void BlockUtility::TGenerateRowStencil(const Epetra_CrsGraph& LocalBlockGraph,
std::vector<int_type> RowIndices,
std::vector< std::vector<int_type> >& RowStencil)
{
// Get row indices
int NumMyRows = LocalBlockGraph.NumMyRows();
RowIndices.resize(NumMyRows);
const Epetra_BlockMap& RowMap = LocalBlockGraph.RowMap();
RowMap.MyGlobalElements(&RowIndices[0]);
// Get stencil
RowStencil.resize(NumMyRows);
if (LocalBlockGraph.IndicesAreGlobal()) {
for (int i=0; i<NumMyRows; i++) {
int_type Row = RowIndices[i];
int NumCols = LocalBlockGraph.NumGlobalIndices(Row);
RowStencil[i].resize(NumCols);
LocalBlockGraph.ExtractGlobalRowCopy(Row, NumCols, NumCols,
&RowStencil[i][0]);
for (int k=0; k<NumCols; k++)
RowStencil[i][k] -= Row;
}
}
else {
for (int i=0; i<NumMyRows; i++) {
int NumCols = LocalBlockGraph.NumMyIndices(i);
std::vector<int> RowStencil_local(NumCols);
RowStencil[i].resize(NumCols);
LocalBlockGraph.ExtractMyRowCopy(i, NumCols, NumCols,
&RowStencil_local[0]);
for (int k=0; k<NumCols; k++)
RowStencil[i][k] = (int_type) ((int) (LocalBlockGraph.GCID64(RowStencil_local[k]) - RowIndices[i]));
}
}
}
Teuchos::RCP<Epetra_CrsGraph> BlockAdjacencyGraph::compute( Epetra_CrsGraph& B, int nbrr, std::vector<int>&r, std::vector<double>& weights, bool verbose)
{
// Check if the graph is on one processor.
int myMatProc = -1, matProc = -1;
int myPID = B.Comm().MyPID();
for (int proc=0; proc<B.Comm().NumProc(); proc++)
{
if (B.NumGlobalEntries() == B.NumMyEntries())
myMatProc = myPID;
}
B.Comm().MaxAll( &myMatProc, &matProc, 1 );
if( matProc == -1)
{ cout << "FAIL for Global! All CrsGraph entries must be on one processor!\n"; abort(); }
int i= 0, j = 0, k, l = 0, p, pm, q = -1, ns;
int tree_height;
int error = -1; /* error detected, possibly a problem with the input */
int nrr; /* number of rows in B */
int nzM = 0; /* number of edges in graph */
int m = 0; /* maximum number of nonzeros in any block row of B */
int* colstack = 0; /* stack used to process each block row */
int* bstree = 0; /* binary search tree */
std::vector<int> Mi, Mj, Mnum(nbrr+1,0);
nrr = B.NumMyRows();
if ( matProc == myPID && verbose )
std::printf(" Matrix Size = %d Number of Blocks = %d\n",nrr, nbrr);
else
nrr = -1; /* Prevent processor from doing any computations */
bstree = csr_bst(nbrr); /* 0 : nbrr-1 */
tree_height = ceil31log2(nbrr) + 1;
error = -1;
l = 0; j = 0; m = 0;
for( i = 0; i < nrr; i++ ){
if( i >= r[l+1] ){
++l; /* new block row */
m = EPETRA_MAX(m,j) ; /* nonzeros in block row */
j = B.NumGlobalIndices(i);
}else{
j += B.NumGlobalIndices(i);
}
}
/* one more time for the final block */
m = EPETRA_MAX(m,j) ; /* nonzeros in block row */
colstack = (int*) malloc( EPETRA_MAX(m,1) * sizeof(int) );
// The compressed graph is actually computed twice,
// due to concerns about memory limitations. First,
// without memory allocation, just nzM is computed.
// Next Mj is allocated. Then, the second time, the
// arrays are actually populated.
nzM = 0; q = -1; l = 0;
int * indices;
int numEntries;
for( i = 0; i <= nrr; i++ ){
if( i >= r[l+1] ){
if( q > 0 ) std::qsort(colstack,q+1,sizeof(int),compare_ints); /* sort stack */
if( q >= 0 ) ns = 1; /* l, colstack[0] M */
for( j=1; j<=q ; j++ ){ /* delete copies */
if( colstack[j] > colstack[j-1] ) ++ns;
}
nzM += ns; /*M->p[l+1] = M->p[l] + ns;*/
++l;
q = -1;
}
if( i < nrr ){
B.ExtractMyRowView( i, numEntries, indices );
for( k = 0; k < numEntries; k++){
j = indices[k]; ns = 0; p = 0;
while( (r[bstree[p]] > j) || (j >= r[bstree[p]+1]) ){
if( r[bstree[p]] > j){
p = 2*p+1;
}else{
if( r[bstree[p]+1] <= j) p = 2*p+2;
}
++ns;
if( p > nbrr || ns > tree_height ) {
error = j;
std::printf("error: p %d nbrr %d ns %d %d\n",p,nbrr,ns,j); break;
}
}
colstack[++q] = bstree[p];
}
//if( error >-1 ){ std::printf("%d\n",error); break; }
// p > nbrr is a fatal error that is ignored
}
}
if ( matProc == myPID && verbose )
std::printf("nzM = %d \n", nzM );
Mi.resize( nzM );
Mj.resize( nzM );
q = -1; l = 0; pm = -1;
for( i = 0; i <= nrr; i++ ){
if( i >= r[l+1] ){
if( q > 0 ) std::qsort(colstack,q+1,sizeof(colstack[0]),compare_ints); /* sort stack */
if( q >= 0 ){
Mi[++pm] = l;
Mj[pm] = colstack[0];
}
for( j=1; j<=q ; j++ ){ /* delete copies */
if( colstack[j] > colstack[j-1] ){ /* l, colstack[j] */
Mi[++pm] = l;
Mj[pm] = colstack[j];
}
}
++l;
Mnum[l] = pm + 1;
/* sparse row format: M->p[l+1] = M->p[l] + ns; */
q = -1;
}
if( i < nrr ){
B.ExtractMyRowView( i, numEntries, indices );
for( k = 0; k < numEntries; k++){
j = indices[k]; ns = 0; p = 0;
while( (r[bstree[p]] > j) || (j >= r[bstree[p]+1]) ){
if( r[bstree[p]] > j){
p = 2*p+1;
}else{
if( r[bstree[p]+1] <= j) p = 2*p+2;
}
++ns;
}
colstack[++q] = bstree[p];
}
}
}
if ( bstree ) free ( bstree );
if ( colstack ) free( colstack );
// Compute weights as number of rows in each block.
weights.resize( nbrr );
for( l=0; l<nbrr; l++) weights[l] = r[l+1] - r[l];
// Compute Epetra_CrsGraph and return
Teuchos::RCP<Epetra_Map> newMap;
if ( matProc == myPID )
newMap = Teuchos::rcp( new Epetra_Map(nbrr, nbrr, 0, B.Comm() ) );
else
newMap = Teuchos::rcp( new Epetra_Map( nbrr, 0, 0, B.Comm() ) );
Teuchos::RCP<Epetra_CrsGraph> newGraph = Teuchos::rcp( new Epetra_CrsGraph( Copy, *newMap, 0 ) );
for( l=0; l<newGraph->NumMyRows(); l++) {
newGraph->InsertGlobalIndices( l, Mnum[l+1]-Mnum[l], &Mj[Mnum[l]] );
}
newGraph->FillComplete();
return (newGraph);
}
