int
hypre_BoomerAMGIndepSetInit( hypre_ParCSRMatrix *S,
double *measure_array ,
int seq_rand)
{
hypre_CSRMatrix *S_diag = hypre_ParCSRMatrixDiag(S);
MPI_Comm comm = hypre_ParCSRMatrixComm(S);
int S_num_nodes = hypre_CSRMatrixNumRows(S_diag);
HYPRE_BigInt i;
int j, my_id;
int ierr = 0;
MPI_Comm_rank(comm,&my_id);
j = 2747+my_id;
if (seq_rand) j = 2747;
hypre_SeedRand(j);
if (seq_rand)
{
for (i = 0; i < hypre_ParCSRMatrixFirstRowIndex(S); i++)
hypre_Rand();
}
for (j = 0; j < S_num_nodes; j++)
{
measure_array[j] += hypre_Rand();
}
return (ierr);
}
HYPRE_Int
hypre_ParKrylovCommInfo( void *A, HYPRE_Int *my_id, HYPRE_Int *num_procs)
{
MPI_Comm comm = hypre_ParCSRMatrixComm ( (hypre_ParCSRMatrix *) A);
hypre_MPI_Comm_size(comm,num_procs);
hypre_MPI_Comm_rank(comm,my_id);
return 0;
}
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 AmgCGCBoundaryFix (hypre_ParCSRMatrix *S,HYPRE_Int *CF_marker,HYPRE_Int *CF_marker_offd)
/* Checks whether an interpolation is possible for a fine grid point with strong couplings.
* Required after CGC coarsening
* ========================================================================================
* S : the strength matrix
* CF_marker, CF_marker_offd : the coarse/fine markers
* ========================================================================================*/
{
HYPRE_Int mpirank,i,j,has_c_pt,ierr=0;
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 added_cpts=0;
MPI_Comm comm = hypre_ParCSRMatrixComm(S);
hypre_MPI_Comm_rank (comm,&mpirank);
if (num_cols_offd) {
S_offd_j = hypre_CSRMatrixJ(S_offd);
}
for (i=0;i<num_variables;i++) {
if (S_offd_i[i]==S_offd_i[i+1] || CF_marker[i] == C_PT) continue;
has_c_pt=0;
/* fine grid point with strong connections across the boundary */
for (j=S_i[i];j<S_i[i+1];j++)
if (CF_marker[S_j[j]] == C_PT) {has_c_pt=1; break;}
if (has_c_pt) continue;
for (j=S_offd_i[i];j<S_offd_i[i+1];j++)
if (CF_marker_offd[S_offd_j[j]] == C_PT) {has_c_pt=1; break;}
if (has_c_pt) continue;
/* all points i is strongly coupled to are fine: make i C_PT */
CF_marker[i] = C_PT;
#if 0
hypre_printf ("Processor %d: added point %d in AmgCGCBoundaryFix\n",mpirank,i);
#endif
added_cpts++;
}
#if 0
if (added_cpts) hypre_printf ("Processor %d: added %d points in AmgCGCBoundaryFix\n",mpirank,added_cpts);
fflush(stdout);
#endif
return(ierr);
}
int HYPRE_ParCSRCotreeSetup(HYPRE_Solver solver, HYPRE_ParCSRMatrix A,
HYPRE_ParVector b, HYPRE_ParVector x)
{
int *partition, *new_partition, nprocs, *tindices, ii;
void *vsolver = (void *) solver;
/*
void *vA = (void *) A;
void *vb = (void *) b;
void *vx = (void *) x;
*/
hypre_CotreeData *cotree_data = (hypre_CotreeData *) vsolver;
hypre_ParCSRMatrix **submatrices;
hypre_ParVector *new_vector;
MPI_Comm comm;
cotree_data->Aee = (hypre_ParCSRMatrix *) A;
hypre_ParCSRMatrixGenSpanningTree(cotree_data->Gen, &tindices, 1);
submatrices = (hypre_ParCSRMatrix **) malloc(sizeof(hypre_ParCSRMatrix *));
hypre_ParCSRMatrixExtractSubmatrices(cotree_data->Aee, tindices,
&submatrices);
cotree_data->Att = submatrices[0];
cotree_data->Atc = submatrices[1];
cotree_data->Act = submatrices[2];
cotree_data->Acc = submatrices[3];
hypre_ParCSRMatrixExtractRowSubmatrices(cotree_data->Gen, tindices,
&submatrices);
cotree_data->Gt = submatrices[0];
cotree_data->Gc = submatrices[1];
free(submatrices);
comm = hypre_ParCSRMatrixComm((hypre_ParCSRMatrix *) A);
MPI_Comm_size(comm, &nprocs);
partition = hypre_ParVectorPartitioning((hypre_ParVector *) b);
new_partition = (int *) malloc((nprocs+1) * sizeof(int));
for (ii = 0; ii <= nprocs; ii++) new_partition[ii] = partition[ii];
/* partition = hypre_ParVectorPartitioning((hypre_ParVector *) b); */
new_vector = hypre_ParVectorCreate(hypre_ParVectorComm((hypre_ParVector *)b),
(int) hypre_ParVectorGlobalSize((hypre_ParVector *) b),
new_partition);
hypre_ParVectorInitialize(new_vector);
cotree_data->w = new_vector;
return 0;
}
HYPRE_Int
hypre_BoomerAMGSolve( void *amg_vdata,
hypre_ParCSRMatrix *A,
hypre_ParVector *f,
hypre_ParVector *u )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
HYPRE_Int amg_print_level;
HYPRE_Int amg_logging;
HYPRE_Int cycle_count;
HYPRE_Int num_levels;
/* HYPRE_Int num_unknowns; */
HYPRE_Real tol;
HYPRE_Int block_mode;
hypre_ParCSRMatrix **A_array;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
hypre_ParCSRBlockMatrix **A_block_array;
/* Local variables */
HYPRE_Int j;
HYPRE_Int Solve_err_flag;
HYPRE_Int min_iter;
HYPRE_Int max_iter;
HYPRE_Int num_procs, my_id;
HYPRE_Int additive;
HYPRE_Int mult_additive;
HYPRE_Int simple;
HYPRE_Real alpha = 1.0;
HYPRE_Real beta = -1.0;
HYPRE_Real cycle_op_count;
HYPRE_Real total_coeffs;
HYPRE_Real total_variables;
HYPRE_Real *num_coeffs;
HYPRE_Real *num_variables;
HYPRE_Real cycle_cmplxty = 0.0;
HYPRE_Real operat_cmplxty;
HYPRE_Real grid_cmplxty;
HYPRE_Real conv_factor = 0.0;
HYPRE_Real resid_nrm = 1.0;
HYPRE_Real resid_nrm_init = 0.0;
HYPRE_Real relative_resid;
HYPRE_Real rhs_norm = 0.0;
HYPRE_Real old_resid;
HYPRE_Real ieee_check = 0.;
hypre_ParVector *Vtemp;
hypre_ParVector *Residual;
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
amg_print_level = hypre_ParAMGDataPrintLevel(amg_data);
amg_logging = hypre_ParAMGDataLogging(amg_data);
if ( amg_logging > 1 )
Residual = hypre_ParAMGDataResidual(amg_data);
/* num_unknowns = hypre_ParAMGDataNumUnknowns(amg_data); */
num_levels = hypre_ParAMGDataNumLevels(amg_data);
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
tol = hypre_ParAMGDataTol(amg_data);
min_iter = hypre_ParAMGDataMinIter(amg_data);
max_iter = hypre_ParAMGDataMaxIter(amg_data);
additive = hypre_ParAMGDataAdditive(amg_data);
simple = hypre_ParAMGDataSimple(amg_data);
mult_additive = hypre_ParAMGDataMultAdditive(amg_data);
A_array[0] = A;
F_array[0] = f;
U_array[0] = u;
block_mode = hypre_ParAMGDataBlockMode(amg_data);
A_block_array = hypre_ParAMGDataABlockArray(amg_data);
/* Vtemp = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(A_array[0]),
hypre_ParCSRMatrixGlobalNumRows(A_array[0]),
hypre_ParCSRMatrixRowStarts(A_array[0]));
hypre_ParVectorInitialize(Vtemp);
hypre_ParVectorSetPartitioningOwner(Vtemp,0);
hypre_ParAMGDataVtemp(amg_data) = Vtemp;
*/
Vtemp = hypre_ParAMGDataVtemp(amg_data);
/*-----------------------------------------------------------------------
* Write the solver parameters
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1)
hypre_BoomerAMGWriteSolverParams(amg_data);
/*-----------------------------------------------------------------------
* Initialize the solver error flag and assorted bookkeeping variables
*-----------------------------------------------------------------------*/
Solve_err_flag = 0;
total_coeffs = 0;
total_variables = 0;
cycle_count = 0;
operat_cmplxty = 0;
grid_cmplxty = 0;
/*-----------------------------------------------------------------------
* write some initial info
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1 && tol > 0.)
hypre_printf("\n\nAMG SOLUTION INFO:\n");
/*-----------------------------------------------------------------------
* Compute initial fine-grid residual and print
*-----------------------------------------------------------------------*/
if (amg_print_level > 1 || amg_logging > 1)
{
if ( amg_logging > 1 ) {
hypre_ParVectorCopy(F_array[0], Residual );
if (tol > 0)
hypre_ParCSRMatrixMatvec(alpha, A_array[0], U_array[0], beta, Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParVectorCopy(F_array[0], Vtemp);
if (tol > 0)
hypre_ParCSRMatrixMatvec(alpha, A_array[0], U_array[0], beta, Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
/* Since it is does not diminish performance, attempt to return an error flag
and notify users when they supply bad input. */
if (resid_nrm != 0.) ieee_check = resid_nrm/resid_nrm; /* INF -> NaN conversion */
if (ieee_check != ieee_check)
{
/* ...INFs or NaNs in input can make ieee_check a NaN. This test
for ieee_check self-equality works on all IEEE-compliant compilers/
machines, c.f. page 8 of "Lecture Notes on the Status of IEEE 754"
by W. Kahan, May 31, 1996. Currently (July 2002) this paper may be
found at http://HTTP.CS.Berkeley.EDU/~wkahan/ieee754status/IEEE754.PDF */
if (amg_print_level > 0)
{
hypre_printf("\n\nERROR detected by Hypre ... BEGIN\n");
hypre_printf("ERROR -- hypre_BoomerAMGSolve: INFs and/or NaNs detected in input.\n");
hypre_printf("User probably placed non-numerics in supplied A, x_0, or b.\n");
hypre_printf("ERROR detected by Hypre ... END\n\n\n");
}
hypre_error(HYPRE_ERROR_GENERIC);
return hypre_error_flag;
}
resid_nrm_init = resid_nrm;
rhs_norm = sqrt(hypre_ParVectorInnerProd(f, f));
if (rhs_norm)
{
relative_resid = resid_nrm_init / rhs_norm;
}
else
{
relative_resid = resid_nrm_init;
}
}
else
{
relative_resid = 1.;
}
if (my_id == 0 && amg_print_level > 1)
{
hypre_printf(" relative\n");
hypre_printf(" residual factor residual\n");
hypre_printf(" -------- ------ --------\n");
hypre_printf(" Initial %e %e\n",resid_nrm_init,
relative_resid);
}
/*-----------------------------------------------------------------------
* Main V-cycle loop
*-----------------------------------------------------------------------*/
while ((relative_resid >= tol || cycle_count < min_iter)
&& cycle_count < max_iter)
{
hypre_ParAMGDataCycleOpCount(amg_data) = 0;
/* Op count only needed for one cycle */
if ((additive < 0 || additive >= num_levels)
&& (mult_additive < 0 || mult_additive >= num_levels)
&& (simple < 0 || simple >= num_levels) )
hypre_BoomerAMGCycle(amg_data, F_array, U_array);
else
hypre_BoomerAMGAdditiveCycle(amg_data);
/*---------------------------------------------------------------
* Compute fine-grid residual and residual norm
*----------------------------------------------------------------*/
if (amg_print_level > 1 || amg_logging > 1 || tol > 0.)
{
old_resid = resid_nrm;
if ( amg_logging > 1 ) {
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[0], U_array[0], beta, F_array[0], Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[0], U_array[0], beta, F_array[0], Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
if (old_resid) conv_factor = resid_nrm / old_resid;
else conv_factor = resid_nrm;
if (rhs_norm)
{
relative_resid = resid_nrm / rhs_norm;
}
else
{
relative_resid = resid_nrm;
}
hypre_ParAMGDataRelativeResidualNorm(amg_data) = relative_resid;
}
++cycle_count;
hypre_ParAMGDataNumIterations(amg_data) = cycle_count;
#ifdef CUMNUMIT
++hypre_ParAMGDataCumNumIterations(amg_data);
#endif
if (my_id == 0 && amg_print_level > 1)
{
hypre_printf(" Cycle %2d %e %f %e \n", cycle_count,
resid_nrm, conv_factor, relative_resid);
}
}
if (cycle_count == max_iter && tol > 0.)
{
Solve_err_flag = 1;
hypre_error(HYPRE_ERROR_CONV);
}
/*-----------------------------------------------------------------------
* Compute closing statistics
*-----------------------------------------------------------------------*/
if (cycle_count > 0 && resid_nrm_init)
conv_factor = pow((resid_nrm/resid_nrm_init),(1.0/(HYPRE_Real) cycle_count));
else
conv_factor = 1.;
if (amg_print_level > 1)
{
num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels);
num_variables = hypre_CTAlloc(HYPRE_Real, num_levels);
num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A);
num_variables[0] = hypre_ParCSRMatrixGlobalNumRows(A);
if (block_mode)
{
for (j = 1; j < num_levels; j++)
{
num_coeffs[j] = (HYPRE_Real) hypre_ParCSRBlockMatrixNumNonzeros(A_block_array[j]);
num_variables[j] = (HYPRE_Real) hypre_ParCSRBlockMatrixGlobalNumRows(A_block_array[j]);
}
num_coeffs[0] = hypre_ParCSRBlockMatrixDNumNonzeros(A_block_array[0]);
num_variables[0] = hypre_ParCSRBlockMatrixGlobalNumRows(A_block_array[0]);
}
else
{
for (j = 1; j < num_levels; j++)
{
num_coeffs[j] = (HYPRE_Real) hypre_ParCSRMatrixNumNonzeros(A_array[j]);
num_variables[j] = (HYPRE_Real) hypre_ParCSRMatrixGlobalNumRows(A_array[j]);
}
}
for (j=0;j<hypre_ParAMGDataNumLevels(amg_data);j++)
{
total_coeffs += num_coeffs[j];
total_variables += num_variables[j];
}
cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data);
if (num_variables[0])
grid_cmplxty = total_variables / num_variables[0];
if (num_coeffs[0])
{
operat_cmplxty = total_coeffs / num_coeffs[0];
cycle_cmplxty = cycle_op_count / num_coeffs[0];
}
if (my_id == 0)
{
if (Solve_err_flag == 1)
{
hypre_printf("\n\n==============================================");
hypre_printf("\n NOTE: Convergence tolerance was not achieved\n");
hypre_printf(" within the allowed %d V-cycles\n",max_iter);
hypre_printf("==============================================");
}
hypre_printf("\n\n Average Convergence Factor = %f",conv_factor);
hypre_printf("\n\n Complexity: grid = %f\n",grid_cmplxty);
hypre_printf(" operator = %f\n",operat_cmplxty);
hypre_printf(" cycle = %f\n\n\n\n",cycle_cmplxty);
}
hypre_TFree(num_coeffs);
hypre_TFree(num_variables);
}
return hypre_error_flag;
}
HYPRE_Int hypre_seqAMGSetup( hypre_ParAMGData *amg_data,
HYPRE_Int p_level,
HYPRE_Int coarse_threshold)
{
/* Par Data Structure variables */
hypre_ParCSRMatrix **Par_A_array = hypre_ParAMGDataAArray(amg_data);
MPI_Comm comm = hypre_ParCSRMatrixComm(Par_A_array[0]);
MPI_Comm new_comm, seq_comm;
hypre_ParCSRMatrix *A_seq = NULL;
hypre_CSRMatrix *A_seq_diag;
hypre_CSRMatrix *A_seq_offd;
hypre_ParVector *F_seq = NULL;
hypre_ParVector *U_seq = NULL;
hypre_ParCSRMatrix *A;
HYPRE_Int **dof_func_array;
HYPRE_Int num_procs, my_id;
HYPRE_Int not_finished_coarsening;
HYPRE_Int level;
HYPRE_Solver coarse_solver;
/* misc */
dof_func_array = hypre_ParAMGDataDofFuncArray(amg_data);
/*MPI Stuff */
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
/*initial */
level = p_level;
not_finished_coarsening = 1;
/* convert A at this level to sequential */
A = Par_A_array[level];
{
double *A_seq_data = NULL;
HYPRE_Int *A_seq_i = NULL;
HYPRE_Int *A_seq_offd_i = NULL;
HYPRE_Int *A_seq_j = NULL;
double *A_tmp_data = NULL;
HYPRE_Int *A_tmp_i = NULL;
HYPRE_Int *A_tmp_j = NULL;
HYPRE_Int *info, *displs, *displs2;
HYPRE_Int i, j, size, num_nonzeros, total_nnz, cnt;
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int num_rows = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int first_row_index = hypre_ParCSRMatrixFirstRowIndex(A);
hypre_MPI_Group orig_group, new_group;
HYPRE_Int *ranks, new_num_procs, *row_starts;
info = hypre_CTAlloc(HYPRE_Int, num_procs);
hypre_MPI_Allgather(&num_rows, 1, HYPRE_MPI_INT, info, 1, HYPRE_MPI_INT, comm);
ranks = hypre_CTAlloc(HYPRE_Int, num_procs);
new_num_procs = 0;
for (i=0; i < num_procs; i++)
if (info[i])
{
ranks[new_num_procs] = i;
info[new_num_procs++] = info[i];
}
MPI_Comm_group(comm, &orig_group);
hypre_MPI_Group_incl(orig_group, new_num_procs, ranks, &new_group);
MPI_Comm_create(comm, new_group, &new_comm);
hypre_MPI_Group_free(&new_group);
hypre_MPI_Group_free(&orig_group);
if (num_rows)
{
/* alloc space in seq data structure only for participating procs*/
HYPRE_BoomerAMGCreate(&coarse_solver);
HYPRE_BoomerAMGSetMaxRowSum(coarse_solver,
hypre_ParAMGDataMaxRowSum(amg_data));
HYPRE_BoomerAMGSetStrongThreshold(coarse_solver,
hypre_ParAMGDataStrongThreshold(amg_data));
HYPRE_BoomerAMGSetCoarsenType(coarse_solver,
hypre_ParAMGDataCoarsenType(amg_data));
HYPRE_BoomerAMGSetInterpType(coarse_solver,
hypre_ParAMGDataInterpType(amg_data));
HYPRE_BoomerAMGSetTruncFactor(coarse_solver,
hypre_ParAMGDataTruncFactor(amg_data));
HYPRE_BoomerAMGSetPMaxElmts(coarse_solver,
hypre_ParAMGDataPMaxElmts(amg_data));
if (hypre_ParAMGDataUserRelaxType(amg_data) > -1)
HYPRE_BoomerAMGSetRelaxType(coarse_solver,
hypre_ParAMGDataUserRelaxType(amg_data));
HYPRE_BoomerAMGSetRelaxOrder(coarse_solver,
hypre_ParAMGDataRelaxOrder(amg_data));
HYPRE_BoomerAMGSetRelaxWt(coarse_solver,
hypre_ParAMGDataUserRelaxWeight(amg_data));
if (hypre_ParAMGDataUserNumSweeps(amg_data) > -1)
HYPRE_BoomerAMGSetNumSweeps(coarse_solver,
hypre_ParAMGDataUserNumSweeps(amg_data));
HYPRE_BoomerAMGSetNumFunctions(coarse_solver,
hypre_ParAMGDataNumFunctions(amg_data));
HYPRE_BoomerAMGSetMaxIter(coarse_solver, 1);
HYPRE_BoomerAMGSetTol(coarse_solver, 0);
/* Create CSR Matrix, will be Diag part of new matrix */
A_tmp_i = hypre_CTAlloc(HYPRE_Int, num_rows+1);
A_tmp_i[0] = 0;
for (i=1; i < num_rows+1; i++)
A_tmp_i[i] = A_diag_i[i]-A_diag_i[i-1]+A_offd_i[i]-A_offd_i[i-1];
num_nonzeros = A_offd_i[num_rows]+A_diag_i[num_rows];
A_tmp_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros);
A_tmp_data = hypre_CTAlloc(double, num_nonzeros);
cnt = 0;
for (i=0; i < num_rows; i++)
{
for (j=A_diag_i[i]; j < A_diag_i[i+1]; j++)
{
A_tmp_j[cnt] = A_diag_j[j]+first_row_index;
A_tmp_data[cnt++] = A_diag_data[j];
}
for (j=A_offd_i[i]; j < A_offd_i[i+1]; j++)
{
A_tmp_j[cnt] = col_map_offd[A_offd_j[j]];
A_tmp_data[cnt++] = A_offd_data[j];
}
}
displs = hypre_CTAlloc(HYPRE_Int, new_num_procs+1);
displs[0] = 0;
for (i=1; i < new_num_procs+1; i++)
displs[i] = displs[i-1]+info[i-1];
size = displs[new_num_procs];
A_seq_i = hypre_CTAlloc(HYPRE_Int, size+1);
A_seq_offd_i = hypre_CTAlloc(HYPRE_Int, size+1);
hypre_MPI_Allgatherv ( &A_tmp_i[1], num_rows, HYPRE_MPI_INT, &A_seq_i[1], info,
displs, HYPRE_MPI_INT, new_comm );
displs2 = hypre_CTAlloc(HYPRE_Int, new_num_procs+1);
A_seq_i[0] = 0;
displs2[0] = 0;
for (j=1; j < displs[1]; j++)
A_seq_i[j] = A_seq_i[j]+A_seq_i[j-1];
for (i=1; i < new_num_procs; i++)
{
for (j=displs[i]; j < displs[i+1]; j++)
{
A_seq_i[j] = A_seq_i[j]+A_seq_i[j-1];
}
}
A_seq_i[size] = A_seq_i[size]+A_seq_i[size-1];
displs2[new_num_procs] = A_seq_i[size];
for (i=1; i < new_num_procs+1; i++)
{
displs2[i] = A_seq_i[displs[i]];
info[i-1] = displs2[i] - displs2[i-1];
}
total_nnz = displs2[new_num_procs];
A_seq_j = hypre_CTAlloc(HYPRE_Int, total_nnz);
A_seq_data = hypre_CTAlloc(double, total_nnz);
hypre_MPI_Allgatherv ( A_tmp_j, num_nonzeros, HYPRE_MPI_INT,
A_seq_j, info, displs2,
HYPRE_MPI_INT, new_comm );
hypre_MPI_Allgatherv ( A_tmp_data, num_nonzeros, hypre_MPI_DOUBLE,
A_seq_data, info, displs2,
hypre_MPI_DOUBLE, new_comm );
hypre_TFree(displs);
hypre_TFree(displs2);
hypre_TFree(A_tmp_i);
hypre_TFree(A_tmp_j);
hypre_TFree(A_tmp_data);
row_starts = hypre_CTAlloc(HYPRE_Int,2);
row_starts[0] = 0;
row_starts[1] = size;
/* Create 1 proc communicator */
seq_comm = hypre_MPI_COMM_SELF;
A_seq = hypre_ParCSRMatrixCreate(seq_comm,size,size,
row_starts, row_starts,
0,total_nnz,0);
A_seq_diag = hypre_ParCSRMatrixDiag(A_seq);
A_seq_offd = hypre_ParCSRMatrixOffd(A_seq);
hypre_CSRMatrixData(A_seq_diag) = A_seq_data;
hypre_CSRMatrixI(A_seq_diag) = A_seq_i;
hypre_CSRMatrixJ(A_seq_diag) = A_seq_j;
hypre_CSRMatrixI(A_seq_offd) = A_seq_offd_i;
F_seq = hypre_ParVectorCreate(seq_comm, size, row_starts);
U_seq = hypre_ParVectorCreate(seq_comm, size, row_starts);
hypre_ParVectorOwnsPartitioning(F_seq) = 0;
hypre_ParVectorOwnsPartitioning(U_seq) = 0;
hypre_ParVectorInitialize(F_seq);
hypre_ParVectorInitialize(U_seq);
hypre_BoomerAMGSetup(coarse_solver,A_seq,F_seq,U_seq);
hypre_ParAMGDataCoarseSolver(amg_data) = coarse_solver;
hypre_ParAMGDataACoarse(amg_data) = A_seq;
hypre_ParAMGDataFCoarse(amg_data) = F_seq;
hypre_ParAMGDataUCoarse(amg_data) = U_seq;
hypre_ParAMGDataNewComm(amg_data) = new_comm;
}
hypre_TFree(info);
hypre_TFree(ranks);
}
return 0;
}
void hypre_ParCSRMatrixSplit(hypre_ParCSRMatrix *A,
HYPRE_Int nr, HYPRE_Int nc,
hypre_ParCSRMatrix **blocks,
int interleaved_rows, int interleaved_cols)
{
HYPRE_Int i, j, k;
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *Adiag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *Aoffd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int global_rows = hypre_ParCSRMatrixGlobalNumRows(A);
HYPRE_Int global_cols = hypre_ParCSRMatrixGlobalNumCols(A);
HYPRE_Int local_rows = hypre_CSRMatrixNumRows(Adiag);
HYPRE_Int local_cols = hypre_CSRMatrixNumCols(Adiag);
HYPRE_Int offd_cols = hypre_CSRMatrixNumCols(Aoffd);
