HYPRE_Int
hypre_MatTCommPkgCreate ( hypre_ParCSRMatrix *A)
{
hypre_ParCSRCommPkg *comm_pkg;
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
/* hypre_MPI_Datatype *recv_mpi_types;
hypre_MPI_Datatype *send_mpi_types;
*/
HYPRE_Int num_sends;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int num_recvs;
HYPRE_Int *recv_procs;
HYPRE_Int *recv_vec_starts;
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(A);
HYPRE_Int *col_starts = hypre_ParCSRMatrixColStarts(A);
HYPRE_Int ierr = 0;
HYPRE_Int num_rows_diag = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A));
HYPRE_Int num_cols_diag = hypre_CSRMatrixNumCols(hypre_ParCSRMatrixDiag(A));
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(hypre_ParCSRMatrixOffd(A));
HYPRE_Int * row_starts = hypre_ParCSRMatrixRowStarts(A);
hypre_MatTCommPkgCreate_core (
comm, col_map_offd, first_col_diag, col_starts,
num_rows_diag, num_cols_diag, num_cols_offd, row_starts,
hypre_ParCSRMatrixFirstColDiag(A),
hypre_ParCSRMatrixColMapOffd(A),
hypre_CSRMatrixI( hypre_ParCSRMatrixDiag(A) ),
hypre_CSRMatrixJ( hypre_ParCSRMatrixDiag(A) ),
hypre_CSRMatrixI( hypre_ParCSRMatrixOffd(A) ),
hypre_CSRMatrixJ( hypre_ParCSRMatrixOffd(A) ),
1,
&num_recvs, &recv_procs, &recv_vec_starts,
&num_sends, &send_procs, &send_map_starts,
&send_map_elmts
);
comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg, 1);
hypre_ParCSRCommPkgComm(comm_pkg) = comm;
hypre_ParCSRCommPkgNumRecvs(comm_pkg) = num_recvs;
hypre_ParCSRCommPkgRecvProcs(comm_pkg) = recv_procs;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg) = recv_vec_starts;
hypre_ParCSRCommPkgNumSends(comm_pkg) = num_sends;
hypre_ParCSRCommPkgSendProcs(comm_pkg) = send_procs;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg) = send_map_starts;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg) = send_map_elmts;
hypre_ParCSRMatrixCommPkgT(A) = comm_pkg;
return ierr;
}
hypre_ParCSRMatrix *
hypre_ParCSRBlockMatrixCompress( hypre_ParCSRBlockMatrix *matrix )
{
MPI_Comm comm = hypre_ParCSRBlockMatrixComm(matrix);
hypre_CSRBlockMatrix *diag = hypre_ParCSRBlockMatrixDiag(matrix);
hypre_CSRBlockMatrix *offd = hypre_ParCSRBlockMatrixOffd(matrix);
HYPRE_Int global_num_rows = hypre_ParCSRBlockMatrixGlobalNumRows(matrix);
HYPRE_Int global_num_cols = hypre_ParCSRBlockMatrixGlobalNumCols(matrix);
HYPRE_Int *row_starts = hypre_ParCSRBlockMatrixRowStarts(matrix);
HYPRE_Int *col_starts = hypre_ParCSRBlockMatrixColStarts(matrix);
HYPRE_Int num_cols_offd = hypre_CSRBlockMatrixNumCols(offd);
HYPRE_Int num_nonzeros_diag = hypre_CSRBlockMatrixNumNonzeros(diag);
HYPRE_Int num_nonzeros_offd = hypre_CSRBlockMatrixNumNonzeros(offd);
hypre_ParCSRMatrix *matrix_C;
HYPRE_Int i;
matrix_C = hypre_ParCSRMatrixCreate(comm, global_num_rows, global_num_cols,
row_starts,col_starts,num_cols_offd,num_nonzeros_diag,num_nonzeros_offd);
hypre_ParCSRMatrixInitialize(matrix_C);
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixDiag(matrix_C));
hypre_ParCSRMatrixDiag(matrix_C) = hypre_CSRBlockMatrixCompress(diag);
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixOffd(matrix_C));
hypre_ParCSRMatrixOffd(matrix_C) = hypre_CSRBlockMatrixCompress(offd);
for(i = 0; i < num_cols_offd; i++)
hypre_ParCSRMatrixColMapOffd(matrix_C)[i] =
hypre_ParCSRBlockMatrixColMapOffd(matrix)[i];
return matrix_C;
}
void RowsWithColumn_original
( HYPRE_Int * rowmin, HYPRE_Int * rowmax, HYPRE_Int column, hypre_ParCSRMatrix * A )
/* Finds rows of A which have a nonzero at the given (global) column number.
Sets rowmin to the minimum (local) row number of such rows, and rowmax
to the max. If there are no such rows, will return rowmax<0<=rowmin */
{
hypre_CSRMatrix * diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix * offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int * mat_i, * mat_j;
HYPRE_Int i, j, num_rows;
HYPRE_Int firstColDiag;
HYPRE_Int * colMapOffd;
mat_i = hypre_CSRMatrixI(diag);
mat_j = hypre_CSRMatrixJ(diag);
num_rows = hypre_CSRMatrixNumRows(diag);
firstColDiag = hypre_ParCSRMatrixFirstColDiag(A);
*rowmin = num_rows;
*rowmax = -1;
for ( i=0; i<num_rows; ++i ) {
/* global number: row = i + firstRowIndex;*/
for ( j=mat_i[i]; j<mat_i[i+1]; ++j ) {
if ( mat_j[j]+firstColDiag==column ) {
/* row i (local row number) has column mat_j[j] (local column number) */
*rowmin = i<*rowmin ? i : *rowmin;
*rowmax = i>*rowmax ? i : *rowmax;
break;
}
}
}
mat_i = hypre_CSRMatrixI(offd);
mat_j = hypre_CSRMatrixJ(offd);
num_rows = hypre_CSRMatrixNumRows(offd);
colMapOffd = hypre_ParCSRMatrixColMapOffd(A);
for ( i=0; i<num_rows; ++i ) {
/* global number: row = i + firstRowIndex;*/
for ( j=mat_i[i]; j<mat_i[i+1]; ++j ) {
if ( colMapOffd[ mat_j[j] ]==column ) {
/* row i (local row number) has column mat_j[j] (local column number) */
*rowmin = i<*rowmin ? i : *rowmin;
*rowmax = i>*rowmax ? i : *rowmax;
break;
}
}
}
/* global col no.: mat_j[j]+hypre_ParCSRMatrixFirstColDiag(A)
or hypre_ParCSRMatrixColMapOffd(A)[ mat_j[j] ]
global row no.: i + hypre_ParCSRMatrixFirstRowIndex(A)
*/
}
void hypre_ParCSRMatrixSplit(hypre_ParCSRMatrix *A,
HYPRE_Int nr, HYPRE_Int nc,
hypre_ParCSRMatrix **blocks,
int interleaved_rows, int interleaved_cols)
{
HYPRE_Int i, j, k;
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *Adiag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *Aoffd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int global_rows = hypre_ParCSRMatrixGlobalNumRows(A);
HYPRE_Int global_cols = hypre_ParCSRMatrixGlobalNumCols(A);
HYPRE_Int local_rows = hypre_CSRMatrixNumRows(Adiag);
HYPRE_Int local_cols = hypre_CSRMatrixNumCols(Adiag);
HYPRE_Int offd_cols = hypre_CSRMatrixNumCols(Aoffd);
hypre_assert(local_rows % nr == 0 && local_cols % nc == 0);
hypre_assert(global_rows % nr == 0 && global_cols % nc == 0);
HYPRE_Int block_rows = local_rows / nr;
HYPRE_Int block_cols = local_cols / nc;
HYPRE_Int num_blocks = nr * nc;
/* mark local rows and columns with block number */
HYPRE_Int *row_block_num = hypre_TAlloc(HYPRE_Int, local_rows);
HYPRE_Int *col_block_num = hypre_TAlloc(HYPRE_Int, local_cols);
for (i = 0; i < local_rows; i++)
{
row_block_num[i] = interleaved_rows ? (i % nr) : (i / block_rows);
}
for (i = 0; i < local_cols; i++)
{
col_block_num[i] = interleaved_cols ? (i % nc) : (i / block_cols);
}
/* determine the block numbers for offd columns */
HYPRE_Int* offd_col_block_num = hypre_TAlloc(HYPRE_Int, offd_cols);
hypre_ParCSRCommHandle *comm_handle;
HYPRE_Int *int_buf_data;
{
/* make sure A has a communication package */
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(A);
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
/* calculate the final global column numbers for each block */
HYPRE_Int *count = hypre_CTAlloc(HYPRE_Int, nc);
HYPRE_Int *block_global_col = hypre_TAlloc(HYPRE_Int, local_cols);
HYPRE_Int first_col = hypre_ParCSRMatrixFirstColDiag(A) / nc;
for (i = 0; i < local_cols; i++)
{
block_global_col[i] = first_col + count[col_block_num[i]]++;
}
hypre_TFree(count);
/* use a Matvec communication pattern to determine offd_col_block_num */
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
int_buf_data = hypre_CTAlloc(HYPRE_Int,
hypre_ParCSRCommPkgSendMapStart(comm_pkg,
num_sends));
HYPRE_Int start, index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
{
k = hypre_ParCSRCommPkgSendMapElmt(comm_pkg, j);
int_buf_data[index++] = col_block_num[k] + nc*block_global_col[k];
}
}
hypre_TFree(block_global_col);
comm_handle = hypre_ParCSRCommHandleCreate(11, comm_pkg, int_buf_data,
offd_col_block_num);
}
/* create the block matrices */
HYPRE_Int num_procs = 1;
if (!hypre_ParCSRMatrixAssumedPartition(A))
{
hypre_MPI_Comm_size(comm, &num_procs);
}
HYPRE_Int *row_starts = hypre_TAlloc(HYPRE_Int, num_procs+1);
HYPRE_Int *col_starts = hypre_TAlloc(HYPRE_Int, num_procs+1);
for (i = 0; i <= num_procs; i++)
{
row_starts[i] = hypre_ParCSRMatrixRowStarts(A)[i] / nr;
col_starts[i] = hypre_ParCSRMatrixColStarts(A)[i] / nc;
}
for (i = 0; i < num_blocks; i++)
{
blocks[i] = hypre_ParCSRMatrixCreate(comm,
global_rows/nr, global_cols/nc,
row_starts, col_starts, 0, 0, 0);
}
/* split diag part */
hypre_CSRMatrix **csr_blocks = hypre_TAlloc(hypre_CSRMatrix*, nr*nc);
hypre_CSRMatrixSplit(Adiag, nr, nc, row_block_num, col_block_num,
csr_blocks);
for (i = 0; i < num_blocks; i++)
{
hypre_TFree(hypre_ParCSRMatrixDiag(blocks[i]));
hypre_ParCSRMatrixDiag(blocks[i]) = csr_blocks[i];
}
/* finish communication, receive offd_col_block_num */
hypre_ParCSRCommHandleDestroy(comm_handle);
hypre_TFree(int_buf_data);
/* decode global offd column numbers */
HYPRE_Int* offd_global_col = hypre_TAlloc(HYPRE_Int, offd_cols);
for (i = 0; i < offd_cols; i++)
{
offd_global_col[i] = offd_col_block_num[i] / nc;
offd_col_block_num[i] %= nc;
}
/* split offd part */
hypre_CSRMatrixSplit(Aoffd, nr, nc, row_block_num, offd_col_block_num,
csr_blocks);
for (i = 0; i < num_blocks; i++)
{
hypre_TFree(hypre_ParCSRMatrixOffd(blocks[i]));
hypre_ParCSRMatrixOffd(blocks[i]) = csr_blocks[i];
}
hypre_TFree(csr_blocks);
hypre_TFree(col_block_num);
hypre_TFree(row_block_num);
/* update block col-maps */
for (int bi = 0; bi < nr; bi++)
{
for (int bj = 0; bj < nc; bj++)
{
hypre_ParCSRMatrix *block = blocks[bi*nc + bj];
hypre_CSRMatrix *block_offd = hypre_ParCSRMatrixOffd(block);
HYPRE_Int block_offd_cols = hypre_CSRMatrixNumCols(block_offd);
HYPRE_Int *block_col_map = hypre_TAlloc(HYPRE_Int, block_offd_cols);
for (i = j = 0; i < offd_cols; i++)
{
HYPRE_Int bn = offd_col_block_num[i];
if (bn == bj) {
block_col_map[j++] = offd_global_col[i];
}
}
hypre_assert(j == block_offd_cols);
hypre_ParCSRMatrixColMapOffd(block) = block_col_map;
}
}
hypre_TFree(offd_global_col);
hypre_TFree(offd_col_block_num);
/* finish the new matrices, make them own all the stuff */
for (i = 0; i < num_blocks; i++)
{
hypre_ParCSRMatrixSetNumNonzeros(blocks[i]);
hypre_MatvecCommPkgCreate(blocks[i]);
hypre_ParCSRMatrixOwnsData(blocks[i]) = 1;
/* only the first block will own the row/col_starts */
hypre_ParCSRMatrixOwnsRowStarts(blocks[i]) = !i;
hypre_ParCSRMatrixOwnsColStarts(blocks[i]) = !i;
}
}
HYPRE_Int AmgCGCGraphAssemble (hypre_ParCSRMatrix *S,HYPRE_Int *vertexrange,HYPRE_Int *CF_marker,HYPRE_Int *CF_marker_offd,HYPRE_Int coarsen_type,
HYPRE_IJMatrix *ijG)
/* assemble a graph representing the connections between the grids
* ================================================================================================
* S : the strength matrix
* vertexrange : the parallel layout of the candidate coarse grid vertices
* CF_marker, CF_marker_offd : the coarse/fine markers
* coarsen_type : the coarsening type
* ijG : the created graph
* ================================================================================================*/
{
HYPRE_Int ierr=0;
HYPRE_Int i,/* ii,*/ip,j,jj,m,n,p;
HYPRE_Int mpisize,mpirank;
HYPRE_Real weight;
MPI_Comm comm = hypre_ParCSRMatrixComm(S);
/* hypre_MPI_Status status; */
HYPRE_IJMatrix ijmatrix;
hypre_CSRMatrix *S_diag = hypre_ParCSRMatrixDiag (S);
hypre_CSRMatrix *S_offd = hypre_ParCSRMatrixOffd (S);
/* HYPRE_Int *S_i = hypre_CSRMatrixI(S_diag); */
/* HYPRE_Int *S_j = hypre_CSRMatrixJ(S_diag); */
HYPRE_Int *S_offd_i = hypre_CSRMatrixI(S_offd);
HYPRE_Int *S_offd_j = NULL;
HYPRE_Int num_variables = hypre_CSRMatrixNumRows (S_diag);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols (S_offd);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd (S);
HYPRE_Int pointrange_start,pointrange_end;
HYPRE_Int *pointrange,*pointrange_nonlocal,*pointrange_strong=NULL;
HYPRE_Int vertexrange_start,vertexrange_end;
HYPRE_Int *vertexrange_strong= NULL;
HYPRE_Int *vertexrange_nonlocal;
HYPRE_Int num_recvs,num_recvs_strong;
HYPRE_Int *recv_procs,*recv_procs_strong=NULL;
HYPRE_Int /* *zeros,*rownz,*/*rownz_diag,*rownz_offd;
HYPRE_Int nz;
HYPRE_Int nlocal;
HYPRE_Int one=1;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg (S);
hypre_MPI_Comm_size (comm,&mpisize);
hypre_MPI_Comm_rank (comm,&mpirank);
/* determine neighbor processors */
num_recvs = hypre_ParCSRCommPkgNumRecvs (comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs (comm_pkg);
pointrange = hypre_ParCSRMatrixRowStarts (S);
pointrange_nonlocal = hypre_CTAlloc (HYPRE_Int, 2*num_recvs);
vertexrange_nonlocal = hypre_CTAlloc (HYPRE_Int, 2*num_recvs);
#ifdef HYPRE_NO_GLOBAL_PARTITION
{
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends (comm_pkg);
HYPRE_Int *send_procs = hypre_ParCSRCommPkgSendProcs (comm_pkg);
HYPRE_Int *int_buf_data = hypre_CTAlloc (HYPRE_Int,4*num_sends);
HYPRE_Int *int_buf_data2 = int_buf_data + 2*num_sends;
hypre_MPI_Request *sendrequest,*recvrequest;
nlocal = vertexrange[1] - vertexrange[0];
pointrange_start = pointrange[0];
pointrange_end = pointrange[1];
vertexrange_start = vertexrange[0];
vertexrange_end = vertexrange[1];
sendrequest = hypre_CTAlloc (hypre_MPI_Request,2*(num_sends+num_recvs));
recvrequest = sendrequest+2*num_sends;
for (i=0;i<num_recvs;i++) {
hypre_MPI_Irecv (pointrange_nonlocal+2*i,2,HYPRE_MPI_INT,recv_procs[i],tag_pointrange,comm,&recvrequest[2*i]);
hypre_MPI_Irecv (vertexrange_nonlocal+2*i,2,HYPRE_MPI_INT,recv_procs[i],tag_vertexrange,comm,&recvrequest[2*i+1]);
}
for (i=0;i<num_sends;i++) {
int_buf_data[2*i] = pointrange_start;
int_buf_data[2*i+1] = pointrange_end;
int_buf_data2[2*i] = vertexrange_start;
int_buf_data2[2*i+1] = vertexrange_end;
hypre_MPI_Isend (int_buf_data+2*i,2,HYPRE_MPI_INT,send_procs[i],tag_pointrange,comm,&sendrequest[2*i]);
hypre_MPI_Isend (int_buf_data2+2*i,2,HYPRE_MPI_INT,send_procs[i],tag_vertexrange,comm,&sendrequest[2*i+1]);
}
hypre_MPI_Waitall (2*(num_sends+num_recvs),sendrequest,hypre_MPI_STATUSES_IGNORE);
hypre_TFree (int_buf_data);
hypre_TFree (sendrequest);
}
#else
nlocal = vertexrange[mpirank+1] - vertexrange[mpirank];
pointrange_start = pointrange[mpirank];
pointrange_end = pointrange[mpirank+1];
vertexrange_start = vertexrange[mpirank];
vertexrange_end = vertexrange[mpirank+1];
for (i=0;i<num_recvs;i++) {
pointrange_nonlocal[2*i] = pointrange[recv_procs[i]];
pointrange_nonlocal[2*i+1] = pointrange[recv_procs[i]+1];
vertexrange_nonlocal[2*i] = vertexrange[recv_procs[i]];
vertexrange_nonlocal[2*i+1] = vertexrange[recv_procs[i]+1];
}
#endif
/* now we have the array recv_procs. However, it may contain too many entries as it is
inherited from A. We now have to determine the subset which contains only the
strongly connected neighbors */
if (num_cols_offd) {
S_offd_j = hypre_CSRMatrixJ(S_offd);
recv_procs_strong = hypre_CTAlloc (HYPRE_Int,num_recvs);
memset (recv_procs_strong,0,num_recvs*sizeof(HYPRE_Int));
/* don't forget to shorten the pointrange and vertexrange arrays accordingly */
pointrange_strong = hypre_CTAlloc (HYPRE_Int,2*num_recvs);
memset (pointrange_strong,0,2*num_recvs*sizeof(HYPRE_Int));
vertexrange_strong = hypre_CTAlloc (HYPRE_Int,2*num_recvs);
memset (vertexrange_strong,0,2*num_recvs*sizeof(HYPRE_Int));
for (i=0;i<num_variables;i++)
for (j=S_offd_i[i];j<S_offd_i[i+1];j++) {
jj = col_map_offd[S_offd_j[j]];
for (p=0;p<num_recvs;p++) /* S_offd_j is NOT sorted! */
if (jj >= pointrange_nonlocal[2*p] && jj < pointrange_nonlocal[2*p+1]) break;
#if 0
hypre_printf ("Processor %d, remote point %d on processor %d\n",mpirank,jj,recv_procs[p]);
#endif
recv_procs_strong [p]=1;
}
for (p=0,num_recvs_strong=0;p<num_recvs;p++) {
if (recv_procs_strong[p]) {
recv_procs_strong[num_recvs_strong]=recv_procs[p];
pointrange_strong[2*num_recvs_strong] = pointrange_nonlocal[2*p];
pointrange_strong[2*num_recvs_strong+1] = pointrange_nonlocal[2*p+1];
vertexrange_strong[2*num_recvs_strong] = vertexrange_nonlocal[2*p];
vertexrange_strong[2*num_recvs_strong+1] = vertexrange_nonlocal[2*p+1];
num_recvs_strong++;
}
}
}
else num_recvs_strong=0;
hypre_TFree (pointrange_nonlocal);
hypre_TFree (vertexrange_nonlocal);
rownz_diag = hypre_CTAlloc (HYPRE_Int,2*nlocal);
rownz_offd = rownz_diag + nlocal;
for (p=0,nz=0;p<num_recvs_strong;p++) {
nz += vertexrange_strong[2*p+1]-vertexrange_strong[2*p];
}
for (m=0;m<nlocal;m++) {
rownz_diag[m]=nlocal-1;
rownz_offd[m]=nz;
}
HYPRE_IJMatrixCreate(comm, vertexrange_start, vertexrange_end-1, vertexrange_start, vertexrange_end-1, &ijmatrix);
HYPRE_IJMatrixSetObjectType(ijmatrix, HYPRE_PARCSR);
HYPRE_IJMatrixSetDiagOffdSizes (ijmatrix, rownz_diag, rownz_offd);
HYPRE_IJMatrixInitialize(ijmatrix);
hypre_TFree (rownz_diag);
/* initialize graph */
weight = -1;
for (m=vertexrange_start;m<vertexrange_end;m++) {
for (p=0;p<num_recvs_strong;p++) {
for (n=vertexrange_strong[2*p];n<vertexrange_strong[2*p+1];n++) {
ierr = HYPRE_IJMatrixAddToValues (ijmatrix,1,&one,&m,&n,&weight);
#if 0
if (ierr) hypre_printf ("Processor %d: error %d while initializing graphs at (%d, %d)\n",mpirank,ierr,m,n);
#endif
}
}
}
/* weight graph */
for (i=0;i<num_variables;i++) {
for (j=S_offd_i[i];j<S_offd_i[i+1];j++) {
jj = S_offd_j[j]; /* jj is not a global index!!! */
/* determine processor */
for (p=0;p<num_recvs_strong;p++)
if (col_map_offd[jj] >= pointrange_strong[2*p] && col_map_offd[jj] < pointrange_strong[2*p+1]) break;
ip=recv_procs_strong[p];
/* loop over all coarse grids constructed on this processor domain */
for (m=vertexrange_start;m<vertexrange_end;m++) {
/* loop over all coarse grids constructed on neighbor processor domain */
for (n=vertexrange_strong[2*p];n<vertexrange_strong[2*p+1];n++) {
/* coarse grid counting inside gridpartition->local/gridpartition->nonlocal starts with one
while counting inside range starts with zero */
if (CF_marker[i]-1==m && CF_marker_offd[jj]-1==n)
/* C-C-coupling */
weight = -1;
else if ( (CF_marker[i]-1==m && (CF_marker_offd[jj]==0 || CF_marker_offd[jj]-1!=n) )
|| ( (CF_marker[i]==0 || CF_marker[i]-1!=m) && CF_marker_offd[jj]-1==n ) )
/* C-F-coupling */
weight = 0;
else weight = -8; /* F-F-coupling */
ierr = HYPRE_IJMatrixAddToValues (ijmatrix,1,&one,&m,&n,&weight);
#if 0
if (ierr) hypre_printf ("Processor %d: error %d while adding %lf to entry (%d, %d)\n",mpirank,ierr,weight,m,n);
#endif
}
}
}
}
/* assemble */
HYPRE_IJMatrixAssemble (ijmatrix);
/*if (num_recvs_strong) {*/
hypre_TFree (recv_procs_strong);
hypre_TFree (pointrange_strong);
hypre_TFree (vertexrange_strong);
/*} */
*ijG = ijmatrix;
return (ierr);
}
HYPRE_Int hypre_seqAMGSetup( hypre_ParAMGData *amg_data,
HYPRE_Int p_level,
HYPRE_Int coarse_threshold)
{
/* Par Data Structure variables */
hypre_ParCSRMatrix **Par_A_array = hypre_ParAMGDataAArray(amg_data);
MPI_Comm comm = hypre_ParCSRMatrixComm(Par_A_array[0]);
MPI_Comm new_comm, seq_comm;
hypre_ParCSRMatrix *A_seq = NULL;
hypre_CSRMatrix *A_seq_diag;
hypre_CSRMatrix *A_seq_offd;
hypre_ParVector *F_seq = NULL;
hypre_ParVector *U_seq = NULL;
hypre_ParCSRMatrix *A;
HYPRE_Int **dof_func_array;
HYPRE_Int num_procs, my_id;
HYPRE_Int not_finished_coarsening;
HYPRE_Int level;
HYPRE_Solver coarse_solver;
/* misc */
dof_func_array = hypre_ParAMGDataDofFuncArray(amg_data);
/*MPI Stuff */
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
/*initial */
level = p_level;
not_finished_coarsening = 1;
/* convert A at this level to sequential */
A = Par_A_array[level];
{
double *A_seq_data = NULL;
HYPRE_Int *A_seq_i = NULL;
HYPRE_Int *A_seq_offd_i = NULL;
HYPRE_Int *A_seq_j = NULL;
double *A_tmp_data = NULL;
HYPRE_Int *A_tmp_i = NULL;
HYPRE_Int *A_tmp_j = NULL;
HYPRE_Int *info, *displs, *displs2;
HYPRE_Int i, j, size, num_nonzeros, total_nnz, cnt;
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int num_rows = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int first_row_index = hypre_ParCSRMatrixFirstRowIndex(A);
