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_Int
main( HYPRE_Int argc,
char *argv[] )
{
HYPRE_Int num_procs, myid;
HYPRE_Int verbose = 0, build_matrix_type = 1;
HYPRE_Int index, matrix_arg_index, commpkg_flag=3;
HYPRE_Int i, k, ierr=0;
HYPRE_Int row_start, row_end;
HYPRE_Int col_start, col_end, global_num_rows;
HYPRE_Int *row_part, *col_part;
char *csrfilename;
HYPRE_Int preload = 0, loop = 0, loop2 = LOOP2;
HYPRE_Int bcast_rows[2], *info;
hypre_ParCSRMatrix *parcsr_A, *small_A;
HYPRE_ParCSRMatrix A_temp, A_temp_small;
hypre_CSRMatrix *A_CSR;
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int nx, ny, nz;
HYPRE_Int P, Q, R;
HYPRE_Int p, q, r;
HYPRE_Real values[4];
hypre_ParVector *x_new;
hypre_ParVector *y_new, *y;
HYPRE_Int *row_starts;
HYPRE_Real ans;
HYPRE_Real start_time, end_time, total_time, *loop_times;
HYPRE_Real T_avg, T_std;
HYPRE_Int noparmprint = 0;
#if mydebug
HYPRE_Int j, tmp_int;
#endif
/*-----------------------------------------------------------
* Initialize MPI
*-----------------------------------------------------------*/
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &num_procs );
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &myid );
/*-----------------------------------------------------------
* default - is 27pt laplace
*-----------------------------------------------------------*/
build_matrix_type = 2;
matrix_arg_index = argc;
/*-----------------------------------------------------------
* Parse command line
*-----------------------------------------------------------*/
index = 1;
while ( index < argc)
{
if ( strcmp(argv[index], "-verbose") == 0 )
{
index++;
verbose = 1;
}
else if ( strcmp(argv[index], "-fromonecsrfile") == 0 )
{
index++;
build_matrix_type = 1;
matrix_arg_index = index; /*this tells where the name is*/
}
else if ( strcmp(argv[index], "-commpkg") == 0 )
{
index++;
commpkg_flag = atoi(argv[index++]);
}
else if ( strcmp(argv[index], "-laplacian") == 0 )
{
index++;
build_matrix_type = 2;
matrix_arg_index = index;
}
else if ( strcmp(argv[index], "-27pt") == 0 )
{
index++;
build_matrix_type = 4;
matrix_arg_index = index;
}
/*
else if ( strcmp(argv[index], "-nopreload") == 0 )
{
index++;
preload = 0;
}
*/
else if ( strcmp(argv[index], "-loop") == 0 )
{
index++;
loop = atoi(argv[index++]);
}
else if ( strcmp(argv[index], "-noparmprint") == 0 )
{
index++;
noparmprint = 1;
}
else
{
index++;
/*hypre_printf("Warning: Unrecogized option '%s'\n",argv[index++] );*/
}
}
/*-----------------------------------------------------------
* Setup the Matrix problem
*-----------------------------------------------------------*/
/*-----------------------------------------------------------
* Get actual partitioning-
* read in an actual csr matrix.
*-----------------------------------------------------------*/
if (build_matrix_type ==1) /*read in a csr matrix from one file */
{
if (matrix_arg_index < argc)
{
csrfilename = argv[matrix_arg_index];
}
else
{
hypre_printf("Error: No filename specified \n");
exit(1);
}
if (myid == 0)
{
/*hypre_printf(" FromFile: %s\n", csrfilename);*/
A_CSR = hypre_CSRMatrixRead(csrfilename);
}
row_part = NULL;
col_part = NULL;
parcsr_A = hypre_CSRMatrixToParCSRMatrix(hypre_MPI_COMM_WORLD, A_CSR,
row_part, col_part);
if (myid == 0) hypre_CSRMatrixDestroy(A_CSR);
}
else if (build_matrix_type ==2)
{
myBuildParLaplacian(argc, argv, matrix_arg_index, &A_temp, !noparmprint);
parcsr_A = (hypre_ParCSRMatrix *) A_temp;
}
else if (build_matrix_type ==4)
{
myBuildParLaplacian27pt(argc, argv, matrix_arg_index, &A_temp, !noparmprint);
parcsr_A = (hypre_ParCSRMatrix *) A_temp;
}
/*-----------------------------------------------------------
* create a small problem so that timings are more accurate -
* code gets run twice (small laplace)
*-----------------------------------------------------------*/
/*this is no longer being used - preload = 0 is set at the beginning */
if (preload == 1)
{
/*hypre_printf("preload!\n");*/
values[1] = -1;
values[2] = -1;
values[3] = -1;
values[0] = - 6.0 ;
nx = 2;
ny = num_procs;
nz = 2;
P = 1;
Q = num_procs;
R = 1;
p = myid % P;
q = (( myid - p)/P) % Q;
r = ( myid - p - P*q)/( P*Q );
A_temp_small = (HYPRE_ParCSRMatrix) GenerateLaplacian(hypre_MPI_COMM_WORLD, nx, ny, nz,
P, Q, R, p, q, r, values);
small_A = (hypre_ParCSRMatrix *) A_temp_small;
/*do comm packages*/
hypre_NewCommPkgCreate(small_A);
hypre_NewCommPkgDestroy(small_A);
hypre_MatvecCommPkgCreate(small_A);
hypre_ParCSRMatrixDestroy(small_A);
}
/*-----------------------------------------------------------
* Prepare for timing
*-----------------------------------------------------------*/
/* instead of preloading, let's not time the first one if more than one*/
if (!loop)
{
loop = 1;
/* and don't do any timings */
}
else
{
loop +=1;
if (loop < 2) loop = 2;
}
loop_times = hypre_CTAlloc(HYPRE_Real, loop);
/******************************************************************************************/
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
if (commpkg_flag == 1 || commpkg_flag ==3 )
{
/*-----------------------------------------------------------
* Create new comm package
*-----------------------------------------------------------*/
if (!myid) hypre_printf("********************************************************\n" );
/*do loop times*/
for (i=0; i< loop; i++)
{
loop_times[i] = 0.0;
for (k=0; k< loop2; k++)
{
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
start_time = hypre_MPI_Wtime();
#if mpip_on
if (i==(loop-1)) hypre_MPI_Pcontrol(1);
#endif
hypre_NewCommPkgCreate(parcsr_A);
#if mpip_on
if (i==(loop-1)) hypre_MPI_Pcontrol(0);
#endif
end_time = hypre_MPI_Wtime();