hypre_assert(local_rows % nr == 0 && local_cols % nc == 0);
hypre_assert(global_rows % nr == 0 && global_cols % nc == 0);
HYPRE_Int block_rows = local_rows / nr;
HYPRE_Int block_cols = local_cols / nc;
HYPRE_Int num_blocks = nr * nc;
/* mark local rows and columns with block number */
HYPRE_Int *row_block_num = hypre_TAlloc(HYPRE_Int, local_rows);
HYPRE_Int *col_block_num = hypre_TAlloc(HYPRE_Int, local_cols);
for (i = 0; i < local_rows; i++)
{
row_block_num[i] = interleaved_rows ? (i % nr) : (i / block_rows);
}
for (i = 0; i < local_cols; i++)
{
col_block_num[i] = interleaved_cols ? (i % nc) : (i / block_cols);
}
/* determine the block numbers for offd columns */
HYPRE_Int* offd_col_block_num = hypre_TAlloc(HYPRE_Int, offd_cols);
hypre_ParCSRCommHandle *comm_handle;
HYPRE_Int *int_buf_data;
{
/* make sure A has a communication package */
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(A);
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
/* calculate the final global column numbers for each block */
HYPRE_Int *count = hypre_CTAlloc(HYPRE_Int, nc);
HYPRE_Int *block_global_col = hypre_TAlloc(HYPRE_Int, local_cols);
HYPRE_Int first_col = hypre_ParCSRMatrixFirstColDiag(A) / nc;
for (i = 0; i < local_cols; i++)
{
block_global_col[i] = first_col + count[col_block_num[i]]++;
}
hypre_TFree(count);
/* use a Matvec communication pattern to determine offd_col_block_num */
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
int_buf_data = hypre_CTAlloc(HYPRE_Int,
hypre_ParCSRCommPkgSendMapStart(comm_pkg,
num_sends));
HYPRE_Int start, index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
{
k = hypre_ParCSRCommPkgSendMapElmt(comm_pkg, j);
int_buf_data[index++] = col_block_num[k] + nc*block_global_col[k];
}
}
hypre_TFree(block_global_col);
comm_handle = hypre_ParCSRCommHandleCreate(11, comm_pkg, int_buf_data,
offd_col_block_num);
}
/* create the block matrices */
HYPRE_Int num_procs = 1;
if (!hypre_ParCSRMatrixAssumedPartition(A))
{
hypre_MPI_Comm_size(comm, &num_procs);
}
HYPRE_Int *row_starts = hypre_TAlloc(HYPRE_Int, num_procs+1);
HYPRE_Int *col_starts = hypre_TAlloc(HYPRE_Int, num_procs+1);
for (i = 0; i <= num_procs; i++)
{
row_starts[i] = hypre_ParCSRMatrixRowStarts(A)[i] / nr;
col_starts[i] = hypre_ParCSRMatrixColStarts(A)[i] / nc;
}
for (i = 0; i < num_blocks; i++)
{
blocks[i] = hypre_ParCSRMatrixCreate(comm,
global_rows/nr, global_cols/nc,
row_starts, col_starts, 0, 0, 0);
}
/* split diag part */
hypre_CSRMatrix **csr_blocks = hypre_TAlloc(hypre_CSRMatrix*, nr*nc);
hypre_CSRMatrixSplit(Adiag, nr, nc, row_block_num, col_block_num,
csr_blocks);
for (i = 0; i < num_blocks; i++)
{
hypre_TFree(hypre_ParCSRMatrixDiag(blocks[i]));
hypre_ParCSRMatrixDiag(blocks[i]) = csr_blocks[i];
}
/* finish communication, receive offd_col_block_num */
hypre_ParCSRCommHandleDestroy(comm_handle);
hypre_TFree(int_buf_data);
/* decode global offd column numbers */
HYPRE_Int* offd_global_col = hypre_TAlloc(HYPRE_Int, offd_cols);
for (i = 0; i < offd_cols; i++)
{
offd_global_col[i] = offd_col_block_num[i] / nc;
offd_col_block_num[i] %= nc;
}
/* split offd part */
hypre_CSRMatrixSplit(Aoffd, nr, nc, row_block_num, offd_col_block_num,
csr_blocks);
for (i = 0; i < num_blocks; i++)
{
hypre_TFree(hypre_ParCSRMatrixOffd(blocks[i]));
hypre_ParCSRMatrixOffd(blocks[i]) = csr_blocks[i];
}
hypre_TFree(csr_blocks);
hypre_TFree(col_block_num);
hypre_TFree(row_block_num);
/* update block col-maps */
for (int bi = 0; bi < nr; bi++)
{
for (int bj = 0; bj < nc; bj++)
{
hypre_ParCSRMatrix *block = blocks[bi*nc + bj];
hypre_CSRMatrix *block_offd = hypre_ParCSRMatrixOffd(block);
HYPRE_Int block_offd_cols = hypre_CSRMatrixNumCols(block_offd);
HYPRE_Int *block_col_map = hypre_TAlloc(HYPRE_Int, block_offd_cols);
for (i = j = 0; i < offd_cols; i++)
{
HYPRE_Int bn = offd_col_block_num[i];
if (bn == bj) {
block_col_map[j++] = offd_global_col[i];
}
}
hypre_assert(j == block_offd_cols);
hypre_ParCSRMatrixColMapOffd(block) = block_col_map;
}
}
hypre_TFree(offd_global_col);
hypre_TFree(offd_col_block_num);
/* finish the new matrices, make them own all the stuff */
for (i = 0; i < num_blocks; i++)
{
hypre_ParCSRMatrixSetNumNonzeros(blocks[i]);
hypre_MatvecCommPkgCreate(blocks[i]);
hypre_ParCSRMatrixOwnsData(blocks[i]) = 1;
/* only the first block will own the row/col_starts */
hypre_ParCSRMatrixOwnsRowStarts(blocks[i]) = !i;
hypre_ParCSRMatrixOwnsColStarts(blocks[i]) = !i;
}
}
HYPRE_Int
hypre_SchwarzSetup(void *schwarz_vdata,
hypre_ParCSRMatrix *A,
hypre_ParVector *f,
hypre_ParVector *u )
{
hypre_SchwarzData *schwarz_data = schwarz_vdata;
HYPRE_Int *dof_func;
double *scale;
hypre_CSRMatrix *domain_structure;
hypre_CSRMatrix *A_boundary;
hypre_ParVector *Vtemp;
HYPRE_Int *pivots = NULL;
HYPRE_Int variant = hypre_SchwarzDataVariant(schwarz_data);
HYPRE_Int domain_type = hypre_SchwarzDataDomainType(schwarz_data);
HYPRE_Int overlap = hypre_SchwarzDataOverlap(schwarz_data);
HYPRE_Int num_functions = hypre_SchwarzDataNumFunctions(schwarz_data);
double relax_weight = hypre_SchwarzDataRelaxWeight(schwarz_data);
HYPRE_Int use_nonsymm = hypre_SchwarzDataUseNonSymm(schwarz_data);
dof_func = hypre_SchwarzDataDofFunc(schwarz_data);
Vtemp = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(A),
hypre_ParCSRMatrixGlobalNumRows(A),
hypre_ParCSRMatrixRowStarts(A));
hypre_ParVectorSetPartitioningOwner(Vtemp,0);
hypre_ParVectorInitialize(Vtemp);
hypre_SchwarzDataVtemp(schwarz_data) = Vtemp;
if (variant > 1)
{
hypre_ParAMGCreateDomainDof(A,
domain_type, overlap,
num_functions, dof_func,
&domain_structure, &pivots, use_nonsymm);
if (variant == 2)
{
hypre_ParGenerateScale(A, domain_structure, relax_weight,
&scale);
hypre_SchwarzDataScale(schwarz_data) = scale;
}
else
{
hypre_ParGenerateHybridScale(A, domain_structure, &A_boundary, &scale);
hypre_SchwarzDataScale(schwarz_data) = scale;
if (hypre_CSRMatrixNumCols(hypre_ParCSRMatrixOffd(A)))
hypre_SchwarzDataABoundary(schwarz_data) = A_boundary;
else
hypre_SchwarzDataABoundary(schwarz_data) = NULL;
}
}
else
{
hypre_AMGCreateDomainDof (hypre_ParCSRMatrixDiag(A),
domain_type, overlap,
num_functions, dof_func,
&domain_structure, &pivots, use_nonsymm);
if (variant == 1)
{
hypre_GenerateScale(domain_structure,
hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A)),
relax_weight, &scale);
hypre_SchwarzDataScale(schwarz_data) = scale;
}
}
hypre_SchwarzDataDomainStructure(schwarz_data) = domain_structure;
hypre_SchwarzDataPivots(schwarz_data) = pivots;
return hypre_error_flag;
}
HYPRE_Int
hypre_BoomerAMGSolveT( void *amg_vdata,
hypre_ParCSRMatrix *A,
hypre_ParVector *f,
hypre_ParVector *u )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
HYPRE_Int amg_print_level;
HYPRE_Int amg_logging;
HYPRE_Real *num_coeffs;
HYPRE_Int *num_variables;
HYPRE_Real cycle_op_count;
HYPRE_Int num_levels;
/* HYPRE_Int num_unknowns; */
HYPRE_Real tol;
char *file_name;
hypre_ParCSRMatrix **A_array;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
/* Local variables */
/*FILE *fp;*/
HYPRE_Int j;
HYPRE_Int Solve_err_flag;
HYPRE_Int min_iter;
HYPRE_Int max_iter;
HYPRE_Int cycle_count;
HYPRE_Real total_coeffs;
HYPRE_Int total_variables;
HYPRE_Int num_procs, my_id;
HYPRE_Real alpha = 1.0;
HYPRE_Real beta = -1.0;
HYPRE_Real cycle_cmplxty = 0.0;
HYPRE_Real operat_cmplxty;
HYPRE_Real grid_cmplxty;
HYPRE_Real conv_factor;
HYPRE_Real resid_nrm;
HYPRE_Real resid_nrm_init;
HYPRE_Real relative_resid;
HYPRE_Real rhs_norm;
HYPRE_Real old_resid;
hypre_ParVector *Vtemp;
hypre_ParVector *Residual;
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
amg_print_level = hypre_ParAMGDataPrintLevel(amg_data);
amg_logging = hypre_ParAMGDataLogging(amg_data);
if ( amg_logging>1 )
Residual = hypre_ParAMGDataResidual(amg_data);
file_name = hypre_ParAMGDataLogFileName(amg_data);
/* num_unknowns = hypre_ParAMGDataNumUnknowns(amg_data); */
num_levels = hypre_ParAMGDataNumLevels(amg_data);
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
tol = hypre_ParAMGDataTol(amg_data);
min_iter = hypre_ParAMGDataMinIter(amg_data);
max_iter = hypre_ParAMGDataMaxIter(amg_data);
num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels);
num_variables = hypre_CTAlloc(HYPRE_Int, num_levels);
num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A_array[0]);
num_variables[0] = hypre_ParCSRMatrixGlobalNumRows(A_array[0]);
A_array[0] = A;
F_array[0] = f;
U_array[0] = u;
/* Vtemp = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(A_array[0]),
hypre_ParCSRMatrixGlobalNumRows(A_array[0]),
hypre_ParCSRMatrixRowStarts(A_array[0]));
hypre_ParVectorInitialize(Vtemp);
hypre_ParVectorSetPartitioningOwner(Vtemp,0);
hypre_ParAMGDataVtemp(amg_data) = Vtemp;
*/
Vtemp = hypre_ParAMGDataVtemp(amg_data);
for (j = 1; j < num_levels; j++)
{
num_coeffs[j] = hypre_ParCSRMatrixDNumNonzeros(A_array[j]);
num_variables[j] = hypre_ParCSRMatrixGlobalNumRows(A_array[j]);
}
/*-----------------------------------------------------------------------
* Write the solver parameters
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1)
hypre_BoomerAMGWriteSolverParams(amg_data);
/*-----------------------------------------------------------------------
* Initialize the solver error flag and assorted bookkeeping variables
*-----------------------------------------------------------------------*/
Solve_err_flag = 0;
total_coeffs = 0;
total_variables = 0;
cycle_count = 0;
operat_cmplxty = 0;
grid_cmplxty = 0;
/*-----------------------------------------------------------------------
* open the log file and write some initial info
*-----------------------------------------------------------------------*/
if (my_id == 0 && amg_print_level > 1)
{
/*fp = fopen(file_name, "a");*/
hypre_printf("\n\nAMG SOLUTION INFO:\n");
}
/*-----------------------------------------------------------------------
* Compute initial fine-grid residual and print to logfile
*-----------------------------------------------------------------------*/
if ( amg_logging > 1 ) {
hypre_ParVectorCopy(F_array[0], Residual );
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParVectorCopy(F_array[0], Vtemp);
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
resid_nrm_init = resid_nrm;
rhs_norm = sqrt(hypre_ParVectorInnerProd(f, f));
relative_resid = 9999;
if (rhs_norm)
{
relative_resid = resid_nrm_init / rhs_norm;
}
if (my_id ==0 && (amg_print_level > 1))
{
hypre_printf(" relative\n");
hypre_printf(" residual factor residual\n");
hypre_printf(" -------- ------ --------\n");
hypre_printf(" Initial %e %e\n",resid_nrm_init,
relative_resid);
}
/*-----------------------------------------------------------------------
* Main V-cycle loop
*-----------------------------------------------------------------------*/
while ((relative_resid >= tol || cycle_count < min_iter)
&& cycle_count < max_iter
&& Solve_err_flag == 0)
{
hypre_ParAMGDataCycleOpCount(amg_data) = 0;
/* Op count only needed for one cycle */
Solve_err_flag = hypre_BoomerAMGCycleT(amg_data, F_array, U_array);
old_resid = resid_nrm;
/*---------------------------------------------------------------
* Compute fine-grid residual and residual norm
*----------------------------------------------------------------*/
if ( amg_logging > 1 ) {
hypre_ParVectorCopy(F_array[0], Residual );
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Residual );
resid_nrm = sqrt(hypre_ParVectorInnerProd( Residual, Residual ));
}
else {
hypre_ParVectorCopy(F_array[0], Vtemp);
hypre_ParCSRMatrixMatvecT(alpha, A_array[0], U_array[0], beta, Vtemp);
resid_nrm = sqrt(hypre_ParVectorInnerProd(Vtemp, Vtemp));
}
conv_factor = resid_nrm / old_resid;
relative_resid = 9999;
if (rhs_norm)
{
relative_resid = resid_nrm / rhs_norm;
}
++cycle_count;
hypre_ParAMGDataRelativeResidualNorm(amg_data) = relative_resid;
hypre_ParAMGDataNumIterations(amg_data) = cycle_count;
if (my_id == 0 && (amg_print_level > 1))
{
hypre_printf(" Cycle %2d %e %f %e \n", cycle_count,
resid_nrm, conv_factor, relative_resid);
}
}
if (cycle_count == max_iter) Solve_err_flag = 1;
/*-----------------------------------------------------------------------
* Compute closing statistics
*-----------------------------------------------------------------------*/
conv_factor = pow((resid_nrm/resid_nrm_init),(1.0/((HYPRE_Real) cycle_count)));
for (j=0;j<hypre_ParAMGDataNumLevels(amg_data);j++)
{
total_coeffs += num_coeffs[j];
total_variables += num_variables[j];
}
cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data);
if (num_variables[0])
grid_cmplxty = ((HYPRE_Real) total_variables) / ((HYPRE_Real) num_variables[0]);
if (num_coeffs[0])
{
operat_cmplxty = total_coeffs / num_coeffs[0];
cycle_cmplxty = cycle_op_count / num_coeffs[0];
}
if (my_id == 0 && amg_print_level > 1)
{
if (Solve_err_flag == 1)
{
hypre_printf("\n\n==============================================");
hypre_printf("\n NOTE: Convergence tolerance was not achieved\n");
hypre_printf(" within the allowed %d V-cycles\n",max_iter);
hypre_printf("==============================================");
}
hypre_printf("\n\n Average Convergence Factor = %f",conv_factor);
hypre_printf("\n\n Complexity: grid = %f\n",grid_cmplxty);
hypre_printf(" operator = %f\n",operat_cmplxty);
hypre_printf(" cycle = %f\n\n",cycle_cmplxty);
}
/*----------------------------------------------------------
* Close the output file (if open)
*----------------------------------------------------------*/
/*if (my_id == 0 && amg_print_level >= 1)
{
fclose(fp);
}*/
hypre_TFree(num_coeffs);
hypre_TFree(num_variables);
return(Solve_err_flag);
}
hypre_ParCSRMatrix * hypre_ParMatmul_FC(
hypre_ParCSRMatrix * A, hypre_ParCSRMatrix * P, HYPRE_Int * CF_marker,
HYPRE_Int * dof_func, HYPRE_Int * dof_func_offd )
/* hypre_parMatmul_FC creates and returns the "Fine"-designated rows of the
matrix product A*P. A's size is (nC+nF)*(nC+nF), P's size is (nC+nF)*nC
where nC is the number of coarse rows/columns, nF the number of fine
rows/columns. The size of C=A*P is (nC+nF)*nC, even though not all rows
of C are actually computed. If we were to construct a matrix consisting
only of the computed rows of C, its size would be nF*nC.
"Fine" is defined solely by the marker array, and for example could be
a proper subset of the fine points of a multigrid hierarchy.
*/
{
/* To compute a submatrix of C containing only the computed data, i.e.
only "Fine" rows, we would have to do a lot of computational work,
with a lot of communication. The communication is because such a
matrix would need global information that depends on which rows are
"Fine".
*/
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *row_starts_A = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int num_rows_diag_A = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_cols_diag_A = hypre_CSRMatrixNumCols(A_diag);
HYPRE_Int num_cols_offd_A = hypre_CSRMatrixNumCols(A_offd);
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
double *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
HYPRE_Int *col_map_offd_P = hypre_ParCSRMatrixColMapOffd(P);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int first_col_diag_P = hypre_ParCSRMatrixFirstColDiag(P);
HYPRE_Int last_col_diag_P;
HYPRE_Int *col_starts_P = hypre_ParCSRMatrixColStarts(P);
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int num_cols_diag_P = hypre_CSRMatrixNumCols(P_diag);
HYPRE_Int num_cols_offd_P = hypre_CSRMatrixNumCols(P_offd);
hypre_ParCSRMatrix *C;
HYPRE_Int *col_map_offd_C;
HYPRE_Int *map_P_to_C;
hypre_CSRMatrix *C_diag;
double *C_diag_data;
HYPRE_Int *C_diag_i;
HYPRE_Int *C_diag_j;
hypre_CSRMatrix *C_offd;
double *C_offd_data=NULL;
HYPRE_Int *C_offd_i=NULL;
HYPRE_Int *C_offd_j=NULL;
HYPRE_Int C_diag_size;
HYPRE_Int C_offd_size;
HYPRE_Int num_cols_offd_C = 0;
hypre_CSRMatrix *Ps_ext;
double *Ps_ext_data;
HYPRE_Int *Ps_ext_i;
HYPRE_Int *Ps_ext_j;
double *P_ext_diag_data;
HYPRE_Int *P_ext_diag_i;
HYPRE_Int *P_ext_diag_j;
HYPRE_Int P_ext_diag_size;
double *P_ext_offd_data;
HYPRE_Int *P_ext_offd_i;
HYPRE_Int *P_ext_offd_j;
HYPRE_Int P_ext_offd_size;
HYPRE_Int *P_marker;
HYPRE_Int *temp;
HYPRE_Int i, j;
HYPRE_Int i1, i2, i3;
HYPRE_Int jj2, jj3;
HYPRE_Int jj_count_diag, jj_count_offd;
HYPRE_Int jj_row_begin_diag, jj_row_begin_offd;
HYPRE_Int start_indexing = 0; /* start indexing for C_data at 0 */
HYPRE_Int n_rows_A_global, n_cols_A_global;
HYPRE_Int n_rows_P_global, n_cols_P_global;
HYPRE_Int allsquare = 0;
HYPRE_Int cnt, cnt_offd, cnt_diag;
HYPRE_Int num_procs;
HYPRE_Int value;
double a_entry;
double a_b_product;
n_rows_A_global = hypre_ParCSRMatrixGlobalNumRows(A);
n_cols_A_global = hypre_ParCSRMatrixGlobalNumCols(A);
n_rows_P_global = hypre_ParCSRMatrixGlobalNumRows(P);
n_cols_P_global = hypre_ParCSRMatrixGlobalNumCols(P);
if (n_cols_A_global != n_rows_P_global || num_cols_diag_A != num_rows_diag_P)
{
hypre_printf(" Error! Incompatible matrix dimensions!\n");
return NULL;
}
/* if (num_rows_A==num_cols_P) allsquare = 1; */
/*-----------------------------------------------------------------------
* Extract P_ext, i.e. portion of P that is stored on neighbor procs
* and needed locally for matrix matrix product
*-----------------------------------------------------------------------*/
hypre_MPI_Comm_size(comm, &num_procs);
if (num_procs > 1)
{
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings within
* hypre_ParCSRMatrixExtractBExt
*--------------------------------------------------------------------*/
Ps_ext = hypre_ParCSRMatrixExtractBExt(P,A,1);
Ps_ext_data = hypre_CSRMatrixData(Ps_ext);
Ps_ext_i = hypre_CSRMatrixI(Ps_ext);
Ps_ext_j = hypre_CSRMatrixJ(Ps_ext);
}
P_ext_diag_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_offd_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_diag_size = 0;
P_ext_offd_size = 0;
last_col_diag_P = first_col_diag_P + num_cols_diag_P -1;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
P_ext_offd_size++;
else
P_ext_diag_size++;
P_ext_diag_i[i+1] = P_ext_diag_size;
P_ext_offd_i[i+1] = P_ext_offd_size;
}
if (P_ext_diag_size)
{
P_ext_diag_j = hypre_CTAlloc(HYPRE_Int, P_ext_diag_size);
P_ext_diag_data = hypre_CTAlloc(double, P_ext_diag_size);
}
if (P_ext_offd_size)
{
P_ext_offd_j = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size);
P_ext_offd_data = hypre_CTAlloc(double, P_ext_offd_size);
}
cnt_offd = 0;
cnt_diag = 0;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
{
P_ext_offd_j[cnt_offd] = Ps_ext_j[j];
P_ext_offd_data[cnt_offd++] = Ps_ext_data[j];
}
else
{
P_ext_diag_j[cnt_diag] = Ps_ext_j[j] - first_col_diag_P;
P_ext_diag_data[cnt_diag++] = Ps_ext_data[j];
}
}
if (num_procs > 1)
{
hypre_CSRMatrixDestroy(Ps_ext);
Ps_ext = NULL;
}
cnt = 0;
if (P_ext_offd_size || num_cols_offd_P)
{
temp = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size+num_cols_offd_P);
for (i=0; i < P_ext_offd_size; i++)
temp[i] = P_ext_offd_j[i];
cnt = P_ext_offd_size;
for (i=0; i < num_cols_offd_P; i++)
temp[cnt++] = col_map_offd_P[i];
}
if (cnt)
{
qsort0(temp, 0, cnt-1);
num_cols_offd_C = 1;
value = temp[0];
for (i=1; i < cnt; i++)
{
if (temp[i] > value)
{
value = temp[i];
temp[num_cols_offd_C++] = value;
}
}
}
if (num_cols_offd_C)
col_map_offd_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_C);
for (i=0; i < num_cols_offd_C; i++)
col_map_offd_C[i] = temp[i];
if (P_ext_offd_size || num_cols_offd_P)
hypre_TFree(temp);
for (i=0 ; i < P_ext_offd_size; i++)
P_ext_offd_j[i] = hypre_BinarySearch(col_map_offd_C,
P_ext_offd_j[i],
num_cols_offd_C);
if (num_cols_offd_P)
{
map_P_to_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_P);
cnt = 0;
for (i=0; i < num_cols_offd_C; i++)
if (col_map_offd_C[i] == col_map_offd_P[cnt])
{
map_P_to_C[cnt++] = i;
if (cnt == num_cols_offd_P) break;
}
}
/*-----------------------------------------------------------------------
* Allocate marker array.