hypre_MPI_Group orig_group, new_group;
HYPRE_Int *ranks, new_num_procs, *row_starts;
info = hypre_CTAlloc(HYPRE_Int, num_procs);
hypre_MPI_Allgather(&num_rows, 1, HYPRE_MPI_INT, info, 1, HYPRE_MPI_INT, comm);
ranks = hypre_CTAlloc(HYPRE_Int, num_procs);
new_num_procs = 0;
for (i=0; i < num_procs; i++)
if (info[i])
{
ranks[new_num_procs] = i;
info[new_num_procs++] = info[i];
}
MPI_Comm_group(comm, &orig_group);
hypre_MPI_Group_incl(orig_group, new_num_procs, ranks, &new_group);
MPI_Comm_create(comm, new_group, &new_comm);
hypre_MPI_Group_free(&new_group);
hypre_MPI_Group_free(&orig_group);
if (num_rows)
{
/* alloc space in seq data structure only for participating procs*/
HYPRE_BoomerAMGCreate(&coarse_solver);
HYPRE_BoomerAMGSetMaxRowSum(coarse_solver,
hypre_ParAMGDataMaxRowSum(amg_data));
HYPRE_BoomerAMGSetStrongThreshold(coarse_solver,
hypre_ParAMGDataStrongThreshold(amg_data));
HYPRE_BoomerAMGSetCoarsenType(coarse_solver,
hypre_ParAMGDataCoarsenType(amg_data));
HYPRE_BoomerAMGSetInterpType(coarse_solver,
hypre_ParAMGDataInterpType(amg_data));
HYPRE_BoomerAMGSetTruncFactor(coarse_solver,
hypre_ParAMGDataTruncFactor(amg_data));
HYPRE_BoomerAMGSetPMaxElmts(coarse_solver,
hypre_ParAMGDataPMaxElmts(amg_data));
if (hypre_ParAMGDataUserRelaxType(amg_data) > -1)
HYPRE_BoomerAMGSetRelaxType(coarse_solver,
hypre_ParAMGDataUserRelaxType(amg_data));
HYPRE_BoomerAMGSetRelaxOrder(coarse_solver,
hypre_ParAMGDataRelaxOrder(amg_data));
HYPRE_BoomerAMGSetRelaxWt(coarse_solver,
hypre_ParAMGDataUserRelaxWeight(amg_data));
if (hypre_ParAMGDataUserNumSweeps(amg_data) > -1)
HYPRE_BoomerAMGSetNumSweeps(coarse_solver,
hypre_ParAMGDataUserNumSweeps(amg_data));
HYPRE_BoomerAMGSetNumFunctions(coarse_solver,
hypre_ParAMGDataNumFunctions(amg_data));
HYPRE_BoomerAMGSetMaxIter(coarse_solver, 1);
HYPRE_BoomerAMGSetTol(coarse_solver, 0);
/* Create CSR Matrix, will be Diag part of new matrix */
A_tmp_i = hypre_CTAlloc(HYPRE_Int, num_rows+1);
A_tmp_i[0] = 0;
for (i=1; i < num_rows+1; i++)
A_tmp_i[i] = A_diag_i[i]-A_diag_i[i-1]+A_offd_i[i]-A_offd_i[i-1];
num_nonzeros = A_offd_i[num_rows]+A_diag_i[num_rows];
A_tmp_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros);
A_tmp_data = hypre_CTAlloc(double, num_nonzeros);
cnt = 0;
for (i=0; i < num_rows; i++)
{
for (j=A_diag_i[i]; j < A_diag_i[i+1]; j++)
{
A_tmp_j[cnt] = A_diag_j[j]+first_row_index;
A_tmp_data[cnt++] = A_diag_data[j];
}
for (j=A_offd_i[i]; j < A_offd_i[i+1]; j++)
{
A_tmp_j[cnt] = col_map_offd[A_offd_j[j]];
A_tmp_data[cnt++] = A_offd_data[j];
}
}
displs = hypre_CTAlloc(HYPRE_Int, new_num_procs+1);
displs[0] = 0;
for (i=1; i < new_num_procs+1; i++)
displs[i] = displs[i-1]+info[i-1];
size = displs[new_num_procs];
A_seq_i = hypre_CTAlloc(HYPRE_Int, size+1);
A_seq_offd_i = hypre_CTAlloc(HYPRE_Int, size+1);
hypre_MPI_Allgatherv ( &A_tmp_i[1], num_rows, HYPRE_MPI_INT, &A_seq_i[1], info,
displs, HYPRE_MPI_INT, new_comm );
displs2 = hypre_CTAlloc(HYPRE_Int, new_num_procs+1);
A_seq_i[0] = 0;
displs2[0] = 0;
for (j=1; j < displs[1]; j++)
A_seq_i[j] = A_seq_i[j]+A_seq_i[j-1];
for (i=1; i < new_num_procs; i++)
{
for (j=displs[i]; j < displs[i+1]; j++)
{
A_seq_i[j] = A_seq_i[j]+A_seq_i[j-1];
}
}
A_seq_i[size] = A_seq_i[size]+A_seq_i[size-1];
displs2[new_num_procs] = A_seq_i[size];
for (i=1; i < new_num_procs+1; i++)
{
displs2[i] = A_seq_i[displs[i]];
info[i-1] = displs2[i] - displs2[i-1];
}
total_nnz = displs2[new_num_procs];
A_seq_j = hypre_CTAlloc(HYPRE_Int, total_nnz);
A_seq_data = hypre_CTAlloc(double, total_nnz);
hypre_MPI_Allgatherv ( A_tmp_j, num_nonzeros, HYPRE_MPI_INT,
A_seq_j, info, displs2,
HYPRE_MPI_INT, new_comm );
hypre_MPI_Allgatherv ( A_tmp_data, num_nonzeros, hypre_MPI_DOUBLE,
A_seq_data, info, displs2,
hypre_MPI_DOUBLE, new_comm );
hypre_TFree(displs);
hypre_TFree(displs2);
hypre_TFree(A_tmp_i);
hypre_TFree(A_tmp_j);
hypre_TFree(A_tmp_data);
row_starts = hypre_CTAlloc(HYPRE_Int,2);
row_starts[0] = 0;
row_starts[1] = size;
/* Create 1 proc communicator */
seq_comm = hypre_MPI_COMM_SELF;
A_seq = hypre_ParCSRMatrixCreate(seq_comm,size,size,
row_starts, row_starts,
0,total_nnz,0);
A_seq_diag = hypre_ParCSRMatrixDiag(A_seq);
A_seq_offd = hypre_ParCSRMatrixOffd(A_seq);
hypre_CSRMatrixData(A_seq_diag) = A_seq_data;
hypre_CSRMatrixI(A_seq_diag) = A_seq_i;
hypre_CSRMatrixJ(A_seq_diag) = A_seq_j;
hypre_CSRMatrixI(A_seq_offd) = A_seq_offd_i;
F_seq = hypre_ParVectorCreate(seq_comm, size, row_starts);
U_seq = hypre_ParVectorCreate(seq_comm, size, row_starts);
hypre_ParVectorOwnsPartitioning(F_seq) = 0;
hypre_ParVectorOwnsPartitioning(U_seq) = 0;
hypre_ParVectorInitialize(F_seq);
hypre_ParVectorInitialize(U_seq);
hypre_BoomerAMGSetup(coarse_solver,A_seq,F_seq,U_seq);
hypre_ParAMGDataCoarseSolver(amg_data) = coarse_solver;
hypre_ParAMGDataACoarse(amg_data) = A_seq;
hypre_ParAMGDataFCoarse(amg_data) = F_seq;
hypre_ParAMGDataUCoarse(amg_data) = U_seq;
hypre_ParAMGDataNewComm(amg_data) = new_comm;
}
hypre_TFree(info);
hypre_TFree(ranks);
}
return 0;
}
HYPRE_Int
hypre_GetCommPkgRTFromCommPkgA( hypre_ParCSRMatrix *RT,
hypre_ParCSRMatrix *A,
HYPRE_Int *fine_to_coarse_offd)
{
MPI_Comm comm = hypre_ParCSRMatrixComm(RT);
hypre_ParCSRCommPkg *comm_pkg_A = hypre_ParCSRMatrixCommPkg(A);
HYPRE_Int num_recvs_A = hypre_ParCSRCommPkgNumRecvs(comm_pkg_A);
HYPRE_Int *recv_procs_A = hypre_ParCSRCommPkgRecvProcs(comm_pkg_A);
HYPRE_Int *recv_vec_starts_A = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg_A);
HYPRE_Int num_sends_A = hypre_ParCSRCommPkgNumSends(comm_pkg_A);
HYPRE_Int *send_procs_A = hypre_ParCSRCommPkgSendProcs(comm_pkg_A);
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int num_recvs_RT;
HYPRE_Int *recv_procs_RT;
HYPRE_Int *recv_vec_starts_RT;
HYPRE_Int num_sends_RT;
HYPRE_Int *send_procs_RT;
HYPRE_Int *send_map_starts_RT;
HYPRE_Int *send_map_elmts_RT;
HYPRE_Int *col_map_offd_RT = hypre_ParCSRMatrixColMapOffd(RT);
HYPRE_Int num_cols_offd_RT = hypre_CSRMatrixNumCols( hypre_ParCSRMatrixOffd(RT));
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(RT);
HYPRE_Int i, j;
HYPRE_Int vec_len, vec_start;
HYPRE_Int num_procs, my_id;
HYPRE_Int ierr = 0;
HYPRE_Int num_requests;
HYPRE_Int offd_col, proc_num;
HYPRE_Int *proc_mark;
HYPRE_Int *change_array;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
/*--------------------------------------------------------------------------
* determine num_recvs, recv_procs and recv_vec_starts for RT
*--------------------------------------------------------------------------*/
proc_mark = hypre_CTAlloc(HYPRE_Int, num_recvs_A);
for (i=0; i < num_recvs_A; i++)
proc_mark[i] = 0;
proc_num = 0;
num_recvs_RT = 0;
if (num_cols_offd_RT)
{
for (i=0; i < num_recvs_A; i++)
{
for (j=recv_vec_starts_A[i]; j<recv_vec_starts_A[i+1]; j++)
{
offd_col = col_map_offd_RT[proc_num];
if (offd_col == j)
{
proc_mark[i]++;
proc_num++;
if (proc_num == num_cols_offd_RT) break;
}
}
if (proc_mark[i]) num_recvs_RT++;
if (proc_num == num_cols_offd_RT) break;
}
}
for (i=0; i < num_cols_offd_RT; i++)
col_map_offd_RT[i] = fine_to_coarse_offd[col_map_offd_RT[i]];
recv_procs_RT = hypre_CTAlloc(HYPRE_Int,num_recvs_RT);
recv_vec_starts_RT = hypre_CTAlloc(HYPRE_Int, num_recvs_RT+1);
j = 0;
recv_vec_starts_RT[0] = 0;
for (i=0; i < num_recvs_A; i++)
if (proc_mark[i])
{
recv_procs_RT[j] = recv_procs_A[i];
recv_vec_starts_RT[j+1] = recv_vec_starts_RT[j]+proc_mark[i];
j++;
}
/*--------------------------------------------------------------------------
* send num_changes to recv_procs_A and receive change_array from send_procs_A
*--------------------------------------------------------------------------*/
num_requests = num_recvs_A+num_sends_A;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
status = hypre_CTAlloc(hypre_MPI_Status, num_requests);
change_array = hypre_CTAlloc(HYPRE_Int, num_sends_A);
j = 0;
for (i=0; i < num_sends_A; i++)
hypre_MPI_Irecv(&change_array[i],1,HYPRE_MPI_INT,send_procs_A[i],0,comm,
&requests[j++]);
for (i=0; i < num_recvs_A; i++)
hypre_MPI_Isend(&proc_mark[i],1,HYPRE_MPI_INT,recv_procs_A[i],0,comm,
&requests[j++]);
hypre_MPI_Waitall(num_requests,requests,status);
hypre_TFree(proc_mark);
/*--------------------------------------------------------------------------
* if change_array[i] is 0 , omit send_procs_A[i] in send_procs_RT
*--------------------------------------------------------------------------*/
num_sends_RT = 0;
for (i=0; i < num_sends_A; i++)
if (change_array[i])
{
num_sends_RT++;
}
send_procs_RT = hypre_CTAlloc(HYPRE_Int, num_sends_RT);
send_map_starts_RT = hypre_CTAlloc(HYPRE_Int, num_sends_RT+1);
j = 0;
send_map_starts_RT[0] = 0;
for (i=0; i < num_sends_A; i++)
if (change_array[i])
{
send_procs_RT[j] = send_procs_A[i];
send_map_starts_RT[j+1] = send_map_starts_RT[j]+change_array[i];
j++;
}
/*--------------------------------------------------------------------------
* generate send_map_elmts
*--------------------------------------------------------------------------*/
send_map_elmts_RT = hypre_CTAlloc(HYPRE_Int,send_map_starts_RT[num_sends_RT]);
j = 0;
for (i=0; i < num_sends_RT; i++)
{
vec_start = send_map_starts_RT[i];
vec_len = send_map_starts_RT[i+1]-vec_start;
hypre_MPI_Irecv(&send_map_elmts_RT[vec_start],vec_len,HYPRE_MPI_INT,
send_procs_RT[i],0,comm,&requests[j++]);
}
for (i=0; i < num_recvs_RT; i++)
{
vec_start = recv_vec_starts_RT[i];
vec_len = recv_vec_starts_RT[i+1] - vec_start;
hypre_MPI_Isend(&col_map_offd_RT[vec_start],vec_len,HYPRE_MPI_INT,
recv_procs_RT[i],0,comm,&requests[j++]);
}
hypre_MPI_Waitall(j,requests,status);
for (i=0; i < send_map_starts_RT[num_sends_RT]; i++)
send_map_elmts_RT[i] -= first_col_diag;
comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg) = comm;
hypre_ParCSRCommPkgNumSends(comm_pkg) = num_sends_RT;
hypre_ParCSRCommPkgNumRecvs(comm_pkg) = num_recvs_RT;
hypre_ParCSRCommPkgSendProcs(comm_pkg) = send_procs_RT;
hypre_ParCSRCommPkgRecvProcs(comm_pkg) = recv_procs_RT;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg) = recv_vec_starts_RT;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg) = send_map_starts_RT;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg) = send_map_elmts_RT;
hypre_TFree(status);
hypre_TFree(requests);
hypre_ParCSRMatrixCommPkg(RT) = comm_pkg;
hypre_TFree(change_array);
return ierr;
}
hypre_ParCSRBlockMatrix *
hypre_ParCSRBlockMatrixConvertFromParCSRMatrix(hypre_ParCSRMatrix *matrix,
HYPRE_Int matrix_C_block_size )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(matrix);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(matrix);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(matrix);
HYPRE_Int global_num_rows = hypre_ParCSRMatrixGlobalNumRows(matrix);
HYPRE_Int global_num_cols = hypre_ParCSRMatrixGlobalNumCols(matrix);
HYPRE_Int *row_starts = hypre_ParCSRMatrixRowStarts(matrix);
HYPRE_Int *col_starts = hypre_ParCSRMatrixColStarts(matrix);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int *col_map_offd = hypre_ParCSRBlockMatrixColMapOffd(matrix);
HYPRE_Int *map_to_node=NULL, *counter=NULL, *col_in_j_map=NULL;
HYPRE_Int *matrix_C_col_map_offd = NULL;
HYPRE_Int matrix_C_num_cols_offd;
HYPRE_Int matrix_C_num_nonzeros_offd;
HYPRE_Int num_rows, num_nodes;
HYPRE_Int *offd_i = hypre_CSRMatrixI(offd);
HYPRE_Int *offd_j = hypre_CSRMatrixJ(offd);
HYPRE_Complex * offd_data = hypre_CSRMatrixData(offd);
hypre_ParCSRBlockMatrix *matrix_C;
HYPRE_Int *matrix_C_row_starts;
HYPRE_Int *matrix_C_col_starts;
hypre_CSRBlockMatrix *matrix_C_diag;
hypre_CSRBlockMatrix *matrix_C_offd;
HYPRE_Int *matrix_C_offd_i=NULL, *matrix_C_offd_j = NULL;
HYPRE_Complex *matrix_C_offd_data = NULL;
HYPRE_Int num_procs, i, j, k, k_map, count, index, start_index, pos, row;
hypre_MPI_Comm_size(comm,&num_procs);
#ifdef HYPRE_NO_GLOBAL_PARTITION
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, 2);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, 2);
for(i = 0; i < 2; i++)
{
matrix_C_row_starts[i] = row_starts[i]/matrix_C_block_size;
matrix_C_col_starts[i] = col_starts[i]/matrix_C_block_size;
}
#else
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
for(i = 0; i < num_procs + 1; i++)
{
matrix_C_row_starts[i] = row_starts[i]/matrix_C_block_size;
matrix_C_col_starts[i] = col_starts[i]/matrix_C_block_size;
}
#endif
/************* create the diagonal part ************/
matrix_C_diag = hypre_CSRBlockMatrixConvertFromCSRMatrix(diag,
matrix_C_block_size);
/******* the offd part *******************/
/* can't use the same function for the offd part - because this isn't square
and the offd j entries aren't global numbering (have to consider the offd
map) - need to look at col_map_offd first */
/* figure out the new number of offd columns (num rows is same as diag) */
num_cols_offd = hypre_CSRMatrixNumCols(offd);
num_rows = hypre_CSRMatrixNumRows(diag);
num_nodes = num_rows/matrix_C_block_size;
matrix_C_offd_i = hypre_CTAlloc(HYPRE_Int, num_nodes + 1);
matrix_C_num_cols_offd = 0;
matrix_C_offd_i[0] = 0;
matrix_C_num_nonzeros_offd = 0;
if (num_cols_offd)
{
map_to_node = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
matrix_C_num_cols_offd = 1;
map_to_node[0] = col_map_offd[0]/matrix_C_block_size;
for (i=1; i < num_cols_offd; i++)
{
map_to_node[i] = col_map_offd[i]/matrix_C_block_size;
if (map_to_node[i] > map_to_node[i-1]) matrix_C_num_cols_offd++;
}
matrix_C_col_map_offd = hypre_CTAlloc(HYPRE_Int, matrix_C_num_cols_offd);
col_in_j_map = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
matrix_C_col_map_offd[0] = map_to_node[0];
col_in_j_map[0] = 0;
count = 1;
j = 1;
/* fill in the col_map_off_d - these are global numbers. Then we need to
map these to j entries (these have local numbers) */
for (i=1; i < num_cols_offd; i++)
{
if (map_to_node[i] > map_to_node[i-1])
{
matrix_C_col_map_offd[count++] = map_to_node[i];
}
col_in_j_map[j++] = count - 1;
}
/* now figure the nonzeros */
matrix_C_num_nonzeros_offd = 0;
counter = hypre_CTAlloc(HYPRE_Int, matrix_C_num_cols_offd);
for (i=0; i < matrix_C_num_cols_offd; i++)
counter[i] = -1;
for (i=0; i < num_nodes; i++) /* for each block row */
{
matrix_C_offd_i[i] = matrix_C_num_nonzeros_offd;
for (j=0; j < matrix_C_block_size; j++)
{
row = i*matrix_C_block_size+j;
for (k=offd_i[row]; k < offd_i[row+1]; k++) /* go through single row */
{
k_map = col_in_j_map[offd_j[k]]; /*nodal col - see if this has
been in this block row (i)
already*/
if (counter[k_map] < i) /* not yet counted for this nodal row */
{
counter[k_map] = i;
matrix_C_num_nonzeros_offd++;
}
}
}
}
/* fill in final i entry */
matrix_C_offd_i[num_nodes] = matrix_C_num_nonzeros_offd;
}
/* create offd matrix */
matrix_C_offd = hypre_CSRBlockMatrixCreate(matrix_C_block_size, num_nodes,
matrix_C_num_cols_offd,
matrix_C_num_nonzeros_offd);
/* assign i */
hypre_CSRBlockMatrixI(matrix_C_offd) = matrix_C_offd_i;
/* create (and allocate j and data) */
if (matrix_C_num_nonzeros_offd)
{
matrix_C_offd_j = hypre_CTAlloc(HYPRE_Int, matrix_C_num_nonzeros_offd);
matrix_C_offd_data =
hypre_CTAlloc(HYPRE_Complex,
matrix_C_num_nonzeros_offd*matrix_C_block_size*
matrix_C_block_size);
hypre_CSRBlockMatrixJ(matrix_C_offd) = matrix_C_offd_j;
hypre_CSRMatrixData(matrix_C_offd) = matrix_C_offd_data;
for (i=0; i < matrix_C_num_cols_offd; i++)
counter[i] = -1;
index = 0; /*keep track of entry in matrix_C_offd_j*/
start_index = 0;
for (i=0; i < num_nodes; i++) /* for each block row */
{
for (j=0; j < matrix_C_block_size; j++) /* for each row in block */
{
row = i*matrix_C_block_size+j;
for (k=offd_i[row]; k < offd_i[row+1]; k++) /* go through single row's cols */
{
k_map = col_in_j_map[offd_j[k]]; /*nodal col for off_d */
if (counter[k_map] < start_index) /* not yet counted for this nodal row */
{
counter[k_map] = index;
matrix_C_offd_j[index] = k_map;
/*copy the data: which position (corresponds to j array) + which row + which col */
pos = (index * matrix_C_block_size * matrix_C_block_size) + (j * matrix_C_block_size) +
col_map_offd[offd_j[k]]%matrix_C_block_size;
matrix_C_offd_data[pos] = offd_data[k];
index ++;
}
else /* this col has already been listed for this row */
{
/*copy the data: which position (corresponds to j array) + which row + which col */
pos = (counter[k_map]* matrix_C_block_size * matrix_C_block_size) + (j * matrix_C_block_size) +
col_map_offd[offd_j[k]]%matrix_C_block_size;
matrix_C_offd_data[pos] = offd_data[k];
}
}
}
start_index = index; /* first index for current nodal row */
}
}
/* *********create the new matrix *************/
matrix_C = hypre_ParCSRBlockMatrixCreate(comm, matrix_C_block_size,
global_num_rows/matrix_C_block_size,
global_num_cols/matrix_C_block_size, matrix_C_row_starts,
matrix_C_col_starts, matrix_C_num_cols_offd,
hypre_CSRBlockMatrixNumNonzeros(matrix_C_diag),
matrix_C_num_nonzeros_offd);
/* use the diag and off diag matrices we have already created */
hypre_CSRBlockMatrixDestroy(hypre_ParCSRMatrixDiag(matrix_C));
hypre_ParCSRBlockMatrixDiag(matrix_C) = matrix_C_diag;
hypre_CSRBlockMatrixDestroy(hypre_ParCSRMatrixOffd(matrix_C));
hypre_ParCSRBlockMatrixOffd(matrix_C) = matrix_C_offd;
hypre_ParCSRMatrixColMapOffd(matrix_C) = matrix_C_col_map_offd;
/* *********don't bother to copy the comm_pkg *************/
hypre_ParCSRBlockMatrixCommPkg(matrix_C) = NULL;
/* CLEAN UP !!!! */
hypre_TFree(map_to_node);
hypre_TFree(col_in_j_map);
hypre_TFree(counter);
return matrix_C;
}
HYPRE_ParCSRMatrix
GenerateRotate7pt( MPI_Comm comm,
HYPRE_Int nx,
HYPRE_Int ny,
HYPRE_Int P,
HYPRE_Int Q,
HYPRE_Int p,
HYPRE_Int q,
HYPRE_Real alpha,
HYPRE_Real eps )
{
hypre_ParCSRMatrix *A;
hypre_CSRMatrix *diag;
hypre_CSRMatrix *offd;
HYPRE_Int *diag_i;
HYPRE_Int *diag_j;
HYPRE_Real *diag_data;
HYPRE_Int *offd_i;
HYPRE_Int *offd_j;
HYPRE_Real *offd_data;
HYPRE_Real *value;
HYPRE_Real ac, bc, cc, s, c, pi, x;
HYPRE_Int *global_part;
HYPRE_Int ix, iy;
HYPRE_Int cnt, o_cnt;
HYPRE_Int local_num_rows;
HYPRE_Int *col_map_offd;
HYPRE_Int *work;
HYPRE_Int row_index;
HYPRE_Int i,j;
HYPRE_Int nx_local, ny_local;
HYPRE_Int nx_size, ny_size;
HYPRE_Int num_cols_offd;
HYPRE_Int grid_size;
HYPRE_Int *nx_part;
HYPRE_Int *ny_part;
HYPRE_Int num_procs, my_id;
HYPRE_Int P_busy, Q_busy;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
grid_size = nx*ny;
value = hypre_CTAlloc(HYPRE_Real,4);
pi = 4.0*atan(1.0);
x = pi*alpha/180.0;
s = sin(x);
c = cos(x);
ac = -(c*c + eps*s*s);
bc = 2.0*(1.0 - eps)*s*c;
cc = -(s*s + eps*c*c);
value[0] = -2*(2*ac+bc+2*cc);
value[1] = 2*ac+bc;
value[2] = bc+2*cc;
value[3] = -bc;
hypre_GeneratePartitioning(nx,P,&nx_part);
hypre_GeneratePartitioning(ny,Q,&ny_part);
global_part = hypre_CTAlloc(HYPRE_Int,P*Q+1);
global_part[0] = 0;
cnt = 1;
for (iy = 0; iy < Q; iy++)
{
ny_size = ny_part[iy+1]-ny_part[iy];
for (ix = 0; ix < P; ix++)
{
nx_size = nx_part[ix+1] - nx_part[ix];
global_part[cnt] = global_part[cnt-1];
global_part[cnt++] += nx_size*ny_size;
}
}
nx_local = nx_part[p+1] - nx_part[p];
ny_local = ny_part[q+1] - ny_part[q];
my_id = q*P + p;
num_procs = P*Q;
local_num_rows = nx_local*ny_local;
diag_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
offd_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
P_busy = hypre_min(nx,P);
Q_busy = hypre_min(ny,Q);
num_cols_offd = 0;
if (p) num_cols_offd += ny_local;
if (p < P_busy-1) num_cols_offd += ny_local;
if (q) num_cols_offd += nx_local;
if (q < Q_busy-1) num_cols_offd += nx_local;
if (p && q) num_cols_offd++;
if (p && q < Q_busy-1 ) num_cols_offd++;
if (p < P_busy-1 && q ) num_cols_offd++;
