void hypre_sort_and_create_inverse_map(
HYPRE_Int *in, HYPRE_Int len, HYPRE_Int **out, hypre_UnorderedIntMap *inverse_map)
{
if (len == 0)
{
return;
}
#ifdef HYPRE_PROFILE
hypre_profile_times[HYPRE_TIMER_ID_MERGE] -= hypre_MPI_Wtime();
#endif
HYPRE_Int *temp = hypre_TAlloc(HYPRE_Int, len);
hypre_merge_sort(in, temp, len, out);
hypre_UnorderedIntMapCreate(inverse_map, 2*len, 16*hypre_NumThreads());
HYPRE_Int i;
#pragma omp parallel for HYPRE_SMP_SCHEDULE
for (i = 0; i < len; i++)
{
HYPRE_Int old = hypre_UnorderedIntMapPutIfAbsent(inverse_map, (*out)[i], i);
assert(old == HYPRE_HOPSCOTCH_HASH_EMPTY);
#ifdef DBG_MERGE_SORT
if (hypre_UnorderedIntMapGet(inverse_map, (*out)[i]) != i)
{
fprintf(stderr, "%d %d\n", i, (*out)[i]);
assert(false);
}
#endif
}
#ifdef DBG_MERGE_SORT
std::unordered_map<HYPRE_Int, HYPRE_Int> inverse_map2(len);
for (HYPRE_Int i = 0; i < len; ++i) {
inverse_map2[(*out)[i]] = i;
if (hypre_UnorderedIntMapGet(inverse_map, (*out)[i]) != i)
{
fprintf(stderr, "%d %d\n", i, (*out)[i]);
assert(false);
}
}
assert(hypre_UnorderedIntMapSize(inverse_map) == len);
#endif
if (*out == in)
{
hypre_TFree(temp);
}
else
{
hypre_TFree(in);
}
#ifdef HYPRE_PROFILE
hypre_profile_times[HYPRE_TIMER_ID_MERGE] += hypre_MPI_Wtime();
#endif
}
HYPRE_Int *hypre_UnorderedIntSetCopyToArray( hypre_UnorderedIntSet *s, HYPRE_Int *len )
{
HYPRE_Int prefix_sum_workspace[hypre_NumThreads() + 1];
HYPRE_Int *ret_array = NULL;
#ifdef HYPRE_CONCURRENT_HOPSCOTCH
#pragma omp parallel
#endif
{
HYPRE_Int n = s->bucketMask + HYPRE_HOPSCOTCH_HASH_INSERT_RANGE;
HYPRE_Int i_begin, i_end;
hypre_GetSimpleThreadPartition(&i_begin, &i_end, n);
HYPRE_Int cnt = 0;
HYPRE_Int i;
for (i = i_begin; i < i_end; i++)
{
if (HYPRE_HOPSCOTCH_HASH_EMPTY != s->hash[i]) cnt++;
}
hypre_prefix_sum(&cnt, len, prefix_sum_workspace);
#ifdef HYPRE_CONCURRENT_HOPSCOTCH
#pragma omp barrier
#pragma omp master
#endif
{
ret_array = hypre_TAlloc(HYPRE_Int, *len);
}
#ifdef HYPRE_CONCURRENT_HOPSCOTCH
#pragma omp barrier
#endif
for (i = i_begin; i < i_end; i++)
{
if (HYPRE_HOPSCOTCH_HASH_EMPTY != s->hash[i]) ret_array[cnt++] = s->key[i];
}
}
return ret_array;
}
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);
}
int
hypre_CSRMatrixMatvecT( double alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
double beta,
hypre_Vector *y )
{
double *A_data = hypre_CSRMatrixData(A);
int *A_i = hypre_CSRMatrixI(A);
int *A_j = hypre_CSRMatrixJ(A);
int num_rows = hypre_CSRMatrixNumRows(A);
int num_cols = hypre_CSRMatrixNumCols(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
int x_size = hypre_VectorSize(x);
int y_size = hypre_VectorSize(y);
int num_vectors = hypre_VectorNumVectors(x);
int idxstride_y = hypre_VectorIndexStride(y);
int vecstride_y = hypre_VectorVectorStride(y);
int idxstride_x = hypre_VectorIndexStride(x);
int vecstride_x = hypre_VectorVectorStride(x);
double temp;
int i, i1, j, jv, jj, ns, ne, size, rest;
int num_threads;
int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. MatvecT returns ierr = 1 if
* length of X doesn't equal the number of rows of A,
* ierr = 2 if the length of Y doesn't equal the number of
* columns of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in MatvecT, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_rows != x_size)
ierr = 1;
if (num_cols != y_size)
ierr = 2;
if (num_rows != x_size && num_cols != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A^T*x
*-----------------------------------------------------------------*/
num_threads = hypre_NumThreads();
if (num_threads > 1)
{
for (i1 = 0; i1 < num_threads; i1++)
{
size = num_cols/num_threads;
rest = num_cols - size*num_threads;
if (i1 < rest)
{
ns = i1*size+i1-1;
ne = (i1+1)*size+i1+1;
}
else
{
ns = i1*size+rest-1;
ne = (i1+1)*size+rest;
}
if ( num_vectors==1 )
{
for (i = 0; i < num_rows; i++)
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
if (j > ns && j < ne)
y_data[j] += A_data[jj] * x_data[i];
}
}
}
else
{
for (i = 0; i < num_rows; i++)
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
if (j > ns && j < ne)
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x];
}
}
}
}
}
}
else
{
for (i = 0; i < num_rows; i++)
{
if ( num_vectors==1 )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[j] += A_data[jj] * x_data[i];
}
}
else
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x ];
}
}
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= alpha;
}
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;
}
int
hypre_CSRMatrixMatvecT( double alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
double beta,
hypre_Vector *y )
{
double *A_data = hypre_CSRMatrixData(A);
int *A_i = hypre_CSRMatrixI(A);
int *A_j = hypre_CSRMatrixJ(A);
int num_rows = hypre_CSRMatrixNumRows(A);
int num_cols = hypre_CSRMatrixNumCols(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
int x_size = hypre_VectorSize(x);
int y_size = hypre_VectorSize(y);
int num_vectors = hypre_VectorNumVectors(x);
int idxstride_y = hypre_VectorIndexStride(y);
int vecstride_y = hypre_VectorVectorStride(y);
int idxstride_x = hypre_VectorIndexStride(x);
int vecstride_x = hypre_VectorVectorStride(x);
double temp;
double *y_data_expand = NULL;
int offset = 0;
#ifdef HYPRE_USING_OPENMP
int my_thread_num = 0;
#endif
int i, j, jv, jj;
int num_threads;
int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. MatvecT returns ierr = 1 if
* length of X doesn't equal the number of rows of A,
* ierr = 2 if the length of Y doesn't equal the number of
* columns of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in MatvecT, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_rows != x_size)
ierr = 1;
if (num_cols != y_size)
ierr = 2;
if (num_rows != x_size && num_cols != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) schedule(static)
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) schedule(static)
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) schedule(static)
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A^T*x
*-----------------------------------------------------------------*/
num_threads = hypre_NumThreads();
if (num_threads > 1)
{
y_data_expand = hypre_CTAlloc(double, num_threads*y_size);
if ( num_vectors==1 )
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel private(i,jj,j, my_thread_num, offset)
{
my_thread_num = omp_get_thread_num();
offset = y_size*my_thread_num;
#pragma omp for schedule(static)
#endif
for (i = 0; i < num_rows; i++)
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data_expand[offset + j] += A_data[jj] * x_data[i];
}
}
#ifdef HYPRE_USING_OPENMP
/* implied barrier */
#pragma omp for schedule(static)
#endif
for (i = 0; i < y_size; i++)
{
for (j = 0; j < num_threads; j++)
{
y_data[i] += y_data_expand[j*y_size + i];
/*y_data_expand[j*y_size + i] = 0; //zero out for next time */
}
}
#ifdef HYPRE_USING_OPENMP
} /* end parallel region */
#endif
hypre_TFree(y_data_expand);
}
else
{
/* MULTIPLE VECTORS NOT THREADED YET */
for (i = 0; i < num_rows; i++)
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x];
}
}
}
}
hypre_TFree(y_data_expand);
}
HYPRE_Int
hypre_CSRMatrixMatvecT( HYPRE_Complex alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
HYPRE_Complex beta,
hypre_Vector *y )
{
HYPRE_Complex *A_data = hypre_CSRMatrixData(A);