*-----------------------------------------------------------------------*/
P_marker = hypre_CTAlloc(HYPRE_Int, num_cols_diag_P+num_cols_offd_C);
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/* no changes for the marked version above this point */
/* This function call is the first pass: */
hypre_ParMatmul_RowSizes_Marked(
&C_diag_i, &C_offd_i, &P_marker,
A_diag_i, A_diag_j, A_offd_i, A_offd_j,
P_diag_i, P_diag_j, P_offd_i, P_offd_j,
P_ext_diag_i, P_ext_diag_j, P_ext_offd_i, P_ext_offd_j,
map_P_to_C,
&C_diag_size, &C_offd_size,
num_rows_diag_A, num_cols_offd_A, allsquare,
num_cols_diag_P, num_cols_offd_P,
num_cols_offd_C, CF_marker, dof_func, dof_func_offd
);
/* The above call of hypre_ParMatmul_RowSizes_Marked computed
two scalars: C_diag_size, C_offd_size,
and two arrays: C_diag_i, C_offd_i
( P_marker is also computed, but only used internally )
*/
/*-----------------------------------------------------------------------
* Allocate C_diag_data and C_diag_j arrays.
* Allocate C_offd_data and C_offd_j arrays.
*-----------------------------------------------------------------------*/
last_col_diag_P = first_col_diag_P + num_cols_diag_P - 1;
C_diag_data = hypre_CTAlloc(double, C_diag_size);
C_diag_j = hypre_CTAlloc(HYPRE_Int, C_diag_size);
if (C_offd_size)
{
C_offd_data = hypre_CTAlloc(double, C_offd_size);
C_offd_j = hypre_CTAlloc(HYPRE_Int, C_offd_size);
}
/*-----------------------------------------------------------------------
* Second Pass: Fill in C_diag_data and C_diag_j.
* Second Pass: Fill in C_offd_data and C_offd_j.
*-----------------------------------------------------------------------*/
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
jj_count_diag = start_indexing;
jj_count_offd = start_indexing;
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/*-----------------------------------------------------------------------
* Loop over interior c-points.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_A; i1++)
{
if ( CF_marker[i1] < 0 ) /* i1 is a fine row */
/* ... This and the coarse row code are the only parts between first pass
and near the end where
hypre_ParMatmul_FC is different from the regular hypre_ParMatmul */
{
/*--------------------------------------------------------------------
* Create diagonal entry, C_{i1,i1}
*--------------------------------------------------------------------*/
jj_row_begin_diag = jj_count_diag;
jj_row_begin_offd = jj_count_offd;
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_offd.
*-----------------------------------------------------------------*/
if (num_cols_offd_A)
{
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func_offd[i2] )
{ /* interpolate only like "functions" */
a_entry = A_offd_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_ext.
*-----------------------------------------------------------*/
for (jj3 = P_ext_offd_i[i2]; jj3 < P_ext_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+P_ext_offd_j[jj3];
a_b_product = a_entry * P_ext_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
C_offd_data[P_marker[i3]] += a_b_product;
}
for (jj3 = P_ext_diag_i[i2]; jj3 < P_ext_diag_i[i2+1]; jj3++)
{
i3 = P_ext_diag_j[jj3];
a_b_product = a_entry * P_ext_diag_data[jj3];
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
C_diag_data[P_marker[i3]] += a_b_product;
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_diag.
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func[i2] )
{ /* interpolate only like "functions" */
a_entry = A_diag_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_diag.
*-----------------------------------------------------------*/
for (jj3 = P_diag_i[i2]; jj3 < P_diag_i[i2+1]; jj3++)
{
i3 = P_diag_j[jj3];
a_b_product = a_entry * P_diag_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[P_marker[i3]] += a_b_product;
}
}
if (num_cols_offd_P)
{
for (jj3 = P_offd_i[i2]; jj3 < P_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+map_P_to_C[P_offd_j[jj3]];
a_b_product = a_entry * P_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
{
C_offd_data[P_marker[i3]] += a_b_product;
}
}
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
else /* i1 is a coarse row.*/
/* Copy P coarse-row values to C. This is useful if C is meant to
become a replacement for P */
{
if (num_cols_offd_P)
{
for (jj2 = P_offd_i[i1]; jj2 < P_offd_i[i1+1]; jj2++)
{
C_offd_j[jj_count_offd] = P_offd_j[jj_count_offd];
C_offd_data[jj_count_offd] = P_offd_data[jj_count_offd];
++jj_count_offd;
}
}
for (jj2 = P_diag_i[i1]; jj2 < P_diag_i[i1+1]; jj2++)
{
C_diag_j[jj_count_diag] = P_diag_j[jj2];
C_diag_data[jj_count_diag] = P_diag_data[jj2];
++jj_count_diag;
}
}
}
C = hypre_ParCSRMatrixCreate(
comm, n_rows_A_global, n_cols_P_global,
row_starts_A, col_starts_P, num_cols_offd_C, C_diag_size, C_offd_size );
/* Note that C does not own the partitionings */
hypre_ParCSRMatrixSetRowStartsOwner(C,0);
hypre_ParCSRMatrixSetColStartsOwner(C,0);
C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrixData(C_diag) = C_diag_data;
hypre_CSRMatrixI(C_diag) = C_diag_i;
hypre_CSRMatrixJ(C_diag) = C_diag_j;
C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrixI(C_offd) = C_offd_i;
hypre_ParCSRMatrixOffd(C) = C_offd;
if (num_cols_offd_C)
{
hypre_CSRMatrixData(C_offd) = C_offd_data;
hypre_CSRMatrixJ(C_offd) = C_offd_j;
hypre_ParCSRMatrixColMapOffd(C) = col_map_offd_C;
}
/*-----------------------------------------------------------------------
* Free various arrays
*-----------------------------------------------------------------------*/
hypre_TFree(P_marker);
hypre_TFree(P_ext_diag_i);
if (P_ext_diag_size)
{
hypre_TFree(P_ext_diag_j);
hypre_TFree(P_ext_diag_data);
}
hypre_TFree(P_ext_offd_i);
if (P_ext_offd_size)
{
hypre_TFree(P_ext_offd_j);
hypre_TFree(P_ext_offd_data);
}
if (num_cols_offd_P) hypre_TFree(map_P_to_C);
return C;
}
/**************************************************************
*
* CGC Coarsening routine
*
**************************************************************/
HYPRE_Int
hypre_BoomerAMGCoarsenCGCb( hypre_ParCSRMatrix *S,
hypre_ParCSRMatrix *A,
HYPRE_Int measure_type,
HYPRE_Int coarsen_type,
HYPRE_Int cgc_its,
HYPRE_Int debug_flag,
HYPRE_Int **CF_marker_ptr)
{
MPI_Comm comm = hypre_ParCSRMatrixComm(S);
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(S);
hypre_ParCSRCommHandle *comm_handle;
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;
HYPRE_Int num_variables = hypre_CSRMatrixNumRows(S_diag);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(S_offd);
hypre_CSRMatrix *S_ext;
HYPRE_Int *S_ext_i;
HYPRE_Int *S_ext_j;
hypre_CSRMatrix *ST;
HYPRE_Int *ST_i;
HYPRE_Int *ST_j;
HYPRE_Int *CF_marker;
HYPRE_Int *CF_marker_offd=NULL;
HYPRE_Int ci_tilde = -1;
HYPRE_Int ci_tilde_mark = -1;
HYPRE_Int *measure_array;
HYPRE_Int *measure_array_master;
HYPRE_Int *graph_array;
HYPRE_Int *int_buf_data=NULL;
/*HYPRE_Int *ci_array=NULL;*/
HYPRE_Int i, j, k, l, jS;
HYPRE_Int ji, jj, index;
HYPRE_Int set_empty = 1;
HYPRE_Int C_i_nonempty = 0;
HYPRE_Int num_nonzeros;
HYPRE_Int num_procs, my_id;
HYPRE_Int num_sends = 0;
HYPRE_Int first_col, start;
HYPRE_Int col_0, col_n;
hypre_LinkList LoL_head;
hypre_LinkList LoL_tail;
HYPRE_Int *lists, *where;
HYPRE_Int measure, new_meas;
HYPRE_Int num_left;
HYPRE_Int nabor, nabor_two;
HYPRE_Int ierr = 0;
HYPRE_Int use_commpkg_A = 0;
HYPRE_Real wall_time;
HYPRE_Int measure_max; /* BM Aug 30, 2006: maximal measure, needed for CGC */
if (coarsen_type < 0) coarsen_type = -coarsen_type;
/*-------------------------------------------------------
* Initialize the C/F marker, LoL_head, LoL_tail arrays
*-------------------------------------------------------*/
LoL_head = NULL;
LoL_tail = NULL;
lists = hypre_CTAlloc(HYPRE_Int, num_variables);
where = hypre_CTAlloc(HYPRE_Int, num_variables);
#if 0 /* debugging */
char filename[256];
FILE *fp;
HYPRE_Int iter = 0;
#endif
/*--------------------------------------------------------------
* Compute a CSR strength matrix, S.
*
* For now, the "strength" of dependence/influence is defined in
* the following way: i depends on j if
* aij > hypre_max (k != i) aik, aii < 0
* or
* aij < hypre_min (k != i) aik, aii >= 0
* Then S_ij = 1, else S_ij = 0.
*
* NOTE: the entries are negative initially, corresponding
* to "unaccounted-for" dependence.
*----------------------------------------------------------------*/
if (debug_flag == 3) wall_time = time_getWallclockSeconds();
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
if (!comm_pkg)
{
use_commpkg_A = 1;
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
if (num_cols_offd) S_offd_j = hypre_CSRMatrixJ(S_offd);
jS = S_i[num_variables];
ST = hypre_CSRMatrixCreate(num_variables, num_variables, jS);
ST_i = hypre_CTAlloc(HYPRE_Int,num_variables+1);
ST_j = hypre_CTAlloc(HYPRE_Int,jS);
hypre_CSRMatrixI(ST) = ST_i;
hypre_CSRMatrixJ(ST) = ST_j;
/*----------------------------------------------------------
* generate transpose of S, ST
*----------------------------------------------------------*/
for (i=0; i <= num_variables; i++)
ST_i[i] = 0;
for (i=0; i < jS; i++)
{
ST_i[S_j[i]+1]++;
}
for (i=0; i < num_variables; i++)
{
ST_i[i+1] += ST_i[i];
}
for (i=0; i < num_variables; i++)
{
for (j=S_i[i]; j < S_i[i+1]; j++)
{
index = S_j[j];
ST_j[ST_i[index]] = i;
ST_i[index]++;
}
}
for (i = num_variables; i > 0; i--)
{
ST_i[i] = ST_i[i-1];
}
ST_i[0] = 0;
/*----------------------------------------------------------
* Compute the measures
*
* The measures are given by the row sums of ST.
* Hence, measure_array[i] is the number of influences
* of variable i.
* correct actual measures through adding influences from
* neighbor processors
*----------------------------------------------------------*/
measure_array_master = hypre_CTAlloc(HYPRE_Int, num_variables);
measure_array = hypre_CTAlloc(HYPRE_Int, num_variables);
for (i = 0; i < num_variables; i++)
{
measure_array_master[i] = ST_i[i+1]-ST_i[i];
}
if ((measure_type || (coarsen_type != 1 && coarsen_type != 11))
&& num_procs > 1)
{
if (use_commpkg_A)
S_ext = hypre_ParCSRMatrixExtractBExt(S,A,0);
else
S_ext = hypre_ParCSRMatrixExtractBExt(S,S,0);
S_ext_i = hypre_CSRMatrixI(S_ext);
S_ext_j = hypre_CSRMatrixJ(S_ext);
num_nonzeros = S_ext_i[num_cols_offd];
first_col = hypre_ParCSRMatrixFirstColDiag(S);
col_0 = first_col-1;
col_n = col_0+num_variables;
if (measure_type)
{
for (i=0; i < num_nonzeros; i++)
{
index = S_ext_j[i] - first_col;
if (index > -1 && index < num_variables)
measure_array_master[index]++;
}
}
}
/*---------------------------------------------------
* Loop until all points are either fine or coarse.
*---------------------------------------------------*/
if (debug_flag == 3) wall_time = time_getWallclockSeconds();
/* first coarsening phase */
/*************************************************************
*
* Initialize the lists
*
*************************************************************/
CF_marker = hypre_CTAlloc(HYPRE_Int, num_variables);
num_left = 0;
for (j = 0; j < num_variables; j++)
{
if ((S_i[j+1]-S_i[j])== 0 &&
(S_offd_i[j+1]-S_offd_i[j]) == 0)
{
CF_marker[j] = SF_PT;
measure_array_master[j] = 0;
}
else
{
CF_marker[j] = UNDECIDED;
/* num_left++; */ /* BM May 19, 2006: see below*/
}
}
if (coarsen_type==22) {
/* BM Sep 8, 2006: allow_emptygrids only if the following holds for all points j:
(a) the point has no strong connections at all, OR
(b) the point has a strong connection across a boundary */
for (j=0;j<num_variables;j++)
if (S_i[j+1]>S_i[j] && S_offd_i[j+1] == S_offd_i[j]) {coarsen_type=21;break;}
}
for (l = 1; l <= cgc_its; l++)
{
LoL_head = NULL;
LoL_tail = NULL;
num_left = 0; /* compute num_left before each RS coarsening loop */
memcpy (measure_array,measure_array_master,num_variables*sizeof(HYPRE_Int));
memset (lists,0,sizeof(HYPRE_Int)*num_variables);
memset (where,0,sizeof(HYPRE_Int)*num_variables);
for (j = 0; j < num_variables; j++)
{
measure = measure_array[j];
if (CF_marker[j] != SF_PT)
{
if (measure > 0)
{
enter_on_lists(&LoL_head, &LoL_tail, measure, j, lists, where);
num_left++; /* compute num_left before each RS coarsening loop */
}
else if (CF_marker[j] == 0) /* increase weight of strongly coupled neighbors only
if j is not conained in a previously constructed coarse grid.
Reason: these neighbors should start with the same initial weight
in each CGC iteration. BM Aug 30, 2006 */
{
if (measure < 0) hypre_printf("negative measure!\n");
/* CF_marker[j] = f_pnt; */
for (k = S_i[j]; k < S_i[j+1]; k++)
{
nabor = S_j[k];
/* if (CF_marker[nabor] != SF_PT) */
if (CF_marker[nabor] == 0) /* BM Aug 30, 2006: don't alter weights of points
contained in other candidate coarse grids */
{
if (nabor < j)
{
new_meas = measure_array[nabor];
if (new_meas > 0)
remove_point(&LoL_head, &LoL_tail, new_meas,
nabor, lists, where);
else num_left++; /* BM Aug 29, 2006 */
new_meas = ++(measure_array[nabor]);
enter_on_lists(&LoL_head, &LoL_tail, new_meas,
nabor, lists, where);
}
else
{
new_meas = ++(measure_array[nabor]);
}
}
}
/* --num_left; */ /* BM May 19, 2006 */
}
}
}
/* BM Aug 30, 2006: first iteration: determine maximal weight */
if (num_left && l==1) measure_max = measure_array[LoL_head->head];
/* BM Aug 30, 2006: break CGC iteration if no suitable
starting point is available any more */
if (!num_left || measure_array[LoL_head->head]<measure_max) {
while (LoL_head) {
hypre_LinkList list_ptr = LoL_head;
LoL_head = LoL_head->next_elt;
dispose_elt (list_ptr);
}
break;
}
/****************************************************************
*
* Main loop of Ruge-Stueben first coloring pass.
*
* WHILE there are still points to classify DO:
* 1) find first point, i, on list with max_measure
* make i a C-point, remove it from the lists
* 2) For each point, j, in S_i^T,
* a) Set j to be an F-point
* b) For each point, k, in S_j
* move k to the list in LoL with measure one
* greater than it occupies (creating new LoL
* entry if necessary)
* 3) For each point, j, in S_i,
* move j to the list in LoL with measure one
* smaller than it occupies (creating new LoL
* entry if necessary)
*
****************************************************************/
while (num_left > 0)
{
index = LoL_head -> head;
/* index = LoL_head -> tail; */
/* CF_marker[index] = C_PT; */
CF_marker[index] = l; /* BM Aug 18, 2006 */
measure = measure_array[index];
measure_array[index] = 0;
measure_array_master[index] = 0; /* BM May 19: for CGC */
--num_left;
remove_point(&LoL_head, &LoL_tail, measure, index, lists, where);
for (j = ST_i[index]; j < ST_i[index+1]; j++)
{
nabor = ST_j[j];
/* if (CF_marker[nabor] == UNDECIDED) */
if (measure_array[nabor]>0) /* undecided point */
{
/* CF_marker[nabor] = F_PT; */ /* BM Aug 18, 2006 */
measure = measure_array[nabor];
measure_array[nabor]=0;
remove_point(&LoL_head, &LoL_tail, measure, nabor, lists, where);
--num_left;
for (k = S_i[nabor]; k < S_i[nabor+1]; k++)
{
nabor_two = S_j[k];
/* if (CF_marker[nabor_two] == UNDECIDED) */
if (measure_array[nabor_two]>0) /* undecided point */
{
measure = measure_array[nabor_two];
remove_point(&LoL_head, &LoL_tail, measure,
nabor_two, lists, where);
new_meas = ++(measure_array[nabor_two]);
enter_on_lists(&LoL_head, &LoL_tail, new_meas,
nabor_two, lists, where);
}
}
}
}
for (j = S_i[index]; j < S_i[index+1]; j++)
{
nabor = S_j[j];
/* if (CF_marker[nabor] == UNDECIDED) */
if (measure_array[nabor]>0) /* undecided point */
{
measure = measure_array[nabor];
remove_point(&LoL_head, &LoL_tail, measure, nabor, lists, where);
measure_array[nabor] = --measure;
if (measure > 0)
enter_on_lists(&LoL_head, &LoL_tail, measure, nabor,
lists, where);
else
{
/* CF_marker[nabor] = F_PT; */ /* BM Aug 18, 2006 */
--num_left;
for (k = S_i[nabor]; k < S_i[nabor+1]; k++)
{
nabor_two = S_j[k];
/* if (CF_marker[nabor_two] == UNDECIDED) */
if (measure_array[nabor_two]>0)
{
new_meas = measure_array[nabor_two];
remove_point(&LoL_head, &LoL_tail, new_meas,
nabor_two, lists, where);
new_meas = ++(measure_array[nabor_two]);
enter_on_lists(&LoL_head, &LoL_tail, new_meas,
nabor_two, lists, where);
}
}
}
}
}
}
if (LoL_head) hypre_printf ("Linked list not empty! head: %d\n",LoL_head->head);
}
l--; /* BM Aug 15, 2006 */
hypre_TFree(measure_array);
hypre_TFree(measure_array_master);
hypre_CSRMatrixDestroy(ST);
if (debug_flag == 3)
{
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf("Proc = %d Coarsen 1st pass = %f\n",
my_id, wall_time);
}
hypre_TFree(lists);
hypre_TFree(where);
if (num_procs>1) {
if (debug_flag == 3) wall_time = time_getWallclockSeconds();
hypre_BoomerAMGCoarsenCGC (S,l,coarsen_type,CF_marker);
if (debug_flag == 3) {
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf("Proc = %d Coarsen CGC = %f\n",
my_id, wall_time);
}
}
else {
/* the first candiate coarse grid is the coarse grid */
for (j=0;j<num_variables;j++) {
if (CF_marker[j]==1) CF_marker[j]=C_PT;
else CF_marker[j]=F_PT;
}
}
/* BM May 19, 2006:
Set all undecided points to be fine grid points. */
for (j=0;j<num_variables;j++)
if (!CF_marker[j]) CF_marker[j]=F_PT;
/*---------------------------------------------------
* Initialize the graph array
*---------------------------------------------------*/
graph_array = hypre_CTAlloc(HYPRE_Int, num_variables);
for (i = 0; i < num_variables; i++)
{
graph_array[i] = -1;
}
if (debug_flag == 3) wall_time = time_getWallclockSeconds();
for (i=0; i < num_variables; i++)
{
if (ci_tilde_mark != i) ci_tilde = -1;
if (CF_marker[i] == -1)
{
for (ji = S_i[i]; ji < S_i[i+1]; ji++)
{
j = S_j[ji];
if (CF_marker[j] > 0)
graph_array[j] = i;
}
for (ji = S_i[i]; ji < S_i[i+1]; ji++)
{
j = S_j[ji];
if (CF_marker[j] == -1)
{
set_empty = 1;
for (jj = S_i[j]; jj < S_i[j+1]; jj++)
{
index = S_j[jj];
if (graph_array[index] == i)
{
set_empty = 0;
break;
}
}
if (set_empty)
{
if (C_i_nonempty)
{
CF_marker[i] = 1;
if (ci_tilde > -1)
{
CF_marker[ci_tilde] = -1;
ci_tilde = -1;
}
C_i_nonempty = 0;
break;
}
else
{
ci_tilde = j;
ci_tilde_mark = i;
CF_marker[j] = 1;
C_i_nonempty = 1;
i--;
break;
}
}
}
}
}
}
if (debug_flag == 3 && coarsen_type != 2)
{
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf("Proc = %d Coarsen 2nd pass = %f\n",
my_id, wall_time);
}
/* third pass, check boundary fine points for coarse neighbors */
/*------------------------------------------------
* Exchange boundary data for CF_marker
*------------------------------------------------*/
if (debug_flag == 3) wall_time = time_getWallclockSeconds();
CF_marker_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
int_buf_data = hypre_CTAlloc(HYPRE_Int, hypre_ParCSRCommPkgSendMapStart(comm_pkg,
num_sends));
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
int_buf_data[index++]
= CF_marker[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)];
}
if (num_procs > 1)
{
comm_handle = hypre_ParCSRCommHandleCreate(11, comm_pkg, int_buf_data,
CF_marker_offd);
hypre_ParCSRCommHandleDestroy(comm_handle);
}
AmgCGCBoundaryFix (S,CF_marker,CF_marker_offd);
if (debug_flag == 3)
{
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf("Proc = %d CGC boundary fix = %f\n",
my_id, wall_time);
}
/*---------------------------------------------------
* Clean up and return
*---------------------------------------------------*/
/*if (coarsen_type != 1)
{ */
if (CF_marker_offd) hypre_TFree(CF_marker_offd); /* BM Aug 21, 2006 */
if (int_buf_data) hypre_TFree(int_buf_data); /* BM Aug 21, 2006 */
/*if (ci_array) hypre_TFree(ci_array);*/ /* BM Aug 21, 2006 */
/*} */
hypre_TFree(graph_array);
if ((measure_type || (coarsen_type != 1 && coarsen_type != 11))
&& num_procs > 1)
hypre_CSRMatrixDestroy(S_ext);
*CF_marker_ptr = CF_marker;
return (ierr);
}
hypre_ParCSRMatrix *A,hypre_ParVector *f,
hypre_ParVector *u), hypre_ParCSRMatrix *A,
{
int *partition, mypid, start_row, end_row;
int row, row_length, *col_indices;
double *col_values;
Matrix *mat;
ParaSails *ps;
MLI_Smoother_ParaSails *smoother;
MPI_Comm comm;
/*-----------------------------------------------------------------
* fetch machine and matrix parameters
*-----------------------------------------------------------------*/
comm = hypre_ParCSRMatrixComm(A);
MPI_Comm_rank(comm,&mypid);
HYPRE_ParCSRMatrixGetRowPartitioning((HYPRE_ParCSRMatrix) A, &partition);
start_row = partition[mypid];
end_row = partition[mypid+1] - 1;
/*-----------------------------------------------------------------
* construct a ParaSails matrix
*-----------------------------------------------------------------*/
mat = MatrixCreate(comm, start_row, end_row);
for (row = start_row; row <= end_row; row++)
{
hypre_ParCSRMatrixGetRow(A, row, &row_length, &col_indices, &col_values);
MatrixSetRow(mat, row, row_length, col_indices, col_values);
hypre_ParCSRMatrixRestoreRow(A,row,&row_length,&col_indices,&col_values);
void hypre_BoomerAMGJacobiInterp_1( hypre_ParCSRMatrix * A,
hypre_ParCSRMatrix ** P,
hypre_ParCSRMatrix * S,
HYPRE_Int * CF_marker, HYPRE_Int level,
HYPRE_Real truncation_threshold,
HYPRE_Real truncation_threshold_minus,
HYPRE_Int * dof_func, HYPRE_Int * dof_func_offd,
HYPRE_Real weight_AF)
/* One step of Jacobi interpolation:
A is the linear system.