if (p < P_busy-1 && q < Q_busy-1 ) num_cols_offd++;
if (!local_num_rows) num_cols_offd = 0;
col_map_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
cnt = 0;
o_cnt = 0;
diag_i[0] = 0;
offd_i[0] = 0;
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
cnt++;
o_cnt++;
diag_i[cnt] = diag_i[cnt-1];
offd_i[o_cnt] = offd_i[o_cnt-1];
diag_i[cnt]++;
if (iy > ny_part[q])
{
diag_i[cnt]++;
if (ix > nx_part[p])
{
diag_i[cnt]++;
}
else
{
if (ix)
offd_i[o_cnt]++;
}
}
else
{
if (iy)
{
offd_i[o_cnt]++;
if (ix > nx_part[p])
{
offd_i[o_cnt]++;
}
else if (ix)
{
offd_i[o_cnt]++;
}
}
}
if (ix > nx_part[p])
diag_i[cnt]++;
else
{
if (ix)
{
offd_i[o_cnt]++;
}
}
if (ix+1 < nx_part[p+1])
diag_i[cnt]++;
else
{
if (ix+1 < nx)
{
offd_i[o_cnt]++;
}
}
if (iy+1 < ny_part[q+1])
{
diag_i[cnt]++;
if (ix < nx_part[p+1]-1)
{
diag_i[cnt]++;
}
else
{
if (ix+1 < nx)
offd_i[o_cnt]++;
}
}
else
{
if (iy+1 < ny)
{
offd_i[o_cnt]++;
if (ix < nx_part[p+1]-1)
{
offd_i[o_cnt]++;
}
else if (ix < nx-1)
{
offd_i[o_cnt]++;
}
}
}
}
}
diag_j = hypre_CTAlloc(HYPRE_Int, diag_i[local_num_rows]);
diag_data = hypre_CTAlloc(HYPRE_Real, diag_i[local_num_rows]);
if (num_procs > 1)
{
offd_j = hypre_CTAlloc(HYPRE_Int, offd_i[local_num_rows]);
offd_data = hypre_CTAlloc(HYPRE_Real, offd_i[local_num_rows]);
}
row_index = 0;
cnt = 0;
o_cnt = 0;
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
diag_j[cnt] = row_index;
diag_data[cnt++] = value[0];
if (iy > ny_part[q])
{
if (ix > nx_part[p])
{
diag_j[cnt] = row_index-nx_local-1 ;
diag_data[cnt++] = value[3];
}
else
{
if (ix)
{
offd_j[o_cnt] = hypre_map2(ix-1,iy-1,p-1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
diag_j[cnt] = row_index-nx_local;
diag_data[cnt++] = value[2];
}
else
{
if (iy)
{
if (ix > nx_part[p])
{
offd_j[o_cnt] = hypre_map2(ix-1,iy-1,p,q-1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
else if (ix)
{
offd_j[o_cnt] = hypre_map2(ix-1,iy-1,p-1,q-1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
offd_j[o_cnt] = hypre_map2(ix,iy-1,p,q-1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[2];
}
}
if (ix > nx_part[p])
{
diag_j[cnt] = row_index-1;
diag_data[cnt++] = value[1];
}
else
{
if (ix)
{
offd_j[o_cnt] = hypre_map2(ix-1,iy,p-1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[1];
}
}
if (ix+1 < nx_part[p+1])
{
diag_j[cnt] = row_index+1;
diag_data[cnt++] = value[1];
}
else
{
if (ix+1 < nx)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy,p+1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[1];
}
}
if (iy+1 < ny_part[q+1])
{
diag_j[cnt] = row_index+nx_local;
diag_data[cnt++] = value[2];
if (ix < nx_part[p+1]-1)
{
diag_j[cnt] = row_index+nx_local+1 ;
diag_data[cnt++] = value[3];
}
else
{
if (ix+1 < nx)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy+1,p+1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
}
else
{
if (iy+1 < ny)
{
offd_j[o_cnt] = hypre_map2(ix,iy+1,p,q+1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[2];
if (ix < nx_part[p+1]-1)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy+1,p,q+1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
else if (ix < nx-1)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy+1,p+1,q+1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
}
row_index++;
}
}
if (num_procs > 1)
{
work = hypre_CTAlloc(HYPRE_Int,o_cnt);
for (i=0; i < o_cnt; i++)
work[i] = offd_j[i];
qsort0(work, 0, o_cnt-1);
col_map_offd[0] = work[0];
cnt = 0;
for (i=0; i < o_cnt; i++)
{
if (work[i] > col_map_offd[cnt])
{
cnt++;
col_map_offd[cnt] = work[i];
}
}
for (i=0; i < o_cnt; i++)
{
for (j=0; j < num_cols_offd; j++)
{
if (offd_j[i] == col_map_offd[j])
{
offd_j[i] = j;
break;
}
}
}
hypre_TFree(work);
}
A = hypre_ParCSRMatrixCreate(comm, grid_size, grid_size,
global_part, global_part, num_cols_offd,
diag_i[local_num_rows],
offd_i[local_num_rows]);
hypre_ParCSRMatrixColMapOffd(A) = col_map_offd;
diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrixI(diag) = diag_i;
hypre_CSRMatrixJ(diag) = diag_j;
hypre_CSRMatrixData(diag) = diag_data;
offd = hypre_ParCSRMatrixOffd(A);
hypre_CSRMatrixI(offd) = offd_i;
if (num_cols_offd)
{
hypre_CSRMatrixJ(offd) = offd_j;
hypre_CSRMatrixData(offd) = offd_data;
}
hypre_TFree(nx_part);
hypre_TFree(ny_part);
hypre_TFree(value);
return (HYPRE_ParCSRMatrix) A;
}
HYPRE_Int
hypre_BoomerAMGCreateNodalA(hypre_ParCSRMatrix *A,
HYPRE_Int num_functions,
HYPRE_Int *dof_func,
HYPRE_Int option,
HYPRE_Int diag_option,
hypre_ParCSRMatrix **AN_ptr)
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *row_starts = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int num_variables = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_nonzeros_offd = 0;
HYPRE_Int num_cols_offd = 0;
hypre_ParCSRMatrix *AN;
hypre_CSRMatrix *AN_diag;
HYPRE_Int *AN_diag_i;
HYPRE_Int *AN_diag_j;
double *AN_diag_data;
hypre_CSRMatrix *AN_offd;
HYPRE_Int *AN_offd_i;
HYPRE_Int *AN_offd_j;
double *AN_offd_data;
HYPRE_Int *col_map_offd_AN;
HYPRE_Int *new_col_map_offd;
HYPRE_Int *row_starts_AN;
HYPRE_Int AN_num_nonzeros_diag = 0;
HYPRE_Int AN_num_nonzeros_offd = 0;
HYPRE_Int num_cols_offd_AN;
HYPRE_Int new_num_cols_offd;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(A);
HYPRE_Int num_sends;
HYPRE_Int num_recvs;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int *new_send_map_elmts;
HYPRE_Int *recv_procs;
HYPRE_Int *recv_vec_starts;
hypre_ParCSRCommPkg *comm_pkg_AN;
HYPRE_Int *send_procs_AN;
HYPRE_Int *send_map_starts_AN;
HYPRE_Int *send_map_elmts_AN;
HYPRE_Int *recv_procs_AN;
HYPRE_Int *recv_vec_starts_AN;
HYPRE_Int i, j, k, k_map;
HYPRE_Int ierr = 0;
HYPRE_Int index, row;
HYPRE_Int start_index;
HYPRE_Int num_procs;
HYPRE_Int node, cnt;
HYPRE_Int mode;
HYPRE_Int new_send_elmts_size;
HYPRE_Int global_num_nodes;
HYPRE_Int num_nodes;
HYPRE_Int num_fun2;
HYPRE_Int *map_to_node;
HYPRE_Int *map_to_map;
HYPRE_Int *counter;
double sum;
double *data;
hypre_MPI_Comm_size(comm,&num_procs);
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
mode = fabs(option);
comm_pkg_AN = NULL;
col_map_offd_AN = NULL;
#ifdef HYPRE_NO_GLOBAL_PARTITION
row_starts_AN = hypre_CTAlloc(HYPRE_Int, 2);
for (i=0; i < 2; i++)
{
row_starts_AN[i] = row_starts[i]/num_functions;
if (row_starts_AN[i]*num_functions < row_starts[i])
{
hypre_printf("nodes not properly aligned or incomplete info!\n");
return (87);
}
}
global_num_nodes = hypre_ParCSRMatrixGlobalNumRows(A)/num_functions;
#else
row_starts_AN = hypre_CTAlloc(HYPRE_Int, num_procs+1);
for (i=0; i < num_procs+1; i++)
{
row_starts_AN[i] = row_starts[i]/num_functions;
if (row_starts_AN[i]*num_functions < row_starts[i])
{
hypre_printf("nodes not properly aligned or incomplete info!\n");
return (87);
}
}
global_num_nodes = row_starts_AN[num_procs];
#endif
num_nodes = num_variables/num_functions;
num_fun2 = num_functions*num_functions;
map_to_node = hypre_CTAlloc(HYPRE_Int, num_variables);
AN_diag_i = hypre_CTAlloc(HYPRE_Int, num_nodes+1);
counter = hypre_CTAlloc(HYPRE_Int, num_nodes);
for (i=0; i < num_variables; i++)
map_to_node[i] = i/num_functions;
for (i=0; i < num_nodes; i++)
counter[i] = -1;
AN_num_nonzeros_diag = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
AN_diag_i[i] = AN_num_nonzeros_diag;
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < i)
{
counter[k_map] = i;
AN_num_nonzeros_diag++;
}
}
row++;
}
}
AN_diag_i[num_nodes] = AN_num_nonzeros_diag;
AN_diag_j = hypre_CTAlloc(HYPRE_Int, AN_num_nonzeros_diag);
AN_diag_data = hypre_CTAlloc(double, AN_num_nonzeros_diag);
AN_diag = hypre_CSRMatrixCreate(num_nodes,num_nodes,AN_num_nonzeros_diag);
hypre_CSRMatrixI(AN_diag) = AN_diag_i;
hypre_CSRMatrixJ(AN_diag) = AN_diag_j;
hypre_CSRMatrixData(AN_diag) = AN_diag_data;
for (i=0; i < num_nodes; i++)
counter[i] = -1;
index = 0;
start_index = 0;
row = 0;
switch (mode)
{
case 1: /* frobenius norm */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = A_diag_data[k]*A_diag_data[k];
index++;
}
else
{
AN_diag_data[counter[k_map]] +=
A_diag_data[k]*A_diag_data[k];
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_diag; i++)
AN_diag_data[i] = sqrt(AN_diag_data[i]);
}
break;
case 2: /* sum of abs. value of all elements in each block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = fabs(A_diag_data[k]);
index++;
}
else
{
AN_diag_data[counter[k_map]] += fabs(A_diag_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_diag; i++)
AN_diag_data[i] /= num_fun2;
}
break;
case 3: /* largest element of each block (sets true value - not abs. value) */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = A_diag_data[k];
index++;
}
else
{
if (fabs(A_diag_data[k]) >
fabs(AN_diag_data[counter[k_map]]))
AN_diag_data[counter[k_map]] = A_diag_data[k];
}
}
row++;
}
start_index = index;
}
}
break;
case 4: /* inf. norm (row-sum) */
{
data = hypre_CTAlloc(double, AN_num_nonzeros_diag*num_functions);
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
data[index*num_functions + j] = fabs(A_diag_data[k]);
index++;
}
else
{
data[(counter[k_map])*num_functions + j] += fabs(A_diag_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_diag; i++)
{
AN_diag_data[i] = data[i*num_functions];
for (j=1; j< num_functions; j++)
{
AN_diag_data[i] = hypre_max( AN_diag_data[i],data[i*num_functions+j]);
}
}
hypre_TFree(data);
}
break;
case 6: /* sum of all elements in each block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = (A_diag_data[k]);
index++;
}
else
{
AN_diag_data[counter[k_map]] += (A_diag_data[k]);
}
}
row++;
}
start_index = index;
}
}
break;
}
if (diag_option ==1 )
{
/* make the diag entry the negative of the sum of off-diag entries (DO MORE BELOW) */
for (i=0; i < num_nodes; i++)
{
index = AN_diag_i[i];
sum = 0.0;
for (k = AN_diag_i[i]+1; k < AN_diag_i[i+1]; k++)
{
sum += AN_diag_data[k];
}
AN_diag_data[index] = -sum;
}
}
else if (diag_option == 2)
{
/* make all diagonal entries negative */
/* the diagonal is the first element listed in each row - */
for (i=0; i < num_nodes; i++)
{
index = AN_diag_i[i];
AN_diag_data[index] = - AN_diag_data[index];
}
}
num_nonzeros_offd = A_offd_i[num_variables];
AN_offd_i = hypre_CTAlloc(HYPRE_Int, num_nodes+1);
num_cols_offd_AN = 0;
if (comm_pkg)
{
comm_pkg_AN = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg_AN) = comm;
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
hypre_ParCSRCommPkgNumSends(comm_pkg_AN) = num_sends;
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
hypre_ParCSRCommPkgNumRecvs(comm_pkg_AN) = num_recvs;
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
send_procs_AN = NULL;
send_map_elmts_AN = NULL;
if (num_sends)
{
send_procs_AN = hypre_CTAlloc(HYPRE_Int,num_sends);
send_map_elmts_AN = hypre_CTAlloc(HYPRE_Int,send_map_starts[num_sends]);
}
send_map_starts_AN = hypre_CTAlloc(HYPRE_Int,num_sends+1);
recv_vec_starts_AN = hypre_CTAlloc(HYPRE_Int,num_recvs+1);
recv_procs_AN = NULL;
if (num_recvs) recv_procs_AN = hypre_CTAlloc(HYPRE_Int,num_recvs);
for (i=0; i < num_sends; i++)
send_procs_AN[i] = send_procs[i];
for (i=0; i < num_recvs; i++)
recv_procs_AN[i] = recv_procs[i];
send_map_starts_AN[0] = 0;
cnt = 0;
for (i=0; i < num_sends; i++)
{
k_map = send_map_starts[i];
if (send_map_starts[i+1]-k_map)
send_map_elmts_AN[cnt++] = send_map_elmts[k_map]/num_functions;
for (j=send_map_starts[i]+1; j < send_map_starts[i+1]; j++)
{
node = send_map_elmts[j]/num_functions;
if (node > send_map_elmts_AN[cnt-1])
send_map_elmts_AN[cnt++] = node;
}
send_map_starts_AN[i+1] = cnt;
}
hypre_ParCSRCommPkgSendProcs(comm_pkg_AN) = send_procs_AN;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg_AN) = send_map_starts_AN;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg_AN) = send_map_elmts_AN;
hypre_ParCSRCommPkgRecvProcs(comm_pkg_AN) = recv_procs_AN;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg_AN) = recv_vec_starts_AN;
}
num_cols_offd = hypre_CSRMatrixNumCols(A_offd);
if (num_cols_offd)
{
if (num_cols_offd > num_variables)
{
hypre_TFree(map_to_node);
map_to_node = hypre_CTAlloc(HYPRE_Int,num_cols_offd);
}
num_cols_offd_AN = 1;
map_to_node[0] = col_map_offd[0]/num_functions;
for (i=1; i < num_cols_offd; i++)
{
map_to_node[i] = col_map_offd[i]/num_functions;
if (map_to_node[i] > map_to_node[i-1]) num_cols_offd_AN++;
}
if (num_cols_offd_AN > num_nodes)
{
hypre_TFree(counter);
counter = hypre_CTAlloc(HYPRE_Int,num_cols_offd_AN);
}
map_to_map = NULL;
col_map_offd_AN = NULL;
map_to_map = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
col_map_offd_AN = hypre_CTAlloc(HYPRE_Int,num_cols_offd_AN);
col_map_offd_AN[0] = map_to_node[0];
recv_vec_starts_AN[0] = 0;
cnt = 1;
for (i=0; i < num_recvs; i++)
{
for (j=recv_vec_starts[i]; j < recv_vec_starts[i+1]; j++)
{
node = map_to_node[j];
if (node > col_map_offd_AN[cnt-1])
{
col_map_offd_AN[cnt++] = node;
}
map_to_map[j] = cnt-1;
}
recv_vec_starts_AN[i+1] = cnt;
}
for (i=0; i < num_cols_offd_AN; i++)
counter[i] = -1;
AN_num_nonzeros_offd = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
AN_offd_i[i] = AN_num_nonzeros_offd;
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < i)
{
counter[k_map] = i;
AN_num_nonzeros_offd++;
}
}
row++;
}
}
AN_offd_i[num_nodes] = AN_num_nonzeros_offd;
}
AN_offd = hypre_CSRMatrixCreate(num_nodes,num_cols_offd_AN,
AN_num_nonzeros_offd);
hypre_CSRMatrixI(AN_offd) = AN_offd_i;
if (AN_num_nonzeros_offd)
{
AN_offd_j = hypre_CTAlloc(HYPRE_Int, AN_num_nonzeros_offd);
AN_offd_data = hypre_CTAlloc(double, AN_num_nonzeros_offd);
hypre_CSRMatrixJ(AN_offd) = AN_offd_j;
hypre_CSRMatrixData(AN_offd) = AN_offd_data;
for (i=0; i < num_cols_offd_AN; i++)
counter[i] = -1;
index = 0;
row = 0;
AN_offd_i[0] = 0;
start_index = 0;
switch (mode)
{
case 1: /* frobenius norm */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = A_offd_data[k]*A_offd_data[k];
index++;
}
else
{
AN_offd_data[counter[k_map]] +=
A_offd_data[k]*A_offd_data[k];
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_offd; i++)
AN_offd_data[i] = sqrt(AN_offd_data[i]);
}
break;
case 2: /* sum of abs. value of all elements in block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = fabs(A_offd_data[k]);
index++;
}
else
{
AN_offd_data[counter[k_map]] += fabs(A_offd_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_offd; i++)
AN_offd_data[i] /= num_fun2;
}
break;
case 3: /* largest element in each block (not abs. value ) */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = A_offd_data[k];
index++;
}
else
{
if (fabs(A_offd_data[k]) >
fabs(AN_offd_data[counter[k_map]]))
AN_offd_data[counter[k_map]] = A_offd_data[k];
}
}
row++;
}
start_index = index;
}
}
break;
case 4: /* inf. norm (row-sum) */
{
data = hypre_CTAlloc(double, AN_num_nonzeros_offd*num_functions);
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
data[index*num_functions + j] = fabs(A_offd_data[k]);
index++;
}
else
{
data[(counter[k_map])*num_functions + j] += fabs(A_offd_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_offd; i++)
{
AN_offd_data[i] = data[i*num_functions];
for (j=1; j< num_functions; j++)
{
AN_offd_data[i] = hypre_max( AN_offd_data[i],data[i*num_functions+j]);
}
}
hypre_TFree(data);
}
break;
case 6: /* sum of value of all elements in block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = (A_offd_data[k]);
index++;
}
else
{
AN_offd_data[counter[k_map]] += (A_offd_data[k]);
}
}
row++;
}
start_index = index;
}
}
break;
}
hypre_TFree(map_to_map);
}
if (diag_option ==1 )
{
/* make the diag entry the negative of the sum of off-diag entries (here we are adding the
off_diag contribution)*/
/* the diagonal is the first element listed in each row of AN_diag_data - */
for (i=0; i < num_nodes; i++)
{
sum = 0.0;
for (k = AN_offd_i[i]; k < AN_offd_i[i+1]; k++)
{
sum += AN_offd_data[k];
}
index = AN_diag_i[i];/* location of diag entry in data */
AN_diag_data[index] -= sum; /* subtract from current value */
}
}
AN = hypre_ParCSRMatrixCreate(comm, global_num_nodes, global_num_nodes,
row_starts_AN, row_starts_AN, num_cols_offd_AN,
AN_num_nonzeros_diag, AN_num_nonzeros_offd);
/* we already created the diag and offd matrices - so we don't need the ones
created above */
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixDiag(AN));
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixOffd(AN));
hypre_ParCSRMatrixDiag(AN) = AN_diag;
hypre_ParCSRMatrixOffd(AN) = AN_offd;
hypre_ParCSRMatrixColMapOffd(AN) = col_map_offd_AN;
hypre_ParCSRMatrixCommPkg(AN) = comm_pkg_AN;
new_num_cols_offd = num_functions*num_cols_offd_AN;
if (new_num_cols_offd > num_cols_offd)
{
new_col_map_offd = hypre_CTAlloc(HYPRE_Int, new_num_cols_offd);
cnt = 0;
for (i=0; i < num_cols_offd_AN; i++)
{
for (j=0; j < num_functions; j++)
{
new_col_map_offd[cnt++] = num_functions*col_map_offd_AN[i]+j;
}
}
cnt = 0;
for (i=0; i < num_cols_offd; i++)
{
while (col_map_offd[i] > new_col_map_offd[cnt])
cnt++;
col_map_offd[i] = cnt++;
}
for (i=0; i < num_recvs+1; i++)
{
recv_vec_starts[i] = num_functions*recv_vec_starts_AN[i];
}
for (i=0; i < num_nonzeros_offd; i++)
{
j = A_offd_j[i];
A_offd_j[i] = col_map_offd[j];
}
hypre_ParCSRMatrixColMapOffd(A) = new_col_map_offd;
hypre_CSRMatrixNumCols(A_offd) = new_num_cols_offd;
hypre_TFree(col_map_offd);
}
hypre_TFree(map_to_node);
new_send_elmts_size = send_map_starts_AN[num_sends]*num_functions;
if (new_send_elmts_size > send_map_starts[num_sends])
{
new_send_map_elmts = hypre_CTAlloc(HYPRE_Int,new_send_elmts_size);
cnt = 0;
send_map_starts[0] = 0;
for (i=0; i < num_sends; i++)
{
send_map_starts[i+1] = send_map_starts_AN[i+1]*num_functions;
for (j=send_map_starts_AN[i]; j < send_map_starts_AN[i+1]; j++)
{
for (k=0; k < num_functions; k++)
new_send_map_elmts[cnt++] = send_map_elmts_AN[j]*num_functions+k;
}
}
hypre_TFree(send_map_elmts);
hypre_ParCSRCommPkgSendMapElmts(comm_pkg) = new_send_map_elmts;
}
*AN_ptr = AN;
hypre_TFree(counter);
return (ierr);
}
/*
Function: hypre_ParCSRMatrixEliminateAAe
(input) (output)
/ A_ii | A_ib \ / A_ii | 0 \
A = | -----+----- | ---> | -----+----- |
\ A_bi | A_bb / \ 0 | I /
/ 0 | A_ib \
Ae = | -----+--------- |
\ A_bi | A_bb - I /
*/
void hypre_ParCSRMatrixEliminateAAe(hypre_ParCSRMatrix *A,
hypre_ParCSRMatrix **Ae,
HYPRE_Int num_rowscols_to_elim,
HYPRE_Int *rowscols_to_elim)
{
HYPRE_Int i, j, k;
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int A_diag_nrows = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int A_offd_ncols = hypre_CSRMatrixNumCols(A_offd);
*Ae = hypre_ParCSRMatrixCreate(hypre_ParCSRMatrixComm(A),
hypre_ParCSRMatrixGlobalNumRows(A),
hypre_ParCSRMatrixGlobalNumCols(A),
hypre_ParCSRMatrixRowStarts(A),
hypre_ParCSRMatrixColStarts(A),
0, 0, 0);
hypre_ParCSRMatrixSetRowStartsOwner(*Ae, 0);
hypre_ParCSRMatrixSetColStartsOwner(*Ae, 0);
hypre_CSRMatrix *Ae_diag = hypre_ParCSRMatrixDiag(*Ae);
hypre_CSRMatrix *Ae_offd = hypre_ParCSRMatrixOffd(*Ae);
HYPRE_Int Ae_offd_ncols;
HYPRE_Int num_offd_cols_to_elim;
HYPRE_Int *offd_cols_to_elim;
HYPRE_Int *A_col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *Ae_col_map_offd;
HYPRE_Int *col_mark;
HYPRE_Int *col_remap;
/* figure out which offd cols should be eliminated */
{
hypre_ParCSRCommHandle *comm_handle;
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int num_sends, *int_buf_data;
HYPRE_Int index, start;
HYPRE_Int *eliminate_row = hypre_CTAlloc(HYPRE_Int, A_diag_nrows);
HYPRE_Int *eliminate_col = hypre_CTAlloc(HYPRE_Int, A_offd_ncols);
/* make sure A has a communication package */
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
/* which of the local rows are to be eliminated */
for (i = 0; i < A_diag_nrows; i++)
{