HYPRE_Int *A_i = hypre_CSRMatrixI(A);
HYPRE_Int *A_j = hypre_CSRMatrixJ(A);
HYPRE_Int num_rows = hypre_CSRMatrixNumRows(A);
HYPRE_Int num_cols = hypre_CSRMatrixNumCols(A);
HYPRE_Complex *x_data = hypre_VectorData(x);
HYPRE_Complex *y_data = hypre_VectorData(y);
HYPRE_Int x_size = hypre_VectorSize(x);
HYPRE_Int y_size = hypre_VectorSize(y);
HYPRE_Int num_vectors = hypre_VectorNumVectors(x);
HYPRE_Int idxstride_y = hypre_VectorIndexStride(y);
HYPRE_Int vecstride_y = hypre_VectorVectorStride(y);
HYPRE_Int idxstride_x = hypre_VectorIndexStride(x);
HYPRE_Int vecstride_x = hypre_VectorVectorStride(x);
HYPRE_Complex temp;
HYPRE_Complex *y_data_expand;
HYPRE_Int my_thread_num = 0, offset = 0;
HYPRE_Int i, j, jv, jj;
HYPRE_Int num_threads;
HYPRE_Int ierr = 0;
hypre_Vector *x_tmp = NULL;
/*---------------------------------------------------------------------
* Check for size compatibility. MatvecT returns ierr = 1 if
* length of X doesn't equal the number of rows of A,
* ierr = 2 if the length of Y doesn't equal the number of
* columns of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in MatvecT, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_rows != x_size)
ierr = 1;
if (num_cols != y_size)
ierr = 2;
if (num_rows != x_size && num_cols != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
if (x == y)
{
x_tmp = hypre_SeqVectorCloneDeep(x);
x_data = hypre_VectorData(x_tmp);
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A^T*x
*-----------------------------------------------------------------*/
num_threads = hypre_NumThreads();
if (num_threads > 1)
{
y_data_expand = hypre_CTAlloc(HYPRE_Complex, num_threads*y_size);
if ( num_vectors==1 )
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel private(i,jj,j,my_thread_num,offset)
#endif
{
my_thread_num = hypre_GetThreadNum();
offset = y_size*my_thread_num;
#ifdef HYPRE_USING_OPENMP
#pragma omp for HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data_expand[offset + j] += A_data[jj] * x_data[i];
}
}
/* implied barrier (for threads)*/
#ifdef HYPRE_USING_OPENMP
#pragma omp for HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < y_size; i++)
{
for (j = 0; j < num_threads; j++)
{
y_data[i] += y_data_expand[j*y_size + i];
}
}
} /* end parallel threaded region */
}
else
{
/* multiple vector case is not threaded */
for (i = 0; i < num_rows; i++)
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x];
}
}
}
}
hypre_TFree(y_data_expand);
}
else
{
for (i = 0; i < num_rows; i++)
{
if ( num_vectors==1 )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[j] += A_data[jj] * x_data[i];
}
}
else
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x ];
}
}
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= alpha;
}
if (x == y) hypre_SeqVectorDestroy(x_tmp);
return ierr;
}
HYPRE_Int
hypre_BoomerAMGCycle( void *amg_vdata,
hypre_ParVector **F_array,
hypre_ParVector **U_array )
{
hypre_ParAMGData *amg_data = amg_vdata;
HYPRE_Solver *smoother;
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParCSRMatrix **P_array;
hypre_ParCSRMatrix **R_array;
hypre_ParVector *Utemp;
hypre_ParVector *Vtemp;
hypre_ParVector *Rtemp;
hypre_ParVector *Ptemp;
hypre_ParVector *Ztemp;
hypre_ParVector *Aux_U;
hypre_ParVector *Aux_F;
hypre_ParCSRBlockMatrix **A_block_array;
hypre_ParCSRBlockMatrix **P_block_array;
hypre_ParCSRBlockMatrix **R_block_array;
HYPRE_Real *Ztemp_data;
HYPRE_Real *Ptemp_data;
HYPRE_Int **CF_marker_array;
/* HYPRE_Int **unknown_map_array;
HYPRE_Int **point_map_array;
HYPRE_Int **v_at_point_array; */
HYPRE_Real cycle_op_count;
HYPRE_Int cycle_type;
HYPRE_Int num_levels;
HYPRE_Int max_levels;
HYPRE_Real *num_coeffs;
HYPRE_Int *num_grid_sweeps;
HYPRE_Int *grid_relax_type;
HYPRE_Int **grid_relax_points;
HYPRE_Int block_mode;
HYPRE_Real *max_eig_est;
HYPRE_Real *min_eig_est;
HYPRE_Int cheby_order;
HYPRE_Real cheby_fraction;
/* Local variables */
HYPRE_Int *lev_counter;
HYPRE_Int Solve_err_flag;
HYPRE_Int k;
HYPRE_Int i, j, jj;
HYPRE_Int level;
HYPRE_Int cycle_param;
HYPRE_Int coarse_grid;
HYPRE_Int fine_grid;
HYPRE_Int Not_Finished;
HYPRE_Int num_sweep;
HYPRE_Int cg_num_sweep = 1;
HYPRE_Int relax_type;
HYPRE_Int relax_points;
HYPRE_Int relax_order;
HYPRE_Int relax_local;
HYPRE_Int old_version = 0;
HYPRE_Real *relax_weight;
HYPRE_Real *omega;
HYPRE_Real alfa, beta, gammaold;
HYPRE_Real gamma = 1.0;
HYPRE_Int local_size;
/* HYPRE_Int *smooth_option; */
HYPRE_Int smooth_type;
HYPRE_Int smooth_num_levels;
HYPRE_Int num_threads, my_id;
HYPRE_Real alpha;
HYPRE_Real **l1_norms = NULL;
HYPRE_Real *l1_norms_level;
HYPRE_Int seq_cg = 0;
MPI_Comm comm;
#if 0
HYPRE_Real *D_mat;
HYPRE_Real *S_vec;
#endif
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
P_array = hypre_ParAMGDataPArray(amg_data);
R_array = hypre_ParAMGDataRArray(amg_data);
CF_marker_array = hypre_ParAMGDataCFMarkerArray(amg_data);
Vtemp = hypre_ParAMGDataVtemp(amg_data);
Rtemp = hypre_ParAMGDataRtemp(amg_data);
Ptemp = hypre_ParAMGDataPtemp(amg_data);
Ztemp = hypre_ParAMGDataZtemp(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
max_levels = hypre_ParAMGDataMaxLevels(amg_data);
cycle_type = hypre_ParAMGDataCycleType(amg_data);
A_block_array = hypre_ParAMGDataABlockArray(amg_data);
P_block_array = hypre_ParAMGDataPBlockArray(amg_data);
R_block_array = hypre_ParAMGDataRBlockArray(amg_data);
block_mode = hypre_ParAMGDataBlockMode(amg_data);
num_grid_sweeps = hypre_ParAMGDataNumGridSweeps(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
grid_relax_points = hypre_ParAMGDataGridRelaxPoints(amg_data);
relax_order = hypre_ParAMGDataRelaxOrder(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
omega = hypre_ParAMGDataOmega(amg_data);
smooth_type = hypre_ParAMGDataSmoothType(amg_data);
smooth_num_levels = hypre_ParAMGDataSmoothNumLevels(amg_data);
l1_norms = hypre_ParAMGDataL1Norms(amg_data);
/* smooth_option = hypre_ParAMGDataSmoothOption(amg_data); */
max_eig_est = hypre_ParAMGDataMaxEigEst(amg_data);
min_eig_est = hypre_ParAMGDataMinEigEst(amg_data);
cheby_order = hypre_ParAMGDataChebyOrder(amg_data);
cheby_fraction = hypre_ParAMGDataChebyFraction(amg_data);
cycle_op_count = hypre_ParAMGDataCycleOpCount(amg_data);
lev_counter = hypre_CTAlloc(HYPRE_Int, num_levels);
if (hypre_ParAMGDataParticipate(amg_data)) seq_cg = 1;
/* Initialize */
Solve_err_flag = 0;
if (grid_relax_points) old_version = 1;
num_coeffs = hypre_CTAlloc(HYPRE_Real, num_levels);
num_coeffs[0] = hypre_ParCSRMatrixDNumNonzeros(A_array[0]);
comm = hypre_ParCSRMatrixComm(A_array[0]);
hypre_MPI_Comm_rank(comm,&my_id);
if (block_mode)
{
for (j = 1; j < num_levels; j++)
num_coeffs[j] = hypre_ParCSRBlockMatrixNumNonzeros(A_block_array[j]);
}
else
{
for (j = 1; j < num_levels; j++)
num_coeffs[j] = hypre_ParCSRMatrixDNumNonzeros(A_array[j]);
}
/*---------------------------------------------------------------------
* Initialize cycling control counter
*
* Cycling is controlled using a level counter: lev_counter[k]
*
* Each time relaxation is performed on level k, the
* counter is decremented by 1. If the counter is then
* negative, we go to the next finer level. If non-
* negative, we go to the next coarser level. The
* following actions control cycling:
*
* a. lev_counter[0] is initialized to 1.