P is an interpolation matrix, input and output
CF_marker identifies coarse and fine points
If we imagine P and A as split into coarse and fine submatrices,
[ AFF AFC ] [ AF ] [ IFC ]
A = [ ] = [ ] , P = [ ]
[ ACF ACC ] [ AC ] [ ICC ]
(note that ICC is an identity matrix, applied to coarse points only)
then this function computes
IFCnew = IFCold - DFF(-1) * ( AFF*IFCold + AFC )
= IFCold - DFF(-1) * AF * Pold)
where DFF is the diagonal of AFF, (-1) represents the inverse, and
where "old" denotes a value on entry to this function, "new" a returned value.
*/
{
hypre_ParCSRMatrix * Pnew;
hypre_ParCSRMatrix * C;
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(*P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(*P);
HYPRE_Real *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
HYPRE_Real *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
hypre_CSRMatrix *C_diag;
hypre_CSRMatrix *C_offd;
hypre_CSRMatrix *Pnew_diag;
hypre_CSRMatrix *Pnew_offd;
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int i;
HYPRE_Int Jnochanges=0, Jchanges, Pnew_num_nonzeros;
HYPRE_Int CF_coarse=0;
HYPRE_Int * J_marker = hypre_CTAlloc( HYPRE_Int, num_rows_diag_P );
HYPRE_Int nc, ncmax, ncmin, nc1;
HYPRE_Int num_procs, my_id;
MPI_Comm comm = hypre_ParCSRMatrixComm( A );
#ifdef HYPRE_JACINT_PRINT_ROW_SUMS
HYPRE_Int m, nmav, npav;
HYPRE_Real PIi, PIimax, PIimin, PIimav, PIipav, randthresh;
HYPRE_Real eps = 1.0e-17;
#endif
#ifdef HYPRE_JACINT_PRINT_MATRICES
char filename[80];
HYPRE_Int i_dummy, j_dummy;
HYPRE_Int *base_i_ptr = &i_dummy;
HYPRE_Int *base_j_ptr = &j_dummy;
#endif
#ifdef HYPRE_JACINT_PRINT_SOME_ROWS
HYPRE_Int sample_rows[50], n_sample_rows=0, isamp;
#endif
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
for ( i=0; i<num_rows_diag_P; ++i )
{
J_marker[i] = CF_marker[i];
if (CF_marker[i]>=0) ++CF_coarse;
}
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i Jacobi_Interp_1, P has %i+%i=%i nonzeros, local sum %e\n", my_id, level,
hypre_CSRMatrixNumNonzeros(P_diag), hypre_CSRMatrixNumNonzeros(P_offd),
hypre_CSRMatrixNumNonzeros(P_diag)+hypre_CSRMatrixNumNonzeros(P_offd),
hypre_ParCSRMatrixLocalSumElts(*P) );
#endif
/* row sum computations, for output */
#ifdef HYPRE_JACINT_PRINT_ROW_SUMS
PIimax=-1.0e12, PIimin=1.0e12, PIimav=0, PIipav=0;
nmav=0, npav=0;
for ( i=0; i<num_rows_diag_P; ++i )
{
PIi = 0; /* i-th value of P*1, i.e. sum of row i of P */
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
PIi += P_diag_data[m];
for ( m=P_offd_i[i]; m<P_offd_i[i+1]; ++m )
PIi += P_offd_data[m];
if (CF_marker[i]<0)
{
PIimax = hypre_max( PIimax, PIi );
PIimin = hypre_min( PIimin, PIi );
if (PIi<=1-eps) { PIimav+=PIi; ++nmav; };
if (PIi>=1+eps) { PIipav+=PIi; ++npav; };
}
}
if ( nmav>0 ) PIimav = PIimav/nmav;
if ( npav>0 ) PIipav = PIipav/npav;
hypre_printf("%i %i P in max,min row sums %e %e\n", my_id, level, PIimax, PIimin );
#endif
ncmax=0; ncmin=num_rows_diag_P; nc1=0;
for ( i=0; i<num_rows_diag_P; ++i )
if (CF_marker[i]<0)
{
nc = P_diag_i[i+1] - P_diag_i[i];
if (nc<=1)
{
++nc1;
}
ncmax = hypre_max( nc, ncmax );
ncmin = hypre_min( nc, ncmin );
}
#if 0
/* a very agressive reduction in how much the Jacobi step does: */
for ( i=0; i<num_rows_diag_P; ++i )
if (CF_marker[i]<0)
{
nc = P_diag_i[i+1] - P_diag_i[i];
if (nc>ncmin+1)
/*if ( nc > ncmin + 0.5*(ncmax-ncmin) )*/
{
J_marker[i] = 1;
++Jnochanges;
}
}
#endif
Jchanges = num_rows_diag_P - Jnochanges - CF_coarse;
#ifdef HYPRE_JACINT_PRINT_SOME_ROWS
hypre_printf("some rows to be changed: ");
randthresh = 15/(HYPRE_Real)Jchanges;
for ( i=0; i<num_rows_diag_P; ++i )
{
if ( J_marker[i]<0 )
{
if ( ((HYPRE_Real)rand())/RAND_MAX < randthresh )
{
hypre_printf( "%i: ", i );
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
hypre_printf( " %i %f, ", P_diag_j[m], P_diag_data[m] );
hypre_printf("; ");
sample_rows[n_sample_rows] = i;
++n_sample_rows;
}
}
}
hypre_printf("\n");
#endif
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i P has %i rows, %i changeable, %i don't change-good, %i coarse\n",
my_id, level, num_rows_diag_P, Jchanges, Jnochanges, CF_coarse );
hypre_printf("%i %i min,max diag cols per row: %i, %i; no.rows w.<=1 col: %i\n", my_id, level, ncmin, ncmax, nc1 );
#endif
#ifdef HYPRE_JACINT_PRINT_MATRICES
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX )
{
hypre_sprintf( filename, "Ain%i", level );
hypre_ParCSRMatrixPrintIJ( A,0,0,filename);
hypre_sprintf( filename, "Sin%i", level );
hypre_ParCSRMatrixPrintIJ( S,0,0,filename);
hypre_sprintf( filename, "Pin%i", level );
hypre_ParCSRMatrixPrintIJ( *P,0,0,filename);
}
#endif
C = hypre_ParMatmul_FC( A, *P, J_marker, dof_func, dof_func_offd );
/* hypre_parMatmul_FC creates and returns C, a variation of the
matrix product A*P in which only the "Fine"-designated rows have
been computed. (all columns are Coarse because all columns of P
are). "Fine" is defined solely by the marker array, and for
example could be a proper subset of the fine points of a
multigrid hierarchy.
As a matrix, C is the size of A*P. But only the marked rows have
been computed.
*/
#ifdef HYPRE_JACINT_PRINT_MATRICES
hypre_sprintf( filename, "C%i", level );
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX ) hypre_ParCSRMatrixPrintIJ( C,0,0,filename);
#endif
C_diag = hypre_ParCSRMatrixDiag(C);
C_offd = hypre_ParCSRMatrixOffd(C);
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i Jacobi_Interp_1 after matmul, C has %i+%i=%i nonzeros, local sum %e\n",
my_id, level, hypre_CSRMatrixNumNonzeros(C_diag),
hypre_CSRMatrixNumNonzeros(C_offd),
hypre_CSRMatrixNumNonzeros(C_diag)+hypre_CSRMatrixNumNonzeros(C_offd),
hypre_ParCSRMatrixLocalSumElts(C) );
#endif
hypre_ParMatScaleDiagInv_F( C, A, weight_AF, J_marker );
/* hypre_ParMatScaleDiagInv scales of its first argument by premultiplying with
a submatrix of the inverse of the diagonal of its second argument.
The marker array determines which diagonal elements are used. The marker
array should select exactly the right number of diagonal elements (the number
of rows of AP_FC).
*/
#ifdef HYPRE_JACINT_PRINT_MATRICES
hypre_sprintf( filename, "Cout%i", level );
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX ) hypre_ParCSRMatrixPrintIJ( C,0,0,filename);
#endif
Pnew = hypre_ParMatMinus_F( *P, C, J_marker );
/* hypre_ParMatMinus_F subtracts rows of its second argument from selected rows
of its first argument. The marker array determines which rows of the first
argument are affected, and they should exactly correspond to all the rows
of the second argument.
*/
Pnew_diag = hypre_ParCSRMatrixDiag(Pnew);
Pnew_offd = hypre_ParCSRMatrixOffd(Pnew);
Pnew_num_nonzeros = hypre_CSRMatrixNumNonzeros(Pnew_diag)+hypre_CSRMatrixNumNonzeros(Pnew_offd);
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i Jacobi_Interp_1 after MatMinus, Pnew has %i+%i=%i nonzeros, local sum %e\n",
my_id, level, hypre_CSRMatrixNumNonzeros(Pnew_diag),
hypre_CSRMatrixNumNonzeros(Pnew_offd), Pnew_num_nonzeros,
hypre_ParCSRMatrixLocalSumElts(Pnew) );
#endif
/* Transfer ownership of col_starts from P to Pnew ... */
if ( hypre_ParCSRMatrixColStarts(*P) &&
hypre_ParCSRMatrixColStarts(*P)==hypre_ParCSRMatrixColStarts(Pnew) )
{
if ( hypre_ParCSRMatrixOwnsColStarts(*P) && !hypre_ParCSRMatrixOwnsColStarts(Pnew) )
{
hypre_ParCSRMatrixSetColStartsOwner(*P,0);
hypre_ParCSRMatrixSetColStartsOwner(Pnew,1);
}
}
hypre_ParCSRMatrixDestroy( C );
hypre_ParCSRMatrixDestroy( *P );
/* Note that I'm truncating all the fine rows, not just the J-marked ones. */
#if 0
if ( Pnew_num_nonzeros < 10000 ) /* a fixed number like this makes it no.procs.-depdendent */
{ /* ad-hoc attempt to reduce zero-matrix problems seen in testing..*/
truncation_threshold = 1.0e-6 * truncation_threshold;
truncation_threshold_minus = 1.0e-6 * truncation_threshold_minus;
}
#endif
hypre_BoomerAMGTruncateInterp( Pnew, truncation_threshold,
truncation_threshold_minus, CF_marker );
hypre_MatvecCommPkgCreate ( Pnew );
*P = Pnew;
P_diag = hypre_ParCSRMatrixDiag(*P);
P_offd = hypre_ParCSRMatrixOffd(*P);
P_diag_data = hypre_CSRMatrixData(P_diag);
P_diag_i = hypre_CSRMatrixI(P_diag);
P_diag_j = hypre_CSRMatrixJ(P_diag);
P_offd_data = hypre_CSRMatrixData(P_offd);
P_offd_i = hypre_CSRMatrixI(P_offd);
/* row sum computations, for output */
#ifdef HYPRE_JACINT_PRINT_ROW_SUMS
PIimax=-1.0e12, PIimin=1.0e12, PIimav=0, PIipav=0;
nmav=0, npav=0;
for ( i=0; i<num_rows_diag_P; ++i )
{
PIi = 0; /* i-th value of P*1, i.e. sum of row i of P */
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
PIi += P_diag_data[m];
for ( m=P_offd_i[i]; m<P_offd_i[i+1]; ++m )
PIi += P_offd_data[m];
if (CF_marker[i]<0)
{
PIimax = hypre_max( PIimax, PIi );
PIimin = hypre_min( PIimin, PIi );
if (PIi<=1-eps) { PIimav+=PIi; ++nmav; };
if (PIi>=1+eps) { PIipav+=PIi; ++npav; };
}
}
if ( nmav>0 ) PIimav = PIimav/nmav;
if ( npav>0 ) PIipav = PIipav/npav;
hypre_printf("%i %i P out max,min row sums %e %e\n", my_id, level, PIimax, PIimin );
#endif
#ifdef HYPRE_JACINT_PRINT_SOME_ROWS
hypre_printf("some changed rows: ");
for ( isamp=0; isamp<n_sample_rows; ++isamp )
{
i = sample_rows[isamp];
hypre_printf( "%i: ", i );
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
hypre_printf( " %i %f, ", P_diag_j[m], P_diag_data[m] );
hypre_printf("; ");
}
hypre_printf("\n");
#endif
ncmax=0; ncmin=num_rows_diag_P; nc1=0;
for ( i=0; i<num_rows_diag_P; ++i )
if (CF_marker[i]<0)
{
nc = P_diag_i[i+1] - P_diag_i[i];
if (nc<=1) ++nc1;
ncmax = hypre_max( nc, ncmax );
ncmin = hypre_min( nc, ncmin );
}
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i P has %i rows, %i changeable, %i too good, %i coarse\n",
my_id, level, num_rows_diag_P, num_rows_diag_P-Jnochanges-CF_coarse, Jnochanges, CF_coarse );
hypre_printf("%i %i min,max diag cols per row: %i, %i; no.rows w.<=1 col: %i\n", my_id, level, ncmin, ncmax, nc1 );
hypre_printf("%i %i Jacobi_Interp_1 after truncation (%e), Pnew has %i+%i=%i nonzeros, local sum %e\n",
my_id, level, truncation_threshold,
hypre_CSRMatrixNumNonzeros(Pnew_diag), hypre_CSRMatrixNumNonzeros(Pnew_offd),
hypre_CSRMatrixNumNonzeros(Pnew_diag)+hypre_CSRMatrixNumNonzeros(Pnew_offd),
hypre_ParCSRMatrixLocalSumElts(Pnew) );
#endif
/* Programming Notes:
1. Judging by around line 299 of par_interp.c, they typical use of CF_marker
is that CF_marker>=0 means Coarse, CF_marker<0 means Fine.
*/
#ifdef HYPRE_JACINT_PRINT_MATRICES
hypre_sprintf( filename, "Pout%i", level );
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX ) hypre_ParCSRMatrixPrintIJ( *P,0,0,filename);
#endif
hypre_TFree( J_marker );
}
/*--------------------------------------------------------------------------
* hypre_SStructSharedDOF_ParcsrMatRowsComm
* Given a sstruct_grid & parcsr matrix with rows corresponding to the
* sstruct_grid, determine and extract the rows that must be communicated.
* These rows are for shared dof that geometrically lie on processor
* boundaries but internally are stored on one processor.
* Algo:
* for each cellbox
* RECVs:
* i) stretch the cellbox to the variable box
* ii) in the appropriate (dof-dependent) direction, take the
* boundary and boxman_intersect to extract boxmanentries
* that contain these boundary edges.
* iii)loop over the boxmanentries and see if they belong
* on this proc or another proc
* a) if belong on another proc, these are the recvs:
* count and prepare the communication buffers and
* values.
*
* SENDs:
* i) form layer of cells that is one layer off cellbox
* (stretches in the appropriate direction)
* ii) boxman_intersect with the cellgrid boxman
* iii)loop over the boxmanentries and see if they belong
* on this proc or another proc
* a) if belong on another proc, these are the sends:
* count and prepare the communication buffers and
* values.
*
* Note: For the recv data, the dof can come from only one processor.
* For the send data, the dof can go to more than one processor
* (the same dof is on the boundary of several cells).
*--------------------------------------------------------------------------*/
HYPRE_Int
hypre_SStructSharedDOF_ParcsrMatRowsComm( hypre_SStructGrid *grid,
hypre_ParCSRMatrix *A,
HYPRE_Int *num_offprocrows_ptr,
hypre_MaxwellOffProcRow ***OffProcRows_ptr)
{
MPI_Comm A_comm= hypre_ParCSRMatrixComm(A);
MPI_Comm grid_comm= hypre_SStructGridComm(grid);
HYPRE_Int matrix_type= HYPRE_PARCSR;
HYPRE_Int nparts= hypre_SStructGridNParts(grid);
HYPRE_Int ndim = hypre_SStructGridNDim(grid);
hypre_SStructGrid *cell_ssgrid;
hypre_SStructPGrid *pgrid;
hypre_StructGrid *cellgrid;
hypre_BoxArray *cellboxes;
hypre_Box *box, *cellbox, vbox, boxman_entry_box;
hypre_Index loop_size, start;
HYPRE_Int loopi, loopj, loopk;
HYPRE_Int start_rank, end_rank, rank;
HYPRE_Int i, j, k, m, n, t, part, var, nvars;
HYPRE_SStructVariable *vartypes;
HYPRE_Int nbdry_slabs;
hypre_BoxArray *recv_slabs, *send_slabs;
hypre_Index varoffset;
hypre_BoxManager **boxmans, *cell_boxman;
hypre_BoxManEntry **boxman_entries, *entry;
HYPRE_Int nboxman_entries;
hypre_Index ishift, jshift, kshift, zero_index;
hypre_Index ilower, iupper, index;
HYPRE_Int proc, nprocs, myproc;
HYPRE_Int *SendToProcs, *RecvFromProcs;
HYPRE_Int **send_RowsNcols; /* buffer for rows & ncols */
HYPRE_Int *send_RowsNcols_alloc;
HYPRE_Int *send_ColsData_alloc;
HYPRE_Int *tot_nsendRowsNcols, *tot_sendColsData;
double **vals; /* buffer for cols & data */
HYPRE_Int *col_inds;
double *values;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
HYPRE_Int **rbuffer_RowsNcols;
double **rbuffer_ColsData;
HYPRE_Int num_sends, num_recvs;
hypre_MaxwellOffProcRow **OffProcRows;
HYPRE_Int *starts;
HYPRE_Int ierr= 0;
hypre_MPI_Comm_rank(A_comm, &myproc);
hypre_MPI_Comm_size(grid_comm, &nprocs);
start_rank= hypre_ParCSRMatrixFirstRowIndex(A);
end_rank = hypre_ParCSRMatrixLastRowIndex(A);
hypre_SetIndex(ishift, 1, 0, 0);
hypre_SetIndex(jshift, 0, 1, 0);
hypre_SetIndex(kshift, 0, 0, 1);
hypre_SetIndex(zero_index, 0, 0, 0);
/* need a cellgrid boxman to determine the send boxes -> only the cell dofs
are unique so a boxman intersect can be used to get the edges that
must be sent. */
HYPRE_SStructGridCreate(grid_comm, ndim, nparts, &cell_ssgrid);
vartypes= hypre_CTAlloc(HYPRE_SStructVariable, 1);
vartypes[0]= HYPRE_SSTRUCT_VARIABLE_CELL;
for (i= 0; i< nparts; i++)
{
pgrid= hypre_SStructGridPGrid(grid, i);
cellgrid= hypre_SStructPGridCellSGrid(pgrid);
cellboxes= hypre_StructGridBoxes(cellgrid);
hypre_ForBoxI(j, cellboxes)
{
box= hypre_BoxArrayBox(cellboxes, j);
HYPRE_SStructGridSetExtents(cell_ssgrid, i,
hypre_BoxIMin(box), hypre_BoxIMax(box));
}
HYPRE_SStructGridSetVariables(cell_ssgrid, i, 1, vartypes);
}
void hypre_BoomerAMGTruncateInterp( hypre_ParCSRMatrix *P,
HYPRE_Real eps, HYPRE_Real dlt,
HYPRE_Int * CF_marker )
/* Truncate the interpolation matrix P, but only in rows for which the
marker is <0. Truncation means that an element P(i,j) is set to 0 if
P(i,j)>0 and P(i,j)<eps*max( P(i,j) ) or if
P(i,j)>0 and P(i,j)<dlt*max( -P(i,j) ) or if
P(i,j)<0 and P(i,j)>dlt*min( -P(i,j) ) or if
P(i,j)<0 and P(i,j)>eps*min( P(i,j) )
( 0<eps,dlt<1, typically 0.1=dlt<eps=0.2, )
The min and max are only computed locally, as I'm guessing that there isn't
usually much to be gained (in the way of improved performance) by getting
them perfectly right.
*/
/* The function hypre_BoomerAMGInterpTruncation in par_interp.c is
very similar. It looks at fabs(value) rather than separately
dealing with value<0 and value>0 as recommended by Klaus Stuben,
thus as this function does. In this function, only "marked" rows
are affected. Lastly, in hypre_BoomerAMGInterpTruncation, if any
element gets discarded, it reallocates arrays to the new size.
*/
{
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
HYPRE_Real *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
HYPRE_Real *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int *new_P_diag_i;
HYPRE_Int *new_P_offd_i;
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int num_rows_offd_P = hypre_CSRMatrixNumRows(P_offd);
HYPRE_Int num_nonzeros_diag = hypre_CSRMatrixNumNonzeros(P_diag);
HYPRE_Int num_nonzeros_offd = hypre_CSRMatrixNumNonzeros(P_offd);
#if 0
MPI_Comm comm = hypre_ParCSRMatrixComm( P );
HYPRE_Real vmax1, vmin1;
#endif
HYPRE_Real vmax = 0.0;
HYPRE_Real vmin = 0.0;
HYPRE_Real v, old_sum, new_sum, scale, wmax, wmin;
HYPRE_Int i1, m, m1d, m1o;
/* compute vmax = eps*max(P(i,j)), vmin = eps*min(P(i,j)) */
for ( i1 = 0; i1 < num_rows_diag_P; i1++ )
{
for ( m=P_diag_i[i1]; m<P_diag_i[i1+1]; ++m )
{
v = P_diag_data[m];
vmax = hypre_max( v, vmax );
vmin = hypre_min( v, vmin );
}
for ( m=P_offd_i[i1]; m<P_offd_i[i1+1]; ++m )
{
v = P_offd_data[m];
vmax = hypre_max( v, vmax );
vmin = hypre_min( v, vmin );
}
}
#if 0
/* This can make max,min global so results don't depend on no. processors
We don't want this except for testing, or maybe this could be put
someplace better. I don't like adding communication here, for a minor reason.
*/
vmax1 = vmax; vmin1 = vmin;
hypre_MPI_Allreduce( &vmax1, &vmax, 1, HYPRE_MPI_REAL, hypre_MPI_MAX, comm );
hypre_MPI_Allreduce( &vmin1, &vmin, 1, HYPRE_MPI_REAL, hypre_MPI_MIN, comm );
#endif
if ( vmax <= 0.0 ) vmax = 1.0; /* make sure no v is v>vmax if no v is v>0 */
if ( vmin >= 0.0 ) vmin = -1.0; /* make sure no v is v<vmin if no v is v<0 */
wmax = - dlt * vmin;
wmin = - dlt * vmax;
vmax *= eps;
vmin *= eps;
/* Repack the i,j,and data arrays so as to discard the small elements of P.
Elements of Coarse rows (CF_marker>=0) are always kept.
