; vID<nvtxs ; vID++){
//printf("test:%d\n",vID+1);fflush(stdout);
if(lastAttractor[vID] && !thisAttractor[vID]){
for(int i=0; i<num_clusters; i++)
countToCluster[i] = 0;
int lastCluster = oldHashtable[lastIndices[vID]];
//printf("lastCluster:cid%d (attractor:%d size:%zu), num clusters:%d\n",lastCluster,vID+1,clusters[lastCluster].size(),num_clusters);fflush(stdout);
for(std::set<idxtype>::iterator nodeIterator = clusters[lastCluster].begin(); nodeIterator != clusters[lastCluster].end(); nodeIterator++){
int currentCluster = newHashtable[thisIndices[*nodeIterator]];
countToCluster[currentCluster] ++;
}
for(idxtype cID = 0; cID < num_clusters; cID++){
if(countToCluster[cID] / clusters[lastCluster].size() >= opt.ratio_nodes_another_cluster && clusters[cID].find(vID)==clusters[cID].end() ){
//clusters[ cID ].insert(vID);
//indices[ vID ].insert(cID);
#ifdef TEST_OUTPUT
printf("^add extra cluster: vid:%d to cid:%d\n",vID+1, cID);fflush(stdout);
#endif
}
}
}
}
free(countToCluster);*/
lastIndices = thisIndices;
lastAttractor = thisAttractor;
oldHashtable = newHashtable;
}
printf("total number of clusters after %d times MLR-MCL:%zu\n", opt.time_rmcl, clusters.size());fflush(stdout);
free(hashtable);
for(int i=0; i <= num_level ; i++)
free(countAttractor[i]);
free(countAttractor);
free(isAttractor);
free(firstIndices);
free(thisIndices);
free(newHashtable);
free(thisAttractor);
freeMatrix(M);
return clusters;
}
void mlmclWithGraph(GraphType *graph,idxtype* indices,Options opt)
{
GraphType *cgraph,*t;
int nvtxs=graph->nvtxs;
CtrlType ctrl;
int levels=0;
Matrix *M=NULL;
int ct = ctrl.CoarsenTo = opt.coarsenTo, runMcl=0;
ctrl.CType=opt.matchType;
/* if ( nvtxs > 50000 )
ctrl.CType = MATCH_POWERLAW_FC; */
ctrl.optype=OP_KMETIS;
// ctrl.maxvwgt = (int) round( ((ct>100)?nvtxs/100: nvtxs/ct));
// ctrl.maxvwgt = (ctrl.maxvwgt > 5000 ) ? 5000 : ctrl.maxvwgt ;
ctrl.dbglvl=1;
if ( ct < 5 )
ctrl.maxvwgt = (int) ceil( (1.5*nvtxs)/ct );
else if ( ct <= 100 )
ctrl.maxvwgt = (int) ceil( (2.0*nvtxs)/ct );
else if ( ct > 100 )
{
// we can allow some imbalance here.
ctrl.maxvwgt = (int) ceil( 10.0 * nvtxs/ct );
}
my_AllocateWorkSpace(&ctrl,graph);
cgraph = Coarsen2Way(&ctrl,graph);
FreeWorkSpace(&ctrl, graph);
// printf( "Coarsen time:%.2f\n",gettimer(ctrl.CoarsenTmr) );
for(levels=0,t=cgraph;t->finer!=NULL;t=t->finer,levels++);
printf("Coarsened %d levels\n", levels);
printf("Number of vertices in coarsest graph:%d\n", cgraph->nvtxs);
// printf("In lib/mlmcl.c: penalty_power:%.1f\n", opt.penalty_power);
fflush(stdout);
while ( cgraph != NULL )
{
Matrix *M0, *Mnew;
int iters;
// dump_graph(cgraph);
M0=setupCanonicalMatrix(cgraph->nvtxs, cgraph->nedges,
cgraph->xadj, cgraph->adjncy, cgraph->adjwgt, opt.ncutify);
if ( cgraph->coarser != NULL )
{
// if this is not the original graph, then we don't
// need the adjacency information in the graph
// anymore. M0 is what we want, and we have it, so we
// can free cgraph->gdata.
GKfree( (void**)&(cgraph->gdata), LTERM);
}
if ( cgraph->coarser == NULL )
{
// if this is the coarsest graph, flow matrix is same as
// the transition matrix.
M=M0;
// dumpMatrix(M);
}
if (cgraph->finer == NULL)
iters=opt.num_last_iter;
else
iters=opt.iter_per_level;
if ( cgraph->coarser == NULL )
{;
//printRow(M,35);
//printRow(M,36);
}
Mnew=dprmcl(M,M0,cgraph, opt, iters, levels);
// M would have been freed in mclForMultiLevel, so no
// need to free it here.
/* if ( M0 != NULL )
{
freeMatrix(M0);
} */
if ( cgraph->coarser == NULL )
{
//printRow(Mnew,35);
//printRow(Mnew,36);
}
if ( runMcl > 0 && cgraph->finer == NULL )
{
/* run original mcl so that the matrix moves closer to
* convergence */
M=dprmcl(Mnew,Mnew,cgraph, opt, iters/2,levels);
/* Mnew would not have been freed in mclForMultiLevel
* above. */
freeMatrix(Mnew);
}
else
{
M=Mnew;
}
if ( cgraph->finer != NULL )
{
int nnzRefinedGraph = 2*M->nnz;
if ( ctrl.CType == MATCH_POWERLAW_FC )
{
int ii;
nnzRefinedGraph = 0;
for ( ii=0; ii<cgraph->finer->nvtxs; ii++ )
{
int tx=cgraph->finer->cmap[ii];
nnzRefinedGraph +=
M->xadj[tx+1]-M->xadj[tx];
}
}
else if ( ctrl.CType == MATCH_HASH && levels <= 3 )
{
int ii;
nnzRefinedGraph=0;
for ( ii=0; ii<cgraph->nvtxs; ii++ )
{
nnzRefinedGraph += cgraph->vwgt[ii] *
(M->xadj[ii+1] - M->xadj[ii]);
}
}
// dumpMatrix(M);
Mnew=propagateFlow(M,cgraph,cgraph->finer,nnzRefinedGraph);
// dumpMatrix(Mnew);
if ( M != NULL )
{
freeMatrix(M);
}
M=Mnew;
}
cgraph=cgraph->finer;
// These two didn't get freed earlier, when we freed
// gdata.
if ( cgraph != NULL )
{
GKfree( (void**)&(cgraph->coarser->vwgt),
(void**)&(cgraph->coarser->rmap1), LTERM);
}
levels--;
printf("Done level %d\n", levels+1);
fflush(stdout);
}
getAttractorsForAll(M);
idxcopy(nvtxs, M->attractors, indices);
/* int nClusters = 0;
wgttype minWgt = 0;
idxtype *ret = getNodesToComponentMap(M, &nClusters, minWgt);
idxcopy(nvtxs, ret, indices);
free(ret);
*/
freeMatrix(M);
}