* b. lev_counter[k] is initialized to cycle_type for k>0.
*
* c. During cycling, when going down to level k, lev_counter[k]
* is set to the max of (lev_counter[k],cycle_type)
*---------------------------------------------------------------------*/
Not_Finished = 1;
lev_counter[0] = 1;
for (k = 1; k < num_levels; ++k)
{
lev_counter[k] = cycle_type;
}
level = 0;
cycle_param = 1;
smoother = hypre_ParAMGDataSmoother(amg_data);
if (smooth_num_levels > 0)
{
if (smooth_type == 7 || smooth_type == 8
|| smooth_type == 17 || smooth_type == 18
|| smooth_type == 9 || smooth_type == 19)
{
HYPRE_Int actual_local_size = hypre_ParVectorActualLocalSize(Vtemp);
Utemp = hypre_ParVectorCreate(comm,hypre_ParVectorGlobalSize(Vtemp),
hypre_ParVectorPartitioning(Vtemp));
hypre_ParVectorOwnsPartitioning(Utemp) = 0;
local_size
= hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp));
if (local_size < actual_local_size)
{
hypre_VectorData(hypre_ParVectorLocalVector(Utemp)) =
hypre_CTAlloc(HYPRE_Complex, actual_local_size);
hypre_ParVectorActualLocalSize(Utemp) = actual_local_size;
}
else
hypre_ParVectorInitialize(Utemp);
}
}
/*---------------------------------------------------------------------
* Main loop of cycling
*--------------------------------------------------------------------*/
while (Not_Finished)
{
if (num_levels > 1)
{
local_size
= hypre_VectorSize(hypre_ParVectorLocalVector(F_array[level]));
hypre_VectorSize(hypre_ParVectorLocalVector(Vtemp)) = local_size;
if (smooth_num_levels <= level)
{
cg_num_sweep = 1;
num_sweep = num_grid_sweeps[cycle_param];
Aux_U = U_array[level];
Aux_F = F_array[level];
}
else if (smooth_type > 9)
{
hypre_VectorSize(hypre_ParVectorLocalVector(Ztemp)) = local_size;
hypre_VectorSize(hypre_ParVectorLocalVector(Rtemp)) = local_size;
hypre_VectorSize(hypre_ParVectorLocalVector(Ptemp)) = local_size;
Ztemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ztemp));
Ptemp_data = hypre_VectorData(hypre_ParVectorLocalVector(Ptemp));
hypre_ParVectorSetConstantValues(Ztemp,0);
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level],
U_array[level], beta, F_array[level], Rtemp);
cg_num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data);
num_sweep = num_grid_sweeps[cycle_param];
Aux_U = Ztemp;
Aux_F = Rtemp;
}
else
{
cg_num_sweep = 1;
num_sweep = hypre_ParAMGDataSmoothNumSweeps(amg_data);
Aux_U = U_array[level];
Aux_F = F_array[level];
}
relax_type = grid_relax_type[cycle_param];
}
else /* AB: 4/08: removed the max_levels > 1 check - should do this when max-levels = 1 also */
{
/* If no coarsening occurred, apply a simple smoother once */
Aux_U = U_array[level];
Aux_F = F_array[level];
num_sweep = 1;
/* TK: Use the user relax type (instead of 0) to allow for setting a
convergent smoother (e.g. in the solution of singular problems). */
relax_type = hypre_ParAMGDataUserRelaxType(amg_data);
}
if (l1_norms != NULL)
l1_norms_level = l1_norms[level];
else
l1_norms_level = NULL;
if (cycle_param == 3 && seq_cg)
{
hypre_seqAMGCycle(amg_data, level, F_array, U_array);
}
else
{
/*------------------------------------------------------------------
* Do the relaxation num_sweep times
*-----------------------------------------------------------------*/
for (jj = 0; jj < cg_num_sweep; jj++)
{
if (smooth_num_levels > level && smooth_type > 9)
hypre_ParVectorSetConstantValues(Aux_U,0);
for (j = 0; j < num_sweep; j++)
{
if (num_levels == 1 && max_levels > 1)
{
relax_points = 0;
relax_local = 0;
}
else
{
if (old_version)
relax_points = grid_relax_points[cycle_param][j];
relax_local = relax_order;
}
/*-----------------------------------------------
* VERY sloppy approximation to cycle complexity
*-----------------------------------------------*/
if (old_version && level < num_levels -1)
{
switch (relax_points)
{
case 1:
cycle_op_count += num_coeffs[level+1];
break;
case -1:
cycle_op_count += (num_coeffs[level]-num_coeffs[level+1]);
break;
}
}
else
{
cycle_op_count += num_coeffs[level];
}
/*-----------------------------------------------
Choose Smoother
-----------------------------------------------*/
if (smooth_num_levels > level &&
(smooth_type == 7 || smooth_type == 8 ||
smooth_type == 9 || smooth_type == 19 ||
smooth_type == 17 || smooth_type == 18))
{
hypre_VectorSize(hypre_ParVectorLocalVector(Utemp)) = local_size;
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[level],
U_array[level], beta, Aux_F, Vtemp);
if (smooth_type == 8 || smooth_type == 18)
HYPRE_ParCSRParaSailsSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
else if (smooth_type == 7 || smooth_type == 17)
HYPRE_ParCSRPilutSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
else if (smooth_type == 9 || smooth_type == 19)
HYPRE_EuclidSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Vtemp,
(HYPRE_ParVector) Utemp);
hypre_ParVectorAxpy(relax_weight[level],Utemp,Aux_U);
}
else if (smooth_num_levels > level &&
(smooth_type == 6 || smooth_type == 16))
{
HYPRE_SchwarzSolve(smoother[level],
(HYPRE_ParCSRMatrix) A_array[level],
(HYPRE_ParVector) Aux_F,
(HYPRE_ParVector) Aux_U);
}
/*else if (relax_type == 99)*/
else if (relax_type == 9 || relax_type == 99)
{ /* Gaussian elimination */
hypre_GaussElimSolve(amg_data, level, relax_type);
}
else if (relax_type == 18)
{ /* L1 - Jacobi*/
if (relax_order == 1 && cycle_param < 3)
{
/* need to do CF - so can't use the AMS one */
HYPRE_Int i;
HYPRE_Int loc_relax_points[2];
if (cycle_type < 2)
{
loc_relax_points[0] = 1;
loc_relax_points[1] = -1;
}
else
{
loc_relax_points[0] = -1;
loc_relax_points[1] = 1;
}
for (i=0; i < 2; i++)
hypre_ParCSRRelax_L1_Jacobi(A_array[level],