end_time = end_time - start_time;
hypre_MPI_Allreduce(&end_time, &total_time, 1,
HYPRE_MPI_REAL, hypre_MPI_MAX, hypre_MPI_COMM_WORLD);
loop_times[i] += total_time;
if ( !((i+1)== loop && (k+1) == loop2)) hypre_NewCommPkgDestroy(parcsr_A);
}/*end of loop2 */
} /*end of loop*/
/* calculate the avg and std. */
if (loop > 1)
{
/* calculate the avg and std. */
stats_mo(loop_times, loop, &T_avg, &T_std);
if (!myid) hypre_printf(" NewCommPkgCreate: AVG. wall clock time = %f seconds\n", T_avg);
if (!myid) hypre_printf(" STD. for %d runs = %f\n", loop-1, T_std);
if (!myid) hypre_printf(" (Note: avg./std. timings exclude run 0.)\n");
if (!myid) hypre_printf("********************************************************\n" );
for (i=0; i< loop; i++)
{
if (!myid) hypre_printf(" run %d = %f sec.\n", i, loop_times[i]);
}
if (!myid) hypre_printf("********************************************************\n" );
}
else
{
if (!myid) hypre_printf("********************************************************\n" );
if (!myid) hypre_printf(" NewCommPkgCreate:\n");
if (!myid) hypre_printf(" run time = %f sec.\n", loop_times[0]);
if (!myid) hypre_printf("********************************************************\n" );
}
/*-----------------------------------------------------------
* Verbose printing
*-----------------------------------------------------------*/
/*some verification*/
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(parcsr_A);
if (verbose)
{
ierr = hypre_ParCSRMatrixGetLocalRange( parcsr_A,
&row_start, &row_end ,
&col_start, &col_end );
comm_pkg = hypre_ParCSRMatrixCommPkg(parcsr_A);
hypre_printf("myid = %i, my ACTUAL local range: [%i, %i]\n", myid,
row_start, row_end);
ierr = hypre_GetAssumedPartitionRowRange( myid, global_num_rows, &row_start,
&row_end);
hypre_printf("myid = %i, my assumed local range: [%i, %i]\n", myid,
row_start, row_end);
hypre_printf("myid = %d, num_recvs = %d\n", myid,
hypre_ParCSRCommPkgNumRecvs(comm_pkg) );
#if mydebug
for (i=0; i < hypre_ParCSRCommPkgNumRecvs(comm_pkg); i++)
{
hypre_printf("myid = %d, recv proc = %d, vec_starts = [%d : %d]\n",
myid, hypre_ParCSRCommPkgRecvProcs(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i+1]-1);
}
#endif
hypre_printf("myid = %d, num_sends = %d\n", myid,
hypre_ParCSRCommPkgNumSends(comm_pkg) );
#if mydebug
for (i=0; i <hypre_ParCSRCommPkgNumSends(comm_pkg) ; i++)
{
tmp_int = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i+1] -
hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
index = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
for (j=0; j< tmp_int; j++)
{
hypre_printf("myid = %d, send proc = %d, send element = %d\n",myid,
hypre_ParCSRCommPkgSendProcs(comm_pkg)[i],
hypre_ParCSRCommPkgSendMapElmts(comm_pkg)[index+j]);
}
}
#endif
}
/*-----------------------------------------------------------
* To verify correctness (if commpkg_flag = 3)
*-----------------------------------------------------------*/
if (commpkg_flag == 3 )
{
/*do a matvec - we are assuming a square matrix */
row_starts = hypre_ParCSRMatrixRowStarts(parcsr_A);
x_new = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD, global_num_rows, row_starts);
hypre_ParVectorSetPartitioningOwner(x_new, 0);
hypre_ParVectorInitialize(x_new);
hypre_ParVectorSetRandomValues(x_new, 1);
y_new = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD, global_num_rows, row_starts);
hypre_ParVectorSetPartitioningOwner(y_new, 0);
hypre_ParVectorInitialize(y_new);
hypre_ParVectorSetConstantValues(y_new, 0.0);
/*y = 1.0*A*x+1.0*y */
hypre_ParCSRMatrixMatvec (1.0, parcsr_A, x_new, 1.0, y_new);
}
/*-----------------------------------------------------------
* Clean up after MyComm
*-----------------------------------------------------------*/
hypre_NewCommPkgDestroy(parcsr_A);
}
/******************************************************************************************/
/******************************************************************************************/
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
if (commpkg_flag > 1 )
{
/*-----------------------------------------------------------
* Set up standard comm package
*-----------------------------------------------------------*/
bcast_rows[0] = 23;
bcast_rows[1] = 1789;
if (!myid) hypre_printf("********************************************************\n" );
/*do loop times*/
for (i=0; i< loop; i++)
{
loop_times[i] = 0.0;
for (k=0; k< loop2; k++)
{
hypre_MPI_Barrier(hypre_MPI_COMM_WORLD);
start_time = hypre_MPI_Wtime();
#if time_gather
info = hypre_CTAlloc(HYPRE_Int, num_procs);
hypre_MPI_Allgather(bcast_rows, 1, HYPRE_MPI_INT, info, 1, HYPRE_MPI_INT, hypre_MPI_COMM_WORLD);
#endif
hypre_MatvecCommPkgCreate(parcsr_A);
end_time = hypre_MPI_Wtime();
end_time = end_time - start_time;
hypre_MPI_Allreduce(&end_time, &total_time, 1,
HYPRE_MPI_REAL, hypre_MPI_MAX, hypre_MPI_COMM_WORLD);
loop_times[i] += total_time;
if ( !((i+1)== loop && (k+1) == loop2)) hypre_MatvecCommPkgDestroy(hypre_ParCSRMatrixCommPkg(parcsr_A));
}/* end of loop 2*/
} /*end of loop*/
/* calculate the avg and std. */
if (loop > 1)
{
stats_mo(loop_times, loop, &T_avg, &T_std);
if (!myid) hypre_printf("Current CommPkgCreate: AVG. wall clock time = %f seconds\n", T_avg);
if (!myid) hypre_printf(" STD. for %d runs = %f\n", loop-1, T_std);
if (!myid) hypre_printf(" (Note: avg./std. timings exclude run 0.)\n");
if (!myid) hypre_printf("********************************************************\n" );
for (i=0; i< loop; i++)
{
if (!myid) hypre_printf(" run %d = %f sec.\n", i, loop_times[i]);
}
if (!myid) hypre_printf("********************************************************\n" );
}
else
{