The arrays are not re-allocated, so there will generally be unused space
at the ends of the arrays. */
new_P_diag_i = hypre_CTAlloc( HYPRE_Int, num_rows_diag_P+1 );
new_P_offd_i = hypre_CTAlloc( HYPRE_Int, num_rows_offd_P+1 );
m1d = P_diag_i[0];
m1o = P_offd_i[0];
for ( i1 = 0; i1 < num_rows_diag_P; i1++ )
{
old_sum = 0;
new_sum = 0;
for ( m=P_diag_i[i1]; m<P_diag_i[i1+1]; ++m )
{
v = P_diag_data[m];
old_sum += v;
if ( CF_marker[i1]>=0 || ( v>=vmax && v>=wmax ) || ( v<=vmin && v<=wmin ) )
{ /* keep v */
new_sum += v;
P_diag_j[m1d] = P_diag_j[m];
P_diag_data[m1d] = P_diag_data[m];
++m1d;
}
else
{ /* discard v */
--num_nonzeros_diag;
}
}
for ( m=P_offd_i[i1]; m<P_offd_i[i1+1]; ++m )
{
v = P_offd_data[m];
old_sum += v;
if ( CF_marker[i1]>=0 || ( v>=vmax && v>=wmax ) || ( v<=vmin && v<=wmin ) )
{ /* keep v */
new_sum += v;
P_offd_j[m1o] = P_offd_j[m];
P_offd_data[m1o] = P_offd_data[m];
++m1o;
}
else
{ /* discard v */
--num_nonzeros_offd;
}
}
new_P_diag_i[i1+1] = m1d;
if ( i1<num_rows_offd_P ) new_P_offd_i[i1+1] = m1o;
/* rescale to keep row sum the same */
if (new_sum!=0) scale = old_sum/new_sum; else scale = 1.0;
for ( m=new_P_diag_i[i1]; m<new_P_diag_i[i1+1]; ++m )
P_diag_data[m] *= scale;
if ( i1<num_rows_offd_P ) /* this test fails when there is no offd block */
for ( m=new_P_offd_i[i1]; m<new_P_offd_i[i1+1]; ++m )
P_offd_data[m] *= scale;
}
for ( i1 = 1; i1 <= num_rows_diag_P; i1++ )
{
P_diag_i[i1] = new_P_diag_i[i1];
if ( i1<=num_rows_offd_P && num_nonzeros_offd>0 ) P_offd_i[i1] = new_P_offd_i[i1];
}
hypre_TFree( new_P_diag_i );
if ( num_rows_offd_P>0 ) hypre_TFree( new_P_offd_i );
hypre_CSRMatrixNumNonzeros(P_diag) = num_nonzeros_diag;
hypre_CSRMatrixNumNonzeros(P_offd) = num_nonzeros_offd;
hypre_ParCSRMatrixSetDNumNonzeros( P );
hypre_ParCSRMatrixSetNumNonzeros( P );
}
hypre_ParCSRBlockMatrix *
hypre_ParCSRBlockMatrixConvertFromParCSRMatrix(hypre_ParCSRMatrix *matrix,
HYPRE_Int matrix_C_block_size )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(matrix);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(matrix);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(matrix);
HYPRE_Int global_num_rows = hypre_ParCSRMatrixGlobalNumRows(matrix);
HYPRE_Int global_num_cols = hypre_ParCSRMatrixGlobalNumCols(matrix);
HYPRE_Int *row_starts = hypre_ParCSRMatrixRowStarts(matrix);
HYPRE_Int *col_starts = hypre_ParCSRMatrixColStarts(matrix);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int *col_map_offd = hypre_ParCSRBlockMatrixColMapOffd(matrix);
HYPRE_Int *map_to_node=NULL, *counter=NULL, *col_in_j_map=NULL;
HYPRE_Int *matrix_C_col_map_offd = NULL;
HYPRE_Int matrix_C_num_cols_offd;
HYPRE_Int matrix_C_num_nonzeros_offd;
HYPRE_Int num_rows, num_nodes;
HYPRE_Int *offd_i = hypre_CSRMatrixI(offd);
HYPRE_Int *offd_j = hypre_CSRMatrixJ(offd);
HYPRE_Complex * offd_data = hypre_CSRMatrixData(offd);
hypre_ParCSRBlockMatrix *matrix_C;
HYPRE_Int *matrix_C_row_starts;
HYPRE_Int *matrix_C_col_starts;
hypre_CSRBlockMatrix *matrix_C_diag;
hypre_CSRBlockMatrix *matrix_C_offd;
HYPRE_Int *matrix_C_offd_i=NULL, *matrix_C_offd_j = NULL;
HYPRE_Complex *matrix_C_offd_data = NULL;
HYPRE_Int num_procs, i, j, k, k_map, count, index, start_index, pos, row;
hypre_MPI_Comm_size(comm,&num_procs);
#ifdef HYPRE_NO_GLOBAL_PARTITION
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, 2);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, 2);
for(i = 0; i < 2; i++)
{
matrix_C_row_starts[i] = row_starts[i]/matrix_C_block_size;
matrix_C_col_starts[i] = col_starts[i]/matrix_C_block_size;
}
#else
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
for(i = 0; i < num_procs + 1; i++)
{
matrix_C_row_starts[i] = row_starts[i]/matrix_C_block_size;
matrix_C_col_starts[i] = col_starts[i]/matrix_C_block_size;
}
#endif
/************* create the diagonal part ************/
matrix_C_diag = hypre_CSRBlockMatrixConvertFromCSRMatrix(diag,
matrix_C_block_size);
/******* the offd part *******************/
/* can't use the same function for the offd part - because this isn't square
and the offd j entries aren't global numbering (have to consider the offd
map) - need to look at col_map_offd first */
/* figure out the new number of offd columns (num rows is same as diag) */
num_cols_offd = hypre_CSRMatrixNumCols(offd);
num_rows = hypre_CSRMatrixNumRows(diag);
num_nodes = num_rows/matrix_C_block_size;
matrix_C_offd_i = hypre_CTAlloc(HYPRE_Int, num_nodes + 1);
matrix_C_num_cols_offd = 0;
matrix_C_offd_i[0] = 0;
matrix_C_num_nonzeros_offd = 0;
if (num_cols_offd)
{
map_to_node = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
matrix_C_num_cols_offd = 1;
map_to_node[0] = col_map_offd[0]/matrix_C_block_size;
for (i=1; i < num_cols_offd; i++)
{
map_to_node[i] = col_map_offd[i]/matrix_C_block_size;
if (map_to_node[i] > map_to_node[i-1]) matrix_C_num_cols_offd++;
}
matrix_C_col_map_offd = hypre_CTAlloc(HYPRE_Int, matrix_C_num_cols_offd);
col_in_j_map = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
matrix_C_col_map_offd[0] = map_to_node[0];
col_in_j_map[0] = 0;
count = 1;
j = 1;
/* fill in the col_map_off_d - these are global numbers. Then we need to
map these to j entries (these have local numbers) */
for (i=1; i < num_cols_offd; i++)
{
if (map_to_node[i] > map_to_node[i-1])
{
matrix_C_col_map_offd[count++] = map_to_node[i];
}
col_in_j_map[j++] = count - 1;
}
/* now figure the nonzeros */
matrix_C_num_nonzeros_offd = 0;
counter = hypre_CTAlloc(HYPRE_Int, matrix_C_num_cols_offd);
for (i=0; i < matrix_C_num_cols_offd; i++)
counter[i] = -1;
for (i=0; i < num_nodes; i++) /* for each block row */
{
matrix_C_offd_i[i] = matrix_C_num_nonzeros_offd;
for (j=0; j < matrix_C_block_size; j++)
{
row = i*matrix_C_block_size+j;
for (k=offd_i[row]; k < offd_i[row+1]; k++) /* go through single row */
{
k_map = col_in_j_map[offd_j[k]]; /*nodal col - see if this has
been in this block row (i)
already*/
if (counter[k_map] < i) /* not yet counted for this nodal row */
{
counter[k_map] = i;
matrix_C_num_nonzeros_offd++;
}
}
}
}
/* fill in final i entry */
matrix_C_offd_i[num_nodes] = matrix_C_num_nonzeros_offd;
}
/* create offd matrix */
matrix_C_offd = hypre_CSRBlockMatrixCreate(matrix_C_block_size, num_nodes,
matrix_C_num_cols_offd,
matrix_C_num_nonzeros_offd);
/* assign i */
hypre_CSRBlockMatrixI(matrix_C_offd) = matrix_C_offd_i;
/* create (and allocate j and data) */
if (matrix_C_num_nonzeros_offd)
{
matrix_C_offd_j = hypre_CTAlloc(HYPRE_Int, matrix_C_num_nonzeros_offd);
matrix_C_offd_data =
hypre_CTAlloc(HYPRE_Complex,
matrix_C_num_nonzeros_offd*matrix_C_block_size*
matrix_C_block_size);
hypre_CSRBlockMatrixJ(matrix_C_offd) = matrix_C_offd_j;
hypre_CSRMatrixData(matrix_C_offd) = matrix_C_offd_data;
for (i=0; i < matrix_C_num_cols_offd; i++)
counter[i] = -1;
index = 0; /*keep track of entry in matrix_C_offd_j*/
start_index = 0;
for (i=0; i < num_nodes; i++) /* for each block row */
{
for (j=0; j < matrix_C_block_size; j++) /* for each row in block */
{
row = i*matrix_C_block_size+j;
for (k=offd_i[row]; k < offd_i[row+1]; k++) /* go through single row's cols */
{
k_map = col_in_j_map[offd_j[k]]; /*nodal col for off_d */
if (counter[k_map] < start_index) /* not yet counted for this nodal row */
{
counter[k_map] = index;
matrix_C_offd_j[index] = k_map;
/*copy the data: which position (corresponds to j array) + which row + which col */
pos = (index * matrix_C_block_size * matrix_C_block_size) + (j * matrix_C_block_size) +
col_map_offd[offd_j[k]]%matrix_C_block_size;
matrix_C_offd_data[pos] = offd_data[k];
index ++;
}
else /* this col has already been listed for this row */
{
/*copy the data: which position (corresponds to j array) + which row + which col */
pos = (counter[k_map]* matrix_C_block_size * matrix_C_block_size) + (j * matrix_C_block_size) +
col_map_offd[offd_j[k]]%matrix_C_block_size;
matrix_C_offd_data[pos] = offd_data[k];
}
}
}
start_index = index; /* first index for current nodal row */
}
}
/* *********create the new matrix *************/
matrix_C = hypre_ParCSRBlockMatrixCreate(comm, matrix_C_block_size,
global_num_rows/matrix_C_block_size,
global_num_cols/matrix_C_block_size, matrix_C_row_starts,
matrix_C_col_starts, matrix_C_num_cols_offd,
hypre_CSRBlockMatrixNumNonzeros(matrix_C_diag),
matrix_C_num_nonzeros_offd);
/* use the diag and off diag matrices we have already created */
hypre_CSRBlockMatrixDestroy(hypre_ParCSRMatrixDiag(matrix_C));
hypre_ParCSRBlockMatrixDiag(matrix_C) = matrix_C_diag;
hypre_CSRBlockMatrixDestroy(hypre_ParCSRMatrixOffd(matrix_C));
hypre_ParCSRBlockMatrixOffd(matrix_C) = matrix_C_offd;
hypre_ParCSRMatrixColMapOffd(matrix_C) = matrix_C_col_map_offd;
/* *********don't bother to copy the comm_pkg *************/
hypre_ParCSRBlockMatrixCommPkg(matrix_C) = NULL;
/* CLEAN UP !!!! */
hypre_TFree(map_to_node);
hypre_TFree(col_in_j_map);
hypre_TFree(counter);
return matrix_C;
}
HYPRE_Int hypre_AMESetup(void *esolver)
{
HYPRE_Int ne, *edge_bc;
hypre_AMEData *ame_data = esolver;
hypre_AMSData *ams_data = ame_data -> precond;
if (ams_data -> beta_is_zero)
{
ame_data -> t1 = hypre_ParVectorInDomainOf(ams_data -> G);
ame_data -> t2 = hypre_ParVectorInDomainOf(ams_data -> G);
}
else
{
ame_data -> t1 = ams_data -> r1;
ame_data -> t2 = ams_data -> g1;
}
ame_data -> t3 = ams_data -> r0;
/* Eliminate boundary conditions in G = [Gii, Gib; 0, Gbb], i.e.,
compute [Gii, 0; 0, 0] */
{
HYPRE_Int i, j, k, nv;
HYPRE_Int *offd_edge_bc;
hypre_ParCSRMatrix *Gt;
nv = hypre_ParCSRMatrixNumCols(ams_data -> G);
ne = hypre_ParCSRMatrixNumRows(ams_data -> G);
edge_bc = hypre_TAlloc(HYPRE_Int, ne);
for (i = 0; i < ne; i++)
edge_bc[i] = 0;
/* Find boundary (eliminated) edges */
{
hypre_CSRMatrix *Ad = hypre_ParCSRMatrixDiag(ams_data -> A);
HYPRE_Int *AdI = hypre_CSRMatrixI(Ad);
HYPRE_Int *AdJ = hypre_CSRMatrixJ(Ad);
HYPRE_Real *AdA = hypre_CSRMatrixData(Ad);
hypre_CSRMatrix *Ao = hypre_ParCSRMatrixOffd(ams_data -> A);
HYPRE_Int *AoI = hypre_CSRMatrixI(Ao);
HYPRE_Real *AoA = hypre_CSRMatrixData(Ao);
HYPRE_Real l1_norm;
/* A row (edge) is boundary if its off-diag l1 norm is less than eps */
HYPRE_Real eps = DBL_EPSILON * 1e+4;
for (i = 0; i < ne; i++)
{
l1_norm = 0.0;
for (j = AdI[i]; j < AdI[i+1]; j++)
if (AdJ[j] != i)
l1_norm += fabs(AdA[j]);
if (AoI)
for (j = AoI[i]; j < AoI[i+1]; j++)
l1_norm += fabs(AoA[j]);
if (l1_norm < eps)
edge_bc[i] = 1;
}
}
hypre_ParCSRMatrixTranspose(ams_data -> G, &Gt, 1);
/* Use a Matvec communication to find which of the edges
connected to local vertices are on the boundary */
{
hypre_ParCSRCommHandle *comm_handle;
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int num_sends, *int_buf_data;
HYPRE_Int index, start;
offd_edge_bc = hypre_CTAlloc(HYPRE_Int, hypre_CSRMatrixNumCols(hypre_ParCSRMatrixOffd(Gt)));
hypre_MatvecCommPkgCreate(Gt);
comm_pkg = hypre_ParCSRMatrixCommPkg(Gt);
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
int_buf_data = hypre_CTAlloc(HYPRE_Int,
hypre_ParCSRCommPkgSendMapStart(comm_pkg,
num_sends));
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
{
k = hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j);
int_buf_data[index++] = edge_bc[k];
}
}
comm_handle = hypre_ParCSRCommHandleCreate(11, comm_pkg,
int_buf_data, offd_edge_bc);
hypre_ParCSRCommHandleDestroy(comm_handle);
hypre_TFree(int_buf_data);
}
/* Eliminate boundary vertex entries in G^t */
{
hypre_CSRMatrix *Gtd = hypre_ParCSRMatrixDiag(Gt);
HYPRE_Int *GtdI = hypre_CSRMatrixI(Gtd);
HYPRE_Int *GtdJ = hypre_CSRMatrixJ(Gtd);
HYPRE_Real *GtdA = hypre_CSRMatrixData(Gtd);
hypre_CSRMatrix *Gto = hypre_ParCSRMatrixOffd(Gt);
HYPRE_Int *GtoI = hypre_CSRMatrixI(Gto);
HYPRE_Int *GtoJ = hypre_CSRMatrixJ(Gto);
HYPRE_Real *GtoA = hypre_CSRMatrixData(Gto);
HYPRE_Int bdr;
for (i = 0; i < nv; i++)
{
bdr = 0;
/* A vertex is boundary if it belongs to a boundary edge */
for (j = GtdI[i]; j < GtdI[i+1]; j++)
if (edge_bc[GtdJ[j]]) { bdr = 1; break; }
if (!bdr && GtoI)
for (j = GtoI[i]; j < GtoI[i+1]; j++)
if (offd_edge_bc[GtoJ[j]]) { bdr = 1; break; }
if (bdr)
{
for (j = GtdI[i]; j < GtdI[i+1]; j++)
/* if (!edge_bc[GtdJ[j]]) */
GtdA[j] = 0.0;
if (GtoI)
for (j = GtoI[i]; j < GtoI[i+1]; j++)
/* if (!offd_edge_bc[GtoJ[j]]) */
GtoA[j] = 0.0;
}
}
}
hypre_ParCSRMatrixTranspose(Gt, &ame_data -> G, 1);
hypre_ParCSRMatrixDestroy(Gt);
hypre_TFree(offd_edge_bc);
}
/* Compute G^t M G */
{
if (!hypre_ParCSRMatrixCommPkg(ame_data -> G))
hypre_MatvecCommPkgCreate(ame_data -> G);
if (!hypre_ParCSRMatrixCommPkg(ame_data -> M))
hypre_MatvecCommPkgCreate(ame_data -> M);
hypre_BoomerAMGBuildCoarseOperator(ame_data -> G,
ame_data -> M,
ame_data -> G,
&ame_data -> A_G);
hypre_ParCSRMatrixFixZeroRows(ame_data -> A_G);
}
/* Create AMG preconditioner and PCG-AMG solver for G^tMG */
{
HYPRE_BoomerAMGCreate(&ame_data -> B1_G);
HYPRE_BoomerAMGSetCoarsenType(ame_data -> B1_G, ams_data -> B_G_coarsen_type);
HYPRE_BoomerAMGSetAggNumLevels(ame_data -> B1_G, ams_data -> B_G_agg_levels);
HYPRE_BoomerAMGSetRelaxType(ame_data -> B1_G, ams_data -> B_G_relax_type);
HYPRE_BoomerAMGSetNumSweeps(ame_data -> B1_G, 1);
HYPRE_BoomerAMGSetMaxLevels(ame_data -> B1_G, 25);
HYPRE_BoomerAMGSetTol(ame_data -> B1_G, 0.0);
HYPRE_BoomerAMGSetMaxIter(ame_data -> B1_G, 1);
HYPRE_BoomerAMGSetStrongThreshold(ame_data -> B1_G, ams_data -> B_G_theta);
/* don't use exact solve on the coarsest level (matrix may be singular) */
HYPRE_BoomerAMGSetCycleRelaxType(ame_data -> B1_G,
ams_data -> B_G_relax_type,
3);
HYPRE_ParCSRPCGCreate(hypre_ParCSRMatrixComm(ame_data->A_G),
&ame_data -> B2_G);
HYPRE_PCGSetPrintLevel(ame_data -> B2_G, 0);
HYPRE_PCGSetTol(ame_data -> B2_G, 1e-12);
HYPRE_PCGSetMaxIter(ame_data -> B2_G, 20);
HYPRE_PCGSetPrecond(ame_data -> B2_G,
(HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSolve,
(HYPRE_PtrToSolverFcn) HYPRE_BoomerAMGSetup,
ame_data -> B1_G);
HYPRE_ParCSRPCGSetup(ame_data -> B2_G,
(HYPRE_ParCSRMatrix)ame_data->A_G,
(HYPRE_ParVector)ame_data->t1,
(HYPRE_ParVector)ame_data->t2);
}
/* Setup LOBPCG */
{
HYPRE_Int seed = 75;
mv_InterfaceInterpreter* interpreter;
mv_MultiVectorPtr eigenvectors;
ame_data -> interpreter = hypre_CTAlloc(mv_InterfaceInterpreter,1);
interpreter = (mv_InterfaceInterpreter*) ame_data -> interpreter;
HYPRE_ParCSRSetupInterpreter(interpreter);
ame_data -> eigenvalues = hypre_CTAlloc(HYPRE_Real, ame_data -> block_size);
ame_data -> eigenvectors =
mv_MultiVectorCreateFromSampleVector(interpreter,
ame_data -> block_size,
ame_data -> t3);
eigenvectors = (mv_MultiVectorPtr) ame_data -> eigenvectors;
mv_MultiVectorSetRandom (eigenvectors, seed);
/* Make the initial vectors discretely divergence free */
{
HYPRE_Int i, j;
HYPRE_Real *data;
mv_TempMultiVector* tmp = mv_MultiVectorGetData(eigenvectors);
HYPRE_ParVector *v = (HYPRE_ParVector*)(tmp -> vector);
hypre_ParVector *vi;
for (i = 0; i < ame_data -> block_size; i++)
{
vi = (hypre_ParVector*) v[i];
data = hypre_VectorData(hypre_ParVectorLocalVector(vi));
for (j = 0; j < ne; j++)
if (edge_bc[j])
data[j] = 0.0;
hypre_AMEDiscrDivFreeComponent(esolver, vi);
}
}
}
hypre_TFree(edge_bc);
return hypre_error_flag;
}
hypre_ParCSRMatrix *hypre_ParCSRAAt( hypre_ParCSRMatrix *A )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Complex *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Complex *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *A_col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *A_ext_row_map;
HYPRE_Int *row_starts_A = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int num_rows_diag_A = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_cols_offd_A = hypre_CSRMatrixNumCols(A_offd);
hypre_ParCSRMatrix *C;
HYPRE_Int *col_map_offd_C;
hypre_CSRMatrix *C_diag;
HYPRE_Complex *C_diag_data;
HYPRE_Int *C_diag_i;
HYPRE_Int *C_diag_j;
hypre_CSRMatrix *C_offd;
HYPRE_Complex *C_offd_data=NULL;
HYPRE_Int *C_offd_i=NULL;
HYPRE_Int *C_offd_j=NULL;
HYPRE_Int *new_C_offd_j;
HYPRE_Int C_diag_size;
HYPRE_Int C_offd_size;
HYPRE_Int last_col_diag_C;
HYPRE_Int num_cols_offd_C;
hypre_CSRMatrix *A_ext;
HYPRE_Complex *A_ext_data;
HYPRE_Int *A_ext_i;
HYPRE_Int *A_ext_j;
HYPRE_Int num_rows_A_ext=0;
HYPRE_Int first_row_index_A = hypre_ParCSRMatrixFirstRowIndex(A);
HYPRE_Int first_col_diag_A = hypre_ParCSRMatrixFirstColDiag(A);
HYPRE_Int *B_marker;
HYPRE_Int i;
HYPRE_Int i1, i2, i3;
HYPRE_Int jj2, jj3;
HYPRE_Int jj_count_diag, jj_count_offd;
HYPRE_Int jj_row_begin_diag, jj_row_begin_offd;
HYPRE_Int start_indexing = 0; /* start indexing for C_data at 0 */
HYPRE_Int count;
HYPRE_Int n_rows_A, n_cols_A;
HYPRE_Complex a_entry;
HYPRE_Complex a_b_product;
HYPRE_Complex zero = 0.0;
n_rows_A = hypre_ParCSRMatrixGlobalNumRows(A);
n_cols_A = hypre_ParCSRMatrixGlobalNumCols(A);
if (n_cols_A != n_rows_A)
{
hypre_printf(" Error! Incompatible matrix dimensions!\n");
return NULL;
}
/*-----------------------------------------------------------------------
* Extract A_ext, i.e. portion of A that is stored on neighbor procs
* and needed locally for A^T in the matrix matrix product A*A^T
*-----------------------------------------------------------------------*/
if (num_rows_diag_A != n_rows_A)
{
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings
*--------------------------------------------------------------------*/
if (!hypre_ParCSRMatrixCommPkg(A))
{
hypre_MatTCommPkgCreate(A);
}
A_ext = hypre_ParCSRMatrixExtractAExt( A, 1, &A_ext_row_map );
A_ext_data = hypre_CSRMatrixData(A_ext);
A_ext_i = hypre_CSRMatrixI(A_ext);
A_ext_j = hypre_CSRMatrixJ(A_ext);
num_rows_A_ext = hypre_CSRMatrixNumRows(A_ext);
}
/*-----------------------------------------------------------------------
* Allocate marker array.
*-----------------------------------------------------------------------*/
B_marker = hypre_CTAlloc(HYPRE_Int, num_rows_diag_A+num_rows_A_ext );
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
for ( i1=0; i1<num_rows_diag_A+num_rows_A_ext; ++i1 )
{
B_marker[i1] = -1;
}
hypre_ParAat_RowSizes(
&C_diag_i, &C_offd_i, B_marker,
A_diag_i, A_diag_j,
A_offd_i, A_offd_j, A_col_map_offd,
A_ext_i, A_ext_j, A_ext_row_map,
&C_diag_size, &C_offd_size,
num_rows_diag_A, num_cols_offd_A,
num_rows_A_ext,
first_col_diag_A, first_row_index_A
);
#if 0
/* debugging output: */
hypre_printf("A_ext_row_map (%i):",num_rows_A_ext);
for ( i1=0; i1<num_rows_A_ext; ++i1 ) hypre_printf(" %i",A_ext_row_map[i1] );
hypre_printf("\nC_diag_i (%i):",C_diag_size);
for ( i1=0; i1<=num_rows_diag_A; ++i1 ) hypre_printf(" %i",C_diag_i[i1] );
hypre_printf("\nC_offd_i (%i):",C_offd_size);
for ( i1=0; i1<=num_rows_diag_A; ++i1 ) hypre_printf(" %i",C_offd_i[i1] );
hypre_printf("\n");
#endif
/*-----------------------------------------------------------------------
* Allocate C_diag_data and C_diag_j arrays.
* Allocate C_offd_data and C_offd_j arrays.
*-----------------------------------------------------------------------*/
last_col_diag_C = first_row_index_A + num_rows_diag_A - 1;
C_diag_data = hypre_CTAlloc(HYPRE_Complex, C_diag_size);
C_diag_j = hypre_CTAlloc(HYPRE_Int, C_diag_size);
if (C_offd_size)
{
C_offd_data = hypre_CTAlloc(HYPRE_Complex, C_offd_size);
C_offd_j = hypre_CTAlloc(HYPRE_Int, C_offd_size);
}
/*-----------------------------------------------------------------------
* Second Pass: Fill in C_diag_data and C_diag_j.
* Second Pass: Fill in C_offd_data and C_offd_j.
*-----------------------------------------------------------------------*/
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
jj_count_diag = start_indexing;
jj_count_offd = start_indexing;
for ( i1=0; i1<num_rows_diag_A+num_rows_A_ext; ++i1 )
{
B_marker[i1] = -1;
}
/*-----------------------------------------------------------------------
* Loop over interior c-points.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_A; i1++)
{
/*--------------------------------------------------------------------
* Create diagonal entry, C_{i1,i1}
*--------------------------------------------------------------------*/
B_marker[i1] = jj_count_diag;
jj_row_begin_diag = jj_count_diag;
jj_row_begin_offd = jj_count_offd;
C_diag_data[jj_count_diag] = zero;
C_diag_j[jj_count_diag] = i1;
jj_count_diag++;
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_offd.
*-----------------------------------------------------------------*/
/* There are 3 CSRMatrix or CSRBooleanMatrix objects here:
ext*ext, ext*diag, and ext*offd belong to another processor.
diag*offd and offd*diag don't count - never share a column by definition.
