CFD/hypre
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

optimization for fffc: squeeze out unused offd columns

Browse Source
This commit is contained in:
Ruipeng Li 2020-05-11 17:18:17 -07:00
parent 4182f90cdb
commit ff79fb0d7e
1 changed files with 53 additions and 41 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
src/parcsr_mv/gen_fffc.c
View File

@ -73,7 +73,7 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
HYPRE_Int i, j, jj;
HYPRE_Int startc, index;
HYPRE_Int cpt, fpt, row;
HYPRE_Int *CF_marker_offd = NULL;
HYPRE_Int *CF_marker_offd = NULL, *marker_offd=NULL;
HYPRE_Int *int_buf_data = NULL;
HYPRE_BigInt *big_convert;
HYPRE_BigInt *big_convert_offd = NULL;
@ -87,7 +87,7 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
HYPRE_Int start, stop;
HYPRE_Int num_threads;
num_threads = hypre_NumThreads();
num_threads = hypre_NumThreads();
/* MPI size and rank*/
hypre_MPI_Comm_size(comm, &num_procs);
@ -125,7 +125,7 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
fpt_array[my_thread_num+1]++;
}
}
#ifdef HYPRE_USING_OPENMP
#pragma omp barrier
#endif
@ -216,7 +216,7 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
#endif
if (my_thread_num == 0)
{
if (num_cols_A_offd)
if (num_cols_A_offd)
{
CF_marker_offd = hypre_CTAlloc(HYPRE_Int, num_cols_A_offd, HYPRE_MEMORY_HOST);
big_convert_offd = hypre_CTAlloc(HYPRE_BigInt, num_cols_A_offd, HYPRE_MEMORY_HOST);
@ -245,18 +245,30 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
hypre_ParCSRCommHandleDestroy(comm_handle);
marker_offd = hypre_CTAlloc(HYPRE_Int, num_cols_A_offd, HYPRE_MEMORY_HOST);
for (i = 0; i < n_fine; i++)
{
if (CF_marker[i] < 0)
{
for (j = S_offd_i[i]; j < S_offd_i[i+1]; j++)
{
marker_offd[S_offd_j[j]] = 1;
}
}
}
num_cols_offd_A_FC = 0;
num_cols_offd_A_FF = 0;
if (num_cols_A_offd)
if (num_cols_A_offd)
{
for (i=0; i < num_cols_A_offd; i++)
{
if (CF_marker_offd[i] > 0)
if (CF_marker_offd[i] > 0 && marker_offd[i] > 0)
{
fine_to_coarse_offd[i] = num_cols_offd_A_FC++;
fine_to_fine_offd[i] = -1;
}
else
else if (CF_marker_offd[i] < 0 && marker_offd[i] > 0)
{
fine_to_fine_offd[i] = num_cols_offd_A_FF++;
fine_to_coarse_offd[i] = -1;
@ -265,18 +277,18 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
col_map_offd_A_FF = hypre_TAlloc(HYPRE_BigInt, num_cols_offd_A_FF, HYPRE_MEMORY_HOST);
col_map_offd_A_FC = hypre_TAlloc(HYPRE_BigInt, num_cols_offd_A_FC, HYPRE_MEMORY_HOST);
cpt = 0;
fpt = 0;
for (i=0; i < num_cols_A_offd; i++)
{
if (CF_marker_offd[i] > 0)
if (CF_marker_offd[i] > 0 && marker_offd[i] > 0)
{
col_map_offd_A_FC[cpt++] = big_convert_offd[i];
}
else
else if (CF_marker_offd[i] < 0 && marker_offd[i] > 0)
{
col_map_offd_A_FF[fpt++] = big_convert_offd[i];
col_map_offd_A_FF[fpt++] = big_convert_offd[i];
}
}
}
@ -304,8 +316,8 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
for (j = S_diag_i[i]; j < S_diag_i[i+1]; j++)
{
jj = S_diag_j[j];
if (CF_marker[jj] > 0)
d_count_FC++;
if (CF_marker[jj] > 0)
d_count_FC++;
else
d_count_FF++;
}
@ -314,8 +326,8 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
for (j = S_offd_i[i]; j < S_offd_i[i+1]; j++)
{
jj = S_offd_j[j];
if (CF_marker_offd[jj] > 0)
o_count_FC++;
if (CF_marker_offd[jj] > 0)
o_count_FC++;
else
o_count_FF++;
}
@ -334,10 +346,10 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
{
fpt = fpt_array[i];
fpt2 = fpt_array[i-1];
A_FC_diag_i[fpt] += A_FC_diag_i[fpt2];
A_FF_diag_i[fpt] += A_FF_diag_i[fpt2];
A_FC_offd_i[fpt] += A_FC_offd_i[fpt2];
A_FF_offd_i[fpt] += A_FF_offd_i[fpt2];
A_FC_diag_i[fpt] += A_FC_diag_i[fpt2];
A_FF_diag_i[fpt] += A_FF_diag_i[fpt2];
A_FC_offd_i[fpt] += A_FC_offd_i[fpt2];
A_FF_offd_i[fpt] += A_FF_offd_i[fpt2];
}
row = fpt_array[num_threads];
d_count_FC = A_FC_diag_i[row];
@ -375,7 +387,7 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
{
jA = A_diag_i[i]+1;