if (!myid) hypre_printf("********************************************************\n" );
if (!myid) hypre_printf(" Current CommPkgCreate:\n");
if (!myid) hypre_printf(" run time = %f sec.\n", loop_times[0]);
if (!myid) hypre_printf("********************************************************\n" );
}
/*-----------------------------------------------------------
* Verbose printing
*-----------------------------------------------------------*/
/*some verification*/
if (verbose)
{
ierr = hypre_ParCSRMatrixGetLocalRange( parcsr_A,
&row_start, &row_end ,
&col_start, &col_end );
comm_pkg = hypre_ParCSRMatrixCommPkg(parcsr_A);
hypre_printf("myid = %i, std - my local range: [%i, %i]\n", myid,
row_start, row_end);
ierr = hypre_ParCSRMatrixGetLocalRange( parcsr_A,
&row_start, &row_end ,
&col_start, &col_end );
hypre_printf("myid = %d, std - num_recvs = %d\n", myid,
hypre_ParCSRCommPkgNumRecvs(comm_pkg) );
#if mydebug
for (i=0; i < hypre_ParCSRCommPkgNumRecvs(comm_pkg); i++)
{
hypre_printf("myid = %d, std - recv proc = %d, vec_starts = [%d : %d]\n",
myid, hypre_ParCSRCommPkgRecvProcs(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i],
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg)[i+1]-1);
}
#endif
hypre_printf("myid = %d, std - num_sends = %d\n", myid,
hypre_ParCSRCommPkgNumSends(comm_pkg));
#if mydebug
for (i=0; i <hypre_ParCSRCommPkgNumSends(comm_pkg) ; i++)
{
tmp_int = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i+1] -
hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
index = hypre_ParCSRCommPkgSendMapStarts(comm_pkg)[i];
for (j=0; j< tmp_int; j++)
{
hypre_printf("myid = %d, std - send proc = %d, send element = %d\n",myid,
hypre_ParCSRCommPkgSendProcs(comm_pkg)[i],
hypre_ParCSRCommPkgSendMapElmts(comm_pkg)[index+j]);
}
}
#endif
}
/*-----------------------------------------------------------
* Verify correctness
*-----------------------------------------------------------*/
if (commpkg_flag == 3 )
{
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(parcsr_A);
row_starts = hypre_ParCSRMatrixRowStarts(parcsr_A);
y = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD, global_num_rows,row_starts);
hypre_ParVectorSetPartitioningOwner(y, 0);
hypre_ParVectorInitialize(y);
hypre_ParVectorSetConstantValues(y, 0.0);
hypre_ParCSRMatrixMatvec (1.0, parcsr_A, x_new, 1.0, y);
}
}
/*-----------------------------------------------------------
* Compare matvecs for both comm packages (3)
*-----------------------------------------------------------*/
if (commpkg_flag == 3 )
{
/*make sure that y and y_new are the same - now y_new should=0*/
hypre_ParVectorAxpy( -1.0, y, y_new );
hypre_ParVectorSetRandomValues(y, 1);
ans = hypre_ParVectorInnerProd( y, y_new );
if (!myid)
{
if ( fabs(ans) > 1e-8 )
{
hypre_printf("!!!!! WARNING !!!!! should be zero if correct = %6.10f\n",
ans);
}
else
{
hypre_printf("Matvecs match ( should be zero = %6.10f )\n",
ans);
}
}
}
/*-----------------------------------------------------------
* Clean up
*-----------------------------------------------------------*/
hypre_ParCSRMatrixDestroy(parcsr_A); /*this calls the standard comm
package destroy - but we'll destroy
ours separately until it is
incorporated */
if (commpkg_flag == 3 )
{
hypre_ParVectorDestroy(x_new);
hypre_ParVectorDestroy(y);
hypre_ParVectorDestroy(y_new);
}
hypre_MPI_Finalize();
return(ierr);
}
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_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);
}
hypre_ParCSRCommMultiHandle *
hypre_ParCSRCommMultiHandleCreate (HYPRE_Int job,
hypre_ParCSRCommPkg *comm_pkg,
void *send_data,
void *recv_data,
HYPRE_Int num_vecs )
{
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
HYPRE_Int num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
MPI_Comm comm = hypre_ParCSRCommPkgComm(comm_pkg);
hypre_ParCSRCommMultiHandle *comm_handle;
HYPRE_Int num_requests;
hypre_MPI_Request *requests;
HYPRE_Int i, j;
HYPRE_Int my_id, num_procs;
HYPRE_Int ip, vec_start, vec_len;
/*--------------------------------------------------------------------
* hypre_Initialize sets up a communication handle,
* posts receives and initiates sends. It always requires num_sends,
* num_recvs, recv_procs and send_procs to be set in comm_pkg.
* There are different options for job:
* job = 1 : is used to initialize communication exchange for the parts
* of vector needed to perform a Matvec, it requires send_data
* and recv_data to be doubles, recv_vec_starts and
* send_map_starts need to be set in comm_pkg.
* job = 2 : is used to initialize communication exchange for the parts
* of vector needed to perform a MatvecT, it requires send_data
* and recv_data to be doubles, recv_vec_starts and
* send_map_starts need to be set in comm_pkg.
*--------------------------------------------------------------------*/
num_requests = num_sends + num_recvs;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
j = 0;
switch (job)
{
case 1:
{
double *d_send_data = (double *) send_data;
double *d_recv_data = (double *) recv_data;
for (i = 0; i < num_recvs; i++)
{
ip = hypre_ParCSRCommPkgRecvProc(comm_pkg, i);
vec_start = hypre_ParCSRCommPkgRecvVecStart(comm_pkg,i);
vec_len = hypre_ParCSRCommPkgRecvVecStart(comm_pkg,i+1)-vec_start;
hypre_MPI_Irecv(&d_recv_data[vec_start*num_vecs], vec_len*num_vecs,
hypre_MPI_DOUBLE, ip, 0, comm, &requests[j++]);
}
for (i = 0; i < num_sends; i++)
{
vec_start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
vec_len = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1)-vec_start;
ip = hypre_ParCSRCommPkgSendProc(comm_pkg, i);
hypre_MPI_Isend(&d_send_data[vec_start*num_vecs], vec_len*num_vecs,
hypre_MPI_DOUBLE, ip, 0, comm, &requests[j++]);
}