Aux_F,
CF_marker_array[level],
loc_relax_points[i],
relax_weight[level],
l1_norms[level],
Aux_U,
Vtemp);
}
else /* not CF - so use through AMS */
{
if (num_threads == 1)
hypre_ParCSRRelax(A_array[level],
Aux_F,
1,
1,
l1_norms_level,
relax_weight[level],
omega[level],0,0,0,0,
Aux_U,
Vtemp,
Ztemp);
else
hypre_ParCSRRelaxThreads(A_array[level],
Aux_F,
1,
1,
l1_norms_level,
relax_weight[level],
omega[level],
Aux_U,
Vtemp,
Ztemp);
}
}
else if (relax_type == 15)
{ /* CG */
if (j ==0) /* do num sweep iterations of CG */
hypre_ParCSRRelax_CG( smoother[level],
A_array[level],
Aux_F,
Aux_U,
num_sweep);
}
else if (relax_type == 16)
{ /* scaled Chebyshev */
HYPRE_Int scale = 1;
HYPRE_Int variant = 0;
hypre_ParCSRRelax_Cheby(A_array[level],
Aux_F,
max_eig_est[level],
min_eig_est[level],
cheby_fraction, cheby_order, scale,
variant, Aux_U, Vtemp, Ztemp );
}
else if (relax_type ==17)
{
hypre_BoomerAMGRelax_FCFJacobi(A_array[level],
Aux_F,
CF_marker_array[level],
relax_weight[level],
Aux_U,
Vtemp);
}
else if (old_version)
{
Solve_err_flag = hypre_BoomerAMGRelax(A_array[level],
Aux_F,
CF_marker_array[level],
relax_type, relax_points,
relax_weight[level],
omega[level],
l1_norms_level,
Aux_U,
Vtemp,
Ztemp);
}
else
{
/* smoother than can have CF ordering */
if (block_mode)
{
Solve_err_flag = hypre_BoomerAMGBlockRelaxIF(A_block_array[level],
Aux_F,
CF_marker_array[level],
relax_type,
relax_local,
cycle_param,
relax_weight[level],
omega[level],
Aux_U,
Vtemp);
}
else
{
Solve_err_flag = hypre_BoomerAMGRelaxIF(A_array[level],
Aux_F,
CF_marker_array[level],
relax_type,
relax_local,
cycle_param,
relax_weight[level],
omega[level],
l1_norms_level,
Aux_U,
Vtemp,
Ztemp);
}
}
if (Solve_err_flag != 0)
return(Solve_err_flag);
}
if (smooth_num_levels > level && smooth_type > 9)
{
gammaold = gamma;
gamma = hypre_ParVectorInnerProd(Rtemp,Ztemp);
if (jj == 0)
hypre_ParVectorCopy(Ztemp,Ptemp);
else
{
beta = gamma/gammaold;
for (i=0; i < local_size; i++)
Ptemp_data[i] = Ztemp_data[i] + beta*Ptemp_data[i];
}
hypre_ParCSRMatrixMatvec(1.0,A_array[level],Ptemp,0.0,Vtemp);
alfa = gamma /hypre_ParVectorInnerProd(Ptemp,Vtemp);
hypre_ParVectorAxpy(alfa,Ptemp,U_array[level]);
hypre_ParVectorAxpy(-alfa,Vtemp,Rtemp);
}
}
}
/*------------------------------------------------------------------
* Decrement the control counter and determine which grid to visit next
*-----------------------------------------------------------------*/
--lev_counter[level];
if (lev_counter[level] >= 0 && level != num_levels-1)
{
/*---------------------------------------------------------------
* Visit coarser level next.
* Compute residual using hypre_ParCSRMatrixMatvec.
* Perform restriction using hypre_ParCSRMatrixMatvecT.
* Reset counters and cycling parameters for coarse level
*--------------------------------------------------------------*/
fine_grid = level;
coarse_grid = level + 1;
hypre_ParVectorSetConstantValues(U_array[coarse_grid], 0.0);
alpha = -1.0;
beta = 1.0;
if (block_mode)
{
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
hypre_ParCSRBlockMatrixMatvec(alpha, A_block_array[fine_grid], U_array[fine_grid],
beta, Vtemp);
}
else
{
// JSP: avoid unnecessary copy using out-of-place version of SpMV
hypre_ParCSRMatrixMatvecOutOfPlace(alpha, A_array[fine_grid], U_array[fine_grid],
beta, F_array[fine_grid], Vtemp);
}
alpha = 1.0;
beta = 0.0;
if (block_mode)
{
hypre_ParCSRBlockMatrixMatvecT(alpha,R_block_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
else
{
hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
++level;
lev_counter[level] = hypre_max(lev_counter[level],cycle_type);
cycle_param = 1;
if (level == num_levels-1) cycle_param = 3;
}
else if (level != 0)
{
/*---------------------------------------------------------------
* Visit finer level next.
* Interpolate and add correction using hypre_ParCSRMatrixMatvec.
* Reset counters and cycling parameters for finer level.
*--------------------------------------------------------------*/
fine_grid = level - 1;
coarse_grid = level;
alpha = 1.0;
beta = 1.0;
if (block_mode)
{
hypre_ParCSRBlockMatrixMatvec(alpha, P_block_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
else
{
hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
--level;
cycle_param = 2;
}
else
{
Not_Finished = 0;
}
}
hypre_ParAMGDataCycleOpCount(amg_data) = cycle_op_count;
hypre_TFree(lev_counter);
hypre_TFree(num_coeffs);
if (smooth_num_levels > 0)
{
if (smooth_type == 7 || smooth_type == 8 || smooth_type == 9 ||
smooth_type == 17 || smooth_type == 18 || smooth_type == 19 )
hypre_ParVectorDestroy(Utemp);
}
return(Solve_err_flag);
}
HYPRE_Int hypre_CreateDinv(void *amg_vdata)
{
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
hypre_ParCSRMatrix *A_tmp;
hypre_CSRMatrix *A_tmp_diag;
hypre_ParVector *Xtilde;
hypre_ParVector *Rtilde;
hypre_Vector *Xtilde_local;
hypre_Vector *Rtilde_local;
HYPRE_Real *x_data;
HYPRE_Real *r_data;
HYPRE_Real *tmp_data;
HYPRE_Real *D_inv = NULL;
HYPRE_Real *relax_weight = NULL;
HYPRE_Real relax_type;
HYPRE_Int addlvl;
HYPRE_Int num_levels;
HYPRE_Int num_add_lvls;
HYPRE_Int num_rows_L;
HYPRE_Int num_rows_A;
HYPRE_Int num_rows_tmp;
HYPRE_Int level, i;
/* Local variables */
HYPRE_Int Solve_err_flag = 0;
HYPRE_Int num_threads;
HYPRE_Real **l1_norms_ptr = NULL;
HYPRE_Real *l1_norms;
HYPRE_Int l1_start;
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
addlvl = hypre_ParAMGDataSimple(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
relax_type = hypre_ParAMGDataGridRelaxType(amg_data)[1];
num_rows_A = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A_array[0]));
