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_ParCSRMatrix * hypre_ParMatmul_FC(
hypre_ParCSRMatrix * A, hypre_ParCSRMatrix * P, HYPRE_Int * CF_marker,
HYPRE_Int * dof_func, HYPRE_Int * dof_func_offd )
/* hypre_parMatmul_FC creates and returns the "Fine"-designated rows of the
matrix product A*P. A's size is (nC+nF)*(nC+nF), P's size is (nC+nF)*nC
where nC is the number of coarse rows/columns, nF the number of fine
rows/columns. The size of C=A*P is (nC+nF)*nC, even though not all rows
of C are actually computed. If we were to construct a matrix consisting
only of the computed rows of C, its size would be nF*nC.
"Fine" is defined solely by the marker array, and for example could be
a proper subset of the fine points of a multigrid hierarchy.
*/
{
/* To compute a submatrix of C containing only the computed data, i.e.
only "Fine" rows, we would have to do a lot of computational work,
with a lot of communication. The communication is because such a
matrix would need global information that depends on which rows are
"Fine".
*/
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *row_starts_A = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int num_rows_diag_A = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_cols_diag_A = hypre_CSRMatrixNumCols(A_diag);
HYPRE_Int num_cols_offd_A = hypre_CSRMatrixNumCols(A_offd);
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
double *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
HYPRE_Int *col_map_offd_P = hypre_ParCSRMatrixColMapOffd(P);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int first_col_diag_P = hypre_ParCSRMatrixFirstColDiag(P);
HYPRE_Int last_col_diag_P;
HYPRE_Int *col_starts_P = hypre_ParCSRMatrixColStarts(P);
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int num_cols_diag_P = hypre_CSRMatrixNumCols(P_diag);
HYPRE_Int num_cols_offd_P = hypre_CSRMatrixNumCols(P_offd);
hypre_ParCSRMatrix *C;
HYPRE_Int *col_map_offd_C;
HYPRE_Int *map_P_to_C;
hypre_CSRMatrix *C_diag;
double *C_diag_data;
HYPRE_Int *C_diag_i;
HYPRE_Int *C_diag_j;
hypre_CSRMatrix *C_offd;
double *C_offd_data=NULL;
HYPRE_Int *C_offd_i=NULL;
HYPRE_Int *C_offd_j=NULL;
HYPRE_Int C_diag_size;
HYPRE_Int C_offd_size;
HYPRE_Int num_cols_offd_C = 0;
hypre_CSRMatrix *Ps_ext;
double *Ps_ext_data;
HYPRE_Int *Ps_ext_i;
HYPRE_Int *Ps_ext_j;
double *P_ext_diag_data;
HYPRE_Int *P_ext_diag_i;
HYPRE_Int *P_ext_diag_j;
HYPRE_Int P_ext_diag_size;
double *P_ext_offd_data;
HYPRE_Int *P_ext_offd_i;
HYPRE_Int *P_ext_offd_j;
HYPRE_Int P_ext_offd_size;
HYPRE_Int *P_marker;
HYPRE_Int *temp;
HYPRE_Int i, j;
HYPRE_Int i1, i2, i3;
HYPRE_Int jj2, jj3;
HYPRE_Int jj_count_diag, jj_count_offd;
HYPRE_Int jj_row_begin_diag, jj_row_begin_offd;
HYPRE_Int start_indexing = 0; /* start indexing for C_data at 0 */
HYPRE_Int n_rows_A_global, n_cols_A_global;
HYPRE_Int n_rows_P_global, n_cols_P_global;
HYPRE_Int allsquare = 0;
HYPRE_Int cnt, cnt_offd, cnt_diag;
HYPRE_Int num_procs;
HYPRE_Int value;
double a_entry;
double a_b_product;
n_rows_A_global = hypre_ParCSRMatrixGlobalNumRows(A);
n_cols_A_global = hypre_ParCSRMatrixGlobalNumCols(A);
n_rows_P_global = hypre_ParCSRMatrixGlobalNumRows(P);