eliminate_row[i] = 0;
}
for (i = 0; i < num_rowscols_to_elim; i++)
{
eliminate_row[rowscols_to_elim[i]] = 1;
}
/* use a Matvec communication pattern to find (in eliminate_col)
which of the local offd columns are to be eliminated */
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
int_buf_data = hypre_CTAlloc(HYPRE_Int,
hypre_ParCSRCommPkgSendMapStart(comm_pkg,
num_sends));
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
{
k = hypre_ParCSRCommPkgSendMapElmt(comm_pkg, j);
int_buf_data[index++] = eliminate_row[k];
}
}
comm_handle = hypre_ParCSRCommHandleCreate(11, comm_pkg,
int_buf_data, eliminate_col);
/* eliminate diagonal part, overlapping it with communication */
hypre_CSRMatrixElimCreate(A_diag, Ae_diag,
num_rowscols_to_elim, rowscols_to_elim,
num_rowscols_to_elim, rowscols_to_elim,
NULL);
hypre_CSRMatrixEliminateRowsCols(A_diag, Ae_diag,
num_rowscols_to_elim, rowscols_to_elim,
num_rowscols_to_elim, rowscols_to_elim,
1, NULL);
hypre_CSRMatrixReorder(Ae_diag);
/* finish the communication */
hypre_ParCSRCommHandleDestroy(comm_handle);
/* received eliminate_col[], count offd columns to eliminate */
num_offd_cols_to_elim = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (eliminate_col[i]) {
num_offd_cols_to_elim++;
}
}
offd_cols_to_elim = hypre_CTAlloc(HYPRE_Int, num_offd_cols_to_elim);
/* get a list of offd column indices and coefs */
num_offd_cols_to_elim = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (eliminate_col[i])
{
offd_cols_to_elim[num_offd_cols_to_elim++] = i;
}
}
hypre_TFree(int_buf_data);
hypre_TFree(eliminate_row);
hypre_TFree(eliminate_col);
}
/* eliminate the off-diagonal part */
col_mark = hypre_CTAlloc(HYPRE_Int, A_offd_ncols);
col_remap = hypre_CTAlloc(HYPRE_Int, A_offd_ncols);
hypre_CSRMatrixElimCreate(A_offd, Ae_offd,
num_rowscols_to_elim, rowscols_to_elim,
num_offd_cols_to_elim, offd_cols_to_elim,
col_mark);
for (i = k = 0; i < A_offd_ncols; i++)
{
if (col_mark[i]) {
col_remap[i] = k++;
}
}
hypre_CSRMatrixEliminateRowsCols(A_offd, Ae_offd,
num_rowscols_to_elim, rowscols_to_elim,
num_offd_cols_to_elim, offd_cols_to_elim,
0, col_remap);
/* create col_map_offd for Ae */
Ae_offd_ncols = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (col_mark[i]) {
Ae_offd_ncols++;
}
}
Ae_col_map_offd = hypre_CTAlloc(HYPRE_Int, Ae_offd_ncols);
Ae_offd_ncols = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (col_mark[i])
{
Ae_col_map_offd[Ae_offd_ncols++] = A_col_map_offd[i];
}
}
hypre_ParCSRMatrixColMapOffd(*Ae) = Ae_col_map_offd;
hypre_CSRMatrixNumCols(Ae_offd) = Ae_offd_ncols;
hypre_TFree(col_remap);
hypre_TFree(col_mark);
hypre_TFree(offd_cols_to_elim);
hypre_ParCSRMatrixSetNumNonzeros(*Ae);
hypre_MatvecCommPkgCreate(*Ae);
}
hypre_ParCSRMatrix * hypre_ParMatmul_FC(
hypre_ParCSRMatrix * A, hypre_ParCSRMatrix * P, HYPRE_Int * CF_marker,
HYPRE_Int * dof_func, HYPRE_Int * dof_func_offd )
/* hypre_parMatmul_FC creates and returns the "Fine"-designated rows of the
matrix product A*P. A's size is (nC+nF)*(nC+nF), P's size is (nC+nF)*nC
where nC is the number of coarse rows/columns, nF the number of fine
rows/columns. The size of C=A*P is (nC+nF)*nC, even though not all rows
of C are actually computed. If we were to construct a matrix consisting
only of the computed rows of C, its size would be nF*nC.
"Fine" is defined solely by the marker array, and for example could be
a proper subset of the fine points of a multigrid hierarchy.
*/
{
/* To compute a submatrix of C containing only the computed data, i.e.
only "Fine" rows, we would have to do a lot of computational work,
with a lot of communication. The communication is because such a
matrix would need global information that depends on which rows are
"Fine".
*/
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *row_starts_A = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int num_rows_diag_A = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_cols_diag_A = hypre_CSRMatrixNumCols(A_diag);
HYPRE_Int num_cols_offd_A = hypre_CSRMatrixNumCols(A_offd);
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
double *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
HYPRE_Int *col_map_offd_P = hypre_ParCSRMatrixColMapOffd(P);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int first_col_diag_P = hypre_ParCSRMatrixFirstColDiag(P);
HYPRE_Int last_col_diag_P;
HYPRE_Int *col_starts_P = hypre_ParCSRMatrixColStarts(P);
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int num_cols_diag_P = hypre_CSRMatrixNumCols(P_diag);
HYPRE_Int num_cols_offd_P = hypre_CSRMatrixNumCols(P_offd);
hypre_ParCSRMatrix *C;
HYPRE_Int *col_map_offd_C;
HYPRE_Int *map_P_to_C;
hypre_CSRMatrix *C_diag;
double *C_diag_data;
HYPRE_Int *C_diag_i;
HYPRE_Int *C_diag_j;
hypre_CSRMatrix *C_offd;
double *C_offd_data=NULL;
HYPRE_Int *C_offd_i=NULL;
HYPRE_Int *C_offd_j=NULL;
HYPRE_Int C_diag_size;
HYPRE_Int C_offd_size;
HYPRE_Int num_cols_offd_C = 0;
hypre_CSRMatrix *Ps_ext;
double *Ps_ext_data;
HYPRE_Int *Ps_ext_i;
HYPRE_Int *Ps_ext_j;
double *P_ext_diag_data;
HYPRE_Int *P_ext_diag_i;
HYPRE_Int *P_ext_diag_j;
HYPRE_Int P_ext_diag_size;
double *P_ext_offd_data;
HYPRE_Int *P_ext_offd_i;
HYPRE_Int *P_ext_offd_j;
HYPRE_Int P_ext_offd_size;
HYPRE_Int *P_marker;
HYPRE_Int *temp;
HYPRE_Int i, j;
HYPRE_Int i1, i2, i3;
HYPRE_Int jj2, jj3;
HYPRE_Int jj_count_diag, jj_count_offd;
HYPRE_Int jj_row_begin_diag, jj_row_begin_offd;
HYPRE_Int start_indexing = 0; /* start indexing for C_data at 0 */
HYPRE_Int n_rows_A_global, n_cols_A_global;
HYPRE_Int n_rows_P_global, n_cols_P_global;
HYPRE_Int allsquare = 0;
HYPRE_Int cnt, cnt_offd, cnt_diag;
HYPRE_Int num_procs;
HYPRE_Int value;
double a_entry;
double a_b_product;
n_rows_A_global = hypre_ParCSRMatrixGlobalNumRows(A);
n_cols_A_global = hypre_ParCSRMatrixGlobalNumCols(A);
n_rows_P_global = hypre_ParCSRMatrixGlobalNumRows(P);
n_cols_P_global = hypre_ParCSRMatrixGlobalNumCols(P);
if (n_cols_A_global != n_rows_P_global || num_cols_diag_A != num_rows_diag_P)
{
hypre_printf(" Error! Incompatible matrix dimensions!\n");
return NULL;
}
/* if (num_rows_A==num_cols_P) allsquare = 1; */
/*-----------------------------------------------------------------------
* Extract P_ext, i.e. portion of P that is stored on neighbor procs
* and needed locally for matrix matrix product
*-----------------------------------------------------------------------*/
hypre_MPI_Comm_size(comm, &num_procs);
if (num_procs > 1)
{
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings within
* hypre_ParCSRMatrixExtractBExt
*--------------------------------------------------------------------*/
Ps_ext = hypre_ParCSRMatrixExtractBExt(P,A,1);
Ps_ext_data = hypre_CSRMatrixData(Ps_ext);
Ps_ext_i = hypre_CSRMatrixI(Ps_ext);
Ps_ext_j = hypre_CSRMatrixJ(Ps_ext);
}
P_ext_diag_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_offd_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_diag_size = 0;
P_ext_offd_size = 0;
last_col_diag_P = first_col_diag_P + num_cols_diag_P -1;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
P_ext_offd_size++;
else
P_ext_diag_size++;
P_ext_diag_i[i+1] = P_ext_diag_size;
P_ext_offd_i[i+1] = P_ext_offd_size;
}
if (P_ext_diag_size)
{
P_ext_diag_j = hypre_CTAlloc(HYPRE_Int, P_ext_diag_size);
P_ext_diag_data = hypre_CTAlloc(double, P_ext_diag_size);
}
if (P_ext_offd_size)
{
P_ext_offd_j = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size);
P_ext_offd_data = hypre_CTAlloc(double, P_ext_offd_size);
}
cnt_offd = 0;
cnt_diag = 0;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
{
P_ext_offd_j[cnt_offd] = Ps_ext_j[j];
P_ext_offd_data[cnt_offd++] = Ps_ext_data[j];
}
else
{
P_ext_diag_j[cnt_diag] = Ps_ext_j[j] - first_col_diag_P;
P_ext_diag_data[cnt_diag++] = Ps_ext_data[j];
}
}
if (num_procs > 1)
{
hypre_CSRMatrixDestroy(Ps_ext);
Ps_ext = NULL;
}
cnt = 0;
if (P_ext_offd_size || num_cols_offd_P)
{
temp = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size+num_cols_offd_P);
for (i=0; i < P_ext_offd_size; i++)
temp[i] = P_ext_offd_j[i];
cnt = P_ext_offd_size;
for (i=0; i < num_cols_offd_P; i++)
temp[cnt++] = col_map_offd_P[i];
}
if (cnt)
{
qsort0(temp, 0, cnt-1);
num_cols_offd_C = 1;
value = temp[0];
for (i=1; i < cnt; i++)
{
if (temp[i] > value)
{
value = temp[i];
temp[num_cols_offd_C++] = value;
}
}
}
if (num_cols_offd_C)
col_map_offd_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_C);
for (i=0; i < num_cols_offd_C; i++)
col_map_offd_C[i] = temp[i];
if (P_ext_offd_size || num_cols_offd_P)
hypre_TFree(temp);
for (i=0 ; i < P_ext_offd_size; i++)
P_ext_offd_j[i] = hypre_BinarySearch(col_map_offd_C,
P_ext_offd_j[i],
num_cols_offd_C);
if (num_cols_offd_P)
{
map_P_to_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_P);
cnt = 0;
for (i=0; i < num_cols_offd_C; i++)
if (col_map_offd_C[i] == col_map_offd_P[cnt])
{
map_P_to_C[cnt++] = i;
if (cnt == num_cols_offd_P) break;
}
}
/*-----------------------------------------------------------------------
* Allocate marker array.
*-----------------------------------------------------------------------*/
P_marker = hypre_CTAlloc(HYPRE_Int, num_cols_diag_P+num_cols_offd_C);
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/* no changes for the marked version above this point */
/* This function call is the first pass: */
hypre_ParMatmul_RowSizes_Marked(
&C_diag_i, &C_offd_i, &P_marker,
A_diag_i, A_diag_j, A_offd_i, A_offd_j,
P_diag_i, P_diag_j, P_offd_i, P_offd_j,
P_ext_diag_i, P_ext_diag_j, P_ext_offd_i, P_ext_offd_j,
map_P_to_C,
&C_diag_size, &C_offd_size,
num_rows_diag_A, num_cols_offd_A, allsquare,
num_cols_diag_P, num_cols_offd_P,
num_cols_offd_C, CF_marker, dof_func, dof_func_offd
);
/* The above call of hypre_ParMatmul_RowSizes_Marked computed
two scalars: C_diag_size, C_offd_size,
and two arrays: C_diag_i, C_offd_i
( P_marker is also computed, but only used internally )
*/
/*-----------------------------------------------------------------------
* Allocate C_diag_data and C_diag_j arrays.
* Allocate C_offd_data and C_offd_j arrays.
*-----------------------------------------------------------------------*/
last_col_diag_P = first_col_diag_P + num_cols_diag_P - 1;
C_diag_data = hypre_CTAlloc(double, C_diag_size);
C_diag_j = hypre_CTAlloc(HYPRE_Int, C_diag_size);
if (C_offd_size)
{
C_offd_data = hypre_CTAlloc(double, C_offd_size);
C_offd_j = hypre_CTAlloc(HYPRE_Int, C_offd_size);
}
/*-----------------------------------------------------------------------
* Second Pass: Fill in C_diag_data and C_diag_j.
* Second Pass: Fill in C_offd_data and C_offd_j.
*-----------------------------------------------------------------------*/
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
jj_count_diag = start_indexing;
jj_count_offd = start_indexing;
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/*-----------------------------------------------------------------------
* Loop over interior c-points.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_A; i1++)
{
if ( CF_marker[i1] < 0 ) /* i1 is a fine row */
/* ... This and the coarse row code are the only parts between first pass
and near the end where
hypre_ParMatmul_FC is different from the regular hypre_ParMatmul */
{
/*--------------------------------------------------------------------
* Create diagonal entry, C_{i1,i1}
*--------------------------------------------------------------------*/
jj_row_begin_diag = jj_count_diag;
jj_row_begin_offd = jj_count_offd;
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_offd.
*-----------------------------------------------------------------*/
if (num_cols_offd_A)
{
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func_offd[i2] )
{ /* interpolate only like "functions" */
a_entry = A_offd_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_ext.
*-----------------------------------------------------------*/
for (jj3 = P_ext_offd_i[i2]; jj3 < P_ext_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+P_ext_offd_j[jj3];
a_b_product = a_entry * P_ext_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
C_offd_data[P_marker[i3]] += a_b_product;
}
for (jj3 = P_ext_diag_i[i2]; jj3 < P_ext_diag_i[i2+1]; jj3++)
{
i3 = P_ext_diag_j[jj3];
a_b_product = a_entry * P_ext_diag_data[jj3];
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
C_diag_data[P_marker[i3]] += a_b_product;
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_diag.
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func[i2] )
{ /* interpolate only like "functions" */
a_entry = A_diag_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_diag.
*-----------------------------------------------------------*/
for (jj3 = P_diag_i[i2]; jj3 < P_diag_i[i2+1]; jj3++)
{
i3 = P_diag_j[jj3];
a_b_product = a_entry * P_diag_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[P_marker[i3]] += a_b_product;
}
}
if (num_cols_offd_P)
{
for (jj3 = P_offd_i[i2]; jj3 < P_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+map_P_to_C[P_offd_j[jj3]];
a_b_product = a_entry * P_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
{
C_offd_data[P_marker[i3]] += a_b_product;
}
}
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
else /* i1 is a coarse row.*/
/* Copy P coarse-row values to C. This is useful if C is meant to
become a replacement for P */
{
if (num_cols_offd_P)
{
for (jj2 = P_offd_i[i1]; jj2 < P_offd_i[i1+1]; jj2++)
{
C_offd_j[jj_count_offd] = P_offd_j[jj_count_offd];
C_offd_data[jj_count_offd] = P_offd_data[jj_count_offd];
++jj_count_offd;
}
}
for (jj2 = P_diag_i[i1]; jj2 < P_diag_i[i1+1]; jj2++)
{
C_diag_j[jj_count_diag] = P_diag_j[jj2];
C_diag_data[jj_count_diag] = P_diag_data[jj2];
++jj_count_diag;
}
}
}
C = hypre_ParCSRMatrixCreate(
comm, n_rows_A_global, n_cols_P_global,
row_starts_A, col_starts_P, num_cols_offd_C, C_diag_size, C_offd_size );
/* Note that C does not own the partitionings */
hypre_ParCSRMatrixSetRowStartsOwner(C,0);
hypre_ParCSRMatrixSetColStartsOwner(C,0);
C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrixData(C_diag) = C_diag_data;
hypre_CSRMatrixI(C_diag) = C_diag_i;
hypre_CSRMatrixJ(C_diag) = C_diag_j;
C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrixI(C_offd) = C_offd_i;
hypre_ParCSRMatrixOffd(C) = C_offd;
if (num_cols_offd_C)
{
hypre_CSRMatrixData(C_offd) = C_offd_data;
hypre_CSRMatrixJ(C_offd) = C_offd_j;
hypre_ParCSRMatrixColMapOffd(C) = col_map_offd_C;
}
/*-----------------------------------------------------------------------
* Free various arrays
*-----------------------------------------------------------------------*/
hypre_TFree(P_marker);
hypre_TFree(P_ext_diag_i);
if (P_ext_diag_size)
{
hypre_TFree(P_ext_diag_j);
hypre_TFree(P_ext_diag_data);
}
hypre_TFree(P_ext_offd_i);
if (P_ext_offd_size)
{
hypre_TFree(P_ext_offd_j);
hypre_TFree(P_ext_offd_data);
}
if (num_cols_offd_P) hypre_TFree(map_P_to_C);
return C;
}
hypre_CSRMatrix *
hypre_ParCSRMatrixExtractAExt( hypre_ParCSRMatrix *A,
HYPRE_Int data,
HYPRE_Int ** pA_ext_row_map )
{
/* Note that A's role as the first factor in A*A^T is used only
through ...CommPkgT(A), which basically says which rows of A
(columns of A^T) are needed. In all the other places where A
serves as an input, it is through its role as A^T, the matrix
whose data needs to be passed between processors. */
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(A);
HYPRE_Int first_row_index = hypre_ParCSRMatrixFirstRowIndex(A);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkgT(A);
/* ... CommPkgT(A) should identify all rows of A^T needed for A*A^T (that is
* generally a bigger set than ...CommPkg(A), the rows of B needed for A*B) */
HYPRE_Int num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
HYPRE_Int *recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
HYPRE_Int *send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
HYPRE_Int *send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Int *diag_i = hypre_CSRMatrixI(diag);
HYPRE_Int *diag_j = hypre_CSRMatrixJ(diag);
HYPRE_Complex *diag_data = hypre_CSRMatrixData(diag);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *offd_i = hypre_CSRMatrixI(offd);
HYPRE_Int *offd_j = hypre_CSRMatrixJ(offd);
HYPRE_Complex *offd_data = hypre_CSRMatrixData(offd);
HYPRE_Int num_cols_A, num_nonzeros;
HYPRE_Int num_rows_A_ext;
hypre_CSRMatrix *A_ext;
HYPRE_Int *A_ext_i;
HYPRE_Int *A_ext_j;
HYPRE_Complex *A_ext_data;
num_cols_A = hypre_ParCSRMatrixGlobalNumCols(A);
num_rows_A_ext = recv_vec_starts[num_recvs];
hypre_ParCSRMatrixExtractBExt_Arrays
( &A_ext_i, &A_ext_j, &A_ext_data, pA_ext_row_map,
&num_nonzeros,
data, 1, comm, comm_pkg,
num_cols_A, num_recvs, num_sends,
first_col_diag, first_row_index,
recv_vec_starts, send_map_starts, send_map_elmts,
diag_i, diag_j, offd_i, offd_j, col_map_offd,
diag_data, offd_data
);
A_ext = hypre_CSRMatrixCreate(num_rows_A_ext,num_cols_A,num_nonzeros);
hypre_CSRMatrixI(A_ext) = A_ext_i;
hypre_CSRMatrixJ(A_ext) = A_ext_j;
if (data) hypre_CSRMatrixData(A_ext) = A_ext_data;
return A_ext;
}
hypre_ParCSRMatrix *hypre_ParCSRAAt( hypre_ParCSRMatrix *A )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Complex *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Complex *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *A_col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *A_ext_row_map;
HYPRE_Int *row_starts_A = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int num_rows_diag_A = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_cols_offd_A = hypre_CSRMatrixNumCols(A_offd);
hypre_ParCSRMatrix *C;
HYPRE_Int *col_map_offd_C;
hypre_CSRMatrix *C_diag;
HYPRE_Complex *C_diag_data;
HYPRE_Int *C_diag_i;
HYPRE_Int *C_diag_j;
hypre_CSRMatrix *C_offd;
HYPRE_Complex *C_offd_data=NULL;
HYPRE_Int *C_offd_i=NULL;
HYPRE_Int *C_offd_j=NULL;
HYPRE_Int *new_C_offd_j;
HYPRE_Int C_diag_size;
HYPRE_Int C_offd_size;
HYPRE_Int last_col_diag_C;
HYPRE_Int num_cols_offd_C;
hypre_CSRMatrix *A_ext;
HYPRE_Complex *A_ext_data;
HYPRE_Int *A_ext_i;
HYPRE_Int *A_ext_j;
HYPRE_Int num_rows_A_ext=0;
HYPRE_Int first_row_index_A = hypre_ParCSRMatrixFirstRowIndex(A);
HYPRE_Int first_col_diag_A = hypre_ParCSRMatrixFirstColDiag(A);
HYPRE_Int *B_marker;
HYPRE_Int i;
HYPRE_Int i1, i2, i3;
HYPRE_Int jj2, jj3;
HYPRE_Int jj_count_diag, jj_count_offd;
HYPRE_Int jj_row_begin_diag, jj_row_begin_offd;
HYPRE_Int start_indexing = 0; /* start indexing for C_data at 0 */
HYPRE_Int count;
HYPRE_Int n_rows_A, n_cols_A;
HYPRE_Complex a_entry;
HYPRE_Complex a_b_product;
HYPRE_Complex zero = 0.0;
n_rows_A = hypre_ParCSRMatrixGlobalNumRows(A);
n_cols_A = hypre_ParCSRMatrixGlobalNumCols(A);
if (n_cols_A != n_rows_A)
{
hypre_printf(" Error! Incompatible matrix dimensions!\n");
return NULL;
}
/*-----------------------------------------------------------------------
* Extract A_ext, i.e. portion of A that is stored on neighbor procs
* and needed locally for A^T in the matrix matrix product A*A^T
*-----------------------------------------------------------------------*/
if (num_rows_diag_A != n_rows_A)
{
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings
*--------------------------------------------------------------------*/
if (!hypre_ParCSRMatrixCommPkg(A))
{
hypre_MatTCommPkgCreate(A);
}
A_ext = hypre_ParCSRMatrixExtractAExt( A, 1, &A_ext_row_map );
A_ext_data = hypre_CSRMatrixData(A_ext);
A_ext_i = hypre_CSRMatrixI(A_ext);
A_ext_j = hypre_CSRMatrixJ(A_ext);
num_rows_A_ext = hypre_CSRMatrixNumRows(A_ext);
}
/*-----------------------------------------------------------------------
* Allocate marker array.