So we have to do 4 cases:
diag*ext, offd*ext, diag*diag, and offd*offd.
*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
a_entry = A_diag_data[jj2];
/* diag*ext */
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of (A_ext)^T
* That is, rows i3 having a column i2 of A_ext.
* For now, for each row i3 of A_ext we crudely check _all_
* columns to see whether one matches i2.
* For each entry (i2,i3) of (A_ext)^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3) . This contributes to both the diag and offd
* blocks of C.
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_A_ext; i3++ ) {
for ( jj3=A_ext_i[i3]; jj3<A_ext_i[i3+1]; jj3++ ) {
if ( A_ext_j[jj3]==i2+first_col_diag_A ) {
/* row i3, column i2 of A_ext; or,
row i2, column i3 of (A_ext)^T */
a_b_product = a_entry * A_ext_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if ( A_ext_row_map[i3] < first_row_index_A ||
A_ext_row_map[i3] > last_col_diag_C ) { /* offd */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_offd) {
B_marker[i3+num_rows_diag_A] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3;
jj_count_offd++;
}
else
C_offd_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
else { /* diag */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_diag) {
B_marker[i3+num_rows_diag_A] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3-first_col_diag_A;
jj_count_diag++;
}
else
C_diag_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
}
}
}
}
if (num_cols_offd_A)
{
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
a_entry = A_offd_data[jj2];
/* offd * ext */
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of (A_ext)^T
* That is, rows i3 having a column i2 of A_ext.
* For now, for each row i3 of A_ext we crudely check _all_
* columns to see whether one matches i2.
* For each entry (i2,i3) of (A_ext)^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3) . This contributes to both the diag and offd
* blocks of C.
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_A_ext; i3++ ) {
for ( jj3=A_ext_i[i3]; jj3<A_ext_i[i3+1]; jj3++ ) {
if ( A_ext_j[jj3]==A_col_map_offd[i2] ) {
/* row i3, column i2 of A_ext; or,
row i2, column i3 of (A_ext)^T */
a_b_product = a_entry * A_ext_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if ( A_ext_row_map[i3] < first_row_index_A ||
A_ext_row_map[i3] > last_col_diag_C ) { /* offd */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_offd) {
B_marker[i3+num_rows_diag_A] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3;
jj_count_offd++;
}
else
C_offd_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
else { /* diag */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_diag) {
B_marker[i3+num_rows_diag_A] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3-first_row_index_A;
jj_count_diag++;
}
else
C_diag_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
}
}
}
}
}
/* diag * diag */
/*-----------------------------------------------------------------
* Loop over entries (columns) i2 in row i1 of A_diag.
* For each such column we will find the contributions of the
* corresponding rows i2 of A^T to C=A*A^T . Now we only look
* at the local part of A^T - with columns (rows of A) living
* on this processor.
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
a_entry = A_diag_data[jj2];
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of A^T
* That is, rows i3 having a column i2 of A (local part).
* For now, for each row i3 of A we crudely check _all_
* columns to see whether one matches i2.
* This i3-loop is for the diagonal block of A.
* It contributes to the diagonal block of C.
* For each entry (i2,i3) of A^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3)
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_diag_A; i3++ ) {
for ( jj3=A_diag_i[i3]; jj3<A_diag_i[i3+1]; jj3++ ) {
if ( A_diag_j[jj3]==i2 ) {
/* row i3, column i2 of A; or,
row i2, column i3 of A^T */
a_b_product = a_entry * A_diag_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, mark it and increment
* counter.
*--------------------------------------------------------*/
if (B_marker[i3] < jj_row_begin_diag)
{
B_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[B_marker[i3]] += a_b_product;
}
}
}
} /* end of i3 loop */
} /* end of third i2 loop */
/* offd * offd */
/*-----------------------------------------------------------
* Loop over offd columns i2 of A in A*A^T. Then
* loop over offd entries (columns) i3 in row i2 of A^T
* That is, rows i3 having a column i2 of A (local part).
* For now, for each row i3 of A we crudely check _all_
* columns to see whether one matches i2.
* This i3-loop is for the off-diagonal block of A.
* It contributes to the diag block of C.
* For each entry (i2,i3) of A^T, add A*A^T to C
*-----------------------------------------------------------*/
if (num_cols_offd_A) {
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
a_entry = A_offd_data[jj2];
for ( i3=0; i3<num_rows_diag_A; i3++ ) {
/* ... note that num_rows_diag_A == num_rows_offd_A */
for ( jj3=A_offd_i[i3]; jj3<A_offd_i[i3+1]; jj3++ ) {
if ( A_offd_j[jj3]==i2 ) {
/* row i3, column i2 of A; or,
row i2, column i3 of A^T */
a_b_product = a_entry * A_offd_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution
*--------------------------------------------------------*/
if (B_marker[i3] < jj_row_begin_diag)
{
B_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[B_marker[i3]] += a_b_product;
}
}
}
} /* end of last i3 loop */
} /* end of if (num_cols_offd_A) */
} /* end of fourth and last i2 loop */
#if 0 /* debugging printout */
hypre_printf("end of i1 loop: i1=%i jj_count_diag=%i\n", i1, jj_count_diag );
hypre_printf(" C_diag_j=");
for ( jj3=0; jj3<jj_count_diag; ++jj3) hypre_printf("%i ",C_diag_j[jj3]);
hypre_printf(" C_diag_data=");
for ( jj3=0; jj3<jj_count_diag; ++jj3) hypre_printf("%f ",C_diag_data[jj3]);
hypre_printf("\n");
hypre_printf(" C_offd_j=");
for ( jj3=0; jj3<jj_count_offd; ++jj3) hypre_printf("%i ",C_offd_j[jj3]);
hypre_printf(" C_offd_data=");
for ( jj3=0; jj3<jj_count_offd; ++jj3) hypre_printf("%f ",C_offd_data[jj3]);
hypre_printf("\n");
hypre_printf( " B_marker =" );
for ( it=0; it<num_rows_diag_A+num_rows_A_ext; ++it )
hypre_printf(" %i", B_marker[it] );
hypre_printf( "\n" );
#endif
} /* end of i1 loop */
/*-----------------------------------------------------------------------
* Delete 0-columns in C_offd, i.e. generate col_map_offd and reset
* C_offd_j. Note that (with the indexing we have coming into this
* block) col_map_offd_C[i3]==A_ext_row_map[i3].
*-----------------------------------------------------------------------*/
for ( i=0; i<num_rows_diag_A+num_rows_A_ext; ++i )
B_marker[i] = -1;
for ( i=0; i<C_offd_size; i++ )
B_marker[ C_offd_j[i] ] = -2;
count = 0;
for (i=0; i < num_rows_diag_A + num_rows_A_ext; i++) {
if (B_marker[i] == -2) {
B_marker[i] = count;
count++;
}
}
num_cols_offd_C = count;
if (num_cols_offd_C) {
col_map_offd_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_C);
new_C_offd_j = hypre_CTAlloc(HYPRE_Int,C_offd_size);
/* ... a bit big, but num_cols_offd_C is too small. It might be worth
computing the correct size, which is sum( no. columns in row i, over all rows i )
*/
for (i=0; i < C_offd_size; i++) {
new_C_offd_j[i] = B_marker[C_offd_j[i]];
col_map_offd_C[ new_C_offd_j[i] ] = A_ext_row_map[ C_offd_j[i] ];
}
hypre_TFree(C_offd_j);
C_offd_j = new_C_offd_j;
}
/*----------------------------------------------------------------
* Create C
*----------------------------------------------------------------*/
C = hypre_ParCSRMatrixCreate(comm, n_rows_A, n_rows_A, row_starts_A,
row_starts_A, num_cols_offd_C,
C_diag_size, C_offd_size);
/* Note that C does not own the partitionings */
hypre_ParCSRMatrixSetRowStartsOwner(C,0);
hypre_ParCSRMatrixSetColStartsOwner(C,0);
C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrixData(C_diag) = C_diag_data;
hypre_CSRMatrixI(C_diag) = C_diag_i;
hypre_CSRMatrixJ(C_diag) = C_diag_j;
if (num_cols_offd_C)
{
C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrixData(C_offd) = C_offd_data;
hypre_CSRMatrixI(C_offd) = C_offd_i;
hypre_CSRMatrixJ(C_offd) = C_offd_j;
hypre_ParCSRMatrixOffd(C) = C_offd;
hypre_ParCSRMatrixColMapOffd(C) = col_map_offd_C;
}
else
hypre_TFree(C_offd_i);
/*-----------------------------------------------------------------------
* Free B_ext and marker array.
*-----------------------------------------------------------------------*/
if (num_cols_offd_A)
{
hypre_CSRMatrixDestroy(A_ext);
A_ext = NULL;
}
hypre_TFree(B_marker);
if ( num_rows_diag_A != n_rows_A )
hypre_TFree(A_ext_row_map);
return C;
}
HYPRE_Int hypre_BoomerAMGCoarsenCGC (hypre_ParCSRMatrix *S,HYPRE_Int numberofgrids,HYPRE_Int coarsen_type,HYPRE_Int *CF_marker)
/* CGC algorithm
* ====================================================================================================
* coupling : the strong couplings
* numberofgrids : the number of grids
* coarsen_type : the coarsening type
* gridpartition : the grid partition
* =====================================================================================================*/
{
HYPRE_Int j,/*p,*/mpisize,mpirank,/*rstart,rend,*/choice,*coarse,ierr=0;
HYPRE_Int *vertexrange = NULL;
HYPRE_Int *vertexrange_all = NULL;
HYPRE_Int *CF_marker_offd = NULL;
HYPRE_Int num_variables = hypre_CSRMatrixNumRows (hypre_ParCSRMatrixDiag(S));
/* HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols (hypre_ParCSRMatrixOffd (S)); */
/* HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd (S); */
/* HYPRE_Real wall_time; */
HYPRE_IJMatrix ijG;
hypre_ParCSRMatrix *G;
hypre_CSRMatrix *Gseq;
MPI_Comm comm = hypre_ParCSRMatrixComm(S);
hypre_MPI_Comm_size (comm,&mpisize);
hypre_MPI_Comm_rank (comm,&mpirank);
#if 0
if (!mpirank) {
wall_time = time_getWallclockSeconds();
hypre_printf ("Starting CGC preparation\n");
}
#endif
AmgCGCPrepare (S,numberofgrids,CF_marker,&CF_marker_offd,coarsen_type,&vertexrange);
#if 0 /* debugging */
if (!mpirank) {
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf ("Finished CGC preparation, wall_time = %f s\n",wall_time);
wall_time = time_getWallclockSeconds();
hypre_printf ("Starting CGC matrix assembly\n");
}
#endif
AmgCGCGraphAssemble (S,vertexrange,CF_marker,CF_marker_offd,coarsen_type,&ijG);
#if 0
HYPRE_IJMatrixPrint (ijG,"graph.txt");
#endif
HYPRE_IJMatrixGetObject (ijG,(void**)&G);
#if 0 /* debugging */
if (!mpirank) {
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf ("Finished CGC matrix assembly, wall_time = %f s\n",wall_time);
wall_time = time_getWallclockSeconds();
hypre_printf ("Starting CGC matrix communication\n");
}
#endif
#ifdef HYPRE_NO_GLOBAL_PARTITION
{
/* classical CGC does not really make sense in combination with HYPRE_NO_GLOBAL_PARTITION,
but anyway, here it is:
*/
HYPRE_Int nlocal = vertexrange[1]-vertexrange[0];
vertexrange_all = hypre_CTAlloc (HYPRE_Int,mpisize+1);
hypre_MPI_Allgather (&nlocal,1,HYPRE_MPI_INT,vertexrange_all+1,1,HYPRE_MPI_INT,comm);
vertexrange_all[0]=0;
for (j=2;j<=mpisize;j++) vertexrange_all[j]+=vertexrange_all[j-1];
}
#else
vertexrange_all = vertexrange;
#endif
Gseq = hypre_ParCSRMatrixToCSRMatrixAll (G);
#if 0 /* debugging */
if (!mpirank) {
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf ("Finished CGC matrix communication, wall_time = %f s\n",wall_time);
}
#endif
if (Gseq) { /* BM Aug 31, 2006: Gseq==NULL if G has no local rows */
#if 0 /* debugging */
if (!mpirank) {
wall_time = time_getWallclockSeconds();
hypre_printf ("Starting CGC election\n");
}
#endif
AmgCGCChoose (Gseq,vertexrange_all,mpisize,&coarse);
#if 0 /* debugging */
if (!mpirank) {
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf ("Finished CGC election, wall_time = %f s\n",wall_time);
}
#endif
#if 0 /* debugging */
if (!mpirank) {
for (j=0;j<mpisize;j++)
hypre_printf ("Processor %d, choice = %d of range %d - %d\n",j,coarse[j],vertexrange_all[j]+1,vertexrange_all[j+1]);
}
fflush(stdout);
#endif
#if 0 /* debugging */
if (!mpirank) {
wall_time = time_getWallclockSeconds();
hypre_printf ("Starting CGC CF assignment\n");
}
#endif
choice = coarse[mpirank];
for (j=0;j<num_variables;j++) {
if (CF_marker[j]==choice)
CF_marker[j] = C_PT;
else
CF_marker[j] = F_PT;
}
hypre_CSRMatrixDestroy (Gseq);
hypre_TFree (coarse);
}
else
for (j=0;j<num_variables;j++) CF_marker[j] = F_PT;
#if 0
if (!mpirank) {
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf ("Finished CGC CF assignment, wall_time = %f s\n",wall_time);
}
#endif
#if 0 /* debugging */
if (!mpirank) {
wall_time = time_getWallclockSeconds();
hypre_printf ("Starting CGC cleanup\n");
}
#endif
HYPRE_IJMatrixDestroy (ijG);
hypre_TFree (vertexrange);
#ifdef HYPRE_NO_GLOBAL_PARTITION
hypre_TFree (vertexrange_all);
#endif
hypre_TFree (CF_marker_offd);
#if 0
if (!mpirank) {
wall_time = time_getWallclockSeconds() - wall_time;
hypre_printf ("Finished CGC cleanup, wall_time = %f s\n",wall_time);
}
#endif
return(ierr);
}
int
hypre_BoomerAMGSetupStats( void *amg_vdata,
hypre_ParCSRMatrix *A )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_ParAMGData *amg_data = (hypre_ParAMGData*)amg_vdata;
/*hypre_SeqAMGData *seq_data = hypre_ParAMGDataSeqData(amg_data);*/
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParCSRMatrix **P_array;
hypre_CSRMatrix *A_diag;
double *A_diag_data;
int *A_diag_i;
hypre_CSRMatrix *A_offd;
double *A_offd_data;
int *A_offd_i;
hypre_CSRMatrix *P_diag;
double *P_diag_data;
int *P_diag_i;
hypre_CSRMatrix *P_offd;
double *P_offd_data;
int *P_offd_i;
int numrows;
HYPRE_BigInt *row_starts;
int num_levels;
int coarsen_type;
int interp_type;
int measure_type;
double global_nonzeros;
double *send_buff;
double *gather_buff;
/* Local variables */
int level;
int j;
HYPRE_BigInt fine_size;
int min_entries;
int max_entries;
int num_procs,my_id, num_threads;
double min_rowsum;
double max_rowsum;
double sparse;
int i;
HYPRE_BigInt coarse_size;
int entries;
double avg_entries;
double rowsum;
double min_weight;
double max_weight;
int global_min_e;
int global_max_e;
double global_min_rsum;
double global_max_rsum;
double global_min_wt;
double global_max_wt;
double *num_coeffs;
double *num_variables;
double total_variables;
double operat_cmplxty;
double grid_cmplxty;
/* amg solve params */
int max_iter;
int cycle_type;
int *num_grid_sweeps;
int *grid_relax_type;
int relax_order;
int **grid_relax_points;
double *relax_weight;
double *omega;
double tol;
int one = 1;
int minus_one = -1;
int zero = 0;
int smooth_type;
int smooth_num_levels;
int agg_num_levels;
/*int seq_cg = 0;*/
/*if (seq_data)
seq_cg = 1;*/
MPI_Comm_size(comm, &num_procs);
MPI_Comm_rank(comm,&my_id);
num_threads = hypre_NumThreads();
if (my_id == 0)
printf("\nNumber of MPI processes: %d , Number of OpenMP threads: %d\n", num_procs, num_threads);
A_array = hypre_ParAMGDataAArray(amg_data);
P_array = hypre_ParAMGDataPArray(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
coarsen_type = hypre_ParAMGDataCoarsenType(amg_data);
interp_type = hypre_ParAMGDataInterpType(amg_data);
measure_type = hypre_ParAMGDataMeasureType(amg_data);
smooth_type = hypre_ParAMGDataSmoothType(amg_data);
smooth_num_levels = hypre_ParAMGDataSmoothNumLevels(amg_data);
agg_num_levels = hypre_ParAMGDataAggNumLevels(amg_data);
/*----------------------------------------------------------
* Get the amg_data data
*----------------------------------------------------------*/
num_levels = hypre_ParAMGDataNumLevels(amg_data);
max_iter = hypre_ParAMGDataMaxIter(amg_data);
cycle_type = hypre_ParAMGDataCycleType(amg_data);
num_grid_sweeps = hypre_ParAMGDataNumGridSweeps(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
grid_relax_points = hypre_ParAMGDataGridRelaxPoints(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
relax_order = hypre_ParAMGDataRelaxOrder(amg_data);
omega = hypre_ParAMGDataOmega(amg_data);
tol = hypre_ParAMGDataTol(amg_data);
/*block_mode = hypre_ParAMGDataBlockMode(amg_data);*/
send_buff = hypre_CTAlloc(double, 6);
#ifdef HYPRE_NO_GLOBAL_PARTITION
gather_buff = hypre_CTAlloc(double,6);
#else
gather_buff = hypre_CTAlloc(double,6*num_procs);
#endif
if (my_id==0)
{
printf("\nBoomerAMG SETUP PARAMETERS:\n\n");
printf(" Max levels = %d\n",hypre_ParAMGDataMaxLevels(amg_data));
printf(" Num levels = %d\n\n",num_levels);
printf(" Strength Threshold = %f\n",
hypre_ParAMGDataStrongThreshold(amg_data));
printf(" Interpolation Truncation Factor = %f\n",
hypre_ParAMGDataTruncFactor(amg_data));
printf(" Maximum Row Sum Threshold for Dependency Weakening = %f\n\n",
hypre_ParAMGDataMaxRowSum(amg_data));
if (coarsen_type == 0)
{
printf(" Coarsening Type = Cleary-Luby-Jones-Plassman\n");
}
else if (abs(coarsen_type) == 1)
{
printf(" Coarsening Type = Ruge\n");
}
else if (abs(coarsen_type) == 2)
{
printf(" Coarsening Type = Ruge2B\n");
}
else if (abs(coarsen_type) == 3)
{
printf(" Coarsening Type = Ruge3\n");
}
else if (abs(coarsen_type) == 4)
{
printf(" Coarsening Type = Ruge 3c \n");
}
else if (abs(coarsen_type) == 5)
{
printf(" Coarsening Type = Ruge relax special points \n");
}
else if (abs(coarsen_type) == 6)
{
printf(" Coarsening Type = Falgout-CLJP \n");
}
else if (abs(coarsen_type) == 8)
{
printf(" Coarsening Type = PMIS \n");
}
else if (abs(coarsen_type) == 10)
{
printf(" Coarsening Type = HMIS \n");
}
else if (abs(coarsen_type) == 11)
{
printf(" Coarsening Type = Ruge 1st pass only \n");
}
else if (abs(coarsen_type) == 9)
{
printf(" Coarsening Type = PMIS fixed random \n");
}
else if (abs(coarsen_type) == 7)
{
printf(" Coarsening Type = CLJP, fixed random \n");
}
if (coarsen_type > 0)
{
printf(" Hybrid Coarsening (switch to CLJP when coarsening slows)\n");
}
if (coarsen_type)
printf(" measures are determined %s\n\n",
(measure_type ? "globally" : "locally"));
if (agg_num_levels)
printf(" no. of levels of aggressive coarsening: %d\n\n", agg_num_levels);
#ifdef HYPRE_NO_GLOBAL_PARTITION
printf( "\n No global partition option chosen.\n\n");
#endif
if (interp_type == 0)
{
printf(" Interpolation = modified classical interpolation\n");
}
else if (interp_type == 1)
{
printf(" Interpolation = LS interpolation \n");
}
else if (interp_type == 2)
{
printf(" Interpolation = modified classical interpolation for hyperbolic PDEs\n");
}
else if (interp_type == 3)
{
printf(" Interpolation = direct interpolation with separation of weights\n");
}
else if (interp_type == 4)
{
printf(" Interpolation = multipass interpolation\n");
}
else if (interp_type == 5)
{
printf(" Interpolation = multipass interpolation with separation of weights\n");
}
else if (interp_type == 6)
{
printf(" Interpolation = extended+i interpolation\n");
}
else if (interp_type == 7)
{
printf(" Interpolation = extended+i interpolation (only when needed)\n");
}
else if (interp_type == 8)
{
printf(" Interpolation = standard interpolation\n");
}
else if (interp_type == 9)
{
printf(" Interpolation = standard interpolation with separation of weights\n");
}
else if (interp_type == 12)
{
printf(" FF interpolation \n");
}
else if (interp_type == 13)
{
printf(" FF1 interpolation \n");
}
{
printf( "\nOperator Matrix Information:\n\n");
}
#if HYPRE_LONG_LONG
printf(" nonzero entries p");
printf("er row row sums\n");
printf("lev rows entries sparse min max ");
printf("avg min max\n");
printf("=======================================");
printf("==================================\n");
#else
printf(" nonzero entries p");
printf("er row row sums\n");
printf("lev rows entries sparse min max ");
printf("avg min max\n");
printf("=======================================");
printf("============================\n");
#endif
}
/*-----------------------------------------------------
* Enter Statistics Loop
*-----------------------------------------------------*/
num_coeffs = hypre_CTAlloc(double,num_levels);
num_variables = hypre_CTAlloc(double,num_levels);
for (level = 0; level < num_levels; level++)
{
{
A_diag = hypre_ParCSRMatrixDiag(A_array[level]);
A_diag_data = hypre_CSRMatrixData(A_diag);
A_diag_i = hypre_CSRMatrixI(A_diag);
A_offd = hypre_ParCSRMatrixOffd(A_array[level]);
A_offd_data = hypre_CSRMatrixData(A_offd);
A_offd_i = hypre_CSRMatrixI(A_offd);
row_starts = hypre_ParCSRMatrixRowStarts(A_array[level]);
fine_size = hypre_ParCSRMatrixGlobalNumRows(A_array[level]);
global_nonzeros = hypre_ParCSRMatrixDNumNonzeros(A_array[level]);
num_coeffs[level] = global_nonzeros;