n_cols_P_global = hypre_ParCSRMatrixGlobalNumCols(P);
if (n_cols_A_global != n_rows_P_global || num_cols_diag_A != num_rows_diag_P)
{
hypre_printf(" Error! Incompatible matrix dimensions!\n");
return NULL;
}
/* if (num_rows_A==num_cols_P) allsquare = 1; */
/*-----------------------------------------------------------------------
* Extract P_ext, i.e. portion of P that is stored on neighbor procs
* and needed locally for matrix matrix product
*-----------------------------------------------------------------------*/
hypre_MPI_Comm_size(comm, &num_procs);
if (num_procs > 1)
{
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings within
* hypre_ParCSRMatrixExtractBExt
*--------------------------------------------------------------------*/
Ps_ext = hypre_ParCSRMatrixExtractBExt(P,A,1);
Ps_ext_data = hypre_CSRMatrixData(Ps_ext);
Ps_ext_i = hypre_CSRMatrixI(Ps_ext);
Ps_ext_j = hypre_CSRMatrixJ(Ps_ext);
}
P_ext_diag_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_offd_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_diag_size = 0;
P_ext_offd_size = 0;
last_col_diag_P = first_col_diag_P + num_cols_diag_P -1;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
P_ext_offd_size++;
else
P_ext_diag_size++;
P_ext_diag_i[i+1] = P_ext_diag_size;
P_ext_offd_i[i+1] = P_ext_offd_size;
}
if (P_ext_diag_size)
{
P_ext_diag_j = hypre_CTAlloc(HYPRE_Int, P_ext_diag_size);
P_ext_diag_data = hypre_CTAlloc(double, P_ext_diag_size);
}
if (P_ext_offd_size)
{
P_ext_offd_j = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size);
P_ext_offd_data = hypre_CTAlloc(double, P_ext_offd_size);
}
cnt_offd = 0;
cnt_diag = 0;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
{
P_ext_offd_j[cnt_offd] = Ps_ext_j[j];
P_ext_offd_data[cnt_offd++] = Ps_ext_data[j];
}
else
{
P_ext_diag_j[cnt_diag] = Ps_ext_j[j] - first_col_diag_P;
P_ext_diag_data[cnt_diag++] = Ps_ext_data[j];
}
}
if (num_procs > 1)
{
hypre_CSRMatrixDestroy(Ps_ext);
Ps_ext = NULL;
}
cnt = 0;
if (P_ext_offd_size || num_cols_offd_P)
{
temp = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size+num_cols_offd_P);
for (i=0; i < P_ext_offd_size; i++)
temp[i] = P_ext_offd_j[i];
cnt = P_ext_offd_size;
for (i=0; i < num_cols_offd_P; i++)
temp[cnt++] = col_map_offd_P[i];
}
if (cnt)
{
qsort0(temp, 0, cnt-1);
num_cols_offd_C = 1;
value = temp[0];
for (i=1; i < cnt; i++)
{
if (temp[i] > value)
{
value = temp[i];
temp[num_cols_offd_C++] = value;
}
}
}
if (num_cols_offd_C)
col_map_offd_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_C);
for (i=0; i < num_cols_offd_C; i++)
col_map_offd_C[i] = temp[i];
if (P_ext_offd_size || num_cols_offd_P)
hypre_TFree(temp);
for (i=0 ; i < P_ext_offd_size; i++)
P_ext_offd_j[i] = hypre_BinarySearch(col_map_offd_C,
P_ext_offd_j[i],
num_cols_offd_C);
if (num_cols_offd_P)
{
map_P_to_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_P);
cnt = 0;
for (i=0; i < num_cols_offd_C; i++)
if (col_map_offd_C[i] == col_map_offd_P[cnt])
{
map_P_to_C[cnt++] = i;
if (cnt == num_cols_offd_P) break;
}
}
/*-----------------------------------------------------------------------
* Allocate marker array.