*-----------------------------------------------------------------------*/
B_marker = hypre_CTAlloc(HYPRE_Int, num_rows_diag_A+num_rows_A_ext );
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
for ( i1=0; i1<num_rows_diag_A+num_rows_A_ext; ++i1 )
{
B_marker[i1] = -1;
}
hypre_ParAat_RowSizes(
&C_diag_i, &C_offd_i, B_marker,
A_diag_i, A_diag_j,
A_offd_i, A_offd_j, A_col_map_offd,
A_ext_i, A_ext_j, A_ext_row_map,
&C_diag_size, &C_offd_size,
num_rows_diag_A, num_cols_offd_A,
num_rows_A_ext,
first_col_diag_A, first_row_index_A
);
#if 0
/* debugging output: */
hypre_printf("A_ext_row_map (%i):",num_rows_A_ext);
for ( i1=0; i1<num_rows_A_ext; ++i1 ) hypre_printf(" %i",A_ext_row_map[i1] );
hypre_printf("\nC_diag_i (%i):",C_diag_size);
for ( i1=0; i1<=num_rows_diag_A; ++i1 ) hypre_printf(" %i",C_diag_i[i1] );
hypre_printf("\nC_offd_i (%i):",C_offd_size);
for ( i1=0; i1<=num_rows_diag_A; ++i1 ) hypre_printf(" %i",C_offd_i[i1] );
hypre_printf("\n");
#endif
/*-----------------------------------------------------------------------
* Allocate C_diag_data and C_diag_j arrays.
* Allocate C_offd_data and C_offd_j arrays.
*-----------------------------------------------------------------------*/
last_col_diag_C = first_row_index_A + num_rows_diag_A - 1;
C_diag_data = hypre_CTAlloc(HYPRE_Complex, C_diag_size);
C_diag_j = hypre_CTAlloc(HYPRE_Int, C_diag_size);
if (C_offd_size)
{
C_offd_data = hypre_CTAlloc(HYPRE_Complex, C_offd_size);
C_offd_j = hypre_CTAlloc(HYPRE_Int, C_offd_size);
}
/*-----------------------------------------------------------------------
* Second Pass: Fill in C_diag_data and C_diag_j.
* Second Pass: Fill in C_offd_data and C_offd_j.
*-----------------------------------------------------------------------*/
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
jj_count_diag = start_indexing;
jj_count_offd = start_indexing;
for ( i1=0; i1<num_rows_diag_A+num_rows_A_ext; ++i1 )
{
B_marker[i1] = -1;
}
/*-----------------------------------------------------------------------
* Loop over interior c-points.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_A; i1++)
{
/*--------------------------------------------------------------------
* Create diagonal entry, C_{i1,i1}
*--------------------------------------------------------------------*/
B_marker[i1] = jj_count_diag;
jj_row_begin_diag = jj_count_diag;
jj_row_begin_offd = jj_count_offd;
C_diag_data[jj_count_diag] = zero;
C_diag_j[jj_count_diag] = i1;
jj_count_diag++;
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_offd.
*-----------------------------------------------------------------*/
/* There are 3 CSRMatrix or CSRBooleanMatrix objects here:
ext*ext, ext*diag, and ext*offd belong to another processor.
diag*offd and offd*diag don't count - never share a column by definition.
So we have to do 4 cases:
diag*ext, offd*ext, diag*diag, and offd*offd.
*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
a_entry = A_diag_data[jj2];
/* diag*ext */
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of (A_ext)^T
* That is, rows i3 having a column i2 of A_ext.
* For now, for each row i3 of A_ext we crudely check _all_
* columns to see whether one matches i2.
* For each entry (i2,i3) of (A_ext)^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3) . This contributes to both the diag and offd
* blocks of C.
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_A_ext; i3++ ) {
for ( jj3=A_ext_i[i3]; jj3<A_ext_i[i3+1]; jj3++ ) {
if ( A_ext_j[jj3]==i2+first_col_diag_A ) {
/* row i3, column i2 of A_ext; or,
row i2, column i3 of (A_ext)^T */
a_b_product = a_entry * A_ext_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if ( A_ext_row_map[i3] < first_row_index_A ||
A_ext_row_map[i3] > last_col_diag_C ) { /* offd */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_offd) {
B_marker[i3+num_rows_diag_A] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3;
jj_count_offd++;
}
else
C_offd_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
else { /* diag */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_diag) {
B_marker[i3+num_rows_diag_A] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3-first_col_diag_A;
jj_count_diag++;
}
else
C_diag_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
}
}
}
}
if (num_cols_offd_A)
{
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
a_entry = A_offd_data[jj2];
/* offd * ext */
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of (A_ext)^T
* That is, rows i3 having a column i2 of A_ext.
* For now, for each row i3 of A_ext we crudely check _all_
* columns to see whether one matches i2.
* For each entry (i2,i3) of (A_ext)^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3) . This contributes to both the diag and offd
* blocks of C.
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_A_ext; i3++ ) {
for ( jj3=A_ext_i[i3]; jj3<A_ext_i[i3+1]; jj3++ ) {
if ( A_ext_j[jj3]==A_col_map_offd[i2] ) {
/* row i3, column i2 of A_ext; or,
row i2, column i3 of (A_ext)^T */
a_b_product = a_entry * A_ext_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if ( A_ext_row_map[i3] < first_row_index_A ||
A_ext_row_map[i3] > last_col_diag_C ) { /* offd */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_offd) {
B_marker[i3+num_rows_diag_A] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3;
jj_count_offd++;
}
else
C_offd_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
else { /* diag */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_diag) {
B_marker[i3+num_rows_diag_A] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3-first_row_index_A;
jj_count_diag++;
}
else
C_diag_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
}
}
}
}
}
/* diag * diag */
/*-----------------------------------------------------------------
* Loop over entries (columns) i2 in row i1 of A_diag.
* For each such column we will find the contributions of the
* corresponding rows i2 of A^T to C=A*A^T . Now we only look
* at the local part of A^T - with columns (rows of A) living
* on this processor.
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
a_entry = A_diag_data[jj2];
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of A^T
* That is, rows i3 having a column i2 of A (local part).
* For now, for each row i3 of A we crudely check _all_
* columns to see whether one matches i2.
* This i3-loop is for the diagonal block of A.
* It contributes to the diagonal block of C.
* For each entry (i2,i3) of A^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3)
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_diag_A; i3++ ) {
for ( jj3=A_diag_i[i3]; jj3<A_diag_i[i3+1]; jj3++ ) {
if ( A_diag_j[jj3]==i2 ) {
/* row i3, column i2 of A; or,
row i2, column i3 of A^T */
a_b_product = a_entry * A_diag_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, mark it and increment
* counter.
*--------------------------------------------------------*/
if (B_marker[i3] < jj_row_begin_diag)
{
B_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[B_marker[i3]] += a_b_product;
}
}
}
} /* end of i3 loop */
} /* end of third i2 loop */
/* offd * offd */
/*-----------------------------------------------------------
* Loop over offd columns i2 of A in A*A^T. Then
* loop over offd entries (columns) i3 in row i2 of A^T
* That is, rows i3 having a column i2 of A (local part).
* For now, for each row i3 of A we crudely check _all_
* columns to see whether one matches i2.
* This i3-loop is for the off-diagonal block of A.
* It contributes to the diag block of C.
* For each entry (i2,i3) of A^T, add A*A^T to C
*-----------------------------------------------------------*/
if (num_cols_offd_A) {
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
a_entry = A_offd_data[jj2];
for ( i3=0; i3<num_rows_diag_A; i3++ ) {
/* ... note that num_rows_diag_A == num_rows_offd_A */
for ( jj3=A_offd_i[i3]; jj3<A_offd_i[i3+1]; jj3++ ) {
if ( A_offd_j[jj3]==i2 ) {
/* row i3, column i2 of A; or,
row i2, column i3 of A^T */
a_b_product = a_entry * A_offd_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution
*--------------------------------------------------------*/
if (B_marker[i3] < jj_row_begin_diag)
{
B_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[B_marker[i3]] += a_b_product;
}
}
}
} /* end of last i3 loop */
} /* end of if (num_cols_offd_A) */
} /* end of fourth and last i2 loop */
#if 0 /* debugging printout */
hypre_printf("end of i1 loop: i1=%i jj_count_diag=%i\n", i1, jj_count_diag );
hypre_printf(" C_diag_j=");
for ( jj3=0; jj3<jj_count_diag; ++jj3) hypre_printf("%i ",C_diag_j[jj3]);
hypre_printf(" C_diag_data=");
for ( jj3=0; jj3<jj_count_diag; ++jj3) hypre_printf("%f ",C_diag_data[jj3]);
hypre_printf("\n");
hypre_printf(" C_offd_j=");
for ( jj3=0; jj3<jj_count_offd; ++jj3) hypre_printf("%i ",C_offd_j[jj3]);
hypre_printf(" C_offd_data=");
for ( jj3=0; jj3<jj_count_offd; ++jj3) hypre_printf("%f ",C_offd_data[jj3]);
hypre_printf("\n");
hypre_printf( " B_marker =" );
for ( it=0; it<num_rows_diag_A+num_rows_A_ext; ++it )
hypre_printf(" %i", B_marker[it] );
hypre_printf( "\n" );
#endif
} /* end of i1 loop */
/*-----------------------------------------------------------------------
* Delete 0-columns in C_offd, i.e. generate col_map_offd and reset
* C_offd_j. Note that (with the indexing we have coming into this
* block) col_map_offd_C[i3]==A_ext_row_map[i3].
*-----------------------------------------------------------------------*/
for ( i=0; i<num_rows_diag_A+num_rows_A_ext; ++i )
B_marker[i] = -1;
for ( i=0; i<C_offd_size; i++ )
B_marker[ C_offd_j[i] ] = -2;
count = 0;
for (i=0; i < num_rows_diag_A + num_rows_A_ext; i++) {
if (B_marker[i] == -2) {
B_marker[i] = count;
count++;
}
}
num_cols_offd_C = count;
if (num_cols_offd_C) {
col_map_offd_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_C);
new_C_offd_j = hypre_CTAlloc(HYPRE_Int,C_offd_size);
/* ... a bit big, but num_cols_offd_C is too small. It might be worth
computing the correct size, which is sum( no. columns in row i, over all rows i )
*/
for (i=0; i < C_offd_size; i++) {
new_C_offd_j[i] = B_marker[C_offd_j[i]];
col_map_offd_C[ new_C_offd_j[i] ] = A_ext_row_map[ C_offd_j[i] ];
}
hypre_TFree(C_offd_j);
C_offd_j = new_C_offd_j;
}
/*----------------------------------------------------------------
* Create C
*----------------------------------------------------------------*/
C = hypre_ParCSRMatrixCreate(comm, n_rows_A, n_rows_A, row_starts_A,
row_starts_A, num_cols_offd_C,
C_diag_size, C_offd_size);
/* Note that C does not own the partitionings */
hypre_ParCSRMatrixSetRowStartsOwner(C,0);
hypre_ParCSRMatrixSetColStartsOwner(C,0);
C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrixData(C_diag) = C_diag_data;
hypre_CSRMatrixI(C_diag) = C_diag_i;
hypre_CSRMatrixJ(C_diag) = C_diag_j;
if (num_cols_offd_C)
{
C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrixData(C_offd) = C_offd_data;
hypre_CSRMatrixI(C_offd) = C_offd_i;
hypre_CSRMatrixJ(C_offd) = C_offd_j;
hypre_ParCSRMatrixOffd(C) = C_offd;
hypre_ParCSRMatrixColMapOffd(C) = col_map_offd_C;
}
else
hypre_TFree(C_offd_i);
/*-----------------------------------------------------------------------
* Free B_ext and marker array.
*-----------------------------------------------------------------------*/
if (num_cols_offd_A)
{
hypre_CSRMatrixDestroy(A_ext);
A_ext = NULL;
}
hypre_TFree(B_marker);
if ( num_rows_diag_A != n_rows_A )
hypre_TFree(A_ext_row_map);
return C;
}
HYPRE_ParCSRMatrix
GenerateSysLaplacianVCoef( MPI_Comm comm,
HYPRE_Int nx,
HYPRE_Int ny,
HYPRE_Int nz,
HYPRE_Int P,
HYPRE_Int Q,
HYPRE_Int R,
HYPRE_Int p,
HYPRE_Int q,
HYPRE_Int r,
HYPRE_Int num_fun,
HYPRE_Real *mtrx,
HYPRE_Real *value )
{
hypre_ParCSRMatrix *A;
hypre_CSRMatrix *diag;
hypre_CSRMatrix *offd;
HYPRE_Int *diag_i;
HYPRE_Int *diag_j;
HYPRE_Real *diag_data;
HYPRE_Int *offd_i;
HYPRE_Int *offd_j;
HYPRE_Real *offd_data;
HYPRE_Int *global_part;
HYPRE_Int ix, iy, iz;
HYPRE_Int cnt, o_cnt;
HYPRE_Int local_num_rows;
HYPRE_Int *col_map_offd;
HYPRE_Int row_index, row, col;
HYPRE_Int index, diag_index;
HYPRE_Int i,j;
HYPRE_Int nx_local, ny_local, nz_local;
HYPRE_Int nx_size, ny_size, nz_size;
HYPRE_Int num_cols_offd;
HYPRE_Int grid_size;
HYPRE_Int local_grid_size;
HYPRE_Int first_j, j_ind;
HYPRE_Int num_coeffs, num_offd_coeffs;
HYPRE_Int *nx_part;
HYPRE_Int *ny_part;
HYPRE_Int *nz_part;
HYPRE_Int num_procs, my_id;
HYPRE_Int P_busy, Q_busy, R_busy;
HYPRE_Real val;
/* for indexing in values */
HYPRE_Int sz = num_fun*num_fun;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
grid_size = nx*ny*nz;
hypre_GeneratePartitioning(nx,P,&nx_part);
hypre_GeneratePartitioning(ny,Q,&ny_part);
hypre_GeneratePartitioning(nz,R,&nz_part);
global_part = hypre_CTAlloc(HYPRE_Int,P*Q*R+1);
global_part[0] = 0;
cnt = 1;
for (iz = 0; iz < R; iz++)
{
nz_size = nz_part[iz+1]-nz_part[iz];
for (iy = 0; iy < Q; iy++)
{
ny_size = ny_part[iy+1]-ny_part[iy];
for (ix = 0; ix < P; ix++)
{
nx_size = nx_part[ix+1] - nx_part[ix];
global_part[cnt] = global_part[cnt-1];
global_part[cnt++] += nx_size*ny_size*nz_size;
}
}
}
nx_local = nx_part[p+1] - nx_part[p];
ny_local = ny_part[q+1] - ny_part[q];
nz_local = nz_part[r+1] - nz_part[r];
my_id = r*(P*Q) + q*P + p;
num_procs = P*Q*R;
local_grid_size = nx_local*ny_local*nz_local;
local_num_rows = num_fun*local_grid_size;
diag_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
offd_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
P_busy = hypre_min(nx,P);
Q_busy = hypre_min(ny,Q);
R_busy = hypre_min(nz,R);
num_cols_offd = 0;
if (p) num_cols_offd += ny_local*nz_local;
if (p < P_busy-1) num_cols_offd += ny_local*nz_local;
if (q) num_cols_offd += nx_local*nz_local;
if (q < Q_busy-1) num_cols_offd += nx_local*nz_local;
if (r) num_cols_offd += nx_local*ny_local;
if (r < R_busy-1) num_cols_offd += nx_local*ny_local;
num_cols_offd *= num_fun;
if (!local_num_rows) num_cols_offd = 0;
col_map_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
cnt = 1;
diag_i[0] = 0;
offd_i[0] = 0;
for (iz = nz_part[r]; iz < nz_part[r+1]; iz++)
{
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
diag_i[cnt] = diag_i[cnt-1];
offd_i[cnt] = offd_i[cnt-1];
diag_i[cnt] += num_fun;
if (iz > nz_part[r])
diag_i[cnt] += num_fun;
else
{
if (iz)
{
offd_i[cnt] += num_fun;
}
}
if (iy > ny_part[q])
diag_i[cnt] += num_fun;
else
{
if (iy)
{
offd_i[cnt] += num_fun;
}
}
if (ix > nx_part[p])
diag_i[cnt] += num_fun;
else
{
if (ix)
{
offd_i[cnt] += num_fun;
}
}
if (ix+1 < nx_part[p+1])
diag_i[cnt] += num_fun;
else
{
if (ix+1 < nx)
{
offd_i[cnt] += num_fun;
}
}
if (iy+1 < ny_part[q+1])
diag_i[cnt] += num_fun;
else
{
if (iy+1 < ny)
{
offd_i[cnt] += num_fun;
}
}
if (iz+1 < nz_part[r+1])
diag_i[cnt] += num_fun;
else
{
if (iz+1 < nz)
{
offd_i[cnt] += num_fun;
}
}
num_coeffs = diag_i[cnt]-diag_i[cnt-1];
num_offd_coeffs = offd_i[cnt]-offd_i[cnt-1];
cnt++;
for (i=1; i < num_fun; i++)
{
diag_i[cnt] = diag_i[cnt-1]+num_coeffs;
offd_i[cnt] = offd_i[cnt-1]+num_offd_coeffs;
cnt++;
}
}
}
}
diag_j = hypre_CTAlloc(HYPRE_Int, diag_i[local_num_rows]);
diag_data = hypre_CTAlloc(HYPRE_Real, diag_i[local_num_rows]);
if (num_procs > 1)
{
offd_j = hypre_CTAlloc(HYPRE_Int, offd_i[local_num_rows]);
offd_data = hypre_CTAlloc(HYPRE_Real, offd_i[local_num_rows]);
}
row_index = 0;
for (iz = nz_part[r]; iz < nz_part[r+1]; iz++)
{
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
cnt = diag_i[row_index];;
o_cnt = offd_i[row_index];;
num_coeffs = diag_i[row_index+1]-diag_i[row_index];
num_offd_coeffs = offd_i[row_index+1]-offd_i[row_index];
first_j = row_index;
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = cnt+i*num_coeffs+j;
diag_j[j_ind] = first_j+j;
diag_data[j_ind] = value[0*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
cnt += num_fun;
if (iz > nz_part[r])
{
first_j = row_index-nx_local*ny_local*num_fun;
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = cnt+i*num_coeffs+j;
diag_j[j_ind] = first_j+j;
diag_data[j_ind] = value[3*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
cnt += num_fun;
}
else
{
if (iz)
{
first_j = num_fun*hypre_map(ix,iy,iz-1,p,q,r-1,P,Q,R,
nx_part,ny_part,nz_part,global_part);
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = o_cnt+i*num_offd_coeffs+j;
offd_j[j_ind] = first_j+j;
offd_data[j_ind] = value[3*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
o_cnt += num_fun;
}
}
if (iy > ny_part[q])
{
first_j = row_index-nx_local*num_fun;
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = cnt+i*num_coeffs+j;
diag_j[j_ind] = first_j+j;
diag_data[j_ind] = value[2*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
cnt += num_fun;
}
else
{
if (iy)
{
first_j = num_fun*hypre_map(ix,iy-1,iz,p,q-1,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = o_cnt+i*num_offd_coeffs+j;
offd_j[j_ind] = first_j+j;
offd_data[j_ind] = value[2*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
o_cnt += num_fun;
}
}
if (ix > nx_part[p])
{
first_j = row_index-num_fun;
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = cnt+i*num_coeffs+j;
diag_j[j_ind] = first_j+j;
diag_data[j_ind] = value[1*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
cnt += num_fun;
}
else
{
if (ix)
{
first_j = num_fun*hypre_map(ix-1,iy,iz,p-1,q,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = o_cnt+i*num_offd_coeffs+j;
offd_j[j_ind] = first_j+j;
offd_data[j_ind] = value[1*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
o_cnt += num_fun;
}
}
if (ix+1 < nx_part[p+1])
{
first_j = row_index+num_fun;
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = cnt+i*num_coeffs+j;
diag_j[j_ind] = first_j+j;
diag_data[j_ind] = value[1*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
cnt += num_fun;
}
else
{
if (ix+1 < nx)
{
first_j = num_fun*hypre_map(ix+1,iy,iz,p+1,q,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = o_cnt+i*num_offd_coeffs+j;
offd_j[j_ind] = first_j+j;
offd_data[j_ind] = value[1*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