jS = S_diag_j[j];
while (A_diag_j[jA] != jS) jA++;
while (A_diag_j[jA] != jS) jA++;
if (CF_marker[S_diag_j[j]] > 0)
{
A_FC_diag_j[d_count_FC] = fine_to_coarse[A_diag_j[jA]];
@ -468,7 +480,7 @@ hypre_ParCSRMatrixGenerateFFFC( hypre_ParCSRMatrix *A,
hypre_TFree(CF_marker_offd, HYPRE_MEMORY_HOST);
hypre_TFree(int_buf_data, HYPRE_MEMORY_HOST);
hypre_TFree(big_buf_data, HYPRE_MEMORY_HOST);
hypre_TFree(marker_offd, HYPRE_MEMORY_HOST);
hypre_TFree(cpt_array, HYPRE_MEMORY_HOST);
hypre_TFree(fpt_array, HYPRE_MEMORY_HOST);
@ -556,7 +568,7 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
HYPRE_Int start, stop;
HYPRE_Int num_threads;
num_threads = hypre_NumThreads();
num_threads = hypre_NumThreads();
/* MPI size and rank*/
hypre_MPI_Comm_size(comm, &num_procs);
@ -595,12 +607,12 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
new_fpt_array[my_thread_num+1]++;
fpt_array[my_thread_num+1]++;
}
else
else
{
fpt_array[my_thread_num+1]++;
}
}
#ifdef HYPRE_USING_OPENMP
#pragma omp barrier
#endif
@ -703,7 +715,7 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
#endif
if (my_thread_num == 0)
{
if (num_cols_A_offd)
if (num_cols_A_offd)
{
CF_marker_offd = hypre_CTAlloc(HYPRE_Int, num_cols_A_offd, HYPRE_MEMORY_HOST);
big_convert_offd = hypre_CTAlloc(HYPRE_BigInt, num_cols_A_offd, HYPRE_MEMORY_HOST);
@ -734,7 +746,7 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
num_cols_offd_A_FC = 0;
num_cols_offd_A_FF = 0;
if (num_cols_A_offd)
if (num_cols_A_offd)
{
for (i=0; i < num_cols_A_offd; i++)
{
@ -752,7 +764,7 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
col_map_offd_A_FF = hypre_TAlloc(HYPRE_BigInt, num_cols_offd_A_FF, HYPRE_MEMORY_HOST);
col_map_offd_A_FC = hypre_TAlloc(HYPRE_BigInt, num_cols_offd_A_FC, HYPRE_MEMORY_HOST);
cpt = 0;
fpt = 0;
for (i=0; i < num_cols_A_offd; i++)
@ -793,8 +805,8 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
for (j = S_diag_i[i]; j < S_diag_i[i+1]; j++)
{
jj = S_diag_j[j];
if (CF_marker[jj] > 0)
d_count_FC++;
if (CF_marker[jj] > 0)
d_count_FC++;
else
d_count_FF++;
}
@ -803,8 +815,8 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
for (j = S_offd_i[i]; j < S_offd_i[i+1]; j++)
{
jj = S_offd_j[j];
if (CF_marker_offd[jj] > 0)
o_count_FC++;
if (CF_marker_offd[jj] > 0)
o_count_FC++;
else
o_count_FF++;
}
@ -817,15 +829,15 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
for (j = S_diag_i[i]; j < S_diag_i[i+1]; j++)
{
jj = S_diag_j[j];
if (CF_marker[jj] > 0)
d_count_FC++;
if (CF_marker[jj] > 0)
d_count_FC++;
}
A_FC_diag_i[rowc] = d_count_FC;
for (j = S_offd_i[i]; j < S_offd_i[i+1]; j++)
{
jj = S_offd_j[j];
if (CF_marker_offd[jj] > 0)
o_count_FC++;
if (CF_marker_offd[jj] > 0)
o_count_FC++;
}
A_FC_offd_i[rowc] = o_count_FC;
}
@ -843,10 +855,10 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
new_fpt = new_fpt_array[i];
fpt2 = fpt_array[i-1];
new_fpt2 = new_fpt_array[i-1];
A_FC_diag_i[fpt] += A_FC_diag_i[fpt2];
A_FF_diag_i[new_fpt] += A_FF_diag_i[new_fpt2];
A_FC_offd_i[fpt] += A_FC_offd_i[fpt2];
A_FF_offd_i[new_fpt] += A_FF_offd_i[new_fpt2];
A_FC_diag_i[fpt] += A_FC_diag_i[fpt2];
A_FF_diag_i[new_fpt] += A_FF_diag_i[new_fpt2];
A_FC_offd_i[fpt] += A_FC_offd_i[fpt2];
A_FF_offd_i[new_fpt] += A_FF_offd_i[new_fpt2];
}
row = new_fpt_array[num_threads];
rowc = fpt_array[num_threads];
@ -887,7 +899,7 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
{
jA = A_diag_i[i]+1;
jS = S_diag_j[j];
while (A_diag_j[jA] != jS) jA++;
while (A_diag_j[jA] != jS) jA++;
if (CF_marker[S_diag_j[j]] > 0)
{
A_FC_diag_j[d_count_FC] = fine_to_coarse[A_diag_j[jA]];
@ -928,7 +940,7 @@ hypre_ParCSRMatrixGenerateFFFC3( hypre_ParCSRMatrix *A,
{
jA = A_diag_i[i]+1;
jS = S_diag_j[j];
while (A_diag_j[jA] != jS) jA++;
while (A_diag_j[jA] != jS) jA++;
if (CF_marker[S_diag_j[j]] > 0)
{
A_FC_diag_j[d_count_FC] = fine_to_coarse[A_diag_j[jA]];