break;
}
case 2:
{
double *d_send_data = (double *) send_data;
double *d_recv_data = (double *) recv_data;
for (i = 0; i < num_sends; i++)
{
vec_start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
vec_len = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1) - vec_start;
ip = hypre_ParCSRCommPkgSendProc(comm_pkg, i);
hypre_MPI_Irecv(&d_recv_data[vec_start*num_vecs], vec_len*num_vecs,
hypre_MPI_DOUBLE, ip, 0, comm, &requests[j++]);
}
for (i = 0; i < num_recvs; i++)
{
ip = hypre_ParCSRCommPkgRecvProc(comm_pkg, i);
vec_start = hypre_ParCSRCommPkgRecvVecStart(comm_pkg,i);
vec_len = hypre_ParCSRCommPkgRecvVecStart(comm_pkg,i+1)-vec_start;
hypre_MPI_Isend(&d_send_data[vec_start*num_vecs], vec_len*num_vecs,
hypre_MPI_DOUBLE, ip, 0, comm, &requests[j++]);
}
break;
}
}
/*--------------------------------------------------------------------
* set up comm_handle and return
*--------------------------------------------------------------------*/
comm_handle = hypre_CTAlloc(hypre_ParCSRCommMultiHandle, 1);
hypre_ParCSRCommMultiHandleCommPkg(comm_handle) = comm_pkg;
hypre_ParCSRCommMultiHandleSendData(comm_handle) = send_data;
hypre_ParCSRCommMultiHandleRecvData(comm_handle) = recv_data;
hypre_ParCSRCommMultiHandleNumRequests(comm_handle) = num_requests;
hypre_ParCSRCommMultiHandleRequests(comm_handle) = requests;
return (comm_handle);
}
hypre_CSRBlockMatrix *
hypre_ParCSRBlockMatrixExtractBExt(hypre_ParCSRBlockMatrix *B,
hypre_ParCSRBlockMatrix *A, HYPRE_Int data)
{
MPI_Comm comm = hypre_ParCSRBlockMatrixComm(B);
HYPRE_Int first_col_diag = hypre_ParCSRBlockMatrixFirstColDiag(B);
HYPRE_Int *col_map_offd = hypre_ParCSRBlockMatrixColMapOffd(B);
HYPRE_Int block_size = hypre_ParCSRBlockMatrixBlockSize(B);
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRBlockMatrixCommPkg(A);
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_ParCSRCommHandle *comm_handle;
hypre_ParCSRCommPkg *tmp_comm_pkg;
hypre_CSRBlockMatrix *diag = hypre_ParCSRBlockMatrixDiag(B);
HYPRE_Int *diag_i = hypre_CSRBlockMatrixI(diag);
HYPRE_Int *diag_j = hypre_CSRBlockMatrixJ(diag);
HYPRE_Complex *diag_data = hypre_CSRBlockMatrixData(diag);
hypre_CSRBlockMatrix *offd = hypre_ParCSRBlockMatrixOffd(B);
HYPRE_Int *offd_i = hypre_CSRBlockMatrixI(offd);
HYPRE_Int *offd_j = hypre_CSRBlockMatrixJ(offd);
HYPRE_Complex *offd_data = hypre_CSRBlockMatrixData(offd);
HYPRE_Int *B_int_i;
HYPRE_Int *B_int_j;
HYPRE_Complex *B_int_data;
HYPRE_Int num_cols_B, num_nonzeros;
HYPRE_Int num_rows_B_ext;
HYPRE_Int num_procs, my_id;
hypre_CSRBlockMatrix *B_ext;
HYPRE_Int *B_ext_i;
HYPRE_Int *B_ext_j;
HYPRE_Complex *B_ext_data;
HYPRE_Int *jdata_recv_vec_starts;
HYPRE_Int *jdata_send_map_starts;
HYPRE_Int i, j, k, l, counter, bnnz;
HYPRE_Int start_index;
HYPRE_Int j_cnt, jrow;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
bnnz = block_size * block_size;
num_cols_B = hypre_ParCSRMatrixGlobalNumCols(B);
num_rows_B_ext = recv_vec_starts[num_recvs];
B_int_i = hypre_CTAlloc(HYPRE_Int, send_map_starts[num_sends]+1);
B_ext_i = hypre_CTAlloc(HYPRE_Int, num_rows_B_ext+1);
/*--------------------------------------------------------------------------
* generate B_int_i through adding number of row-elements of offd and diag
* for corresponding rows. B_int_i[j+1] contains the number of elements of
* a row j (which is determined through send_map_elmts)
*--------------------------------------------------------------------------*/
B_int_i[0] = 0;
j_cnt = 0;
num_nonzeros = 0;
for (i=0; i < num_sends; i++)
{
for (j = send_map_starts[i]; j < send_map_starts[i+1]; j++)
{
jrow = send_map_elmts[j];
B_int_i[++j_cnt] = offd_i[jrow+1] - offd_i[jrow]
+ diag_i[jrow+1] - diag_i[jrow];
num_nonzeros += B_int_i[j_cnt];
}
}
/*--------------------------------------------------------------------------
* initialize communication
*--------------------------------------------------------------------------*/
comm_handle = hypre_ParCSRCommHandleCreate(11,comm_pkg,
&B_int_i[1],&B_ext_i[1]);
B_int_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros);
if (data) B_int_data = hypre_CTAlloc(HYPRE_Complex, num_nonzeros*bnnz);
jdata_send_map_starts = hypre_CTAlloc(HYPRE_Int, num_sends+1);
jdata_recv_vec_starts = hypre_CTAlloc(HYPRE_Int, num_recvs+1);
start_index = B_int_i[0];
jdata_send_map_starts[0] = start_index;
counter = 0;
for (i=0; i < num_sends; i++)
{
num_nonzeros = counter;
for (j = send_map_starts[i]; j < send_map_starts[i+1]; j++)
{
jrow = send_map_elmts[j];
for (k=diag_i[jrow]; k < diag_i[jrow+1]; k++)
{
B_int_j[counter] = diag_j[k]+first_col_diag;
if (data) {
for(l = 0; l < bnnz; l++)
B_int_data[counter*bnnz+ l] = diag_data[k*bnnz+ l];
}
counter++;
}
for (k=offd_i[jrow]; k < offd_i[jrow+1]; k++)
{
B_int_j[counter] = col_map_offd[offd_j[k]];
if (data) {
for(l = 0; l < bnnz; l++)
B_int_data[counter*bnnz+ l] =
offd_data[k*bnnz+ l];
}
counter++;
}
}
num_nonzeros = counter - num_nonzeros;
start_index += num_nonzeros;
jdata_send_map_starts[i+1] = start_index;
}
tmp_comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(tmp_comm_pkg) = comm;
hypre_ParCSRCommPkgNumSends(tmp_comm_pkg) = num_sends;
hypre_ParCSRCommPkgNumRecvs(tmp_comm_pkg) = num_recvs;
hypre_ParCSRCommPkgSendProcs(tmp_comm_pkg) = hypre_ParCSRCommPkgSendProcs(comm_pkg);
hypre_ParCSRCommPkgRecvProcs(tmp_comm_pkg) = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
hypre_ParCSRCommPkgSendMapStarts(tmp_comm_pkg) = jdata_send_map_starts;
hypre_ParCSRCommHandleDestroy(comm_handle);
comm_handle = NULL;
/*--------------------------------------------------------------------------
* after communication exchange B_ext_i[j+1] contains the number of elements
* of a row j !