o_cnt += num_fun;
}
}
if (iy+1 < ny_part[q+1])
{
first_j = row_index+nx_local*num_fun;
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = cnt+i*num_coeffs+j;
diag_j[j_ind] = first_j+j;
diag_data[j_ind] = value[2*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
cnt += num_fun;
}
else
{
if (iy+1 < ny)
{
first_j = num_fun*hypre_map(ix,iy+1,iz,p,q+1,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = o_cnt+i*num_offd_coeffs+j;
offd_j[j_ind] = first_j+j;
offd_data[j_ind] = value[2*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
o_cnt += num_fun;
}
}
if (iz+1 < nz_part[r+1])
{
first_j = row_index+nx_local*ny_local*num_fun;
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = cnt+i*num_coeffs+j;
diag_j[j_ind] = first_j+j;
diag_data[j_ind] = value[3*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
cnt += num_fun;
}
else
{
if (iz+1 < nz)
{
first_j = num_fun*hypre_map(ix,iy,iz+1,p,q,r+1,P,Q,R,
nx_part,ny_part,nz_part,global_part);
for (i=0; i < num_fun; i++)
{
for (j=0; j < num_fun; j++)
{
j_ind = o_cnt+i*num_offd_coeffs+j;
offd_j[j_ind] = first_j+j;
offd_data[j_ind] = value[3*sz + i*num_fun+j]*mtrx[i*num_fun+j];
}
}
o_cnt += num_fun;
}
}
row_index += num_fun;
}
}
}
if (num_procs > 1)
{
cnt = 0;
for (i=0; i < local_num_rows; i+=num_fun)
{
for (j=offd_i[i]; j < offd_i[i+1]; j++)
{
col_map_offd[cnt++] = offd_j[j];
}
}
hypre_qsort0(col_map_offd, 0, num_cols_offd-1);
for (i=0; i < num_fun*num_cols_offd; i++)
for (j=hypre_min(0,abs(i-num_fun)); j < num_cols_offd; j++)
if (offd_j[i] == col_map_offd[j])
{
offd_j[i] = j;
break;
}
}
for (i=0; i < num_procs+1; i++)
global_part[i] *= num_fun;
for (j=1; j< num_fun; j++)
{
for (i=0; i<local_grid_size; i++)
{
row = i*num_fun+j;
diag_index = diag_i[row];
index = diag_index+j;
val = diag_data[diag_index];
col = diag_j[diag_index];
diag_data[diag_index] = diag_data[index];
diag_j[diag_index] = diag_j[index];
diag_data[index] = val;
diag_j[index] = col;
}
}
#ifdef HYPRE_NO_GLOBAL_PARTITION
/* ideally we would use less storage earlier in this function, but this is fine
for testing */
{
HYPRE_Int tmp1, tmp2;
tmp1 = global_part[my_id];
tmp2 = global_part[my_id + 1];
hypre_TFree(global_part);
global_part = hypre_CTAlloc(HYPRE_Int, 2);
global_part[0] = tmp1;
global_part[1] = tmp2;
}
#endif
A = hypre_ParCSRMatrixCreate(comm, num_fun*grid_size, num_fun*grid_size,
global_part, global_part, num_cols_offd,
diag_i[local_num_rows],
offd_i[local_num_rows]);
hypre_ParCSRMatrixColMapOffd(A) = col_map_offd;
diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrixI(diag) = diag_i;
hypre_CSRMatrixJ(diag) = diag_j;
hypre_CSRMatrixData(diag) = diag_data;
offd = hypre_ParCSRMatrixOffd(A);
hypre_CSRMatrixI(offd) = offd_i;
if (num_cols_offd)
{
hypre_CSRMatrixJ(offd) = offd_j;
hypre_CSRMatrixData(offd) = offd_data;
}
hypre_TFree(nx_part);
hypre_TFree(ny_part);
hypre_TFree(nz_part);
return (HYPRE_ParCSRMatrix) A;
}
HYPRE_ParCSRMatrix
GenerateLaplacian( MPI_Comm comm,
HYPRE_Int nx,
HYPRE_Int ny,
HYPRE_Int nz,
HYPRE_Int P,
HYPRE_Int Q,
HYPRE_Int R,
HYPRE_Int p,
HYPRE_Int q,
HYPRE_Int r,
HYPRE_Real *value )
{
hypre_ParCSRMatrix *A;
hypre_CSRMatrix *diag;
hypre_CSRMatrix *offd;
HYPRE_Int *diag_i;
HYPRE_Int *diag_j;
HYPRE_Real *diag_data;
HYPRE_Int *offd_i;
HYPRE_Int *offd_j = NULL;
HYPRE_Real *offd_data = NULL;
HYPRE_Int *global_part;
HYPRE_Int ix, iy, iz;
HYPRE_Int cnt, o_cnt;
HYPRE_Int local_num_rows;
HYPRE_Int *col_map_offd;
HYPRE_Int row_index;
HYPRE_Int i,j;
HYPRE_Int nx_local, ny_local, nz_local;
HYPRE_Int nx_size, ny_size, nz_size;
HYPRE_Int num_cols_offd;
HYPRE_Int grid_size;
HYPRE_Int *nx_part;
HYPRE_Int *ny_part;
HYPRE_Int *nz_part;
HYPRE_Int num_procs, my_id;
HYPRE_Int P_busy, Q_busy, R_busy;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
grid_size = nx*ny*nz;
hypre_GeneratePartitioning(nx,P,&nx_part);
hypre_GeneratePartitioning(ny,Q,&ny_part);
hypre_GeneratePartitioning(nz,R,&nz_part);
global_part = hypre_CTAlloc(HYPRE_Int,P*Q*R+1);
global_part[0] = 0;
cnt = 1;
for (iz = 0; iz < R; iz++)
{
nz_size = nz_part[iz+1]-nz_part[iz];
for (iy = 0; iy < Q; iy++)
{
ny_size = ny_part[iy+1]-ny_part[iy];
for (ix = 0; ix < P; ix++)
{
nx_size = nx_part[ix+1] - nx_part[ix];
global_part[cnt] = global_part[cnt-1];
global_part[cnt++] += nx_size*ny_size*nz_size;
}
}
}
nx_local = nx_part[p+1] - nx_part[p];
ny_local = ny_part[q+1] - ny_part[q];
nz_local = nz_part[r+1] - nz_part[r];
my_id = r*(P*Q) + q*P + p;
num_procs = P*Q*R;
local_num_rows = nx_local*ny_local*nz_local;
diag_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
offd_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
P_busy = hypre_min(nx,P);
Q_busy = hypre_min(ny,Q);
R_busy = hypre_min(nz,R);
num_cols_offd = 0;
if (p) num_cols_offd += ny_local*nz_local;
if (p < P_busy-1) num_cols_offd += ny_local*nz_local;
if (q) num_cols_offd += nx_local*nz_local;
if (q < Q_busy-1) num_cols_offd += nx_local*nz_local;
if (r) num_cols_offd += nx_local*ny_local;
if (r < R_busy-1) num_cols_offd += nx_local*ny_local;
if (!local_num_rows) num_cols_offd = 0;
col_map_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
cnt = 1;
o_cnt = 1;
diag_i[0] = 0;
offd_i[0] = 0;
for (iz = nz_part[r]; iz < nz_part[r+1]; iz++)
{
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
diag_i[cnt] = diag_i[cnt-1];
offd_i[o_cnt] = offd_i[o_cnt-1];
diag_i[cnt]++;
if (iz > nz_part[r])
diag_i[cnt]++;
else
{
if (iz)
{
offd_i[o_cnt]++;
}
}
if (iy > ny_part[q])
diag_i[cnt]++;
else
{
if (iy)
{
offd_i[o_cnt]++;
}
}
if (ix > nx_part[p])
diag_i[cnt]++;
else
{
if (ix)
{
offd_i[o_cnt]++;
}
}
if (ix+1 < nx_part[p+1])
diag_i[cnt]++;
else
{
if (ix+1 < nx)
{
offd_i[o_cnt]++;
}
}
if (iy+1 < ny_part[q+1])
diag_i[cnt]++;
else
{
if (iy+1 < ny)
{
offd_i[o_cnt]++;
}
}
if (iz+1 < nz_part[r+1])
diag_i[cnt]++;
else
{
if (iz+1 < nz)
{
offd_i[o_cnt]++;
}
}
cnt++;
o_cnt++;
}
}
}
diag_j = hypre_CTAlloc(HYPRE_Int, diag_i[local_num_rows]);
diag_data = hypre_CTAlloc(HYPRE_Real, diag_i[local_num_rows]);
if (num_procs > 1)
{
offd_j = hypre_CTAlloc(HYPRE_Int, offd_i[local_num_rows]);
offd_data = hypre_CTAlloc(HYPRE_Real, offd_i[local_num_rows]);
}
row_index = 0;
cnt = 0;
o_cnt = 0;
for (iz = nz_part[r]; iz < nz_part[r+1]; iz++)
{
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
diag_j[cnt] = row_index;
diag_data[cnt++] = value[0];
if (iz > nz_part[r])
{
diag_j[cnt] = row_index-nx_local*ny_local;
diag_data[cnt++] = value[3];
}
else
{
if (iz)
{
offd_j[o_cnt] = hypre_map(ix,iy,iz-1,p,q,r-1,P,Q,R,
nx_part,ny_part,nz_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
if (iy > ny_part[q])
{
diag_j[cnt] = row_index-nx_local;
diag_data[cnt++] = value[2];
}
else
{
if (iy)
{
offd_j[o_cnt] = hypre_map(ix,iy-1,iz,p,q-1,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
offd_data[o_cnt++] = value[2];
}
}
if (ix > nx_part[p])
{
diag_j[cnt] = row_index-1;
diag_data[cnt++] = value[1];
}
else
{
if (ix)
{
offd_j[o_cnt] = hypre_map(ix-1,iy,iz,p-1,q,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
offd_data[o_cnt++] = value[1];
}
}
if (ix+1 < nx_part[p+1])
{
diag_j[cnt] = row_index+1;
diag_data[cnt++] = value[1];
}
else
{
if (ix+1 < nx)
{
offd_j[o_cnt] = hypre_map(ix+1,iy,iz,p+1,q,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
offd_data[o_cnt++] = value[1];
}
}
if (iy+1 < ny_part[q+1])
{
diag_j[cnt] = row_index+nx_local;
diag_data[cnt++] = value[2];
}
else
{
if (iy+1 < ny)
{
offd_j[o_cnt] = hypre_map(ix,iy+1,iz,p,q+1,r,P,Q,R,
nx_part,ny_part,nz_part,global_part);
offd_data[o_cnt++] = value[2];
}
}
if (iz+1 < nz_part[r+1])
{
diag_j[cnt] = row_index+nx_local*ny_local;
diag_data[cnt++] = value[3];
}
else
{
if (iz+1 < nz)
{
offd_j[o_cnt] = hypre_map(ix,iy,iz+1,p,q,r+1,P,Q,R,
nx_part,ny_part,nz_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
row_index++;
}
}
}
if (num_procs > 1)
{
for (i=0; i < num_cols_offd; i++)
col_map_offd[i] = offd_j[i];
hypre_qsort0(col_map_offd, 0, num_cols_offd-1);
for (i=0; i < num_cols_offd; i++)
for (j=0; j < num_cols_offd; j++)
if (offd_j[i] == col_map_offd[j])
{
offd_j[i] = j;
break;
}
}
#ifdef HYPRE_NO_GLOBAL_PARTITION
/* ideally we would use less storage earlier in this function, but this is fine
for testing */
{
HYPRE_Int tmp1, tmp2;
tmp1 = global_part[my_id];
tmp2 = global_part[my_id + 1];
hypre_TFree(global_part);
global_part = hypre_CTAlloc(HYPRE_Int, 2);
global_part[0] = tmp1;
global_part[1] = tmp2;
}
#endif
A = hypre_ParCSRMatrixCreate(comm, grid_size, grid_size,
global_part, global_part, num_cols_offd,
diag_i[local_num_rows],
offd_i[local_num_rows]);
hypre_ParCSRMatrixColMapOffd(A) = col_map_offd;
diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrixI(diag) = diag_i;
hypre_CSRMatrixJ(diag) = diag_j;
hypre_CSRMatrixData(diag) = diag_data;
offd = hypre_ParCSRMatrixOffd(A);
hypre_CSRMatrixI(offd) = offd_i;
if (num_cols_offd)
{
hypre_CSRMatrixJ(offd) = offd_j;
hypre_CSRMatrixData(offd) = offd_data;
}
hypre_TFree(nx_part);
hypre_TFree(ny_part);
hypre_TFree(nz_part);
return (HYPRE_ParCSRMatrix) A;
}
HYPRE_Int
hypre_GenerateSendMapAndCommPkg(MPI_Comm comm, HYPRE_Int num_sends, HYPRE_Int num_recvs,
HYPRE_Int *recv_procs, HYPRE_Int *send_procs,
HYPRE_Int *recv_vec_starts, hypre_ParCSRMatrix *A)
{
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int i, j;
HYPRE_Int num_requests = num_sends+num_recvs;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
HYPRE_Int vec_len, vec_start;
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(A);
/*--------------------------------------------------------------------------
* generate send_map_starts and send_map_elmts
*--------------------------------------------------------------------------*/
requests = hypre_CTAlloc(hypre_MPI_Request,num_requests);
status = hypre_CTAlloc(hypre_MPI_Status,num_requests);
send_map_starts = hypre_CTAlloc(HYPRE_Int, num_sends+1);
j = 0;
for (i=0; i < num_sends; i++)
hypre_MPI_Irecv(&send_map_starts[i+1],1,HYPRE_MPI_INT,send_procs[i],0,comm,
&requests[j++]);
for (i=0; i < num_recvs; i++)
{
vec_len = recv_vec_starts[i+1] - recv_vec_starts[i];
hypre_MPI_Isend(&vec_len,1,HYPRE_MPI_INT, recv_procs[i],0,comm,&requests[j++]);
}
hypre_MPI_Waitall(j,requests,status);
send_map_starts[0] = 0;
for (i=0; i < num_sends; i++)
send_map_starts[i+1] += send_map_starts[i];
send_map_elmts = hypre_CTAlloc(HYPRE_Int,send_map_starts[num_sends]);
j = 0;
for (i=0; i < num_sends; i++)
{
vec_start = send_map_starts[i];
vec_len = send_map_starts[i+1]-vec_start;
hypre_MPI_Irecv(&send_map_elmts[vec_start],vec_len,HYPRE_MPI_INT,
send_procs[i],0,comm,&requests[j++]);
}
for (i=0; i < num_recvs; i++)
{
vec_start = recv_vec_starts[i];
vec_len = recv_vec_starts[i+1] - vec_start;
hypre_MPI_Isend(&col_map_offd[vec_start],vec_len,HYPRE_MPI_INT,
recv_procs[i],0,comm,&requests[j++]);
}
hypre_MPI_Waitall(j,requests,status);
for (i=0; i < send_map_starts[num_sends]; i++)
send_map_elmts[i] -= first_col_diag;
comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg) = comm;
hypre_ParCSRCommPkgNumSends(comm_pkg) = num_sends;
hypre_ParCSRCommPkgNumRecvs(comm_pkg) = num_recvs;
hypre_ParCSRCommPkgSendProcs(comm_pkg) = send_procs;
hypre_ParCSRCommPkgRecvProcs(comm_pkg) = recv_procs;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg) = recv_vec_starts;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg) = send_map_starts;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg) = send_map_elmts;
hypre_TFree(status);
hypre_TFree(requests);
hypre_ParCSRMatrixCommPkg(A) = comm_pkg;
return 0;
}
hypre_ParCSRMatrix * hypre_ParMatMinus_F(
hypre_ParCSRMatrix * P, hypre_ParCSRMatrix * C, HYPRE_Int * CF_marker )
/* hypre_ParMatMinus_F subtracts selected rows of its second argument
from selected rows of its first argument. The marker array
determines which rows are affected - those for which CF_marker<0.
The result is returned as a new matrix.
*/
{
/*
If P=(Pik),C=(Cik), i in Fine+Coarse, k in Coarse, we want
new Pik = Pik - Cik, for Fine i only, all k.
This computation is purely local.
*/
/* This is _not_ a general-purpose matrix subtraction function.
This is written for an interpolation problem where it is known that C(i,k)
exists whenever P(i,k) does (because C=A*P where A has nonzero diagonal elements).
*/
hypre_ParCSRMatrix *Pnew;
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
hypre_CSRMatrix *C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrix *C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrix *Pnew_diag;
hypre_CSRMatrix *Pnew_offd;
double *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int *P_col_map_offd = hypre_ParCSRMatrixColMapOffd( P );
double *C_diag_data = hypre_CSRMatrixData(C_diag);
HYPRE_Int *C_diag_i = hypre_CSRMatrixI(C_diag);
HYPRE_Int *C_diag_j = hypre_CSRMatrixJ(C_diag);
double *C_offd_data = hypre_CSRMatrixData(C_offd);
HYPRE_Int *C_offd_i = hypre_CSRMatrixI(C_offd);
HYPRE_Int *C_offd_j = hypre_CSRMatrixJ(C_offd);
HYPRE_Int *C_col_map_offd = hypre_ParCSRMatrixColMapOffd( C );
HYPRE_Int *Pnew_diag_i;
HYPRE_Int *Pnew_diag_j;
double *Pnew_diag_data;
HYPRE_Int *Pnew_offd_i;
HYPRE_Int *Pnew_offd_j;
double *Pnew_offd_data;
HYPRE_Int *Pnew_j2m;
HYPRE_Int *Pnew_col_map_offd;
HYPRE_Int num_rows_diag_C = hypre_CSRMatrixNumRows(C_diag);
/* HYPRE_Int num_rows_offd_C = hypre_CSRMatrixNumRows(C_offd); */
HYPRE_Int num_cols_offd_C = hypre_CSRMatrixNumCols(C_offd);
HYPRE_Int num_cols_offd_P = hypre_CSRMatrixNumCols(P_offd);
HYPRE_Int num_cols_offd_Pnew, num_rows_offd_Pnew;
HYPRE_Int i1, jmin, jmax, jrange, jrangem1;
HYPRE_Int j, m, mc, mp, jc, jp, jP, jC, jg, jCg, jPg;
double dc, dp;
/* Pnew = hypre_ParCSRMatrixCompleteClone( C );*/
Pnew = hypre_ParCSRMatrixUnion( C, P );
;
hypre_ParCSRMatrixZero_F( Pnew, CF_marker ); /* fine rows of Pnew set to 0 */
hypre_ParCSRMatrixCopy_C( Pnew, C, CF_marker ); /* coarse rows of Pnew copied from C (or P) */
/* ...Zero_F may not be needed depending on how Pnew is made */
Pnew_diag = hypre_ParCSRMatrixDiag(Pnew);
Pnew_offd = hypre_ParCSRMatrixOffd(Pnew);
Pnew_diag_i = hypre_CSRMatrixI(Pnew_diag);
Pnew_diag_j = hypre_CSRMatrixJ(Pnew_diag);
Pnew_offd_i = hypre_CSRMatrixI(Pnew_offd);
Pnew_offd_j = hypre_CSRMatrixJ(Pnew_offd);
Pnew_diag_data = hypre_CSRMatrixData(Pnew_diag);
Pnew_offd_data = hypre_CSRMatrixData(Pnew_offd);
Pnew_col_map_offd = hypre_ParCSRMatrixColMapOffd( Pnew );
num_rows_offd_Pnew = hypre_CSRMatrixNumRows(Pnew_offd);
num_cols_offd_Pnew = hypre_CSRMatrixNumCols(Pnew_offd);
/* Find the j-ranges, needed to allocate a "reverse lookup" array. */
/* This is the max j - min j over P and Pnew (which here is a copy of C).
Only the diag block is considered. */
/* For scalability reasons (jrange can get big) this won't work for the offd block.
Also, indexing is more complicated in the offd block (c.f. col_map_offd).
It's not clear, though whether the "quadratic" algorithm I'm using for the offd
block is really any slower than the more complicated "linear" algorithm here. */
jrange = 0;
jrangem1=-1;
for ( i1 = 0; i1 < num_rows_diag_C; i1++ )
{
if ( CF_marker[i1]<0 && hypre_CSRMatrixNumNonzeros(Pnew_diag)>0 ) /* only Fine rows matter */
{
jmin = Pnew_diag_j[ Pnew_diag_i[i1] ];
jmax = Pnew_diag_j[ Pnew_diag_i[i1+1]-1 ];
jrangem1 = jmax-jmin;
jrange = hypre_max(jrange,jrangem1+1);
/* If columns (of a given row) were in increasing order, the above would be sufficient.
If not, the following would be necessary (and sufficient) */
jmin = Pnew_diag_j[ Pnew_diag_i[i1] ];
jmax = Pnew_diag_j[ Pnew_diag_i[i1] ];
for ( m=Pnew_diag_i[i1]+1; m<Pnew_diag_i[i1+1]; ++m )
{
j = Pnew_diag_j[m];
jmin = hypre_min( jmin, j );
jmax = hypre_max( jmax, j );
}
for ( m=P_diag_i[i1]; m<P_diag_i[i1+1]; ++m )
{
j = P_diag_j[m];
jmin = hypre_min( jmin, j );
jmax = hypre_max( jmax, j );
}
jrangem1 = jmax-jmin;
jrange = hypre_max(jrange,jrangem1+1);
}
}
/*-----------------------------------------------------------------------
* Loop over Pnew_diag rows. Construct a temporary reverse array:
* If j is a column number, Pnew_j2m[j] is the array index for j, i.e.
* Pnew_diag_j[ Pnew_j2m[j] ] = j
*-----------------------------------------------------------------------*/
Pnew_j2m = hypre_CTAlloc( HYPRE_Int, jrange );
for ( i1 = 0; i1 < num_rows_diag_C; i1++ )
{
if ( CF_marker[i1]<0 && hypre_CSRMatrixNumNonzeros(Pnew_diag)>0 ) /* Fine data only */
{
/* just needed for an assertion below... */
for ( j=0; j<jrange; ++j ) Pnew_j2m[j] = -1;
jmin = Pnew_diag_j[ Pnew_diag_i[i1] ];
/* If columns (of a given row) were in increasing order, the above line would be sufficient.
If not, the following loop would have to be added (or store the jmin computed above )*/
for ( m=Pnew_diag_i[i1]+1; m<Pnew_diag_i[i1+1]; ++m )
{
j = Pnew_diag_j[m];
jmin = hypre_min( jmin, j );
}
for ( m=P_diag_i[i1]; m<P_diag_i[i1+1]; ++m )
{
j = P_diag_j[m];
jmin = hypre_min( jmin, j );
}
for ( m = Pnew_diag_i[i1]; m<Pnew_diag_i[i1+1]; ++m )
{
j = Pnew_diag_j[m];
hypre_assert( j-jmin>=0 );
hypre_assert( j-jmin<jrange );
Pnew_j2m[ j-jmin ] = m;
}
/*-----------------------------------------------------------------------
* Loop over C_diag data for the current row.
* Subtract each C data entry from the corresponding Pnew entry.
*-----------------------------------------------------------------------*/
for ( mc=C_diag_i[i1]; mc<C_diag_i[i1+1]; ++mc )
{
jc = C_diag_j[mc];
dc = C_diag_data[mc];
m = Pnew_j2m[jc-jmin];
hypre_assert( m>=0 );
Pnew_diag_data[m] -= dc;
}
/*-----------------------------------------------------------------------
* Loop over P_diag data for the current row.
* Add each P data entry from the corresponding Pnew entry.
*-----------------------------------------------------------------------*/
for ( mp=P_diag_i[i1]; mp<P_diag_i[i1+1]; ++mp )
{
jp = P_diag_j[mp];
dp = P_diag_data[mp];
m = Pnew_j2m[jp-jmin];
hypre_assert( m>=0 );
Pnew_diag_data[m] += dp;
}
}
}
/*-----------------------------------------------------------------------
* Repeat for the offd block.