l1_norms_ptr = hypre_ParAMGDataL1Norms(amg_data);
/* smooth_option = hypre_ParAMGDataSmoothOption(amg_data); */
num_add_lvls = num_levels+1-addlvl;
num_rows_L = 0;
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
A_tmp_diag = hypre_ParCSRMatrixDiag(A_tmp);
num_rows_tmp = hypre_CSRMatrixNumRows(A_tmp_diag);
num_rows_L += num_rows_tmp;
}
Rtilde = hypre_CTAlloc(hypre_ParVector, 1);
Rtilde_local = hypre_SeqVectorCreate(num_rows_L);
hypre_SeqVectorInitialize(Rtilde_local);
hypre_ParVectorLocalVector(Rtilde) = Rtilde_local;
hypre_ParVectorOwnsData(Rtilde) = 1;
Xtilde = hypre_CTAlloc(hypre_ParVector, 1);
Xtilde_local = hypre_SeqVectorCreate(num_rows_L);
hypre_SeqVectorInitialize(Xtilde_local);
hypre_ParVectorLocalVector(Xtilde) = Xtilde_local;
hypre_ParVectorOwnsData(Xtilde) = 1;
x_data = hypre_VectorData(hypre_ParVectorLocalVector(Xtilde));
r_data = hypre_VectorData(hypre_ParVectorLocalVector(Rtilde));
D_inv = hypre_CTAlloc(HYPRE_Real, num_rows_L);
l1_start = 0;
for (level=addlvl; level < num_levels; level++)
{
if (level != 0)
{
tmp_data = hypre_VectorData(hypre_ParVectorLocalVector(F_array[level]));
if (tmp_data) hypre_TFree(tmp_data);
hypre_VectorData(hypre_ParVectorLocalVector(F_array[level])) = &r_data[l1_start];
hypre_VectorOwnsData(hypre_ParVectorLocalVector(F_array[level])) = 0;
tmp_data = hypre_VectorData(hypre_ParVectorLocalVector(U_array[level]));
if (tmp_data) hypre_TFree(tmp_data);
hypre_VectorData(hypre_ParVectorLocalVector(U_array[level])) = &x_data[l1_start];
hypre_VectorOwnsData(hypre_ParVectorLocalVector(U_array[level])) = 0;
}
A_tmp = A_array[level];
A_tmp_diag = hypre_ParCSRMatrixDiag(A_tmp);
num_rows_tmp = hypre_CSRMatrixNumRows(A_tmp_diag);
if (relax_type == 0)
{
HYPRE_Real rlx_wt = relax_weight[level];
HYPRE_Int *A_tmp_diag_i = hypre_CSRMatrixI(A_tmp_diag);
HYPRE_Real *A_tmp_diag_data = hypre_CSRMatrixData(A_tmp_diag);
#ifdef HYPRE_USING_OPENMP
#pragma omp for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < num_rows_tmp; i++)
D_inv[l1_start+i] = rlx_wt/A_tmp_diag_data[A_tmp_diag_i[i]];
}
else
{
l1_norms = l1_norms_ptr[level];
#ifdef HYPRE_USING_OPENMP
#pragma omp for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < num_rows_tmp; i++)
D_inv[l1_start+i] = 1.0/l1_norms[i];
}
l1_start += num_rows_tmp;
}
hypre_ParAMGDataDinv(amg_data) = D_inv;
hypre_ParAMGDataRtilde(amg_data) = Rtilde;
hypre_ParAMGDataXtilde(amg_data) = Xtilde;
return Solve_err_flag;
}
HYPRE_Int
hypre_BoomerAMGAdditiveCycle( void *amg_vdata)
{
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParCSRMatrix **P_array;
hypre_ParCSRMatrix **R_array;
hypre_ParCSRMatrix *Lambda;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
hypre_ParVector *Vtemp;
hypre_ParVector *Ztemp;
hypre_ParVector *Xtilde, *Rtilde;
HYPRE_Int **CF_marker_array;
HYPRE_Int num_levels;
HYPRE_Int addlvl;
HYPRE_Int additive;
HYPRE_Int mult_additive;
HYPRE_Int simple;
HYPRE_Int i, num_rows;
HYPRE_Int n_global;
HYPRE_Int rlx_order;
/* Local variables */
HYPRE_Int Solve_err_flag = 0;
HYPRE_Int level;
HYPRE_Int coarse_grid;
HYPRE_Int fine_grid;
HYPRE_Int relax_type;
HYPRE_Int rlx_down;
HYPRE_Int rlx_up;
HYPRE_Int *grid_relax_type;
HYPRE_Real **l1_norms;
HYPRE_Real alpha, beta;
HYPRE_Int num_threads;
HYPRE_Real *u_data;
HYPRE_Real *f_data;
HYPRE_Real *v_data;
HYPRE_Real *l1_norms_lvl;
HYPRE_Real *D_inv;
HYPRE_Real *x_global;
HYPRE_Real *r_global;
HYPRE_Real *relax_weight;
HYPRE_Real *omega;
#if 0
HYPRE_Real *D_mat;
HYPRE_Real *S_vec;
#endif
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
P_array = hypre_ParAMGDataPArray(amg_data);
R_array = hypre_ParAMGDataRArray(amg_data);
CF_marker_array = hypre_ParAMGDataCFMarkerArray(amg_data);
Vtemp = hypre_ParAMGDataVtemp(amg_data);
Ztemp = hypre_ParAMGDataZtemp(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
additive = hypre_ParAMGDataAdditive(amg_data);
mult_additive = hypre_ParAMGDataMultAdditive(amg_data);
simple = hypre_ParAMGDataSimple(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
Lambda = hypre_ParAMGDataLambda(amg_data);
Xtilde = hypre_ParAMGDataXtilde(amg_data);
Rtilde = hypre_ParAMGDataRtilde(amg_data);
l1_norms = hypre_ParAMGDataL1Norms(amg_data);
D_inv = hypre_ParAMGDataDinv(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
omega = hypre_ParAMGDataOmega(amg_data);
rlx_order = hypre_ParAMGDataRelaxOrder(amg_data);
/* Initialize */
addlvl = hypre_max(additive, mult_additive);
addlvl = hypre_max(addlvl, simple);
Solve_err_flag = 0;
/*---------------------------------------------------------------------
* Main loop of cycling --- multiplicative version --- V-cycle
*--------------------------------------------------------------------*/
/* down cycle */
relax_type = grid_relax_type[1];
rlx_down = grid_relax_type[1];
rlx_up = grid_relax_type[2];
for (level = 0; level < num_levels-1; level++)
{
fine_grid = level;
coarse_grid = level + 1;
u_data = hypre_VectorData(hypre_ParVectorLocalVector(U_array[fine_grid]));
f_data = hypre_VectorData(hypre_ParVectorLocalVector(F_array[fine_grid]));
v_data = hypre_VectorData(hypre_ParVectorLocalVector(Vtemp));
l1_norms_lvl = l1_norms[level];
hypre_ParVectorSetConstantValues(U_array[coarse_grid], 0.0);
if (level < addlvl) /* multiplicative version */
{
/* smoothing step */
if (rlx_down == 0)
{
HYPRE_Real *A_data = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
HYPRE_Int *A_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
num_rows = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
u_data[i] = relax_weight[level]*v_data[i] / A_data[A_i[i]];
}
else if (rlx_down != 18)
{