* evaluate B_ext_i and compute num_nonzeros for B_ext
*--------------------------------------------------------------------------*/
for (i=0; i < num_recvs; i++)
for (j = recv_vec_starts[i]; j < recv_vec_starts[i+1]; j++)
B_ext_i[j+1] += B_ext_i[j];
num_nonzeros = B_ext_i[num_rows_B_ext];
B_ext = hypre_CSRBlockMatrixCreate(block_size, num_rows_B_ext, num_cols_B,
num_nonzeros);
B_ext_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros);
if (data) B_ext_data = hypre_CTAlloc(HYPRE_Complex, num_nonzeros*bnnz);
for (i=0; i < num_recvs; i++)
{
start_index = B_ext_i[recv_vec_starts[i]];
num_nonzeros = B_ext_i[recv_vec_starts[i+1]]-start_index;
jdata_recv_vec_starts[i+1] = B_ext_i[recv_vec_starts[i+1]];
}
hypre_ParCSRCommPkgRecvVecStarts(tmp_comm_pkg) = jdata_recv_vec_starts;
comm_handle = hypre_ParCSRCommHandleCreate(11,tmp_comm_pkg,B_int_j,B_ext_j);
hypre_ParCSRCommHandleDestroy(comm_handle);
comm_handle = NULL;
if (data)
{
comm_handle = hypre_ParCSRBlockCommHandleCreate(1, bnnz,tmp_comm_pkg,
B_int_data, B_ext_data);
hypre_ParCSRBlockCommHandleDestroy(comm_handle);
comm_handle = NULL;
}
hypre_CSRBlockMatrixI(B_ext) = B_ext_i;
hypre_CSRBlockMatrixJ(B_ext) = B_ext_j;
if (data) hypre_CSRBlockMatrixData(B_ext) = B_ext_data;
hypre_TFree(B_int_i);
hypre_TFree(B_int_j);
if (data) hypre_TFree(B_int_data);
hypre_TFree(jdata_send_map_starts);
hypre_TFree(jdata_recv_vec_starts);
hypre_TFree(tmp_comm_pkg);
return B_ext;
}
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_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
hypre_BlockMatvecCommPkgCreate(hypre_ParCSRBlockMatrix *A)
{
#ifdef HYPRE_NO_GLOBAL_PARTITION
HYPRE_Int col_start = 0, col_end = 0;
HYPRE_Int num_recvs, *recv_procs, *recv_vec_starts;
HYPRE_Int num_sends, *send_procs, *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int num_cols_off_d;
HYPRE_Int *col_map_off_d;
HYPRE_Int first_col_diag;
HYPRE_Int global_num_cols;
MPI_Comm comm;
hypre_ParCSRCommPkg *comm_pkg;
hypre_IJAssumedPart *apart;
/*-----------------------------------------------------------
* get parcsr_A information
*----------------------------------------------------------*/
col_start = hypre_ParCSRBlockMatrixFirstColDiag(A);
col_end = hypre_ParCSRBlockMatrixLastColDiag(A);
col_map_off_d = hypre_ParCSRBlockMatrixColMapOffd(A);
num_cols_off_d = hypre_CSRBlockMatrixNumCols(hypre_ParCSRBlockMatrixOffd(A));
global_num_cols = hypre_ParCSRBlockMatrixGlobalNumCols(A);
comm = hypre_ParCSRBlockMatrixComm(A);
first_col_diag = hypre_ParCSRBlockMatrixFirstColDiag(A);
/* Create the assumed partition */
if (hypre_ParCSRBlockMatrixAssumedPartition(A) == NULL)
{
hypre_ParCSRBlockMatrixCreateAssumedPartition(A);
}
apart = hypre_ParCSRBlockMatrixAssumedPartition(A);
/*-----------------------------------------------------------
* get commpkg info information
*----------------------------------------------------------*/
hypre_NewCommPkgCreate_core( comm, col_map_off_d, first_col_diag,
col_start, col_end,
num_cols_off_d, global_num_cols,
&num_recvs, &recv_procs, &recv_vec_starts,
&num_sends, &send_procs, &send_map_starts,
&send_map_elmts, apart);
if (!num_recvs)
{
hypre_TFree(recv_procs);
recv_procs = NULL;
}
if (!num_sends)
{
hypre_TFree(send_procs);
hypre_TFree(send_map_elmts);
send_procs = NULL;
send_map_elmts = NULL;
}
#else
hypre_ParCSRCommPkg *comm_pkg;
MPI_Comm comm = hypre_ParCSRBlockMatrixComm(A);
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_ParCSRBlockMatrixColMapOffd(A);
HYPRE_Int first_col_diag = hypre_ParCSRBlockMatrixFirstColDiag(A);
HYPRE_Int *col_starts = hypre_ParCSRBlockMatrixColStarts(A);
HYPRE_Int num_cols_diag = hypre_CSRBlockMatrixNumCols(hypre_ParCSRBlockMatrixDiag(A));
HYPRE_Int num_cols_offd = hypre_CSRBlockMatrixNumCols(hypre_ParCSRBlockMatrixOffd(A));
hypre_MatvecCommPkgCreate_core(comm, col_map_offd, first_col_diag,
col_starts, num_cols_diag, num_cols_offd, first_col_diag,
col_map_offd, 1, &num_recvs, &recv_procs, &recv_vec_starts,
&num_sends, &send_procs, &send_map_starts, &send_map_elmts);
#endif
/*-----------------------------------------------------------
* setup commpkg
*----------------------------------------------------------*/
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_ParCSRBlockMatrixCommPkg(A) = comm_pkg;
return hypre_error_flag;
}
/* Based on hypre_ParCSRCommHandleCreate in hypre's par_csr_communication.c. The
input variable job controls the communication type: 1=Matvec, 2=MatvecT. */
hypre_ParCSRCommHandle *
hypre_ParCSRCommHandleCreate_bool(HYPRE_Int job,
hypre_ParCSRCommPkg *comm_pkg,
HYPRE_Bool *send_data,
HYPRE_Bool *recv_data)
{
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
HYPRE_Int num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
MPI_Comm comm = hypre_ParCSRCommPkgComm(comm_pkg);
hypre_ParCSRCommHandle *comm_handle;
HYPRE_Int num_requests;
hypre_MPI_Request *requests;
HYPRE_Int i, j;
HYPRE_Int my_id, num_procs;
HYPRE_Int ip, vec_start, vec_len;
num_requests = num_sends + num_recvs;
requests = hypre_CTAlloc(hypre_MPI_Request, num_requests);
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm, &my_id);
j = 0;
switch (job)
{
case 1:
{
HYPRE_Bool *d_send_data = (HYPRE_Bool *) send_data;
HYPRE_Bool *d_recv_data = (HYPRE_Bool *) recv_data;