num_variables[level] = (double) fine_size;
sparse = global_nonzeros /((double) fine_size * (double) fine_size);
min_entries = 0;
max_entries = 0;
min_rowsum = 0.0;
max_rowsum = 0.0;
if (hypre_CSRMatrixNumRows(A_diag))
{
min_entries = (A_diag_i[1]-A_diag_i[0])+(A_offd_i[1]-A_offd_i[0]);
for (j = A_diag_i[0]; j < A_diag_i[1]; j++)
min_rowsum += A_diag_data[j];
for (j = A_offd_i[0]; j < A_offd_i[1]; j++)
min_rowsum += A_offd_data[j];
max_rowsum = min_rowsum;
for (j = 0; j < hypre_CSRMatrixNumRows(A_diag); j++)
{
entries = (A_diag_i[j+1]-A_diag_i[j])+(A_offd_i[j+1]-A_offd_i[j]);
min_entries = hypre_min(entries, min_entries);
max_entries = hypre_max(entries, max_entries);
rowsum = 0.0;
for (i = A_diag_i[j]; i < A_diag_i[j+1]; i++)
rowsum += A_diag_data[i];
for (i = A_offd_i[j]; i < A_offd_i[j+1]; i++)
rowsum += A_offd_data[i];
min_rowsum = hypre_min(rowsum, min_rowsum);
max_rowsum = hypre_max(rowsum, max_rowsum);
}
}
avg_entries = global_nonzeros / ((double) fine_size);
}
#ifdef HYPRE_NO_GLOBAL_PARTITION
numrows = (int)(row_starts[1]-row_starts[0]);
if (!numrows) /* if we don't have any rows, then don't have this count toward
min row sum or min num entries */
{
min_entries = 1000000;
min_rowsum = 1.0e7;
}
send_buff[0] = - (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = - min_rowsum;
send_buff[3] = max_rowsum;
MPI_Reduce(send_buff, gather_buff, 4, MPI_DOUBLE, MPI_MAX, 0, comm);
if (my_id ==0)
{
global_min_e = - gather_buff[0];
global_max_e = gather_buff[1];
global_min_rsum = - gather_buff[2];
global_max_rsum = gather_buff[3];
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#else
printf( "%2d %7d %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#endif
printf(" %4.1f %10.3e %10.3e\n", avg_entries,
global_min_rsum, global_max_rsum);
}
#else
send_buff[0] = (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = min_rowsum;
send_buff[3] = max_rowsum;
MPI_Gather(send_buff,4,MPI_DOUBLE,gather_buff,4,MPI_DOUBLE,0,comm);
if (my_id == 0)
{
global_min_e = 1000000;
global_max_e = 0;
global_min_rsum = 1.0e7;
global_max_rsum = 0.0;
for (j = 0; j < num_procs; j++)
{
numrows = row_starts[j+1]-row_starts[j];
if (numrows)
{
global_min_e = hypre_min(global_min_e, (int) gather_buff[j*4]);
global_min_rsum = hypre_min(global_min_rsum, gather_buff[j*4 +2]);
}
global_max_e = hypre_max(global_max_e, (int) gather_buff[j*4 +1]);
global_max_rsum = hypre_max(global_max_rsum, gather_buff[j*4 +3]);
}
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#else
printf( "%2d %7d %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#endif
printf(" %4.1f %10.3e %10.3e\n", avg_entries,
global_min_rsum, global_max_rsum);
}
#endif
}
if (my_id == 0)
{
{
printf( "\n\nInterpolation Matrix Information:\n\n");
}
#if HYPRE_LONG_LONG
printf(" entries/row min max");
printf(" row sums\n");
printf("lev rows x cols min max ");
printf(" weight weight min max \n");
printf("=======================================");
printf("======================================\n");
#else
printf(" entries/row min max");
printf(" row sums\n");
printf("lev rows cols min max ");
printf(" weight weight min max \n");
printf("=======================================");
printf("==========================\n");
#endif
}
/*-----------------------------------------------------
* Enter Statistics Loop
*-----------------------------------------------------*/
for (level = 0; level < num_levels-1; level++)
{
{
P_diag = hypre_ParCSRMatrixDiag(P_array[level]);
P_diag_data = hypre_CSRMatrixData(P_diag);
P_diag_i = hypre_CSRMatrixI(P_diag);
P_offd = hypre_ParCSRMatrixOffd(P_array[level]);
P_offd_data = hypre_CSRMatrixData(P_offd);
P_offd_i = hypre_CSRMatrixI(P_offd);
row_starts = hypre_ParCSRMatrixRowStarts(P_array[level]);
fine_size = hypre_ParCSRMatrixGlobalNumRows(P_array[level]);
coarse_size = hypre_ParCSRMatrixGlobalNumCols(P_array[level]);
global_nonzeros = hypre_ParCSRMatrixNumNonzeros(P_array[level]);
min_weight = 1.0;
max_weight = 0.0;
max_rowsum = 0.0;
min_rowsum = 0.0;
min_entries = 0;
max_entries = 0;
if (hypre_CSRMatrixNumRows(P_diag))
{
if (hypre_CSRMatrixNumCols(P_diag)) min_weight = P_diag_data[0];
for (j = P_diag_i[0]; j < P_diag_i[1]; j++)
{
min_weight = hypre_min(min_weight, P_diag_data[j]);
if (P_diag_data[j] != 1.0)
max_weight = hypre_max(max_weight, P_diag_data[j]);
min_rowsum += P_diag_data[j];
}
for (j = P_offd_i[0]; j < P_offd_i[1]; j++)
{
min_weight = hypre_min(min_weight, P_offd_data[j]);
if (P_offd_data[j] != 1.0)
max_weight = hypre_max(max_weight, P_offd_data[j]);
min_rowsum += P_offd_data[j];
}
max_rowsum = min_rowsum;
min_entries = (P_diag_i[1]-P_diag_i[0])+(P_offd_i[1]-P_offd_i[0]);
max_entries = 0;
for (j = 0; j < hypre_CSRMatrixNumRows(P_diag); j++)
{
entries = (P_diag_i[j+1]-P_diag_i[j])+(P_offd_i[j+1]-P_offd_i[j]);
min_entries = hypre_min(entries, min_entries);
max_entries = hypre_max(entries, max_entries);
rowsum = 0.0;
for (i = P_diag_i[j]; i < P_diag_i[j+1]; i++)
{
min_weight = hypre_min(min_weight, P_diag_data[i]);
if (P_diag_data[i] != 1.0)
max_weight = hypre_max(max_weight, P_diag_data[i]);
rowsum += P_diag_data[i];
}
for (i = P_offd_i[j]; i < P_offd_i[j+1]; i++)
{
min_weight = hypre_min(min_weight, P_offd_data[i]);
if (P_offd_data[i] != 1.0)
max_weight = hypre_max(max_weight, P_offd_data[i]);
rowsum += P_offd_data[i];
}
min_rowsum = hypre_min(rowsum, min_rowsum);
max_rowsum = hypre_max(rowsum, max_rowsum);
}
}
avg_entries = ((double) global_nonzeros) / ((double) fine_size);
}
#ifdef HYPRE_NO_GLOBAL_PARTITION
numrows = (int)(row_starts[1]-row_starts[0]);
if (!numrows) /* if we don't have any rows, then don't have this count toward
min row sum or min num entries */
{
min_entries = 1000000;
min_rowsum = 1.0e7;
min_weight = 1.0e7;
}
send_buff[0] = - (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = - min_rowsum;
send_buff[3] = max_rowsum;
send_buff[4] = - min_weight;
send_buff[5] = max_weight;
MPI_Reduce(send_buff, gather_buff, 6, MPI_DOUBLE, MPI_MAX, 0, comm);
if (my_id == 0)
{
global_min_e = - gather_buff[0];
global_max_e = gather_buff[1];
global_min_rsum = -gather_buff[2];
global_max_rsum = gather_buff[3];
global_min_wt = -gather_buff[4];
global_max_wt = gather_buff[5];
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld x %-12lld %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#else
printf( "%2d %5d x %-5d %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#endif
printf(" %10.3e %9.3e %9.3e %9.3e\n",
global_min_wt, global_max_wt,
global_min_rsum, global_max_rsum);
}
#else
send_buff[0] = (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = min_rowsum;
send_buff[3] = max_rowsum;
send_buff[4] = min_weight;
send_buff[5] = max_weight;
MPI_Gather(send_buff,6,MPI_DOUBLE,gather_buff,6,MPI_DOUBLE,0,comm);
if (my_id == 0)
{
global_min_e = 1000000;
global_max_e = 0;
global_min_rsum = 1.0e7;
global_max_rsum = 0.0;
global_min_wt = 1.0e7;
global_max_wt = 0.0;
for (j = 0; j < num_procs; j++)
{
numrows = row_starts[j+1] - row_starts[j];
if (numrows)
{
global_min_e = hypre_min(global_min_e, (int) gather_buff[j*6]);
global_min_rsum = hypre_min(global_min_rsum, gather_buff[j*6+2]);
global_min_wt = hypre_min(global_min_wt, gather_buff[j*6+4]);
}
global_max_e = hypre_max(global_max_e, (int) gather_buff[j*6+1]);
global_max_rsum = hypre_max(global_max_rsum, gather_buff[j*6+3]);
global_max_wt = hypre_max(global_max_wt, gather_buff[j*6+5]);
}
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld x %-12lld %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#else
printf( "%2d %5d x %-5d %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#endif
printf(" %10.3e %9.3e %9.3e %9.3e\n",
global_min_wt, global_max_wt,
global_min_rsum, global_max_rsum);
}
#endif
}
total_variables = 0;
operat_cmplxty = 0;
for (j=0;j<hypre_ParAMGDataNumLevels(amg_data);j++)
{
operat_cmplxty += num_coeffs[j] / num_coeffs[0];
total_variables += num_variables[j];
}
if (num_variables[0] != 0)
grid_cmplxty = total_variables / num_variables[0];
if (my_id == 0 )
{
printf("\n\n Complexity: grid = %f\n",grid_cmplxty);
printf(" operator = %f\n",operat_cmplxty);
}
if (my_id == 0) printf("\n\n");
if (my_id == 0)
{
printf("\n\nBoomerAMG SOLVER PARAMETERS:\n\n");
printf( " Maximum number of cycles: %d \n",max_iter);
printf( " Stopping Tolerance: %e \n",tol);
printf( " Cycle type (1 = V, 2 = W, etc.): %d\n\n", cycle_type);
printf( " Relaxation Parameters:\n");
printf( " Visiting Grid: down up coarse\n");
printf( " Number of partial sweeps: %4d %2d %4d \n",
num_grid_sweeps[1],
num_grid_sweeps[2],num_grid_sweeps[3]);
printf( " Type 0=Jac, 3=hGS, 6=hSGS, 9=GE: %4d %2d %4d \n",
grid_relax_type[1],
grid_relax_type[2],grid_relax_type[3]);
#if 1 /* TO DO: may not want this to print if CG in the coarse grid */
printf( " Point types, partial sweeps (1=C, -1=F):\n");
if (grid_relax_points)
{
printf( " Pre-CG relaxation (down):");
for (j = 0; j < num_grid_sweeps[1]; j++)
printf(" %2d", grid_relax_points[1][j]);
printf( "\n");
printf( " Post-CG relaxation (up):");
for (j = 0; j < num_grid_sweeps[2]; j++)
printf(" %2d", grid_relax_points[2][j]);
printf( "\n");
printf( " Coarsest grid:");
for (j = 0; j < num_grid_sweeps[3]; j++)
printf(" %2d", grid_relax_points[3][j]);
printf( "\n\n");
}
else if (relax_order == 1)
{
printf( " Pre-CG relaxation (down):");
for (j = 0; j < num_grid_sweeps[1]; j++)
printf(" %2d %2d", one, minus_one);
printf( "\n");
printf( " Post-CG relaxation (up):");
for (j = 0; j < num_grid_sweeps[2]; j++)
printf(" %2d %2d", minus_one, one);
printf( "\n");
printf( " Coarsest grid:");
for (j = 0; j < num_grid_sweeps[3]; j++)
printf(" %2d", zero);
printf( "\n\n");
}
else
{
printf( " Pre-CG relaxation (down):");
for (j = 0; j < num_grid_sweeps[1]; j++)
printf(" %2d", zero);
printf( "\n");
printf( " Post-CG relaxation (up):");
for (j = 0; j < num_grid_sweeps[2]; j++)
printf(" %2d", zero);
printf( "\n");
printf( " Coarsest grid:");
for (j = 0; j < num_grid_sweeps[3]; j++)
printf(" %2d", zero);
printf( "\n\n");
}
#endif
if (smooth_type == 6)
for (j=0; j < smooth_num_levels; j++)
printf( " Schwarz Relaxation Weight %f level %d\n",
hypre_ParAMGDataSchwarzRlxWeight(amg_data),j);
for (j=0; j < num_levels; j++)
if (relax_weight[j] != 1)
printf( " Relaxation Weight %f level %d\n",relax_weight[j],j);
for (j=0; j < num_levels; j++)
if (omega[j] != 1)
printf( " Outer relaxation weight %f level %d\n",omega[j],j);
}
/*if (seq_cg)
{
hypre_seqAMGSetupStats(amg_data,num_coeffs[0],num_variables[0],
operat_cmplxty, grid_cmplxty );
}*/
hypre_TFree(num_coeffs);
hypre_TFree(num_variables);
hypre_TFree(send_buff);
hypre_TFree(gather_buff);
return(0);
}
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);
}
/*
Assume that we are given a fine and coarse topology and the
coarse degrees of freedom (DOFs) have been chosen. Assume also,
that the global interpolation matrix dof_DOF has a prescribed
nonzero pattern. Then, the fine degrees of freedom can be split
into 4 groups (here "i" stands for "interior"):
NODEidof - dofs which are interpolated only from the DOF
in one coarse vertex
EDGEidof - dofs which are interpolated only from the DOFs
in one coarse edge
FACEidof - dofs which are interpolated only from the DOFs
in one coarse face
ELEMidof - dofs which are interpolated only from the DOFs
in one coarse element
The interpolation operator dof_DOF can be build in 4 steps, by
consequently filling-in the rows corresponding to the above groups.
The code below uses harmonic extension to extend the interpolation
from one group to the next.
*/
HYPRE_Int hypre_ND1AMGeInterpolation (hypre_ParCSRMatrix * Aee,
hypre_ParCSRMatrix * ELEM_idof,
hypre_ParCSRMatrix * FACE_idof,
hypre_ParCSRMatrix * EDGE_idof,
hypre_ParCSRMatrix * ELEM_FACE,
hypre_ParCSRMatrix * ELEM_EDGE,
HYPRE_Int num_OffProcRows,
hypre_MaxwellOffProcRow ** OffProcRows,
hypre_IJMatrix * IJ_dof_DOF)
{
HYPRE_Int ierr = 0;
HYPRE_Int i, j, k;
HYPRE_Int *offproc_rnums, *swap;
hypre_ParCSRMatrix * dof_DOF = hypre_IJMatrixObject(IJ_dof_DOF);
hypre_ParCSRMatrix * ELEM_DOF = ELEM_EDGE;
hypre_ParCSRMatrix * ELEM_FACEidof;
hypre_ParCSRMatrix * ELEM_EDGEidof;
hypre_CSRMatrix *A, *P;
HYPRE_Int numELEM = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(ELEM_EDGE));
HYPRE_Int getrow_ierr;
HYPRE_Int three_dimensional_problem;
MPI_Comm comm= hypre_ParCSRMatrixComm(Aee);
HYPRE_Int myproc;
hypre_MPI_Comm_rank(comm, &myproc);
#if 0
hypre_IJMatrix * ij_dof_DOF = hypre_CTAlloc(hypre_IJMatrix, 1);
/* Convert dof_DOF to IJ matrix, so we can use AddToValues */
hypre_IJMatrixComm(ij_dof_DOF) = hypre_ParCSRMatrixComm(dof_DOF);
hypre_IJMatrixRowPartitioning(ij_dof_DOF) =
hypre_ParCSRMatrixRowStarts(dof_DOF);
hypre_IJMatrixColPartitioning(ij_dof_DOF) =
hypre_ParCSRMatrixColStarts(dof_DOF);
hypre_IJMatrixObject(ij_dof_DOF) = dof_DOF;
hypre_IJMatrixAssembleFlag(ij_dof_DOF) = 1;
#endif
/* sort the offproc rows to get quicker comparison for later */
if (num_OffProcRows)
{
offproc_rnums= hypre_TAlloc(HYPRE_Int, num_OffProcRows);
swap = hypre_TAlloc(HYPRE_Int, num_OffProcRows);
for (i= 0; i< num_OffProcRows; i++)
{
offproc_rnums[i]=(OffProcRows[i] -> row);
swap[i] = i;
}
}
if (num_OffProcRows > 1)
{
hypre_qsort2i(offproc_rnums, swap, 0, num_OffProcRows-1);
}
if (FACE_idof == EDGE_idof)
three_dimensional_problem = 0;
else
three_dimensional_problem = 1;
/* ELEM_FACEidof = ELEM_FACE x FACE_idof */
if (three_dimensional_problem)
ELEM_FACEidof = hypre_ParMatmul(ELEM_FACE, FACE_idof);
/* ELEM_EDGEidof = ELEM_EDGE x EDGE_idof */
ELEM_EDGEidof = hypre_ParMatmul(ELEM_EDGE, EDGE_idof);
/* Loop over local coarse elements */
k = hypre_ParCSRMatrixFirstRowIndex(ELEM_EDGE);
for (i = 0; i < numELEM; i++, k++)
{
HYPRE_Int size1, size2;
HYPRE_Int *col_ind0, *col_ind1, *col_ind2;
HYPRE_Int num_DOF, *DOF0, *DOF;
HYPRE_Int num_idof, *idof0, *idof;
HYPRE_Int num_bdof, *bdof;
double *boolean_data;
/* Determine the coarse DOFs */
hypre_ParCSRMatrixGetRow (ELEM_DOF, k, &num_DOF, &DOF0, &boolean_data);
DOF= hypre_TAlloc(HYPRE_Int, num_DOF);
for (j= 0; j< num_DOF; j++)
{
DOF[j]= DOF0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_DOF, k, &num_DOF, &DOF0, &boolean_data);
qsort0(DOF,0,num_DOF-1);
/* Find the fine dofs interior for the current coarse element */
hypre_ParCSRMatrixGetRow (ELEM_idof, k, &num_idof, &idof0, &boolean_data);
idof= hypre_TAlloc(HYPRE_Int, num_idof);
for (j= 0; j< num_idof; j++)
{
idof[j]= idof0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_idof, k, &num_idof, &idof0, &boolean_data);
/* Sort the interior dofs according to their global number */
qsort0(idof,0,num_idof-1);
/* Find the fine dofs on the boundary of the current coarse element */
if (three_dimensional_problem)
{
hypre_ParCSRMatrixGetRow (ELEM_FACEidof, k, &size1, &col_ind0, &boolean_data);
col_ind1= hypre_TAlloc(HYPRE_Int, size1);
for (j= 0; j< size1; j++)
{
col_ind1[j]= col_ind0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_FACEidof, k, &size1, &col_ind0, &boolean_data);
}
else
size1 = 0;
hypre_ParCSRMatrixGetRow (ELEM_EDGEidof, k, &size2, &col_ind0, &boolean_data);
col_ind2= hypre_TAlloc(HYPRE_Int, size2);
for (j= 0; j< size2; j++)
{
col_ind2[j]= col_ind0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_EDGEidof, k, &size2, &col_ind0, &boolean_data);
/* Merge and sort the boundary dofs according to their global number */
num_bdof = size1 + size2;
bdof = hypre_CTAlloc(HYPRE_Int, num_bdof);
if (three_dimensional_problem)
memcpy(bdof, col_ind1, size1*sizeof(HYPRE_Int));
memcpy(bdof+size1, col_ind2, size2*sizeof(HYPRE_Int));
qsort0(bdof,0,num_bdof-1);
/* A = extract_rows(Aee, idof) */
A = hypre_CSRMatrixCreate (num_idof, num_idof + num_bdof,
num_idof * (num_idof + num_bdof));
hypre_CSRMatrixInitialize(A);
{
HYPRE_Int *I = hypre_CSRMatrixI(A);
HYPRE_Int *J = hypre_CSRMatrixJ(A);
double *data = hypre_CSRMatrixData(A);
HYPRE_Int *tmp_J;
double *tmp_data;
I[0] = 0;
for (j = 0; j < num_idof; j++)
{
getrow_ierr= hypre_ParCSRMatrixGetRow (Aee, idof[j], &I[j+1], &tmp_J, &tmp_data);
if (getrow_ierr <0)
hypre_printf("getrow Aee off proc[%d] = \n",myproc);
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
hypre_ParCSRMatrixRestoreRow (Aee, idof[j], &I[j+1], &tmp_J, &tmp_data);
I[j+1] += I[j];
}
}
/* P = extract_rows(dof_DOF, idof+bdof) */
P = hypre_CSRMatrixCreate (num_idof + num_bdof, num_DOF,
(num_idof + num_bdof) * num_DOF);
hypre_CSRMatrixInitialize(P);
{
HYPRE_Int *I = hypre_CSRMatrixI(P);
HYPRE_Int *J = hypre_CSRMatrixJ(P);
double *data = hypre_CSRMatrixData(P);
HYPRE_Int m;
HYPRE_Int *tmp_J;
double *tmp_data;
I[0] = 0;
for (j = 0; j < num_idof; j++)
{
getrow_ierr= hypre_ParCSRMatrixGetRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data);
if (getrow_ierr >= 0)
{
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
hypre_ParCSRMatrixRestoreRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data);
I[j+1] += I[j];
}
else /* row offproc */
{
hypre_ParCSRMatrixRestoreRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data);
/* search for OffProcRows */
m= 0;
while (m < num_OffProcRows)
{
if (offproc_rnums[m] == idof[j])
{
break;
}
else
{
m++;
}
}
I[j+1]= (OffProcRows[swap[m]] -> ncols);
tmp_J = (OffProcRows[swap[m]] -> cols);
tmp_data= (OffProcRows[swap[m]] -> data);
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
I[j+1] += I[j];
}
}
for ( ; j < num_idof + num_bdof; j++)
{
getrow_ierr= hypre_ParCSRMatrixGetRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data);
if (getrow_ierr >= 0)
{
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
hypre_ParCSRMatrixRestoreRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data);
I[j+1] += I[j];
}
else /* row offproc */
{
hypre_ParCSRMatrixRestoreRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data);
/* search for OffProcRows */
m= 0;
while (m < num_OffProcRows)
{
if (offproc_rnums[m] == bdof[j-num_idof])
{
break;
}
else
{
m++;
}
}
if (m>= num_OffProcRows)hypre_printf("here the mistake\n");
I[j+1]= (OffProcRows[swap[m]] -> ncols);
tmp_J = (OffProcRows[swap[m]] -> cols);
tmp_data= (OffProcRows[swap[m]] -> data);
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
I[j+1] += I[j];
}
}
}
/* Pi = Aii^{-1} Aib Pb */
hypre_HarmonicExtension (A, P, num_DOF, DOF,
num_idof, idof, num_bdof, bdof);
/* Insert Pi in dof_DOF */
{
HYPRE_Int * ncols = hypre_CTAlloc(HYPRE_Int, num_idof);
for (j = 0; j < num_idof; j++)
ncols[j] = num_DOF;
hypre_IJMatrixAddToValuesParCSR (IJ_dof_DOF,
num_idof, ncols, idof,
hypre_CSRMatrixJ(P),
hypre_CSRMatrixData(P));
hypre_TFree(ncols);
}
hypre_TFree(DOF);
hypre_TFree(idof);
if (three_dimensional_problem)
{
hypre_TFree(col_ind1);
}
hypre_TFree(col_ind2);
hypre_TFree(bdof);
hypre_CSRMatrixDestroy(A);
hypre_CSRMatrixDestroy(P);
}
#if 0
hypre_TFree(ij_dof_DOF);
#endif
if (three_dimensional_problem)
hypre_ParCSRMatrixDestroy(ELEM_FACEidof);
hypre_ParCSRMatrixDestroy(ELEM_EDGEidof);
if (num_OffProcRows)
{
hypre_TFree(offproc_rnums);
hypre_TFree(swap);
}
return ierr;
}
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_ParCSRMatrixMatvec_FF
*--------------------------------------------------------------------------*/
HYPRE_Int
hypre_ParCSRMatrixMatvec_FF( HYPRE_Complex alpha,
hypre_ParCSRMatrix *A,
hypre_ParVector *x,
HYPRE_Complex beta,
hypre_ParVector *y,
HYPRE_Int *CF_marker,
HYPRE_Int fpt )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_ParCSRCommHandle *comm_handle;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(A);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(A);
hypre_Vector *x_local = hypre_ParVectorLocalVector(x);
hypre_Vector *y_local = hypre_ParVectorLocalVector(y);
HYPRE_Int num_rows = hypre_ParCSRMatrixGlobalNumRows(A);
HYPRE_Int num_cols = hypre_ParCSRMatrixGlobalNumCols(A);
hypre_Vector *x_tmp;
HYPRE_Int x_size = hypre_ParVectorGlobalSize(x);
HYPRE_Int y_size = hypre_ParVectorGlobalSize(y);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int ierr = 0;
HYPRE_Int num_sends, i, j, index, start, num_procs;
HYPRE_Int *int_buf_data = NULL;
HYPRE_Int *CF_marker_offd = NULL;
HYPRE_Complex *x_tmp_data = NULL;
HYPRE_Complex *x_buf_data = NULL;
HYPRE_Complex *x_local_data = hypre_VectorData(x_local);
/*---------------------------------------------------------------------
* Check for size compatibility. ParMatvec returns ierr = 11 if
* length of X doesn't equal the number of columns of A,
* ierr = 12 if the length of Y doesn't equal the number of rows
* of A, and ierr = 13 if both are true.