*-----------------------------------------------------------------------*/
P_marker = hypre_CTAlloc(HYPRE_Int, num_cols_diag_P+num_cols_offd_C);
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/* no changes for the marked version above this point */
/* This function call is the first pass: */
hypre_ParMatmul_RowSizes_Marked(
&C_diag_i, &C_offd_i, &P_marker,
A_diag_i, A_diag_j, A_offd_i, A_offd_j,
P_diag_i, P_diag_j, P_offd_i, P_offd_j,
P_ext_diag_i, P_ext_diag_j, P_ext_offd_i, P_ext_offd_j,
map_P_to_C,
&C_diag_size, &C_offd_size,
num_rows_diag_A, num_cols_offd_A, allsquare,
num_cols_diag_P, num_cols_offd_P,
num_cols_offd_C, CF_marker, dof_func, dof_func_offd
);
/* The above call of hypre_ParMatmul_RowSizes_Marked computed
two scalars: C_diag_size, C_offd_size,
and two arrays: C_diag_i, C_offd_i
( P_marker is also computed, but only used internally )
*/
/*-----------------------------------------------------------------------
* Allocate C_diag_data and C_diag_j arrays.
* Allocate C_offd_data and C_offd_j arrays.
*-----------------------------------------------------------------------*/
last_col_diag_P = first_col_diag_P + num_cols_diag_P - 1;
C_diag_data = hypre_CTAlloc(double, C_diag_size);
C_diag_j = hypre_CTAlloc(HYPRE_Int, C_diag_size);
if (C_offd_size)
{
C_offd_data = hypre_CTAlloc(double, C_offd_size);
C_offd_j = hypre_CTAlloc(HYPRE_Int, C_offd_size);
}
/*-----------------------------------------------------------------------
* Second Pass: Fill in C_diag_data and C_diag_j.
* Second Pass: Fill in C_offd_data and C_offd_j.
*-----------------------------------------------------------------------*/
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
jj_count_diag = start_indexing;
jj_count_offd = start_indexing;
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/*-----------------------------------------------------------------------
* Loop over interior c-points.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_A; i1++)
{
if ( CF_marker[i1] < 0 ) /* i1 is a fine row */
/* ... This and the coarse row code are the only parts between first pass
and near the end where
hypre_ParMatmul_FC is different from the regular hypre_ParMatmul */
{
/*--------------------------------------------------------------------
* Create diagonal entry, C_{i1,i1}
*--------------------------------------------------------------------*/
jj_row_begin_diag = jj_count_diag;
jj_row_begin_offd = jj_count_offd;
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_offd.
*-----------------------------------------------------------------*/
if (num_cols_offd_A)
{
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func_offd[i2] )
{ /* interpolate only like "functions" */
a_entry = A_offd_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_ext.
*-----------------------------------------------------------*/
for (jj3 = P_ext_offd_i[i2]; jj3 < P_ext_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+P_ext_offd_j[jj3];
a_b_product = a_entry * P_ext_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
C_offd_data[P_marker[i3]] += a_b_product;
}
for (jj3 = P_ext_diag_i[i2]; jj3 < P_ext_diag_i[i2+1]; jj3++)
{
i3 = P_ext_diag_j[jj3];
a_b_product = a_entry * P_ext_diag_data[jj3];
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
C_diag_data[P_marker[i3]] += a_b_product;
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_diag.
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func[i2] )
{ /* interpolate only like "functions" */
a_entry = A_diag_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_diag.