for (i = 0; i < num_recvs; i++)
{
ip = hypre_ParCSRCommPkgRecvProc(comm_pkg, i);
vec_start = hypre_ParCSRCommPkgRecvVecStart(comm_pkg, i);
vec_len = hypre_ParCSRCommPkgRecvVecStart(comm_pkg, i+1)-vec_start;
hypre_MPI_Irecv(&d_recv_data[vec_start], vec_len, HYPRE_MPI_BOOL,
ip, 0, comm, &requests[j++]);
}
for (i = 0; i < num_sends; i++)
{
vec_start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
vec_len = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1)-vec_start;
ip = hypre_ParCSRCommPkgSendProc(comm_pkg, i);
hypre_MPI_Isend(&d_send_data[vec_start], vec_len, HYPRE_MPI_BOOL,
ip, 0, comm, &requests[j++]);
}
break;
}
case 2:
{
HYPRE_Bool *d_send_data = (HYPRE_Bool *) send_data;
HYPRE_Bool *d_recv_data = (HYPRE_Bool *) recv_data;
for (i = 0; i < num_sends; i++)
{
vec_start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
vec_len = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1)-vec_start;
ip = hypre_ParCSRCommPkgSendProc(comm_pkg, i);
hypre_MPI_Irecv(&d_recv_data[vec_start], vec_len, HYPRE_MPI_BOOL,
ip, 0, comm, &requests[j++]);
}
for (i = 0; i < num_recvs; i++)
{
ip = hypre_ParCSRCommPkgRecvProc(comm_pkg, i);
vec_start = hypre_ParCSRCommPkgRecvVecStart(comm_pkg, i);
vec_len = hypre_ParCSRCommPkgRecvVecStart(comm_pkg, i+1)-vec_start;
hypre_MPI_Isend(&d_send_data[vec_start], vec_len, HYPRE_MPI_BOOL,
ip, 0, comm, &requests[j++]);
}
break;
}
}
/*--------------------------------------------------------------------
* set up comm_handle and return
*--------------------------------------------------------------------*/
comm_handle = hypre_CTAlloc(hypre_ParCSRCommHandle, 1);
hypre_ParCSRCommHandleCommPkg(comm_handle) = comm_pkg;
hypre_ParCSRCommHandleSendData(comm_handle) = send_data;
hypre_ParCSRCommHandleRecvData(comm_handle) = recv_data;
hypre_ParCSRCommHandleNumRequests(comm_handle) = num_requests;
hypre_ParCSRCommHandleRequests(comm_handle) = requests;
return comm_handle;
}
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_Int hypre_CreateLambda(void *amg_vdata)
{
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
MPI_Comm comm;
hypre_ParCSRMatrix **A_array;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
hypre_ParCSRMatrix *A_tmp;
hypre_ParCSRMatrix *Lambda;
hypre_CSRMatrix *L_diag;
hypre_CSRMatrix *L_offd;
hypre_CSRMatrix *A_tmp_diag;
hypre_CSRMatrix *A_tmp_offd;
hypre_ParVector *Xtilde;
hypre_ParVector *Rtilde;
hypre_Vector *Xtilde_local;
hypre_Vector *Rtilde_local;
hypre_ParCSRCommPkg *comm_pkg;
hypre_ParCSRCommPkg *L_comm_pkg = NULL;
hypre_ParCSRCommHandle *comm_handle;
HYPRE_Real *L_diag_data;
HYPRE_Real *L_offd_data;
HYPRE_Real *buf_data = NULL;
HYPRE_Real *tmp_data;
HYPRE_Real *x_data;
HYPRE_Real *r_data;
HYPRE_Real *l1_norms;
HYPRE_Real *A_tmp_diag_data;
HYPRE_Real *A_tmp_offd_data;
HYPRE_Real *D_data = NULL;
HYPRE_Real *D_data_offd = NULL;
HYPRE_Int *L_diag_i;
HYPRE_Int *L_diag_j;
HYPRE_Int *L_offd_i;
HYPRE_Int *L_offd_j;
HYPRE_Int *A_tmp_diag_i;
HYPRE_Int *A_tmp_offd_i;
HYPRE_Int *A_tmp_diag_j;
HYPRE_Int *A_tmp_offd_j;
HYPRE_Int *L_recv_ptr = NULL;
HYPRE_Int *L_send_ptr = NULL;
HYPRE_Int *L_recv_procs = NULL;
HYPRE_Int *L_send_procs = NULL;
HYPRE_Int *L_send_map_elmts = NULL;
HYPRE_Int *recv_procs;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_elmts;
HYPRE_Int *send_map_starts;
HYPRE_Int *recv_vec_starts;
HYPRE_Int *all_send_procs = NULL;
HYPRE_Int *all_recv_procs = NULL;
HYPRE_Int *remap = NULL;
HYPRE_Int *level_start;
HYPRE_Int addlvl;
HYPRE_Int additive;
HYPRE_Int mult_additive;
HYPRE_Int num_levels;
HYPRE_Int num_add_lvls;
HYPRE_Int num_procs;
HYPRE_Int num_sends, num_recvs;
HYPRE_Int num_sends_L = 0;
HYPRE_Int num_recvs_L = 0;
HYPRE_Int send_data_L = 0;
HYPRE_Int num_rows_L = 0;
HYPRE_Int num_rows_tmp = 0;
HYPRE_Int num_cols_offd_L = 0;
HYPRE_Int num_cols_offd = 0;
HYPRE_Int level, i, j, k;
HYPRE_Int this_proc, cnt, cnt_diag, cnt_offd;
HYPRE_Int cnt_recv, cnt_send, cnt_row, row_start;
HYPRE_Int start_diag, start_offd, indx, cnt_map;
HYPRE_Int start, j_indx, index, cnt_level;
HYPRE_Int max_sends, max_recvs;
/* Local variables */
HYPRE_Int Solve_err_flag = 0;
HYPRE_Int num_threads;
HYPRE_Int num_nonzeros_diag;
HYPRE_Int num_nonzeros_offd;
HYPRE_Real **l1_norms_ptr = NULL;
HYPRE_Real *relax_weight = NULL;
HYPRE_Real relax_type;
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
additive = hypre_ParAMGDataAdditive(amg_data);
mult_additive = hypre_ParAMGDataMultAdditive(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
relax_type = hypre_ParAMGDataGridRelaxType(amg_data)[1];
comm = hypre_ParCSRMatrixComm(A_array[0]);
hypre_MPI_Comm_size(comm,&num_procs);
l1_norms_ptr = hypre_ParAMGDataL1Norms(amg_data);
addlvl = hypre_max(additive, mult_additive);
num_add_lvls = num_levels+1-addlvl;
level_start = hypre_CTAlloc(HYPRE_Int, num_add_lvls+1);
send_data_L = 0;
num_rows_L = 0;
num_cols_offd_L = 0;
num_nonzeros_diag = 0;
num_nonzeros_offd = 0;
level_start[0] = 0;
cnt = 1;
max_sends = 0;
max_recvs = 0;
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
A_tmp_diag = hypre_ParCSRMatrixDiag(A_tmp);
A_tmp_offd = hypre_ParCSRMatrixOffd(A_tmp);
A_tmp_diag_i = hypre_CSRMatrixI(A_tmp_diag);
A_tmp_offd_i = hypre_CSRMatrixI(A_tmp_offd);