/*hypre_BoomerAMGRelax(A_array[fine_grid],F_array[fine_grid],NULL,rlx_down,0,*/
hypre_BoomerAMGRelaxIF(A_array[fine_grid],F_array[fine_grid],
CF_marker_array[fine_grid], rlx_down,rlx_order,1,
relax_weight[fine_grid], omega[fine_grid],
l1_norms_lvl, U_array[fine_grid], Vtemp, Ztemp);
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
}
else
{
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
num_rows = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(A_array[fine_grid]));
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i = 0; i < num_rows; i++)
u_data[i] += v_data[i] / l1_norms_lvl[i];
}
alpha = -1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvec(alpha, A_array[fine_grid], U_array[fine_grid],
beta, Vtemp);
alpha = 1.0;
beta = 0.0;
hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
else /* additive version */
{
hypre_ParVectorCopy(F_array[fine_grid],Vtemp);
if (level == 0) /* compute residual */
{
hypre_ParVectorCopy(Vtemp, Rtilde);
hypre_ParVectorCopy(U_array[fine_grid],Xtilde);
}
alpha = 1.0;
beta = 0.0;
hypre_ParCSRMatrixMatvecT(alpha,R_array[fine_grid],Vtemp,
beta,F_array[coarse_grid]);
}
}
/* solve coarse grid */
if (addlvl < num_levels)
{
if (simple > -1)
{
x_global = hypre_VectorData(hypre_ParVectorLocalVector(Xtilde));
r_global = hypre_VectorData(hypre_ParVectorLocalVector(Rtilde));
n_global = hypre_VectorSize(hypre_ParVectorLocalVector(Xtilde));
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < n_global; i++)
x_global[i] += D_inv[i]*r_global[i];
}
else
hypre_ParCSRMatrixMatvec(1.0, Lambda, Rtilde, 1.0, Xtilde);
if (addlvl == 0) hypre_ParVectorCopy(Xtilde, U_array[0]);
}
else
{
fine_grid = num_levels -1;
hypre_ParCSRRelax(A_array[fine_grid], F_array[fine_grid],
1, 1, l1_norms[fine_grid],
1.0, 1.0 ,0,0,0,0,
U_array[fine_grid], Vtemp, Ztemp);
}
/* up cycle */
relax_type = grid_relax_type[2];
for (level = num_levels-1; level > 0; level--)
{
fine_grid = level - 1;
coarse_grid = level;
if (level <= addlvl) /* multiplicative version */
{
alpha = 1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
if (rlx_up != 18)
/*hypre_BoomerAMGRelax(A_array[fine_grid],F_array[fine_grid],NULL,rlx_up,0,*/
hypre_BoomerAMGRelaxIF(A_array[fine_grid],F_array[fine_grid],
CF_marker_array[fine_grid],
rlx_up,rlx_order,2,
relax_weight[fine_grid], omega[fine_grid],
l1_norms[fine_grid], U_array[fine_grid], Vtemp, Ztemp);
else if (rlx_order)
{
HYPRE_Int loc_relax_points[2];
loc_relax_points[0] = -1;
loc_relax_points[1] = 1;
for (i=0; i < 2; i++)
hypre_ParCSRRelax_L1_Jacobi(A_array[fine_grid],F_array[fine_grid],
CF_marker_array[fine_grid],
loc_relax_points[i],
1.0, l1_norms[fine_grid],
U_array[fine_grid], Vtemp);
}
else
hypre_ParCSRRelax(A_array[fine_grid], F_array[fine_grid],
1, 1, l1_norms[fine_grid],
1.0, 1.0 ,0,0,0,0,
U_array[fine_grid], Vtemp, Ztemp);
}
else /* additive version */
{
alpha = 1.0;
beta = 1.0;
hypre_ParCSRMatrixMatvec(alpha, P_array[fine_grid],
U_array[coarse_grid],
beta, U_array[fine_grid]);
}
}
return(Solve_err_flag);
}
HYPRE_Int hypre_CreateLambda(void *amg_vdata)
{
hypre_ParAMGData *amg_data = amg_vdata;
/* Data Structure variables */
MPI_Comm comm;
hypre_ParCSRMatrix **A_array;
hypre_ParVector **F_array;
hypre_ParVector **U_array;
hypre_ParCSRMatrix *A_tmp;
hypre_ParCSRMatrix *Lambda;
hypre_CSRMatrix *L_diag;
hypre_CSRMatrix *L_offd;
hypre_CSRMatrix *A_tmp_diag;
hypre_CSRMatrix *A_tmp_offd;
hypre_ParVector *Xtilde;
hypre_ParVector *Rtilde;
hypre_Vector *Xtilde_local;
hypre_Vector *Rtilde_local;
hypre_ParCSRCommPkg *comm_pkg;
hypre_ParCSRCommPkg *L_comm_pkg = NULL;
hypre_ParCSRCommHandle *comm_handle;
HYPRE_Real *L_diag_data;
HYPRE_Real *L_offd_data;
HYPRE_Real *buf_data = NULL;
HYPRE_Real *tmp_data;
HYPRE_Real *x_data;
HYPRE_Real *r_data;
HYPRE_Real *l1_norms;
HYPRE_Real *A_tmp_diag_data;
HYPRE_Real *A_tmp_offd_data;
HYPRE_Real *D_data = NULL;
HYPRE_Real *D_data_offd = NULL;
HYPRE_Int *L_diag_i;
HYPRE_Int *L_diag_j;
HYPRE_Int *L_offd_i;
HYPRE_Int *L_offd_j;
HYPRE_Int *A_tmp_diag_i;
HYPRE_Int *A_tmp_offd_i;
HYPRE_Int *A_tmp_diag_j;
HYPRE_Int *A_tmp_offd_j;
HYPRE_Int *L_recv_ptr = NULL;
HYPRE_Int *L_send_ptr = NULL;
HYPRE_Int *L_recv_procs = NULL;
HYPRE_Int *L_send_procs = NULL;
HYPRE_Int *L_send_map_elmts = NULL;
HYPRE_Int *recv_procs;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_elmts;
HYPRE_Int *send_map_starts;
HYPRE_Int *recv_vec_starts;
HYPRE_Int *all_send_procs = NULL;
HYPRE_Int *all_recv_procs = NULL;
HYPRE_Int *remap = NULL;
HYPRE_Int *level_start;
HYPRE_Int addlvl;
HYPRE_Int additive;
HYPRE_Int mult_additive;
HYPRE_Int num_levels;
HYPRE_Int num_add_lvls;
HYPRE_Int num_procs;
HYPRE_Int num_sends, num_recvs;
HYPRE_Int num_sends_L = 0;
HYPRE_Int num_recvs_L = 0;
HYPRE_Int send_data_L = 0;
HYPRE_Int num_rows_L = 0;
HYPRE_Int num_rows_tmp = 0;
HYPRE_Int num_cols_offd_L = 0;
HYPRE_Int num_cols_offd = 0;
HYPRE_Int level, i, j, k;
HYPRE_Int this_proc, cnt, cnt_diag, cnt_offd;
HYPRE_Int cnt_recv, cnt_send, cnt_row, row_start;
HYPRE_Int start_diag, start_offd, indx, cnt_map;
HYPRE_Int start, j_indx, index, cnt_level;
HYPRE_Int max_sends, max_recvs;
/* Local variables */
HYPRE_Int Solve_err_flag = 0;
HYPRE_Int num_threads;
HYPRE_Int num_nonzeros_diag;
HYPRE_Int num_nonzeros_offd;
HYPRE_Real **l1_norms_ptr = NULL;
HYPRE_Real *relax_weight = NULL;
HYPRE_Real relax_type;
/* Acquire data and allocate storage */
num_threads = hypre_NumThreads();
A_array = hypre_ParAMGDataAArray(amg_data);
F_array = hypre_ParAMGDataFArray(amg_data);
U_array = hypre_ParAMGDataUArray(amg_data);
additive = hypre_ParAMGDataAdditive(amg_data);