*-----------------------------------------------------------------------*/
for ( i1 = 0; i1 < num_rows_offd_Pnew; i1++ )
{
if ( CF_marker[i1]<0 && hypre_CSRMatrixNumNonzeros(Pnew_offd)>0 ) /* Fine data only */
{
if ( num_cols_offd_Pnew )
{
/* This is a simple quadratic algorithm. If necessary I may try
to implement the ideas used on the diag block later. */
for ( m = Pnew_offd_i[i1]; m<Pnew_offd_i[i1+1]; ++m )
{
j = Pnew_offd_j[m];
jg = Pnew_col_map_offd[j];
Pnew_offd_data[m] = 0;
if ( num_cols_offd_C )
for ( mc=C_offd_i[i1]; mc<C_offd_i[i1+1]; ++mc )
{
jC = C_offd_j[mc];
jCg = C_col_map_offd[jC];
if ( jCg==jg ) Pnew_offd_data[m] -= C_offd_data[mc];
}
if ( num_cols_offd_P )
for ( mp=P_offd_i[i1]; mp<P_offd_i[i1+1]; ++mp )
{
jP = P_offd_j[mp];
jPg = P_col_map_offd[jP];
if ( jPg==jg ) Pnew_offd_data[m] += P_offd_data[mp];
}
}
}
}
}
hypre_TFree(Pnew_j2m);
return Pnew;
}
HYPRE_Int
hypre_BoomerAMGBlockCreateNodalA(hypre_ParCSRBlockMatrix *A,
HYPRE_Int option,
HYPRE_Int diag_option,
hypre_ParCSRMatrix **AN_ptr)
{
MPI_Comm comm = hypre_ParCSRBlockMatrixComm(A);
hypre_CSRBlockMatrix *A_diag = hypre_ParCSRBlockMatrixDiag(A);
HYPRE_Int *A_diag_i = hypre_CSRBlockMatrixI(A_diag);
HYPRE_Real *A_diag_data = hypre_CSRBlockMatrixData(A_diag);
HYPRE_Int block_size = hypre_CSRBlockMatrixBlockSize(A_diag);
HYPRE_Int bnnz = block_size*block_size;
hypre_CSRBlockMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *A_offd_i = hypre_CSRBlockMatrixI(A_offd);
HYPRE_Real *A_offd_data = hypre_CSRBlockMatrixData(A_offd);
HYPRE_Int *A_diag_j = hypre_CSRBlockMatrixJ(A_diag);
HYPRE_Int *A_offd_j = hypre_CSRBlockMatrixJ(A_offd);
HYPRE_Int *row_starts = hypre_ParCSRBlockMatrixRowStarts(A);
HYPRE_Int *col_map_offd = hypre_ParCSRBlockMatrixColMapOffd(A);
HYPRE_Int num_nonzeros_diag;
HYPRE_Int num_nonzeros_offd = 0;
HYPRE_Int num_cols_offd = 0;
hypre_ParCSRMatrix *AN;
hypre_CSRMatrix *AN_diag;
HYPRE_Int *AN_diag_i;
HYPRE_Int *AN_diag_j=NULL;
HYPRE_Real *AN_diag_data = NULL;
hypre_CSRMatrix *AN_offd;
HYPRE_Int *AN_offd_i;
HYPRE_Int *AN_offd_j = NULL;
HYPRE_Real *AN_offd_data = NULL;
HYPRE_Int *col_map_offd_AN = NULL;
HYPRE_Int *row_starts_AN;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRBlockMatrixCommPkg(A);
HYPRE_Int num_sends;
HYPRE_Int num_recvs;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int *recv_procs;
HYPRE_Int *recv_vec_starts;
hypre_ParCSRCommPkg *comm_pkg_AN = NULL;
HYPRE_Int *send_procs_AN = NULL;
HYPRE_Int *send_map_starts_AN = NULL;
HYPRE_Int *send_map_elmts_AN = NULL;
HYPRE_Int *recv_procs_AN = NULL;
HYPRE_Int *recv_vec_starts_AN = NULL;
HYPRE_Int i;
HYPRE_Int ierr = 0;
HYPRE_Int num_procs;
HYPRE_Int cnt;
HYPRE_Int norm_type;
HYPRE_Int global_num_nodes;
HYPRE_Int num_nodes;
HYPRE_Int index, k;
HYPRE_Real tmp;
HYPRE_Real sum;
hypre_MPI_Comm_size(comm,&num_procs);
if (!comm_pkg)
{
hypre_BlockMatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRBlockMatrixCommPkg(A);
}
norm_type = fabs(option);
/* Set up the new matrix AN */
#ifdef HYPRE_NO_GLOBAL_PARTITION
row_starts_AN = hypre_CTAlloc(HYPRE_Int, 2);
for (i=0; i < 2; i++)
{
row_starts_AN[i] = row_starts[i];
}
#else
row_starts_AN = hypre_CTAlloc(HYPRE_Int, num_procs+1);
for (i=0; i < num_procs+1; i++)
{
row_starts_AN[i] = row_starts[i];
}
#endif
global_num_nodes = hypre_ParCSRBlockMatrixGlobalNumRows(A);
num_nodes = hypre_CSRBlockMatrixNumRows(A_diag);
/* the diag part */
num_nonzeros_diag = A_diag_i[num_nodes];
AN_diag_i = hypre_CTAlloc(HYPRE_Int, num_nodes+1);
for (i=0; i <= num_nodes; i++)
{
AN_diag_i[i] = A_diag_i[i];
}
AN_diag_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros_diag);
AN_diag_data = hypre_CTAlloc(HYPRE_Real, num_nonzeros_diag);
AN_diag = hypre_CSRMatrixCreate(num_nodes, num_nodes, num_nonzeros_diag);
hypre_CSRMatrixI(AN_diag) = AN_diag_i;
hypre_CSRMatrixJ(AN_diag) = AN_diag_j;
hypre_CSRMatrixData(AN_diag) = AN_diag_data;
for (i=0; i< num_nonzeros_diag; i++)
{
AN_diag_j[i] = A_diag_j[i];
hypre_CSRBlockMatrixBlockNorm(norm_type, &A_diag_data[i*bnnz],
&tmp, block_size);
AN_diag_data[i] = tmp;
}
if (diag_option ==1 )
{
/* make the diag entry the negative of the sum of off-diag entries (NEED
* to get more below!)*/
/* the diagonal is the first element listed in each row - */
for (i=0; i < num_nodes; i++)
{
index = AN_diag_i[i];
sum = 0.0;
for (k = AN_diag_i[i]+1; k < AN_diag_i[i+1]; k++)
{
sum += AN_diag_data[k];
}
AN_diag_data[index] = -sum;
}
}
else if (diag_option == 2)
{
/* make all diagonal entries negative */
/* the diagonal is the first element listed in each row - */
for (i=0; i < num_nodes; i++)
{
index = AN_diag_i[i];
AN_diag_data[index] = -AN_diag_data[index];
}
}
/* copy the commpkg */
if (comm_pkg)
{
comm_pkg_AN = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg_AN) = comm;
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
hypre_ParCSRCommPkgNumSends(comm_pkg_AN) = num_sends;
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
hypre_ParCSRCommPkgNumRecvs(comm_pkg_AN) = num_recvs;
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
if (num_sends)
{
send_procs_AN = hypre_CTAlloc(HYPRE_Int, num_sends);
send_map_elmts_AN = hypre_CTAlloc(HYPRE_Int, send_map_starts[num_sends]);
}
send_map_starts_AN = hypre_CTAlloc(HYPRE_Int, num_sends+1);
send_map_starts_AN[0] = 0;
for (i=0; i < num_sends; i++)
{
send_procs_AN[i] = send_procs[i];
send_map_starts_AN[i+1] = send_map_starts[i+1];
}
cnt = send_map_starts_AN[num_sends];
for (i=0; i< cnt; i++)
{
send_map_elmts_AN[i] = send_map_elmts[i];
}
hypre_ParCSRCommPkgSendProcs(comm_pkg_AN) = send_procs_AN;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg_AN) = send_map_starts_AN;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg_AN) = send_map_elmts_AN;
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
recv_vec_starts_AN = hypre_CTAlloc(HYPRE_Int, num_recvs+1);
if (num_recvs) recv_procs_AN = hypre_CTAlloc(HYPRE_Int, num_recvs);
recv_vec_starts_AN[0] = recv_vec_starts[0];
for (i=0; i < num_recvs; i++)
{
recv_procs_AN[i] = recv_procs[i];
recv_vec_starts_AN[i+1] = recv_vec_starts[i+1];
}
hypre_ParCSRCommPkgRecvProcs(comm_pkg_AN) = recv_procs_AN;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg_AN) = recv_vec_starts_AN;
}
/* the off-diag part */
num_cols_offd = hypre_CSRBlockMatrixNumCols(A_offd);
col_map_offd_AN = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
for (i=0; i < num_cols_offd; i++)
{
col_map_offd_AN[i] = col_map_offd[i];
}
num_nonzeros_offd = A_offd_i[num_nodes];
AN_offd_i = hypre_CTAlloc(HYPRE_Int, num_nodes+1);
for (i=0; i <= num_nodes; i++)
{
AN_offd_i[i] = A_offd_i[i];
}
AN_offd_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros_offd);
AN_offd_data = hypre_CTAlloc(HYPRE_Real, num_nonzeros_offd);
for (i=0; i< num_nonzeros_offd; i++)
{
AN_offd_j[i] = A_offd_j[i];
hypre_CSRBlockMatrixBlockNorm(norm_type, &A_offd_data[i*bnnz],
&tmp, block_size);
AN_offd_data[i] = tmp;
}
AN_offd = hypre_CSRMatrixCreate(num_nodes, num_cols_offd, num_nonzeros_offd);
hypre_CSRMatrixI(AN_offd) = AN_offd_i;
hypre_CSRMatrixJ(AN_offd) = AN_offd_j;
hypre_CSRMatrixData(AN_offd) = AN_offd_data;
if (diag_option ==1 )
{
/* make the diag entry the negative of the sum of off-diag entries (here
we are adding the off_diag contribution)*/
/* the diagonal is the first element listed in each row of AN_diag_data - */
for (i=0; i < num_nodes; i++)
{
sum = 0.0;
for (k = AN_offd_i[i]; k < AN_offd_i[i+1]; k++)
{
sum += AN_offd_data[k];
}
index = AN_diag_i[i];/* location of diag entry in data */
AN_diag_data[index] -= sum; /* subtract from current value */
}
}
/* now create AN */
AN = hypre_ParCSRMatrixCreate(comm, global_num_nodes, global_num_nodes,
row_starts_AN, row_starts_AN, num_cols_offd,
num_nonzeros_diag, num_nonzeros_offd);
/* we already created the diag and offd matrices - so we don't need the ones
created above */
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixDiag(AN));
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixOffd(AN));
hypre_ParCSRMatrixDiag(AN) = AN_diag;
hypre_ParCSRMatrixOffd(AN) = AN_offd;
hypre_ParCSRMatrixColMapOffd(AN) = col_map_offd_AN;
hypre_ParCSRMatrixCommPkg(AN) = comm_pkg_AN;
*AN_ptr = AN;
return (ierr);
}
HYPRE_Int main( HYPRE_Int argc, char *argv[] )
{
hypre_ParCSRMatrix *par_matrix, *g_matrix, **submatrices;
hypre_CSRMatrix *A_diag, *A_offd;
hypre_CSRBlockMatrix *diag;
hypre_CSRBlockMatrix *offd;
hypre_ParCSRBlockMatrix *par_blk_matrix, *par_blk_matrixT, *rap_matrix;
hypre_Vector *x_local;
hypre_Vector *y_local;
hypre_ParVector *x;
hypre_ParVector *y;
HYPRE_Solver gmres_solver, precon;
HYPRE_Int *diag_i, *diag_j, *offd_i, *offd_j;
HYPRE_Int *diag_i2, *diag_j2, *offd_i2, *offd_j2;
double *diag_d, *diag_d2, *offd_d, *offd_d2;
HYPRE_Int mypid, local_size, nprocs;
HYPRE_Int global_num_rows, global_num_cols, num_cols_offd;
HYPRE_Int num_nonzeros_diag, num_nonzeros_offd, *colMap;
HYPRE_Int ii, jj, kk, row, col, nnz, *indices, *colMap2;
double *data, ddata, *y_data;
HYPRE_Int *row_starts, *col_starts, *rstarts, *cstarts;
HYPRE_Int *row_starts2, *col_starts2;
HYPRE_Int block_size=2, bnnz=4, *index_set;
FILE *fp;
/* --------------------------------------------- */
/* Initialize MPI */
/* --------------------------------------------- */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &mypid);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &nprocs);
/* build and fetch matrix */
MyBuildParLaplacian9pt((HYPRE_ParCSRMatrix *) &par_matrix);
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(par_matrix);
global_num_cols = hypre_ParCSRMatrixGlobalNumCols(par_matrix);
row_starts = hypre_ParCSRMatrixRowStarts(par_matrix);
col_starts = hypre_ParCSRMatrixColStarts(par_matrix);
A_diag = hypre_ParCSRMatrixDiag(par_matrix);
A_offd = hypre_ParCSRMatrixOffd(par_matrix);
num_cols_offd = hypre_CSRMatrixNumCols(A_offd);
num_nonzeros_diag = hypre_CSRMatrixNumNonzeros(A_diag);
num_nonzeros_offd = hypre_CSRMatrixNumNonzeros(A_offd);
/* --------------------------------------------- */
/* build vector and apply matvec */
/* --------------------------------------------- */
x = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_cols,col_starts);
hypre_ParVectorSetPartitioningOwner(x,0);
hypre_ParVectorInitialize(x);
x_local = hypre_ParVectorLocalVector(x);
data = hypre_VectorData(x_local);
local_size = col_starts[mypid+1] - col_starts[mypid];
for (ii = 0; ii < local_size; ii++) data[ii] = 1.0;
y = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_rows,row_starts);
hypre_ParVectorSetPartitioningOwner(y,0);
hypre_ParVectorInitialize(y);
hypre_ParCSRMatrixMatvec (1.0, par_matrix, x, 0.0, y);
ddata = hypre_ParVectorInnerProd(y, y);
if (mypid == 0) hypre_printf("y inner product = %e\n", ddata);
hypre_ParVectorDestroy(x);
hypre_ParVectorDestroy(y);
/* --------------------------------------------- */
/* build block matrix */
/* --------------------------------------------- */
rstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) rstarts[ii] = row_starts[ii];
cstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) cstarts[ii] = col_starts[ii];
par_blk_matrix = hypre_ParCSRBlockMatrixCreate(hypre_MPI_COMM_WORLD,block_size,
global_num_rows, global_num_cols, rstarts,
cstarts, num_cols_offd, num_nonzeros_diag,
num_nonzeros_offd);
colMap = hypre_ParCSRMatrixColMapOffd(par_matrix);
if (num_cols_offd > 0) colMap2 = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
else colMap2 = NULL;
for (ii = 0; ii < num_cols_offd; ii++) colMap2[ii] = colMap[ii];
hypre_ParCSRBlockMatrixColMapOffd(par_blk_matrix) = colMap2;
diag_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(par_matrix));
diag_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixDiag(par_matrix));
diag_d = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(par_matrix));
diag = hypre_ParCSRBlockMatrixDiag(par_blk_matrix);
diag_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
diag_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_diag);
diag_d2 = hypre_CTAlloc(double, num_nonzeros_diag*bnnz);
for (ii = 0; ii <= local_size; ii++) diag_i2[ii] = diag_i[ii];
for (ii = 0; ii < num_nonzeros_diag; ii++) diag_j2[ii] = diag_j[ii];
hypre_CSRBlockMatrixI(diag) = diag_i2;
hypre_CSRBlockMatrixJ(diag) = diag_j2;
for (ii = 0; ii < num_nonzeros_diag; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
diag_d2[ii*bnnz+jj*block_size+kk] = diag_d[ii];
else
diag_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(diag) = diag_d2;
offd_i = hypre_CSRMatrixI(hypre_ParCSRMatrixOffd(par_matrix));
offd_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixOffd(par_matrix));
offd_d = hypre_CSRMatrixData(hypre_ParCSRMatrixOffd(par_matrix));
offd = hypre_ParCSRBlockMatrixOffd(par_blk_matrix);
offd_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
for (ii = 0; ii <= local_size; ii++) offd_i2[ii] = offd_i[ii];
hypre_CSRBlockMatrixI(offd) = offd_i2;
if (num_cols_offd)
{
offd_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_offd);
for (ii = 0; ii < num_nonzeros_offd; ii++) offd_j2[ii] = offd_j[ii];
hypre_CSRBlockMatrixJ(offd) = offd_j2;
offd_d2 = hypre_CTAlloc(double, num_nonzeros_offd*bnnz);
for (ii = 0; ii < num_nonzeros_offd; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
offd_d2[ii*bnnz+jj*block_size+kk] = offd_d[ii];
else
offd_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(offd) = offd_d2;
}
else
{
hypre_ParCSRMatrix *
hypre_ParCSRBlockMatrixConvertToParCSRMatrix(hypre_ParCSRBlockMatrix *matrix)
{
MPI_Comm comm = hypre_ParCSRBlockMatrixComm(matrix);
hypre_CSRBlockMatrix *diag = hypre_ParCSRBlockMatrixDiag(matrix);
hypre_CSRBlockMatrix *offd = hypre_ParCSRBlockMatrixOffd(matrix);
HYPRE_Int block_size = hypre_ParCSRBlockMatrixBlockSize(matrix);
HYPRE_Int global_num_rows = hypre_ParCSRBlockMatrixGlobalNumRows(matrix);
HYPRE_Int global_num_cols = hypre_ParCSRBlockMatrixGlobalNumCols(matrix);
HYPRE_Int *row_starts = hypre_ParCSRBlockMatrixRowStarts(matrix);
HYPRE_Int *col_starts = hypre_ParCSRBlockMatrixColStarts(matrix);
HYPRE_Int num_cols_offd = hypre_CSRBlockMatrixNumCols(offd);
HYPRE_Int num_nonzeros_diag = hypre_CSRBlockMatrixNumNonzeros(diag);
HYPRE_Int num_nonzeros_offd = hypre_CSRBlockMatrixNumNonzeros(offd);
hypre_ParCSRMatrix *matrix_C;
HYPRE_Int *matrix_C_row_starts;
HYPRE_Int *matrix_C_col_starts;
HYPRE_Int *counter, *new_j_map;
HYPRE_Int size_j, size_map, index, new_num_cols, removed = 0;
HYPRE_Int *offd_j, *col_map_offd, *new_col_map_offd;
HYPRE_Int num_procs, i, j;
hypre_CSRMatrix *diag_nozeros, *offd_nozeros;
hypre_MPI_Comm_size(comm,&num_procs);
#ifdef HYPRE_NO_GLOBAL_PARTITION
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, 2);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, 2);
for(i = 0; i < 2; i++)
{
matrix_C_row_starts[i] = row_starts[i]*block_size;
matrix_C_col_starts[i] = col_starts[i]*block_size;
}
#else
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
for(i = 0; i < num_procs + 1; i++)
{
matrix_C_row_starts[i] = row_starts[i]*block_size;
matrix_C_col_starts[i] = col_starts[i]*block_size;
}
#endif
matrix_C = hypre_ParCSRMatrixCreate(comm, global_num_rows*block_size,
global_num_cols*block_size, matrix_C_row_starts,
matrix_C_col_starts, num_cols_offd*block_size,
num_nonzeros_diag*block_size*block_size,
num_nonzeros_offd*block_size*block_size);
hypre_ParCSRMatrixInitialize(matrix_C);
/* DIAG */
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixDiag(matrix_C));
hypre_ParCSRMatrixDiag(matrix_C) =
hypre_CSRBlockMatrixConvertToCSRMatrix(diag);
/* AB - added to delete zeros */
diag_nozeros = hypre_CSRMatrixDeleteZeros(
hypre_ParCSRMatrixDiag(matrix_C), 1e-14);
if(diag_nozeros)
{
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixDiag(matrix_C));
hypre_ParCSRMatrixDiag(matrix_C) = diag_nozeros;
}
/* OFF-DIAG */
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixOffd(matrix_C));
hypre_ParCSRMatrixOffd(matrix_C) =
hypre_CSRBlockMatrixConvertToCSRMatrix(offd);
/* AB - added to delete zeros - this just deletes from data and j arrays */
offd_nozeros = hypre_CSRMatrixDeleteZeros(
hypre_ParCSRMatrixOffd(matrix_C), 1e-14);
if(offd_nozeros)
{
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixOffd(matrix_C));
hypre_ParCSRMatrixOffd(matrix_C) = offd_nozeros;
removed = 1;
}
/* now convert the col_map_offd */
for (i = 0; i < num_cols_offd; i++)
for (j = 0; j < block_size; j++)
hypre_ParCSRMatrixColMapOffd(matrix_C)[i*block_size + j] =
hypre_ParCSRBlockMatrixColMapOffd(matrix)[i]*block_size + j;
/* if we deleted zeros, then it is possible that col_map_offd can be
compressed as well - this requires some amount of work that could be skipped... */
if (removed)
{
size_map = num_cols_offd*block_size;
counter = hypre_CTAlloc(HYPRE_Int, size_map);
new_j_map = hypre_CTAlloc(HYPRE_Int, size_map);
offd_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixOffd(matrix_C));
col_map_offd = hypre_ParCSRMatrixColMapOffd(matrix_C);
size_j = hypre_CSRMatrixNumNonzeros(hypre_ParCSRMatrixOffd(matrix_C));
/* mark which off_d entries are found in j */
for (i=0; i < size_j; i++)
{
counter[offd_j[i]] = 1;
}
/*now find new numbering for columns (we will delete the
cols where counter = 0*/
index = 0;
for (i=0; i < size_map; i++)
{
if (counter[i]) new_j_map[i] = index++;
}
new_num_cols = index;
/* if there are some col entries to remove: */
if (!(index == size_map))
{
/* go thru j and adjust entries */
for (i=0; i < size_j; i++)
{
offd_j[i] = new_j_map[offd_j[i]];
}
/*now go thru col map and get rid of non-needed entries */
new_col_map_offd = hypre_CTAlloc(HYPRE_Int, new_num_cols);
index = 0;
for (i=0; i < size_map; i++)
{
if (counter[i])
{
new_col_map_offd[index++] = col_map_offd[i];
}
}
/* set the new col map */
hypre_TFree(col_map_offd);
hypre_ParCSRMatrixColMapOffd(matrix_C) = new_col_map_offd;
/* modify the number of cols */
hypre_CSRMatrixNumCols(hypre_ParCSRMatrixOffd(matrix_C)) = new_num_cols;
}
hypre_TFree(new_j_map);
hypre_TFree(counter);
}
hypre_ParCSRMatrixSetNumNonzeros( matrix_C );
hypre_ParCSRMatrixSetDNumNonzeros( matrix_C );
/* we will not copy the comm package */
hypre_ParCSRMatrixCommPkg(matrix_C) = NULL;
return matrix_C;
}
HYPRE_Int
hypre_BoomerAMGCreateScalarCFS(hypre_ParCSRMatrix *SN,
HYPRE_Int *CFN_marker,
HYPRE_Int *col_offd_SN_to_AN,
HYPRE_Int num_functions,
HYPRE_Int nodal,
HYPRE_Int data,
HYPRE_Int **dof_func_ptr,
HYPRE_Int **CF_marker_ptr,
HYPRE_Int **col_offd_S_to_A_ptr,
hypre_ParCSRMatrix **S_ptr)
{
MPI_Comm comm = hypre_ParCSRMatrixComm(SN);
hypre_ParCSRMatrix *S;
hypre_CSRMatrix *S_diag;
HYPRE_Int *S_diag_i;
HYPRE_Int *S_diag_j;
double *S_diag_data;
hypre_CSRMatrix *S_offd;
HYPRE_Int *S_offd_i;
HYPRE_Int *S_offd_j;
double *S_offd_data;
HYPRE_Int *row_starts_S;
HYPRE_Int *col_starts_S;
HYPRE_Int *row_starts_SN = hypre_ParCSRMatrixRowStarts(SN);
HYPRE_Int *col_starts_SN = hypre_ParCSRMatrixColStarts(SN);
hypre_CSRMatrix *SN_diag = hypre_ParCSRMatrixDiag(SN);
HYPRE_Int *SN_diag_i = hypre_CSRMatrixI(SN_diag);
HYPRE_Int *SN_diag_j = hypre_CSRMatrixJ(SN_diag);
double *SN_diag_data;
hypre_CSRMatrix *SN_offd = hypre_ParCSRMatrixOffd(SN);
HYPRE_Int *SN_offd_i = hypre_CSRMatrixI(SN_offd);
HYPRE_Int *SN_offd_j = hypre_CSRMatrixJ(SN_offd);
double *SN_offd_data;
HYPRE_Int *CF_marker;
HYPRE_Int *col_map_offd_SN = hypre_ParCSRMatrixColMapOffd(SN);
HYPRE_Int *col_map_offd_S;
HYPRE_Int *dof_func;