num_rows_tmp = hypre_CSRMatrixNumRows(A_tmp_diag);
num_cols_offd = hypre_CSRMatrixNumCols(A_tmp_offd);
num_rows_L += num_rows_tmp;
level_start[cnt] = level_start[cnt-1] + num_rows_tmp;
cnt++;
num_cols_offd_L += num_cols_offd;
num_nonzeros_diag += A_tmp_diag_i[num_rows_tmp];
num_nonzeros_offd += A_tmp_offd_i[num_rows_tmp];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
max_sends += num_sends;
if (num_sends)
send_data_L += hypre_ParCSRCommPkgSendMapStart(comm_pkg,num_sends);
max_recvs += hypre_ParCSRCommPkgNumRecvs(comm_pkg);
}
}
if (max_sends >= num_procs ||max_recvs >= num_procs)
{
max_sends = num_procs;
max_recvs = num_procs;
}
if (max_sends) all_send_procs = hypre_CTAlloc(HYPRE_Int, max_sends);
if (max_recvs) all_recv_procs = hypre_CTAlloc(HYPRE_Int, max_recvs);
cnt_send = 0;
cnt_recv = 0;
if (max_sends || max_recvs)
{
if (max_sends < num_procs && max_recvs < num_procs)
{
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
for (j = 0; j < num_sends; j++)
all_send_procs[cnt_send++] = send_procs[j];
for (j = 0; j < num_recvs; j++)
all_recv_procs[cnt_recv++] = recv_procs[j];
}
}
if (max_sends)
{
qsort0(all_send_procs, 0, max_sends-1);
num_sends_L = 1;
this_proc = all_send_procs[0];
for (i=1; i < max_sends; i++)
{
if (all_send_procs[i] > this_proc)
{
this_proc = all_send_procs[i];
all_send_procs[num_sends_L++] = this_proc;
}
}
L_send_procs = hypre_CTAlloc(HYPRE_Int, num_sends_L);
for (j=0; j < num_sends_L; j++)
L_send_procs[j] = all_send_procs[j];
hypre_TFree(all_send_procs);
}
if (max_recvs)
{
qsort0(all_recv_procs, 0, max_recvs-1);
num_recvs_L = 1;
this_proc = all_recv_procs[0];
for (i=1; i < max_recvs; i++)
{
if (all_recv_procs[i] > this_proc)
{
this_proc = all_recv_procs[i];
all_recv_procs[num_recvs_L++] = this_proc;
}
}
L_recv_procs = hypre_CTAlloc(HYPRE_Int, num_recvs_L);
for (j=0; j < num_recvs_L; j++)
L_recv_procs[j] = all_recv_procs[j];
hypre_TFree(all_recv_procs);
}
L_recv_ptr = hypre_CTAlloc(HYPRE_Int, num_recvs_L+1);
L_send_ptr = hypre_CTAlloc(HYPRE_Int, num_sends_L+1);
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
}
else
{
num_sends = 0;
num_recvs = 0;
}
for (k = 0; k < num_sends; k++)
{
this_proc = hypre_BinarySearch(L_send_procs,send_procs[k],num_sends_L);
L_send_ptr[this_proc+1] += send_map_starts[k+1]-send_map_starts[k];
}
for (k = 0; k < num_recvs; k++)
{
this_proc = hypre_BinarySearch(L_recv_procs,recv_procs[k],num_recvs_L);
L_recv_ptr[this_proc+1] += recv_vec_starts[k+1]-recv_vec_starts[k];
}
}
L_recv_ptr[0] = 0;
for (i=1; i < num_recvs_L; i++)
L_recv_ptr[i+1] += L_recv_ptr[i];
L_send_ptr[0] = 0;
for (i=1; i < num_sends_L; i++)
L_send_ptr[i+1] += L_send_ptr[i];
}
else
{
num_recvs_L = 0;
num_sends_L = 0;
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
for (j = 0; j < num_sends; j++)
{
this_proc = send_procs[j];
if (all_send_procs[this_proc] == 0)
num_sends_L++;
all_send_procs[this_proc] += send_map_starts[j+1]-send_map_starts[j];
}
for (j = 0; j < num_recvs; j++)
{
this_proc = recv_procs[j];
if (all_recv_procs[this_proc] == 0)
num_recvs_L++;
all_recv_procs[this_proc] += recv_vec_starts[j+1]-recv_vec_starts[j];
}
}
}
if (max_sends)
{
L_send_procs = hypre_CTAlloc(HYPRE_Int, num_sends_L);
L_send_ptr = hypre_CTAlloc(HYPRE_Int, num_sends_L+1);
num_sends_L = 0;
for (j=0; j < num_procs; j++)
{
this_proc = all_send_procs[j];
if (this_proc)
{
L_send_procs[num_sends_L++] = j;
L_send_ptr[num_sends_L] = this_proc + L_send_ptr[num_sends_L-1];
}
}
}
if (max_recvs)
{
L_recv_procs = hypre_CTAlloc(HYPRE_Int, num_recvs_L);
L_recv_ptr = hypre_CTAlloc(HYPRE_Int, num_recvs_L+1);
num_recvs_L = 0;
for (j=0; j < num_procs; j++)
{
this_proc = all_recv_procs[j];
if (this_proc)
{
L_recv_procs[num_recvs_L++] = j;
L_recv_ptr[num_recvs_L] = this_proc + L_recv_ptr[num_recvs_L-1];
}
}
}
}
}
if (max_sends) hypre_TFree(all_send_procs);
if (max_recvs) hypre_TFree(all_recv_procs);
L_diag = hypre_CSRMatrixCreate(num_rows_L, num_rows_L, num_nonzeros_diag);
L_offd = hypre_CSRMatrixCreate(num_rows_L, num_cols_offd_L, num_nonzeros_offd);
hypre_CSRMatrixInitialize(L_diag);
hypre_CSRMatrixInitialize(L_offd);
if (num_nonzeros_diag)
{
L_diag_data = hypre_CSRMatrixData(L_diag);
L_diag_j = hypre_CSRMatrixJ(L_diag);
}
L_diag_i = hypre_CSRMatrixI(L_diag);
if (num_nonzeros_offd)
{
L_offd_data = hypre_CSRMatrixData(L_offd);
L_offd_j = hypre_CSRMatrixJ(L_offd);
}
L_offd_i = hypre_CSRMatrixI(L_offd);
if (num_rows_L) D_data = hypre_CTAlloc(HYPRE_Real,num_rows_L);
if (send_data_L)
{
L_send_map_elmts = hypre_CTAlloc(HYPRE_Int, send_data_L);
buf_data = hypre_CTAlloc(HYPRE_Real,send_data_L);
}
if (num_cols_offd_L)
{
D_data_offd = hypre_CTAlloc(HYPRE_Real,num_cols_offd_L);
/*L_col_map_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd_L);*/
remap = hypre_CTAlloc(HYPRE_Int, num_cols_offd_L);
}
Rtilde = hypre_CTAlloc(hypre_ParVector, 1);
Rtilde_local = hypre_SeqVectorCreate(num_rows_L);
hypre_SeqVectorInitialize(Rtilde_local);
hypre_ParVectorLocalVector(Rtilde) = Rtilde_local;
hypre_ParVectorOwnsData(Rtilde) = 1;
Xtilde = hypre_CTAlloc(hypre_ParVector, 1);
Xtilde_local = hypre_SeqVectorCreate(num_rows_L);
hypre_SeqVectorInitialize(Xtilde_local);
hypre_ParVectorLocalVector(Xtilde) = Xtilde_local;
hypre_ParVectorOwnsData(Xtilde) = 1;