mult_additive = hypre_ParAMGDataMultAdditive(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
relax_type = hypre_ParAMGDataGridRelaxType(amg_data)[1];
comm = hypre_ParCSRMatrixComm(A_array[0]);
hypre_MPI_Comm_size(comm,&num_procs);
l1_norms_ptr = hypre_ParAMGDataL1Norms(amg_data);
addlvl = hypre_max(additive, mult_additive);
num_add_lvls = num_levels+1-addlvl;
level_start = hypre_CTAlloc(HYPRE_Int, num_add_lvls+1);
send_data_L = 0;
num_rows_L = 0;
num_cols_offd_L = 0;
num_nonzeros_diag = 0;
num_nonzeros_offd = 0;
level_start[0] = 0;
cnt = 1;
max_sends = 0;
max_recvs = 0;
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
A_tmp_diag = hypre_ParCSRMatrixDiag(A_tmp);
A_tmp_offd = hypre_ParCSRMatrixOffd(A_tmp);
A_tmp_diag_i = hypre_CSRMatrixI(A_tmp_diag);
A_tmp_offd_i = hypre_CSRMatrixI(A_tmp_offd);
num_rows_tmp = hypre_CSRMatrixNumRows(A_tmp_diag);
num_cols_offd = hypre_CSRMatrixNumCols(A_tmp_offd);
num_rows_L += num_rows_tmp;
level_start[cnt] = level_start[cnt-1] + num_rows_tmp;
cnt++;
num_cols_offd_L += num_cols_offd;
num_nonzeros_diag += A_tmp_diag_i[num_rows_tmp];
num_nonzeros_offd += A_tmp_offd_i[num_rows_tmp];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
max_sends += num_sends;
if (num_sends)
send_data_L += hypre_ParCSRCommPkgSendMapStart(comm_pkg,num_sends);
max_recvs += hypre_ParCSRCommPkgNumRecvs(comm_pkg);
}
}
if (max_sends >= num_procs ||max_recvs >= num_procs)
{
max_sends = num_procs;
max_recvs = num_procs;
}
if (max_sends) all_send_procs = hypre_CTAlloc(HYPRE_Int, max_sends);
if (max_recvs) all_recv_procs = hypre_CTAlloc(HYPRE_Int, max_recvs);
cnt_send = 0;
cnt_recv = 0;
if (max_sends || max_recvs)
{
if (max_sends < num_procs && max_recvs < num_procs)
{
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
for (j = 0; j < num_sends; j++)
all_send_procs[cnt_send++] = send_procs[j];
for (j = 0; j < num_recvs; j++)
all_recv_procs[cnt_recv++] = recv_procs[j];
}
}
if (max_sends)
{
qsort0(all_send_procs, 0, max_sends-1);
num_sends_L = 1;
this_proc = all_send_procs[0];
for (i=1; i < max_sends; i++)
{
if (all_send_procs[i] > this_proc)
{
this_proc = all_send_procs[i];
all_send_procs[num_sends_L++] = this_proc;
}
}
L_send_procs = hypre_CTAlloc(HYPRE_Int, num_sends_L);
for (j=0; j < num_sends_L; j++)
L_send_procs[j] = all_send_procs[j];
hypre_TFree(all_send_procs);
}
if (max_recvs)
{
qsort0(all_recv_procs, 0, max_recvs-1);
num_recvs_L = 1;
this_proc = all_recv_procs[0];
for (i=1; i < max_recvs; i++)
{
if (all_recv_procs[i] > this_proc)
{
this_proc = all_recv_procs[i];
all_recv_procs[num_recvs_L++] = this_proc;
}
}
L_recv_procs = hypre_CTAlloc(HYPRE_Int, num_recvs_L);
for (j=0; j < num_recvs_L; j++)
L_recv_procs[j] = all_recv_procs[j];
hypre_TFree(all_recv_procs);
}
L_recv_ptr = hypre_CTAlloc(HYPRE_Int, num_recvs_L+1);
L_send_ptr = hypre_CTAlloc(HYPRE_Int, num_sends_L+1);
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
}
else
{
num_sends = 0;
num_recvs = 0;
}
for (k = 0; k < num_sends; k++)
{
this_proc = hypre_BinarySearch(L_send_procs,send_procs[k],num_sends_L);
L_send_ptr[this_proc+1] += send_map_starts[k+1]-send_map_starts[k];
}
for (k = 0; k < num_recvs; k++)
{
this_proc = hypre_BinarySearch(L_recv_procs,recv_procs[k],num_recvs_L);
L_recv_ptr[this_proc+1] += recv_vec_starts[k+1]-recv_vec_starts[k];
}
}
L_recv_ptr[0] = 0;
for (i=1; i < num_recvs_L; i++)
L_recv_ptr[i+1] += L_recv_ptr[i];
L_send_ptr[0] = 0;
for (i=1; i < num_sends_L; i++)
L_send_ptr[i+1] += L_send_ptr[i];
}
else
{
num_recvs_L = 0;
num_sends_L = 0;
for (i=addlvl; i < num_levels; i++)
{
A_tmp = A_array[i];
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
for (j = 0; j < num_sends; j++)
{
this_proc = send_procs[j];
if (all_send_procs[this_proc] == 0)
num_sends_L++;
all_send_procs[this_proc] += send_map_starts[j+1]-send_map_starts[j];
}
for (j = 0; j < num_recvs; j++)
{
this_proc = recv_procs[j];
if (all_recv_procs[this_proc] == 0)
num_recvs_L++;
all_recv_procs[this_proc] += recv_vec_starts[j+1]-recv_vec_starts[j];
}
}
}
if (max_sends)
{
L_send_procs = hypre_CTAlloc(HYPRE_Int, num_sends_L);
L_send_ptr = hypre_CTAlloc(HYPRE_Int, num_sends_L+1);
num_sends_L = 0;
for (j=0; j < num_procs; j++)
{
this_proc = all_send_procs[j];
if (this_proc)
{
L_send_procs[num_sends_L++] = j;
L_send_ptr[num_sends_L] = this_proc + L_send_ptr[num_sends_L-1];
}
}
}
if (max_recvs)
{
L_recv_procs = hypre_CTAlloc(HYPRE_Int, num_recvs_L);
L_recv_ptr = hypre_CTAlloc(HYPRE_Int, num_recvs_L+1);
num_recvs_L = 0;
for (j=0; j < num_procs; j++)
{
this_proc = all_recv_procs[j];
if (this_proc)
{
L_recv_procs[num_recvs_L++] = j;
L_recv_ptr[num_recvs_L] = this_proc + L_recv_ptr[num_recvs_L-1];
}
}
}
}
}
if (max_sends) hypre_TFree(all_send_procs);
if (max_recvs) hypre_TFree(all_recv_procs);
L_diag = hypre_CSRMatrixCreate(num_rows_L, num_rows_L, num_nonzeros_diag);
L_offd = hypre_CSRMatrixCreate(num_rows_L, num_cols_offd_L, num_nonzeros_offd);
hypre_CSRMatrixInitialize(L_diag);
hypre_CSRMatrixInitialize(L_offd);
if (num_nonzeros_diag)
{
L_diag_data = hypre_CSRMatrixData(L_diag);
L_diag_j = hypre_CSRMatrixJ(L_diag);
}
L_diag_i = hypre_CSRMatrixI(L_diag);
if (num_nonzeros_offd)
{
L_offd_data = hypre_CSRMatrixData(L_offd);
L_offd_j = hypre_CSRMatrixJ(L_offd);
}
L_offd_i = hypre_CSRMatrixI(L_offd);
if (num_rows_L) D_data = hypre_CTAlloc(HYPRE_Real,num_rows_L);
if (send_data_L)
{
L_send_map_elmts = hypre_CTAlloc(HYPRE_Int, send_data_L);
buf_data = hypre_CTAlloc(HYPRE_Real,send_data_L);
}
if (num_cols_offd_L)
{
D_data_offd = hypre_CTAlloc(HYPRE_Real,num_cols_offd_L);
/*L_col_map_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd_L);*/