HYPRE_Int num_nodes = hypre_CSRMatrixNumRows(SN_diag);
HYPRE_Int num_variables;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(SN);
HYPRE_Int num_sends;
HYPRE_Int num_recvs;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int *recv_procs;
HYPRE_Int *recv_vec_starts;
hypre_ParCSRCommPkg *comm_pkg_S;
HYPRE_Int *send_procs_S;
HYPRE_Int *send_map_starts_S;
HYPRE_Int *send_map_elmts_S;
HYPRE_Int *recv_procs_S;
HYPRE_Int *recv_vec_starts_S;
HYPRE_Int *col_offd_S_to_A = NULL;
HYPRE_Int num_coarse_nodes;
HYPRE_Int i,j,k,k1,jj,cnt;
HYPRE_Int row, start, end;
HYPRE_Int num_procs;
HYPRE_Int num_cols_offd_SN = hypre_CSRMatrixNumCols(SN_offd);
HYPRE_Int num_cols_offd_S;
HYPRE_Int SN_num_nonzeros_diag;
HYPRE_Int SN_num_nonzeros_offd;
HYPRE_Int S_num_nonzeros_diag;
HYPRE_Int S_num_nonzeros_offd;
HYPRE_Int global_num_vars;
HYPRE_Int global_num_cols;
HYPRE_Int global_num_nodes;
HYPRE_Int ierr = 0;
hypre_MPI_Comm_size(comm, &num_procs);
num_variables = num_functions*num_nodes;
CF_marker = hypre_CTAlloc(HYPRE_Int, num_variables);
if (nodal < 0)
{
cnt = 0;
num_coarse_nodes = 0;
for (i=0; i < num_nodes; i++)
{
if (CFN_marker[i] == 1) num_coarse_nodes++;
for (j=0; j < num_functions; j++)
CF_marker[cnt++] = CFN_marker[i];
}
dof_func = hypre_CTAlloc(HYPRE_Int,num_coarse_nodes*num_functions);
cnt = 0;
for (i=0; i < num_nodes; i++)
{
if (CFN_marker[i] == 1)
{
for (k=0; k < num_functions; k++)
dof_func[cnt++] = k;
}
}
*dof_func_ptr = dof_func;
}
else
{
cnt = 0;
for (i=0; i < num_nodes; i++)
for (j=0; j < num_functions; j++)
CF_marker[cnt++] = CFN_marker[i];
}
*CF_marker_ptr = CF_marker;
#ifdef HYPRE_NO_GLOBAL_PARTITION
row_starts_S = hypre_CTAlloc(HYPRE_Int,2);
for (i=0; i < 2; i++)
row_starts_S[i] = num_functions*row_starts_SN[i];
if (row_starts_SN != col_starts_SN)
{
col_starts_S = hypre_CTAlloc(HYPRE_Int,2);
for (i=0; i < 2; i++)
col_starts_S[i] = num_functions*col_starts_SN[i];
}
else
{
col_starts_S = row_starts_S;
}
#else
row_starts_S = hypre_CTAlloc(HYPRE_Int,num_procs+1);
for (i=0; i < num_procs+1; i++)
row_starts_S[i] = num_functions*row_starts_SN[i];
if (row_starts_SN != col_starts_SN)
{
col_starts_S = hypre_CTAlloc(HYPRE_Int,num_procs+1);
for (i=0; i < num_procs+1; i++)
col_starts_S[i] = num_functions*col_starts_SN[i];
}
else
{
col_starts_S = row_starts_S;
}
#endif
SN_num_nonzeros_diag = SN_diag_i[num_nodes];
SN_num_nonzeros_offd = SN_offd_i[num_nodes];
global_num_nodes = hypre_ParCSRMatrixGlobalNumRows(SN);
global_num_cols = hypre_ParCSRMatrixGlobalNumCols(SN)*num_functions;
global_num_vars = global_num_nodes*num_functions;
S_num_nonzeros_diag = num_functions*SN_num_nonzeros_diag;
S_num_nonzeros_offd = num_functions*SN_num_nonzeros_offd;
num_cols_offd_S = num_functions*num_cols_offd_SN;
S = hypre_ParCSRMatrixCreate(comm, global_num_vars, global_num_cols,
row_starts_S, col_starts_S, num_cols_offd_S,
S_num_nonzeros_diag, S_num_nonzeros_offd);
S_diag = hypre_ParCSRMatrixDiag(S);
S_offd = hypre_ParCSRMatrixOffd(S);
S_diag_i = hypre_CTAlloc(HYPRE_Int, num_variables+1);
S_offd_i = hypre_CTAlloc(HYPRE_Int, num_variables+1);
S_diag_j = hypre_CTAlloc(HYPRE_Int, S_num_nonzeros_diag);
hypre_CSRMatrixI(S_diag) = S_diag_i;
hypre_CSRMatrixJ(S_diag) = S_diag_j;
if (data)
{
SN_diag_data = hypre_CSRMatrixData(SN_diag);
S_diag_data = hypre_CTAlloc(double, S_num_nonzeros_diag);
hypre_CSRMatrixData(S_diag) = S_diag_data;
if (num_cols_offd_S)
{
SN_offd_data = hypre_CSRMatrixData(SN_offd);
S_offd_data = hypre_CTAlloc(double, S_num_nonzeros_offd);
hypre_CSRMatrixData(S_offd) = S_offd_data;
}
}
hypre_CSRMatrixI(S_offd) = S_offd_i;
if (comm_pkg)
{
comm_pkg_S = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg_S) = comm;
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
hypre_ParCSRCommPkgNumSends(comm_pkg_S) = num_sends;
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
hypre_ParCSRCommPkgNumRecvs(comm_pkg_S) = num_recvs;
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
send_procs_S = NULL;
send_map_elmts_S = NULL;
if (num_sends)
{
send_procs_S = hypre_CTAlloc(HYPRE_Int,num_sends);
send_map_elmts_S = hypre_CTAlloc(HYPRE_Int,
num_functions*send_map_starts[num_sends]);
}
send_map_starts_S = hypre_CTAlloc(HYPRE_Int,num_sends+1);
recv_vec_starts_S = hypre_CTAlloc(HYPRE_Int,num_recvs+1);
recv_procs_S = NULL;
if (num_recvs) recv_procs_S = hypre_CTAlloc(HYPRE_Int,num_recvs);
send_map_starts_S[0] = 0;
for (i=0; i < num_sends; i++)
{
send_procs_S[i] = send_procs[i];
send_map_starts_S[i+1] = num_functions*send_map_starts[i+1];
}
recv_vec_starts_S[0] = 0;
for (i=0; i < num_recvs; i++)
{
recv_procs_S[i] = recv_procs[i];
recv_vec_starts_S[i+1] = num_functions*recv_vec_starts[i+1];
}
cnt = 0;
for (i=0; i < send_map_starts[num_sends]; i++)
{
k1 = num_functions*send_map_elmts[i];
for (j=0; j < num_functions; j++)
{
send_map_elmts_S[cnt++] = k1+j;
}
}
hypre_ParCSRCommPkgSendProcs(comm_pkg_S) = send_procs_S;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg_S) = send_map_starts_S;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg_S) = send_map_elmts_S;
hypre_ParCSRCommPkgRecvProcs(comm_pkg_S) = recv_procs_S;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg_S) = recv_vec_starts_S;
hypre_ParCSRMatrixCommPkg(S) = comm_pkg_S;
}
if (num_cols_offd_S)
{
S_offd_j = hypre_CTAlloc(HYPRE_Int, S_num_nonzeros_offd);
hypre_CSRMatrixJ(S_offd) = S_offd_j;
col_map_offd_S = hypre_CTAlloc(HYPRE_Int, num_cols_offd_S);
cnt = 0;
for (i=0; i < num_cols_offd_SN; i++)
{
k1 = col_map_offd_SN[i]*num_functions;
for (j=0; j < num_functions; j++)
col_map_offd_S[cnt++] = k1+j;
}
hypre_ParCSRMatrixColMapOffd(S) = col_map_offd_S;
}
if (col_offd_SN_to_AN)
{
col_offd_S_to_A = hypre_CTAlloc(HYPRE_Int, num_cols_offd_S);
cnt = 0;
for (i=0; i < num_cols_offd_SN; i++)
{
k1 = col_offd_SN_to_AN[i]*num_functions;
for (j=0; j < num_functions; j++)
col_offd_S_to_A[cnt++] = k1+j;
}
*col_offd_S_to_A_ptr = col_offd_S_to_A;
}
cnt = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
row++;
start = cnt;
for (j=SN_diag_i[i]; j < SN_diag_i[i+1]; j++)
{
jj = SN_diag_j[j];
if (data) S_diag_data[cnt] = SN_diag_data[j];
S_diag_j[cnt++] = jj*num_functions;
}
end = cnt;
S_diag_i[row] = cnt;
for (k1=1; k1 < num_functions; k1++)
{
row++;
for (k=start; k < end; k++)
{
if (data) S_diag_data[cnt] = S_diag_data[k];
S_diag_j[cnt++] = S_diag_j[k]+k1;
}
S_diag_i[row] = cnt;
}
}
cnt = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
row++;
start = cnt;
for (j=SN_offd_i[i]; j < SN_offd_i[i+1]; j++)
{
jj = SN_offd_j[j];
if (data) S_offd_data[cnt] = SN_offd_data[j];
S_offd_j[cnt++] = jj*num_functions;
}
end = cnt;
S_offd_i[row] = cnt;
for (k1=1; k1 < num_functions; k1++)
{
row++;
for (k=start; k < end; k++)
{
if (data) S_offd_data[cnt] = S_offd_data[k];
S_offd_j[cnt++] = S_offd_j[k]+k1;
}
S_offd_i[row] = cnt;
}
}
*S_ptr = S;
return (ierr);
}
HYPRE_Int
hypre_ParCSRMatrixToParChordMatrix(
hypre_ParCSRMatrix *Ap,
MPI_Comm comm,
hypre_ParChordMatrix **pAc )
{
HYPRE_Int * row_starts = hypre_ParCSRMatrixRowStarts(Ap);
HYPRE_Int * col_starts = hypre_ParCSRMatrixColStarts(Ap);
hypre_CSRMatrix * diag = hypre_ParCSRMatrixDiag(Ap);
hypre_CSRMatrix * offd = hypre_ParCSRMatrixOffd(Ap);
HYPRE_Int * offd_j = hypre_CSRMatrixJ(offd);
HYPRE_Int * diag_j = hypre_CSRMatrixJ(diag);
HYPRE_Int * col_map_offd = hypre_ParCSRMatrixColMapOffd(Ap);
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(Ap);
hypre_ParChordMatrix * Ac;
hypre_NumbersNode * rdofs, * offd_cols_me;
hypre_NumbersNode ** offd_cols;
HYPRE_Int ** offd_col_array;
HYPRE_Int * len_offd_col_array, * offd_col_array_me;
HYPRE_Int len_offd_col_array_me;
HYPRE_Int num_idofs, num_rdofs, j_local, j_global, row_global;
HYPRE_Int i, j, jj, p, pto, q, qto, my_id, my_q, row, ireq;
HYPRE_Int num_inprocessors, num_toprocessors, num_procs, len_num_rdofs_toprocessor;
HYPRE_Int *inprocessor, *toprocessor, *pcr, *qcr, *num_inchords, *chord, *chordto;
HYPRE_Int *inproc, *toproc, *num_rdofs_toprocessor;
HYPRE_Int **inchord_idof, **inchord_rdof, **rdof_toprocessor;
double **inchord_data;
double data;
HYPRE_Int *first_index_idof, *first_index_rdof;
hypre_MPI_Request * request;
hypre_MPI_Status * status;
hypre_MPI_Comm_rank(comm, &my_id);
hypre_MPI_Comm_size(comm, &num_procs);
num_idofs = row_starts[my_id+1] - row_starts[my_id];
num_rdofs = col_starts[my_id+1] - col_starts[my_id];
hypre_ParChordMatrixCreate( pAc, comm, num_idofs, num_rdofs );
Ac = *pAc;
/* The following block sets Inprocessor:
On each proc. my_id, we find the columns in the offd and diag blocks
(global no.s). The columns are rdofs (contrary to what I wrote in
ChordMatrix.txt).
For each such col/rdof r, find the proc. p which owns row/idof r.
We set the temporary array pcr[p]=1 for such p.
An MPI all-to-all will exchange such arrays so my_id's array qcr has
qcr[q]=1 iff, on proc. q, pcr[my_id]=1. In other words, qcr[q]=1 if
my_id owns a row/idof i which is the same as a col/rdof owned by q.
Collect all such q's into in the array Inprocessor.
While constructing pcr, we also construct pj such that for any index jj
into offd_j,offd_data, pj[jj] is the processor which owns jj as a row (idof)
(the number jj is local to this processor).
*/
pcr = hypre_CTAlloc( HYPRE_Int, num_procs );
qcr = hypre_CTAlloc( HYPRE_Int, num_procs );
for ( p=0; p<num_procs; ++p ) pcr[p]=0;
for ( jj=0; jj<hypre_CSRMatrixNumNonzeros(offd); ++jj ) {
j_local = offd_j[jj];
j_global = col_map_offd[j_local];
for ( p=0; p<num_procs; ++p ) {
if ( j_global >= row_starts[p] && j_global<row_starts[p+1] ) {
pcr[p]=1;
/* not used yet... pj[jj] = p;*/
break;
}
}
}
/* jjd = jj; ...not used yet */
/* pcr[my_id] = 1; ...for square matrices (with nonzero diag block)
this one line would do the job of the following nested loop.
For non-square matrices, the data distribution is too arbitrary. */
for ( jj=0; jj<hypre_CSRMatrixNumNonzeros(diag); ++jj ) {
j_local = diag_j[jj];
j_global = j_local + first_col_diag;
for ( p=0; p<num_procs; ++p ) {
if ( j_global >= row_starts[p] && j_global<row_starts[p+1] ) {
pcr[p]=1;
/* not used yet... pj[jj+jjd] = p;*/
break;
}
}
}
/* Now pcr[p]=1 iff my_id owns a col/rdof r which proc. p owns as a row/idof */
hypre_MPI_Alltoall( pcr, 1, HYPRE_MPI_INT, qcr, 1, HYPRE_MPI_INT, comm );
/* Now qcr[q]==1 if my_id owns a row/idof i which is a col/rdof of proc. q
The array of such q's is the array Inprocessor. */
num_inprocessors = 0;
for ( q=0; q<num_procs; ++q ) if ( qcr[q]==1 ) ++num_inprocessors;
inprocessor = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
p = 0;
for ( q=0; q<num_procs; ++q ) if ( qcr[q]==1 ) inprocessor[ p++ ] = q;
num_toprocessors = 0;
for ( q=0; q<num_procs; ++q ) if ( pcr[q]==1 ) ++num_toprocessors;
toprocessor = hypre_CTAlloc( HYPRE_Int, num_toprocessors );
p = 0;
for ( q=0; q<num_procs; ++q ) if ( pcr[q]==1 ) toprocessor[ p++ ] = q;
hypre_ParChordMatrixNumInprocessors(Ac) = num_inprocessors;
hypre_ParChordMatrixInprocessor(Ac) = inprocessor;
hypre_ParChordMatrixNumToprocessors(Ac) = num_toprocessors;
hypre_ParChordMatrixToprocessor(Ac) = toprocessor;
hypre_TFree( qcr );
/* FirstIndexIdof[p] is the global index of proc. p's row 0 */
/* FirstIndexRdof[p] is the global index of proc. p's col 0 */
/* Fir FirstIndexIdof, we copy the array row_starts rather than its pointers,
because the chord matrix will think it's free to delete FirstIndexIdof */
/* col_starts[p] contains the global index of the first column
in the diag block of p. But for first_index_rdof we want the global
index of the first column in p (whether that's in the diag or offd block).
So it's more involved than row/idof: we also check the offd block, and
have to do a gather to get first_index_rdof for every proc. on every proc. */
first_index_idof = hypre_CTAlloc( HYPRE_Int, num_procs+1 );
first_index_rdof = hypre_CTAlloc( HYPRE_Int, num_procs+1 );
for ( p=0; p<=num_procs; ++p ) {
first_index_idof[p] = row_starts[p];
first_index_rdof[p] = col_starts[p];
};
if ( hypre_CSRMatrixNumRows(offd) > 0 && hypre_CSRMatrixNumCols(offd) > 0 )
first_index_rdof[my_id] =
col_starts[my_id]<col_map_offd[0] ? col_starts[my_id] : col_map_offd[0];
hypre_MPI_Allgather( &first_index_rdof[my_id], 1, HYPRE_MPI_INT,
first_index_rdof, 1, HYPRE_MPI_INT, comm );
/* Set num_inchords: num_inchords[p] is no. chords on my_id connected to p.
Set each chord (idof,jdof,data).
We go through each matrix element in the diag block, find what processor
owns its column no. as a row, then update num_inchords[p], inchord_idof[p],
inchord_rdof[p], inchord_data[p].
*/
inchord_idof = hypre_CTAlloc( HYPRE_Int*, num_inprocessors );
inchord_rdof = hypre_CTAlloc( HYPRE_Int*, num_inprocessors );
inchord_data = hypre_CTAlloc( double*, num_inprocessors );
num_inchords = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
chord = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
chordto = hypre_CTAlloc( HYPRE_Int, num_toprocessors );
num_rdofs = 0;
for ( q=0; q<num_inprocessors; ++q ) num_inchords[q] = 0;
my_q = -1;
for ( q=0; q<num_inprocessors; ++q ) if ( inprocessor[q]==my_id ) my_q = q;
hypre_assert( my_q>=0 );
/* diag block: first count chords (from my_id to my_id),
then set them from diag block's CSR data structure */
num_idofs = hypre_CSRMatrixNumRows(diag);
rdofs = hypre_NumbersNewNode();
for ( row=0; row<hypre_CSRMatrixNumRows(diag); ++row ) {
for ( i=hypre_CSRMatrixI(diag)[row]; i<hypre_CSRMatrixI(diag)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(diag)[i];
hypre_NumbersEnter( rdofs, j_local );
++num_inchords[my_q];
}
};
num_rdofs = hypre_NumbersNEntered( rdofs );
inchord_idof[my_q] = hypre_CTAlloc( HYPRE_Int, num_inchords[my_q] );
inchord_rdof[my_q] = hypre_CTAlloc( HYPRE_Int, num_inchords[my_q] );
inchord_data[my_q] = hypre_CTAlloc( double, num_inchords[my_q] );
chord[0] = 0;
for ( row=0; row<hypre_CSRMatrixNumRows(diag); ++row ) {
for ( i=hypre_CSRMatrixI(diag)[row]; i<hypre_CSRMatrixI(diag)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(diag)[i];
data = hypre_CSRMatrixData(diag)[i];
inchord_idof[my_q][chord[0]] = row;
/* Here We need to convert from j_local - a column local to
the diag of this proc., to a j which is local only to this
processor - a column (rdof) numbering scheme to be shared by the
diag and offd blocks... */
j_global = j_local + hypre_ParCSRMatrixColStarts(Ap)[my_q];
j = j_global - first_index_rdof[my_q];
inchord_rdof[my_q][chord[0]] = j;
inchord_data[my_q][chord[0]] = data;
hypre_assert( chord[0] < num_inchords[my_q] );
++chord[0];
}
};
hypre_NumbersDeleteNode(rdofs);
/* offd block: */
/* >>> offd_cols_me duplicates rdofs */
offd_cols_me = hypre_NumbersNewNode();
for ( row=0; row<hypre_CSRMatrixNumRows(offd); ++row ) {
for ( i=hypre_CSRMatrixI(offd)[row]; i<hypre_CSRMatrixI(offd)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(offd)[i];
j_global = col_map_offd[j_local];
hypre_NumbersEnter( offd_cols_me, j_global );
}
}
offd_col_array = hypre_CTAlloc( HYPRE_Int*, num_inprocessors );
len_offd_col_array = hypre_CTAlloc( HYPRE_Int, num_inprocessors );
offd_col_array_me = hypre_NumbersArray( offd_cols_me );
len_offd_col_array_me = hypre_NumbersNEntered( offd_cols_me );
request = hypre_CTAlloc(hypre_MPI_Request, 2*num_procs );
ireq = 0;
for ( q=0; q<num_inprocessors; ++q )
hypre_MPI_Irecv( &len_offd_col_array[q], 1, HYPRE_MPI_INT,
inprocessor[q], 0, comm, &request[ireq++] );
for ( p=0; p<num_procs; ++p ) if ( pcr[p]==1 ) {
hypre_MPI_Isend( &len_offd_col_array_me, 1, HYPRE_MPI_INT, p, 0, comm, &request[ireq++] );
}
status = hypre_CTAlloc(hypre_MPI_Status, ireq );
hypre_MPI_Waitall( ireq, request, status );
hypre_TFree(status);
ireq = 0;
for ( q=0; q<num_inprocessors; ++q )
offd_col_array[q] = hypre_CTAlloc( HYPRE_Int, len_offd_col_array[q] );
for ( q=0; q<num_inprocessors; ++q )
hypre_MPI_Irecv( offd_col_array[q], len_offd_col_array[q], HYPRE_MPI_INT,
inprocessor[q], 0, comm, &request[ireq++] );
for ( p=0; p<num_procs; ++p ) if ( pcr[p]==1 ) {
hypre_MPI_Isend( offd_col_array_me, len_offd_col_array_me,
HYPRE_MPI_INT, p, 0, comm, &request[ireq++] );
}
status = hypre_CTAlloc(hypre_MPI_Status, ireq );
hypre_MPI_Waitall( ireq, request, status );
hypre_TFree(request);
hypre_TFree(status);
offd_cols = hypre_CTAlloc( hypre_NumbersNode *, num_inprocessors );
for ( q=0; q<num_inprocessors; ++q ) {
offd_cols[q] = hypre_NumbersNewNode();
for ( i=0; i<len_offd_col_array[q]; ++i )
hypre_NumbersEnter( offd_cols[q], offd_col_array[q][i] );
}
len_num_rdofs_toprocessor = 1 + hypre_CSRMatrixI(offd)
[hypre_CSRMatrixNumRows(offd)];
inproc = hypre_CTAlloc( HYPRE_Int, len_num_rdofs_toprocessor );
toproc = hypre_CTAlloc( HYPRE_Int, len_num_rdofs_toprocessor );
num_rdofs_toprocessor = hypre_CTAlloc( HYPRE_Int, len_num_rdofs_toprocessor );
for ( qto=0; qto<len_num_rdofs_toprocessor; ++qto ) {
inproc[qto] = -1;
toproc[qto] = -1;
num_rdofs_toprocessor[qto] = 0;
};
rdofs = hypre_NumbersNewNode();
for ( row=0; row<hypre_CSRMatrixNumRows(offd); ++row ) {
for ( i=hypre_CSRMatrixI(offd)[row]; i<hypre_CSRMatrixI(offd)[row+1]; ++i ) {
j_local = hypre_CSRMatrixJ(offd)[i];
j_global = col_map_offd[j_local];
hypre_NumbersEnter( rdofs, j_local );
/* TO DO: find faster ways to do the two processor lookups below.*/
/* Find a processor p (local index q) from the inprocessor list,
which owns the column(rdof) whichis the same as this processor's
row(idof) row. Update num_inchords for p.
Save q as inproc[i] for quick recall later. It represents
an inprocessor (not unique) connected to a chord i.
*/
inproc[i] = -1;
for ( q=0; q<num_inprocessors; ++q ) if (q!=my_q) {
p = inprocessor[q];
if ( hypre_NumbersQuery( offd_cols[q],
row+hypre_ParCSRMatrixFirstRowIndex(Ap) )
== 1 ) {
/* row is one of the offd columns of p */
++num_inchords[q];
inproc[i] = q;
break;
}
}
if ( inproc[i]<0 ) {
/* For square matrices, we would have found the column in some
other processor's offd. But for non-square matrices it could
exist only in some other processor's diag...*/
/* Note that all data in a diag block is stored. We don't check
whether the value of a data entry is zero. */
for ( q=0; q<num_inprocessors; ++q ) if (q!=my_q) {
p = inprocessor[q];
row_global = row+hypre_ParCSRMatrixFirstRowIndex(Ap);
if ( row_global>=col_starts[p] &&
row_global< col_starts[p+1] ) {
/* row is one of the diag columns of p */
++num_inchords[q];
inproc[i] = q;
break;
}
}
}
hypre_assert( inproc[i]>=0 );
/* Find the processor pto (local index qto) from the toprocessor list,
which owns the row(idof) which is the same as this processor's
column(rdof) j_global. Update num_rdofs_toprocessor for pto.
Save pto as toproc[i] for quick recall later. It represents
the toprocessor connected to a chord i. */
for ( qto=0; qto<num_toprocessors; ++qto ) {
pto = toprocessor[qto];
if ( j_global >= row_starts[pto] && j_global<row_starts[pto+1] ) {
hypre_assert( qto < len_num_rdofs_toprocessor );
++num_rdofs_toprocessor[qto];
/* ... an overestimate, as if two chords share an rdof, that
rdof will be counted twice in num_rdofs_toprocessor.
It can be fixed up later.*/
toproc[i] = qto;
break;
}
}
}
};
num_rdofs += hypre_NumbersNEntered(rdofs);
hypre_NumbersDeleteNode(rdofs);
for ( q=0; q<num_inprocessors; ++q ) if (q!=my_q) {
inchord_idof[q] = hypre_CTAlloc( HYPRE_Int, num_inchords[q] );
inchord_rdof[q] = hypre_CTAlloc( HYPRE_Int, num_inchords[q] );
inchord_data[q] = hypre_CTAlloc( double, num_inchords[q] );
chord[q] = 0;
};