*
* Because temporary vectors are often used in ParMatvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_MPI_Comm_size(comm,&num_procs);
if (num_cols != x_size)
ierr = 11;
if (num_rows != y_size)
ierr = 12;
if (num_cols != x_size && num_rows != y_size)
ierr = 13;
if (num_procs > 1)
{
if (num_cols_offd)
{
x_tmp = hypre_SeqVectorCreate( num_cols_offd );
hypre_SeqVectorInitialize(x_tmp);
x_tmp_data = hypre_VectorData(x_tmp);
}
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings
*--------------------------------------------------------------------*/
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
if (num_sends)
x_buf_data = hypre_CTAlloc(HYPRE_Complex, hypre_ParCSRCommPkgSendMapStart
(comm_pkg, num_sends));
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
x_buf_data[index++]
= x_local_data[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)];
}
comm_handle =
hypre_ParCSRCommHandleCreate ( 1, comm_pkg, x_buf_data, x_tmp_data );
}
hypre_CSRMatrixMatvec_FF( alpha, diag, x_local, beta, y_local, CF_marker,
CF_marker, fpt);
if (num_procs > 1)
{
hypre_ParCSRCommHandleDestroy(comm_handle);
comm_handle = NULL;
if (num_sends)
int_buf_data = hypre_CTAlloc(HYPRE_Int, hypre_ParCSRCommPkgSendMapStart
(comm_pkg, num_sends));
if (num_cols_offd) CF_marker_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
int_buf_data[index++]
= CF_marker[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)];
}
comm_handle =
hypre_ParCSRCommHandleCreate(11,comm_pkg,int_buf_data,CF_marker_offd );
hypre_ParCSRCommHandleDestroy(comm_handle);
comm_handle = NULL;
if (num_cols_offd) hypre_CSRMatrixMatvec_FF( alpha, offd, x_tmp, 1.0, y_local,
CF_marker, CF_marker_offd, fpt);
hypre_SeqVectorDestroy(x_tmp);
x_tmp = NULL;
hypre_TFree(x_buf_data);
hypre_TFree(int_buf_data);
hypre_TFree(CF_marker_offd);
}
return ierr;
}
HYPRE_Int
hypre_MaxwellSolve2( void * maxwell_vdata,
hypre_SStructMatrix * A_in,
hypre_SStructVector * f,
hypre_SStructVector * u )
{
hypre_MaxwellData *maxwell_data = maxwell_vdata;
hypre_ParVector *f_edge;
hypre_ParVector *u_edge;
HYPRE_Int max_iter = maxwell_data-> max_iter;
double tol = maxwell_data-> tol;
HYPRE_Int rel_change = maxwell_data-> rel_change;
HYPRE_Int zero_guess = maxwell_data-> zero_guess;
HYPRE_Int npre_relax = maxwell_data-> num_pre_relax;
HYPRE_Int npost_relax = maxwell_data-> num_post_relax;
hypre_ParCSRMatrix **Ann_l = maxwell_data-> Ann_l;
hypre_ParCSRMatrix **Pn_l = maxwell_data-> Pn_l;
hypre_ParCSRMatrix **RnT_l = maxwell_data-> RnT_l;
hypre_ParVector **bn_l = maxwell_data-> bn_l;
hypre_ParVector **xn_l = maxwell_data-> xn_l;
hypre_ParVector **resn_l = maxwell_data-> resn_l;
hypre_ParVector **en_l = maxwell_data-> en_l;
hypre_ParVector **nVtemp2_l = maxwell_data-> nVtemp2_l;
HYPRE_Int **nCF_marker_l = maxwell_data-> nCF_marker_l;
double *nrelax_weight= maxwell_data-> nrelax_weight;
double *nomega = maxwell_data-> nomega;
HYPRE_Int nrelax_type = maxwell_data-> nrelax_type;
HYPRE_Int node_numlevs = maxwell_data-> node_numlevels;
hypre_ParCSRMatrix *Tgrad = maxwell_data-> Tgrad;
hypre_ParCSRMatrix *T_transpose = maxwell_data-> T_transpose;
hypre_ParCSRMatrix **Aee_l = maxwell_data-> Aee_l;
hypre_IJMatrix **Pe_l = maxwell_data-> Pe_l;
hypre_IJMatrix **ReT_l = maxwell_data-> ReT_l;
hypre_ParVector **be_l = maxwell_data-> be_l;
hypre_ParVector **xe_l = maxwell_data-> xe_l;
hypre_ParVector **rese_l = maxwell_data-> rese_l;
hypre_ParVector **ee_l = maxwell_data-> ee_l;
hypre_ParVector **eVtemp2_l = maxwell_data-> eVtemp2_l;
HYPRE_Int **eCF_marker_l = maxwell_data-> eCF_marker_l;
double *erelax_weight= maxwell_data-> erelax_weight;
double *eomega = maxwell_data-> eomega;
HYPRE_Int erelax_type = maxwell_data-> erelax_type;
HYPRE_Int edge_numlevs = maxwell_data-> edge_numlevels;
HYPRE_Int **BdryRanks_l = maxwell_data-> BdryRanks_l;
HYPRE_Int *BdryRanksCnts_l= maxwell_data-> BdryRanksCnts_l;
HYPRE_Int logging = maxwell_data-> logging;
double *norms = maxwell_data-> norms;
double *rel_norms = maxwell_data-> rel_norms;
HYPRE_Int Solve_err_flag;
HYPRE_Int relax_local, cycle_param;
double b_dot_b = 0, r_dot_r, eps = 0;
double e_dot_e, x_dot_x;
HYPRE_Int i, j;
HYPRE_Int level;
HYPRE_Int ierr= 0;
/* added for the relaxation routines */
hypre_ParVector *ze = NULL;
if (hypre_NumThreads() > 1)
{
/* Aee is always bigger than Ann */
ze = hypre_ParVectorCreate(hypre_ParCSRMatrixComm(Aee_l[0]),
hypre_ParCSRMatrixGlobalNumRows(Aee_l[0]),
hypre_ParCSRMatrixRowStarts(Aee_l[0]));
hypre_ParVectorInitialize(ze);
hypre_ParVectorSetPartitioningOwner(ze,0);
}
hypre_BeginTiming(maxwell_data-> time_index);
hypre_SStructVectorConvert(f, &f_edge);
hypre_SStructVectorConvert(u, &u_edge);
hypre_ParVectorZeroBCValues(f_edge, BdryRanks_l[0], BdryRanksCnts_l[0]);
hypre_ParVectorZeroBCValues(u_edge, BdryRanks_l[0], BdryRanksCnts_l[0]);
be_l[0]= f_edge;
xe_l[0]= u_edge;
/* the nodal fine vectors: xn= 0. bn= T'*(be- Aee*xe) is updated in the cycle. */
hypre_ParVectorSetConstantValues(xn_l[0], 0.0);
relax_local= 0;
cycle_param= 0;
(maxwell_data-> num_iterations) = 0;
/* if max_iter is zero, return */
if (max_iter == 0)
{
/* if using a zero initial guess, return zero */
if (zero_guess)
{
hypre_ParVectorSetConstantValues(xe_l[0], 0.0);
}
hypre_EndTiming(maxwell_data -> time_index);
return ierr;
}
/* part of convergence check */
if (tol > 0.0)
{
/* eps = (tol^2) */
b_dot_b= hypre_ParVectorInnerProd(be_l[0], be_l[0]);
eps = tol*tol;
/* if rhs is zero, return a zero solution */
if (b_dot_b == 0.0)
{
hypre_ParVectorSetConstantValues(xe_l[0], 0.0);
if (logging > 0)
{
norms[0] = 0.0;
rel_norms[0] = 0.0;
}
hypre_EndTiming(maxwell_data -> time_index);
return ierr;
}
}
/*-----------------------------------------------------
* Do V-cycles:
* For each index l, "fine" = l, "coarse" = (l-1)
*
* solution update:
* edge_sol= edge_sol + T*node_sol
*-----------------------------------------------------*/
for (i = 0; i < max_iter; i++)
{
/* compute fine grid residual & nodal rhs. */
hypre_ParVectorCopy(be_l[0], rese_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]);
hypre_ParVectorZeroBCValues(rese_l[0], BdryRanks_l[0], BdryRanksCnts_l[0]);
hypre_ParCSRMatrixMatvec(1.0, T_transpose, rese_l[0], 0.0, bn_l[0]);
/* convergence check */
if (tol > 0.0)
{
r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]);
if (logging > 0)
{
norms[i] = sqrt(r_dot_r);
if (b_dot_b > 0)
rel_norms[i] = sqrt(r_dot_r/b_dot_b);
else
rel_norms[i] = 0.0;
}
/* always do at least 1 V-cycle */
if ((r_dot_r/b_dot_b < eps) && (i > 0))
{
if (rel_change)
{
if ((e_dot_e/x_dot_x) < eps)
break;
}
else
{
break;
}
}
}
hypre_ParVectorCopy(bn_l[0], resn_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]);
r_dot_r= hypre_ParVectorInnerProd(resn_l[0], resn_l[0]);
for (level= 0; level<= node_numlevs-2; level++)
{
/*-----------------------------------------------
* Down cycle
*-----------------------------------------------*/
for (j= 0; j< npre_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level],
bn_l[level],
nCF_marker_l[level],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[level],
nomega[level],
NULL,
xn_l[level],
nVtemp2_l[level],
ze);
} /*for (j= 0; j< npre_relax; j++) */
/* compute residuals */
hypre_ParVectorCopy(bn_l[level], resn_l[level]);
hypre_ParCSRMatrixMatvec(-1.0, Ann_l[level], xn_l[level],
1.0, resn_l[level]);
/* restrict residuals */
hypre_ParCSRMatrixMatvecT(1.0, RnT_l[level], resn_l[level],
0.0, bn_l[level+1]);
/* zero off initial guess for the next level */
hypre_ParVectorSetConstantValues(xn_l[level+1], 0.0);
} /* for (level= 0; level<= node_numlevs-2; level++) */
/* coarsest node solve */
level= node_numlevs-1;
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level],
bn_l[level],
nCF_marker_l[level],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[level],
nomega[level],
NULL,
xn_l[level],
nVtemp2_l[level],
ze);
/*---------------------------------------------------------------------
* Cycle up the levels.
*---------------------------------------------------------------------*/
for (level= (node_numlevs - 2); level>= 1; level--)
{
hypre_ParCSRMatrixMatvec(1.0, Pn_l[level], xn_l[level+1], 0.0,
en_l[level]);
hypre_ParVectorAxpy(1.0, en_l[level], xn_l[level]);
/* post smooth */
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[level],
bn_l[level],
nCF_marker_l[level],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[level],
nomega[level],
NULL,
xn_l[level],
nVtemp2_l[level],
ze);
}
} /* for (level= (en_numlevs - 2); level>= 1; level--) */
/* interpolate error and correct on finest grids */
hypre_ParCSRMatrixMatvec(1.0, Pn_l[0], xn_l[1], 0.0, en_l[0]);
hypre_ParVectorAxpy(1.0, en_l[0], xn_l[0]);
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Ann_l[0],
bn_l[0],
nCF_marker_l[0],
nrelax_type,
relax_local,
cycle_param,
nrelax_weight[0],
nomega[0],
NULL,
xn_l[0],
nVtemp2_l[0],
ze);
} /* for (j= 0; j< npost_relax; j++) */
hypre_ParVectorCopy(bn_l[0], resn_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Ann_l[0], xn_l[0], 1.0, resn_l[0]);
/* add the gradient solution component to xe_l[0] */
hypre_ParCSRMatrixMatvec(1.0, Tgrad, xn_l[0], 1.0, xe_l[0]);
hypre_ParVectorCopy(be_l[0], rese_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]);
r_dot_r= hypre_ParVectorInnerProd(rese_l[0], rese_l[0]);
for (level= 0; level<= edge_numlevs-2; level++)
{
/*-----------------------------------------------
* Down cycle
*-----------------------------------------------*/
for (j= 0; j< npre_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level],
be_l[level],
eCF_marker_l[level],
erelax_type,
relax_local,
cycle_param,
erelax_weight[level],
eomega[level],
NULL,
xe_l[level],
eVtemp2_l[level],
ze);
} /*for (j= 0; j< npre_relax; j++) */
/* compute residuals */
hypre_ParVectorCopy(be_l[level], rese_l[level]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[level], xe_l[level],
1.0, rese_l[level]);
/* restrict residuals */
hypre_ParCSRMatrixMatvecT(1.0,
(hypre_ParCSRMatrix *) hypre_IJMatrixObject(ReT_l[level]),
rese_l[level], 0.0, be_l[level+1]);
hypre_ParVectorZeroBCValues(be_l[level+1], BdryRanks_l[level+1],
BdryRanksCnts_l[level+1]);
/* zero off initial guess for the next level */
hypre_ParVectorSetConstantValues(xe_l[level+1], 0.0);
} /* for (level= 1; level<= edge_numlevels-2; level++) */
/* coarsest edge solve */
level= edge_numlevs-1;
for (j= 0; j< npre_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level],
be_l[level],
eCF_marker_l[level],
erelax_type,
relax_local,
cycle_param,
erelax_weight[level],
eomega[level],
NULL,
xe_l[level],
eVtemp2_l[level],
ze);
}
/*---------------------------------------------------------------------
* Up cycle.
*---------------------------------------------------------------------*/
for (level= (edge_numlevs - 2); level>= 1; level--)
{
hypre_ParCSRMatrixMatvec(1.0,
(hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[level]),
xe_l[level+1], 0.0, ee_l[level]);
hypre_ParVectorZeroBCValues(ee_l[level], BdryRanks_l[level],
BdryRanksCnts_l[level]);
hypre_ParVectorAxpy(1.0, ee_l[level], xe_l[level]);
/* post smooth */
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[level],
be_l[level],
eCF_marker_l[level],
erelax_type,
relax_local,
cycle_param,
erelax_weight[level],
eomega[level],
NULL,
xe_l[level],
eVtemp2_l[level],
ze);
}
} /* for (level= (edge_numlevs - 2); level>= 1; level--) */
/* interpolate error and correct on finest grids */
hypre_ParCSRMatrixMatvec(1.0,
(hypre_ParCSRMatrix *) hypre_IJMatrixObject(Pe_l[0]),
xe_l[1], 0.0, ee_l[0]);
hypre_ParVectorZeroBCValues(ee_l[0], BdryRanks_l[0],
BdryRanksCnts_l[0]);
hypre_ParVectorAxpy(1.0, ee_l[0], xe_l[0]);
for (j= 0; j< npost_relax; j++)
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(Aee_l[0],
be_l[0],
eCF_marker_l[0],
erelax_type,
relax_local,
cycle_param,
erelax_weight[0],
eomega[0],
NULL,
xe_l[0],
eVtemp2_l[0],
ze);
} /* for (j= 0; j< npost_relax; j++) */
e_dot_e= hypre_ParVectorInnerProd(ee_l[0], ee_l[0]);
x_dot_x= hypre_ParVectorInnerProd(xe_l[0], xe_l[0]);
hypre_ParVectorCopy(be_l[0], rese_l[0]);
hypre_ParCSRMatrixMatvec(-1.0, Aee_l[0], xe_l[0], 1.0, rese_l[0]);
(maxwell_data -> num_iterations) = (i + 1);
}
hypre_EndTiming(maxwell_data -> time_index);
if (ze)
hypre_ParVectorDestroy(ze);
return ierr;
}
HYPRE_Int AmgCGCPrepare (hypre_ParCSRMatrix *S,HYPRE_Int nlocal,HYPRE_Int *CF_marker,HYPRE_Int **CF_marker_offd,HYPRE_Int coarsen_type,HYPRE_Int **vrange)
/* assemble a graph representing the connections between the grids
* ================================================================================================
* S : the strength matrix
* nlocal : the number of locally created coarse grids
* CF_marker, CF_marker_offd : the coare/fine markers
* coarsen_type : the coarsening type
* vrange : the ranges of the vertices representing coarse grids
* ================================================================================================*/
{
HYPRE_Int ierr=0;
HYPRE_Int mpisize,mpirank;
HYPRE_Int num_sends;
HYPRE_Int *vertexrange=NULL;
HYPRE_Int vstart,vend;
HYPRE_Int *int_buf_data;
HYPRE_Int start;
HYPRE_Int i,ii,j;
HYPRE_Int num_variables = hypre_CSRMatrixNumRows (hypre_ParCSRMatrixDiag(S));
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols (hypre_ParCSRMatrixOffd (S));
MPI_Comm comm = hypre_ParCSRMatrixComm(S);
/* hypre_MPI_Status status; */
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg (S);
hypre_ParCSRCommHandle *comm_handle;
hypre_MPI_Comm_size (comm,&mpisize);
hypre_MPI_Comm_rank (comm,&mpirank);
if (!comm_pkg) {
hypre_MatvecCommPkgCreate (S);
comm_pkg = hypre_ParCSRMatrixCommPkg (S);
}
num_sends = hypre_ParCSRCommPkgNumSends (comm_pkg);
if (coarsen_type % 2 == 0) nlocal++; /* even coarsen_type means allow_emptygrids */
#ifdef HYPRE_NO_GLOBAL_PARTITION
{
HYPRE_Int scan_recv;
vertexrange = hypre_CTAlloc(HYPRE_Int,2);
hypre_MPI_Scan(&nlocal, &scan_recv, 1, HYPRE_MPI_INT, hypre_MPI_SUM, comm);
/* first point in my range */
vertexrange[0] = scan_recv - nlocal;
/* first point in next proc's range */
vertexrange[1] = scan_recv;
vstart = vertexrange[0];
vend = vertexrange[1];
}
#else
vertexrange = hypre_CTAlloc (HYPRE_Int,mpisize+1);
hypre_MPI_Allgather (&nlocal,1,HYPRE_MPI_INT,vertexrange+1,1,HYPRE_MPI_INT,comm);
vertexrange[0]=0;
for (i=2;i<=mpisize;i++) vertexrange[i]+=vertexrange[i-1];
vstart = vertexrange[mpirank];
vend = vertexrange[mpirank+1];
#endif
/* Note: vstart uses 0-based indexing, while CF_marker uses 1-based indexing */
if (coarsen_type % 2 == 1) { /* see above */
for (i=0;i<num_variables;i++)
if (CF_marker[i]>0)
CF_marker[i]+=vstart;
}
else {
/* hypre_printf ("processor %d: empty grid allowed\n",mpirank); */
for (i=0;i<num_variables;i++) {
if (CF_marker[i]>0)
CF_marker[i]+=vstart+1; /* add one because vertexrange[mpirank]+1 denotes the empty grid.
Hence, vertexrange[mpirank]+2 is the first coarse grid denoted in
global indices, ... */
}
}
/* exchange data */
*CF_marker_offd = hypre_CTAlloc (HYPRE_Int,num_cols_offd);
int_buf_data = hypre_CTAlloc (HYPRE_Int,hypre_ParCSRCommPkgSendMapStart (comm_pkg,num_sends));
for (i=0,ii=0;i<num_sends;i++) {
start = hypre_ParCSRCommPkgSendMapStart (comm_pkg,i);
for (j=start;j<hypre_ParCSRCommPkgSendMapStart (comm_pkg,i+1);j++)
int_buf_data [ii++] = CF_marker[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)];
}
if (mpisize>1) {
comm_handle = hypre_ParCSRCommHandleCreate (11,comm_pkg,int_buf_data,*CF_marker_offd);
hypre_ParCSRCommHandleDestroy (comm_handle);
}
hypre_TFree (int_buf_data);
*vrange=vertexrange;
return (ierr);
}
/*
Function: hypre_ParCSRMatrixEliminateAAe
(input) (output)
/ A_ii | A_ib \ / A_ii | 0 \
A = | -----+----- | ---> | -----+----- |
\ A_bi | A_bb / \ 0 | I /
/ 0 | A_ib \
Ae = | -----+--------- |
\ A_bi | A_bb - I /
*/
void hypre_ParCSRMatrixEliminateAAe(hypre_ParCSRMatrix *A,
hypre_ParCSRMatrix **Ae,
HYPRE_Int num_rowscols_to_elim,
HYPRE_Int *rowscols_to_elim)
{
HYPRE_Int i, j, k;
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int A_diag_nrows = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int A_offd_ncols = hypre_CSRMatrixNumCols(A_offd);
*Ae = hypre_ParCSRMatrixCreate(hypre_ParCSRMatrixComm(A),
hypre_ParCSRMatrixGlobalNumRows(A),
hypre_ParCSRMatrixGlobalNumCols(A),
hypre_ParCSRMatrixRowStarts(A),
hypre_ParCSRMatrixColStarts(A),
0, 0, 0);
hypre_ParCSRMatrixSetRowStartsOwner(*Ae, 0);
hypre_ParCSRMatrixSetColStartsOwner(*Ae, 0);
hypre_CSRMatrix *Ae_diag = hypre_ParCSRMatrixDiag(*Ae);
hypre_CSRMatrix *Ae_offd = hypre_ParCSRMatrixOffd(*Ae);
HYPRE_Int Ae_offd_ncols;
HYPRE_Int num_offd_cols_to_elim;
HYPRE_Int *offd_cols_to_elim;
HYPRE_Int *A_col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *Ae_col_map_offd;
HYPRE_Int *col_mark;
HYPRE_Int *col_remap;
/* figure out which offd cols should be eliminated */
{
hypre_ParCSRCommHandle *comm_handle;
hypre_ParCSRCommPkg *comm_pkg;
HYPRE_Int num_sends, *int_buf_data;
HYPRE_Int index, start;
HYPRE_Int *eliminate_row = hypre_CTAlloc(HYPRE_Int, A_diag_nrows);
HYPRE_Int *eliminate_col = hypre_CTAlloc(HYPRE_Int, A_offd_ncols);
/* make sure A has a communication package */
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
/* which of the local rows are to be eliminated */
for (i = 0; i < A_diag_nrows; i++)
{
eliminate_row[i] = 0;
}
for (i = 0; i < num_rowscols_to_elim; i++)
{
eliminate_row[rowscols_to_elim[i]] = 1;
}
/* use a Matvec communication pattern to find (in eliminate_col)
which of the local offd columns are to be eliminated */
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
int_buf_data = hypre_CTAlloc(HYPRE_Int,
hypre_ParCSRCommPkgSendMapStart(comm_pkg,
num_sends));
index = 0;
for (i = 0; i < num_sends; i++)
{
start = hypre_ParCSRCommPkgSendMapStart(comm_pkg, i);
for (j = start; j < hypre_ParCSRCommPkgSendMapStart(comm_pkg, i+1); j++)
{
k = hypre_ParCSRCommPkgSendMapElmt(comm_pkg, j);
int_buf_data[index++] = eliminate_row[k];
}
}
comm_handle = hypre_ParCSRCommHandleCreate(11, comm_pkg,
int_buf_data, eliminate_col);
/* eliminate diagonal part, overlapping it with communication */
hypre_CSRMatrixElimCreate(A_diag, Ae_diag,
num_rowscols_to_elim, rowscols_to_elim,
num_rowscols_to_elim, rowscols_to_elim,
NULL);
hypre_CSRMatrixEliminateRowsCols(A_diag, Ae_diag,
num_rowscols_to_elim, rowscols_to_elim,
num_rowscols_to_elim, rowscols_to_elim,
1, NULL);
hypre_CSRMatrixReorder(Ae_diag);
/* finish the communication */
hypre_ParCSRCommHandleDestroy(comm_handle);
/* received eliminate_col[], count offd columns to eliminate */
num_offd_cols_to_elim = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (eliminate_col[i]) {
num_offd_cols_to_elim++;
}
}
offd_cols_to_elim = hypre_CTAlloc(HYPRE_Int, num_offd_cols_to_elim);
/* get a list of offd column indices and coefs */
num_offd_cols_to_elim = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (eliminate_col[i])
{
offd_cols_to_elim[num_offd_cols_to_elim++] = i;
}
}
hypre_TFree(int_buf_data);
hypre_TFree(eliminate_row);
hypre_TFree(eliminate_col);
}
/* eliminate the off-diagonal part */
col_mark = hypre_CTAlloc(HYPRE_Int, A_offd_ncols);
col_remap = hypre_CTAlloc(HYPRE_Int, A_offd_ncols);
hypre_CSRMatrixElimCreate(A_offd, Ae_offd,
num_rowscols_to_elim, rowscols_to_elim,
num_offd_cols_to_elim, offd_cols_to_elim,
col_mark);
for (i = k = 0; i < A_offd_ncols; i++)
{
if (col_mark[i]) {
col_remap[i] = k++;
}
}
hypre_CSRMatrixEliminateRowsCols(A_offd, Ae_offd,
num_rowscols_to_elim, rowscols_to_elim,
num_offd_cols_to_elim, offd_cols_to_elim,
0, col_remap);
/* create col_map_offd for Ae */
Ae_offd_ncols = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (col_mark[i]) {
Ae_offd_ncols++;
}
}
Ae_col_map_offd = hypre_CTAlloc(HYPRE_Int, Ae_offd_ncols);
Ae_offd_ncols = 0;
for (i = 0; i < A_offd_ncols; i++)
{
if (col_mark[i])
{
Ae_col_map_offd[Ae_offd_ncols++] = A_col_map_offd[i];
}
}
hypre_ParCSRMatrixColMapOffd(*Ae) = Ae_col_map_offd;
hypre_CSRMatrixNumCols(Ae_offd) = Ae_offd_ncols;
hypre_TFree(col_remap);
hypre_TFree(col_mark);
hypre_TFree(offd_cols_to_elim);
hypre_ParCSRMatrixSetNumNonzeros(*Ae);
hypre_MatvecCommPkgCreate(*Ae);
}
HYPRE_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 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);
}