x_data = hypre_VectorData(hypre_ParVectorLocalVector(Xtilde));
r_data = hypre_VectorData(hypre_ParVectorLocalVector(Rtilde));
cnt = 0;
cnt_level = 0;
cnt_diag = 0;
cnt_offd = 0;
cnt_row = 1;
L_diag_i[0] = 0;
L_offd_i[0] = 0;
for (level=addlvl; level < num_levels; level++)
{
row_start = level_start[cnt_level];
if (level != 0)
{
tmp_data = hypre_VectorData(hypre_ParVectorLocalVector(F_array[level]));
if (tmp_data) hypre_TFree(tmp_data);
hypre_VectorData(hypre_ParVectorLocalVector(F_array[level])) = &r_data[row_start];
hypre_VectorOwnsData(hypre_ParVectorLocalVector(F_array[level])) = 0;
tmp_data = hypre_VectorData(hypre_ParVectorLocalVector(U_array[level]));
if (tmp_data) hypre_TFree(tmp_data);
hypre_VectorData(hypre_ParVectorLocalVector(U_array[level])) = &x_data[row_start];
hypre_VectorOwnsData(hypre_ParVectorLocalVector(U_array[level])) = 0;
}
cnt_level++;
start_diag = L_diag_i[cnt_row-1];
start_offd = L_offd_i[cnt_row-1];
A_tmp = A_array[level];
A_tmp_diag = hypre_ParCSRMatrixDiag(A_tmp);
A_tmp_offd = hypre_ParCSRMatrixOffd(A_tmp);
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
A_tmp_diag_i = hypre_CSRMatrixI(A_tmp_diag);
A_tmp_offd_i = hypre_CSRMatrixI(A_tmp_offd);
A_tmp_diag_j = hypre_CSRMatrixJ(A_tmp_diag);
A_tmp_offd_j = hypre_CSRMatrixJ(A_tmp_offd);
A_tmp_diag_data = hypre_CSRMatrixData(A_tmp_diag);
A_tmp_offd_data = hypre_CSRMatrixData(A_tmp_offd);
num_rows_tmp = hypre_CSRMatrixNumRows(A_tmp_diag);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
}
else
{
num_sends = 0;
num_recvs = 0;
}
/* Compute new combined communication package */
for (i=0; i < num_sends; i++)
{
this_proc = hypre_BinarySearch(L_send_procs,send_procs[i],num_sends_L);
indx = L_send_ptr[this_proc];
for (j=send_map_starts[i]; j < send_map_starts[i+1]; j++)
{
L_send_map_elmts[indx++] = row_start + send_map_elmts[j];
}
L_send_ptr[this_proc] = indx;
}
cnt_map = 0;
for (i = 0; i < num_recvs; i++)
{
this_proc = hypre_BinarySearch(L_recv_procs,recv_procs[i],num_recvs_L);
indx = L_recv_ptr[this_proc];
for (j=recv_vec_starts[i]; j < recv_vec_starts[i+1]; j++)
{
remap[cnt_map++] = indx++;
}
L_recv_ptr[this_proc] = indx;
}
/* Compute Lambda */
if (relax_type == 0)
{
HYPRE_Real rlx_wt = relax_weight[level];
#ifdef HYPRE_USING_OPENMP
#pragma omp for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < num_rows_tmp; i++)
{
D_data[i] = rlx_wt/A_tmp_diag_data[A_tmp_diag_i[i]];
L_diag_i[cnt_row+i] = start_diag + A_tmp_diag_i[i+1];
L_offd_i[cnt_row+i] = start_offd + A_tmp_offd_i[i+1];
}
}
else
{
l1_norms = l1_norms_ptr[level];
#ifdef HYPRE_USING_OPENMP
#pragma omp for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < num_rows_tmp; i++)
{
D_data[i] = 1.0/l1_norms[i];
L_diag_i[cnt_row+i] = start_diag + A_tmp_diag_i[i+1];
L_offd_i[cnt_row+i] = start_offd + A_tmp_offd_i[i+1];
}
}
if (num_procs > 1)
{
index = 0;
for (i=0; i < num_sends; i++)
{
start = send_map_starts[i];
for (j=start; j < send_map_starts[i+1]; j++)
buf_data[index++] = D_data[send_map_elmts[j]];
}
comm_handle = hypre_ParCSRCommHandleCreate(1, comm_pkg,
buf_data, D_data_offd);
hypre_ParCSRCommHandleDestroy(comm_handle);
}
for (i = 0; i < num_rows_tmp; i++)
{
j_indx = A_tmp_diag_i[i];
L_diag_data[cnt_diag] = (2.0 - A_tmp_diag_data[j_indx]*D_data[i])*D_data[i];
L_diag_j[cnt_diag++] = i+row_start;
for (j=A_tmp_diag_i[i]+1; j < A_tmp_diag_i[i+1]; j++)
{
j_indx = A_tmp_diag_j[j];
L_diag_data[cnt_diag] = (- A_tmp_diag_data[j]*D_data[j_indx])*D_data[i];
L_diag_j[cnt_diag++] = j_indx+row_start;
}
for (j=A_tmp_offd_i[i]; j < A_tmp_offd_i[i+1]; j++)
{
j_indx = A_tmp_offd_j[j];
L_offd_data[cnt_offd] = (- A_tmp_offd_data[j]*D_data_offd[j_indx])*D_data[i];
L_offd_j[cnt_offd++] = remap[j_indx];
}
}
cnt_row += num_rows_tmp;
}
if (L_send_ptr)
{
for (i=num_sends_L-1; i > 0; i--)
L_send_ptr[i] = L_send_ptr[i-1];
L_send_ptr[0] = 0;
}
else
L_send_ptr = hypre_CTAlloc(HYPRE_Int,1);
if (L_recv_ptr)
{
for (i=num_recvs_L-1; i > 0; i--)
L_recv_ptr[i] = L_recv_ptr[i-1];
L_recv_ptr[0] = 0;
}
else
L_recv_ptr = hypre_CTAlloc(HYPRE_Int,1);
L_comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgNumRecvs(L_comm_pkg) = num_recvs_L;
hypre_ParCSRCommPkgNumSends(L_comm_pkg) = num_sends_L;
hypre_ParCSRCommPkgRecvProcs(L_comm_pkg) = L_recv_procs;
hypre_ParCSRCommPkgSendProcs(L_comm_pkg) = L_send_procs;
hypre_ParCSRCommPkgRecvVecStarts(L_comm_pkg) = L_recv_ptr;
hypre_ParCSRCommPkgSendMapStarts(L_comm_pkg) = L_send_ptr;
hypre_ParCSRCommPkgSendMapElmts(L_comm_pkg) = L_send_map_elmts;
hypre_ParCSRCommPkgComm(L_comm_pkg) = comm;
Lambda = hypre_CTAlloc(hypre_ParCSRMatrix, 1);
hypre_ParCSRMatrixDiag(Lambda) = L_diag;
hypre_ParCSRMatrixOffd(Lambda) = L_offd;
hypre_ParCSRMatrixCommPkg(Lambda) = L_comm_pkg;
hypre_ParCSRMatrixComm(Lambda) = comm;
hypre_ParCSRMatrixOwnsData(Lambda) = 1;
hypre_ParAMGDataLambda(amg_data) = Lambda;
hypre_ParAMGDataRtilde(amg_data) = Rtilde;
hypre_ParAMGDataXtilde(amg_data) = Xtilde;
hypre_TFree(D_data_offd);
hypre_TFree(D_data);
if (num_procs > 1) hypre_TFree(buf_data);
hypre_TFree(remap);
hypre_TFree(buf_data);
hypre_TFree(level_start);
return Solve_err_flag;
}
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_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;
}