*-----------------------------------------------------------*/
for (jj3 = P_diag_i[i2]; jj3 < P_diag_i[i2+1]; jj3++)
{
i3 = P_diag_j[jj3];
a_b_product = a_entry * P_diag_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[P_marker[i3]] += a_b_product;
}
}
if (num_cols_offd_P)
{
for (jj3 = P_offd_i[i2]; jj3 < P_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+map_P_to_C[P_offd_j[jj3]];
a_b_product = a_entry * P_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
{
C_offd_data[P_marker[i3]] += a_b_product;
}
}
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
else /* i1 is a coarse row.*/
/* Copy P coarse-row values to C. This is useful if C is meant to
become a replacement for P */
{
if (num_cols_offd_P)
{
for (jj2 = P_offd_i[i1]; jj2 < P_offd_i[i1+1]; jj2++)
{
C_offd_j[jj_count_offd] = P_offd_j[jj_count_offd];
C_offd_data[jj_count_offd] = P_offd_data[jj_count_offd];
++jj_count_offd;
}
}
for (jj2 = P_diag_i[i1]; jj2 < P_diag_i[i1+1]; jj2++)
{
C_diag_j[jj_count_diag] = P_diag_j[jj2];
C_diag_data[jj_count_diag] = P_diag_data[jj2];
++jj_count_diag;
}
}
}
C = hypre_ParCSRMatrixCreate(
comm, n_rows_A_global, n_cols_P_global,
row_starts_A, col_starts_P, num_cols_offd_C, C_diag_size, C_offd_size );
/* Note that C does not own the partitionings */
hypre_ParCSRMatrixSetRowStartsOwner(C,0);
hypre_ParCSRMatrixSetColStartsOwner(C,0);
C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrixData(C_diag) = C_diag_data;
hypre_CSRMatrixI(C_diag) = C_diag_i;
hypre_CSRMatrixJ(C_diag) = C_diag_j;
C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrixI(C_offd) = C_offd_i;
hypre_ParCSRMatrixOffd(C) = C_offd;
if (num_cols_offd_C)
{
hypre_CSRMatrixData(C_offd) = C_offd_data;
hypre_CSRMatrixJ(C_offd) = C_offd_j;
hypre_ParCSRMatrixColMapOffd(C) = col_map_offd_C;
}
/*-----------------------------------------------------------------------
* Free various arrays
*-----------------------------------------------------------------------*/
hypre_TFree(P_marker);
hypre_TFree(P_ext_diag_i);
if (P_ext_diag_size)
{
hypre_TFree(P_ext_diag_j);
hypre_TFree(P_ext_diag_data);
}
hypre_TFree(P_ext_offd_i);
if (P_ext_offd_size)
{
hypre_TFree(P_ext_offd_j);
hypre_TFree(P_ext_offd_data);
}
if (num_cols_offd_P) hypre_TFree(map_P_to_C);
return C;
}
HYPRE_Int main( HYPRE_Int argc, char *argv[] )
{
hypre_ParCSRMatrix *par_matrix, *g_matrix, **submatrices;
hypre_CSRMatrix *A_diag, *A_offd;
hypre_CSRBlockMatrix *diag;
hypre_CSRBlockMatrix *offd;
hypre_ParCSRBlockMatrix *par_blk_matrix, *par_blk_matrixT, *rap_matrix;
hypre_Vector *x_local;
hypre_Vector *y_local;
hypre_ParVector *x;
hypre_ParVector *y;
HYPRE_Solver gmres_solver, precon;
HYPRE_Int *diag_i, *diag_j, *offd_i, *offd_j;
HYPRE_Int *diag_i2, *diag_j2, *offd_i2, *offd_j2;
double *diag_d, *diag_d2, *offd_d, *offd_d2;
HYPRE_Int mypid, local_size, nprocs;
HYPRE_Int global_num_rows, global_num_cols, num_cols_offd;
HYPRE_Int num_nonzeros_diag, num_nonzeros_offd, *colMap;
HYPRE_Int ii, jj, kk, row, col, nnz, *indices, *colMap2;
double *data, ddata, *y_data;
HYPRE_Int *row_starts, *col_starts, *rstarts, *cstarts;
HYPRE_Int *row_starts2, *col_starts2;
HYPRE_Int block_size=2, bnnz=4, *index_set;
FILE *fp;
/* --------------------------------------------- */
/* Initialize MPI */
/* --------------------------------------------- */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &mypid);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &nprocs);
/* build and fetch matrix */
MyBuildParLaplacian9pt((HYPRE_ParCSRMatrix *) &par_matrix);
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(par_matrix);
global_num_cols = hypre_ParCSRMatrixGlobalNumCols(par_matrix);
row_starts = hypre_ParCSRMatrixRowStarts(par_matrix);