remap = hypre_CTAlloc(HYPRE_Int, num_cols_offd_L);
}
Rtilde = hypre_CTAlloc(hypre_ParVector, 1);
Rtilde_local = hypre_SeqVectorCreate(num_rows_L);
hypre_SeqVectorInitialize(Rtilde_local);
hypre_ParVectorLocalVector(Rtilde) = Rtilde_local;
hypre_ParVectorOwnsData(Rtilde) = 1;
Xtilde = hypre_CTAlloc(hypre_ParVector, 1);
Xtilde_local = hypre_SeqVectorCreate(num_rows_L);
hypre_SeqVectorInitialize(Xtilde_local);
hypre_ParVectorLocalVector(Xtilde) = Xtilde_local;
hypre_ParVectorOwnsData(Xtilde) = 1;
x_data = hypre_VectorData(hypre_ParVectorLocalVector(Xtilde));
r_data = hypre_VectorData(hypre_ParVectorLocalVector(Rtilde));
cnt = 0;
cnt_level = 0;
cnt_diag = 0;
cnt_offd = 0;
cnt_row = 1;
L_diag_i[0] = 0;
L_offd_i[0] = 0;
for (level=addlvl; level < num_levels; level++)
{
row_start = level_start[cnt_level];
if (level != 0)
{
tmp_data = hypre_VectorData(hypre_ParVectorLocalVector(F_array[level]));
if (tmp_data) hypre_TFree(tmp_data);
hypre_VectorData(hypre_ParVectorLocalVector(F_array[level])) = &r_data[row_start];
hypre_VectorOwnsData(hypre_ParVectorLocalVector(F_array[level])) = 0;
tmp_data = hypre_VectorData(hypre_ParVectorLocalVector(U_array[level]));
if (tmp_data) hypre_TFree(tmp_data);
hypre_VectorData(hypre_ParVectorLocalVector(U_array[level])) = &x_data[row_start];
hypre_VectorOwnsData(hypre_ParVectorLocalVector(U_array[level])) = 0;
}
cnt_level++;
start_diag = L_diag_i[cnt_row-1];
start_offd = L_offd_i[cnt_row-1];
A_tmp = A_array[level];
A_tmp_diag = hypre_ParCSRMatrixDiag(A_tmp);
A_tmp_offd = hypre_ParCSRMatrixOffd(A_tmp);
comm_pkg = hypre_ParCSRMatrixCommPkg(A_tmp);
A_tmp_diag_i = hypre_CSRMatrixI(A_tmp_diag);
A_tmp_offd_i = hypre_CSRMatrixI(A_tmp_offd);
A_tmp_diag_j = hypre_CSRMatrixJ(A_tmp_diag);
A_tmp_offd_j = hypre_CSRMatrixJ(A_tmp_offd);
A_tmp_diag_data = hypre_CSRMatrixData(A_tmp_diag);
A_tmp_offd_data = hypre_CSRMatrixData(A_tmp_offd);
num_rows_tmp = hypre_CSRMatrixNumRows(A_tmp_diag);
if (comm_pkg)
{
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
}
else
{
num_sends = 0;
num_recvs = 0;
}
/* Compute new combined communication package */
for (i=0; i < num_sends; i++)
{
this_proc = hypre_BinarySearch(L_send_procs,send_procs[i],num_sends_L);
indx = L_send_ptr[this_proc];
for (j=send_map_starts[i]; j < send_map_starts[i+1]; j++)
{
L_send_map_elmts[indx++] = row_start + send_map_elmts[j];
}
L_send_ptr[this_proc] = indx;
}
cnt_map = 0;
for (i = 0; i < num_recvs; i++)
{
this_proc = hypre_BinarySearch(L_recv_procs,recv_procs[i],num_recvs_L);
indx = L_recv_ptr[this_proc];
for (j=recv_vec_starts[i]; j < recv_vec_starts[i+1]; j++)
{
remap[cnt_map++] = indx++;
}
L_recv_ptr[this_proc] = indx;
}
/* Compute Lambda */
if (relax_type == 0)
{
HYPRE_Real rlx_wt = relax_weight[level];
#ifdef HYPRE_USING_OPENMP
#pragma omp for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < num_rows_tmp; i++)
{
D_data[i] = rlx_wt/A_tmp_diag_data[A_tmp_diag_i[i]];
L_diag_i[cnt_row+i] = start_diag + A_tmp_diag_i[i+1];
L_offd_i[cnt_row+i] = start_offd + A_tmp_offd_i[i+1];
}
}
else
{
l1_norms = l1_norms_ptr[level];
#ifdef HYPRE_USING_OPENMP
#pragma omp for private(i) HYPRE_SMP_SCHEDULE
#endif
for (i=0; i < num_rows_tmp; i++)
{
D_data[i] = 1.0/l1_norms[i];
L_diag_i[cnt_row+i] = start_diag + A_tmp_diag_i[i+1];
L_offd_i[cnt_row+i] = start_offd + A_tmp_offd_i[i+1];
}
}
if (num_procs > 1)
{
index = 0;
for (i=0; i < num_sends; i++)
{
start = send_map_starts[i];
for (j=start; j < send_map_starts[i+1]; j++)
buf_data[index++] = D_data[send_map_elmts[j]];
}
comm_handle = hypre_ParCSRCommHandleCreate(1, comm_pkg,
buf_data, D_data_offd);
hypre_ParCSRCommHandleDestroy(comm_handle);
}
for (i = 0; i < num_rows_tmp; i++)
{
j_indx = A_tmp_diag_i[i];
L_diag_data[cnt_diag] = (2.0 - A_tmp_diag_data[j_indx]*D_data[i])*D_data[i];
L_diag_j[cnt_diag++] = i+row_start;
for (j=A_tmp_diag_i[i]+1; j < A_tmp_diag_i[i+1]; j++)
{
j_indx = A_tmp_diag_j[j];
L_diag_data[cnt_diag] = (- A_tmp_diag_data[j]*D_data[j_indx])*D_data[i];
L_diag_j[cnt_diag++] = j_indx+row_start;
}
for (j=A_tmp_offd_i[i]; j < A_tmp_offd_i[i+1]; j++)
{
j_indx = A_tmp_offd_j[j];
L_offd_data[cnt_offd] = (- A_tmp_offd_data[j]*D_data_offd[j_indx])*D_data[i];
L_offd_j[cnt_offd++] = remap[j_indx];
}
}
cnt_row += num_rows_tmp;
}
if (L_send_ptr)
{
for (i=num_sends_L-1; i > 0; i--)
L_send_ptr[i] = L_send_ptr[i-1];
L_send_ptr[0] = 0;
}
else
L_send_ptr = hypre_CTAlloc(HYPRE_Int,1);
if (L_recv_ptr)
{
for (i=num_recvs_L-1; i > 0; i--)
L_recv_ptr[i] = L_recv_ptr[i-1];
L_recv_ptr[0] = 0;
}
else
L_recv_ptr = hypre_CTAlloc(HYPRE_Int,1);
L_comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgNumRecvs(L_comm_pkg) = num_recvs_L;
hypre_ParCSRCommPkgNumSends(L_comm_pkg) = num_sends_L;
hypre_ParCSRCommPkgRecvProcs(L_comm_pkg) = L_recv_procs;
hypre_ParCSRCommPkgSendProcs(L_comm_pkg) = L_send_procs;
hypre_ParCSRCommPkgRecvVecStarts(L_comm_pkg) = L_recv_ptr;
hypre_ParCSRCommPkgSendMapStarts(L_comm_pkg) = L_send_ptr;
hypre_ParCSRCommPkgSendMapElmts(L_comm_pkg) = L_send_map_elmts;
hypre_ParCSRCommPkgComm(L_comm_pkg) = comm;
Lambda = hypre_CTAlloc(hypre_ParCSRMatrix, 1);
hypre_ParCSRMatrixDiag(Lambda) = L_diag;
hypre_ParCSRMatrixOffd(Lambda) = L_offd;
hypre_ParCSRMatrixCommPkg(Lambda) = L_comm_pkg;
hypre_ParCSRMatrixComm(Lambda) = comm;
hypre_ParCSRMatrixOwnsData(Lambda) = 1;
hypre_ParAMGDataLambda(amg_data) = Lambda;
hypre_ParAMGDataRtilde(amg_data) = Rtilde;
hypre_ParAMGDataXtilde(amg_data) = Xtilde;
hypre_TFree(D_data_offd);
hypre_TFree(D_data);
if (num_procs > 1) hypre_TFree(buf_data);
hypre_TFree(remap);
hypre_TFree(buf_data);
hypre_TFree(level_start);
return Solve_err_flag;
}