col_starts = hypre_ParCSRMatrixColStarts(par_matrix);
A_diag = hypre_ParCSRMatrixDiag(par_matrix);
A_offd = hypre_ParCSRMatrixOffd(par_matrix);
num_cols_offd = hypre_CSRMatrixNumCols(A_offd);
num_nonzeros_diag = hypre_CSRMatrixNumNonzeros(A_diag);
num_nonzeros_offd = hypre_CSRMatrixNumNonzeros(A_offd);
/* --------------------------------------------- */
/* build vector and apply matvec */
/* --------------------------------------------- */
x = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_cols,col_starts);
hypre_ParVectorSetPartitioningOwner(x,0);
hypre_ParVectorInitialize(x);
x_local = hypre_ParVectorLocalVector(x);
data = hypre_VectorData(x_local);
local_size = col_starts[mypid+1] - col_starts[mypid];
for (ii = 0; ii < local_size; ii++) data[ii] = 1.0;
y = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_rows,row_starts);
hypre_ParVectorSetPartitioningOwner(y,0);
hypre_ParVectorInitialize(y);
hypre_ParCSRMatrixMatvec (1.0, par_matrix, x, 0.0, y);
ddata = hypre_ParVectorInnerProd(y, y);
if (mypid == 0) hypre_printf("y inner product = %e\n", ddata);
hypre_ParVectorDestroy(x);
hypre_ParVectorDestroy(y);
/* --------------------------------------------- */
/* build block matrix */
/* --------------------------------------------- */
rstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) rstarts[ii] = row_starts[ii];
cstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) cstarts[ii] = col_starts[ii];
par_blk_matrix = hypre_ParCSRBlockMatrixCreate(hypre_MPI_COMM_WORLD,block_size,
global_num_rows, global_num_cols, rstarts,
cstarts, num_cols_offd, num_nonzeros_diag,
num_nonzeros_offd);
colMap = hypre_ParCSRMatrixColMapOffd(par_matrix);
if (num_cols_offd > 0) colMap2 = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
else colMap2 = NULL;
for (ii = 0; ii < num_cols_offd; ii++) colMap2[ii] = colMap[ii];
hypre_ParCSRBlockMatrixColMapOffd(par_blk_matrix) = colMap2;
diag_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(par_matrix));
diag_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixDiag(par_matrix));
diag_d = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(par_matrix));
diag = hypre_ParCSRBlockMatrixDiag(par_blk_matrix);
diag_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
diag_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_diag);
diag_d2 = hypre_CTAlloc(double, num_nonzeros_diag*bnnz);
for (ii = 0; ii <= local_size; ii++) diag_i2[ii] = diag_i[ii];
for (ii = 0; ii < num_nonzeros_diag; ii++) diag_j2[ii] = diag_j[ii];
hypre_CSRBlockMatrixI(diag) = diag_i2;
hypre_CSRBlockMatrixJ(diag) = diag_j2;
for (ii = 0; ii < num_nonzeros_diag; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
diag_d2[ii*bnnz+jj*block_size+kk] = diag_d[ii];
else
diag_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(diag) = diag_d2;
offd_i = hypre_CSRMatrixI(hypre_ParCSRMatrixOffd(par_matrix));
offd_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixOffd(par_matrix));
offd_d = hypre_CSRMatrixData(hypre_ParCSRMatrixOffd(par_matrix));
offd = hypre_ParCSRBlockMatrixOffd(par_blk_matrix);
offd_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
for (ii = 0; ii <= local_size; ii++) offd_i2[ii] = offd_i[ii];
hypre_CSRBlockMatrixI(offd) = offd_i2;
if (num_cols_offd)
{
offd_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_offd);
for (ii = 0; ii < num_nonzeros_offd; ii++) offd_j2[ii] = offd_j[ii];
hypre_CSRBlockMatrixJ(offd) = offd_j2;
offd_d2 = hypre_CTAlloc(double, num_nonzeros_offd*bnnz);
for (ii = 0; ii < num_nonzeros_offd; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
offd_d2[ii*bnnz+jj*block_size+kk] = offd_d[ii];
else
offd_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(offd) = offd_d2;
}
else
{
HYPRE_Int
hypre_ParCSRMatrixMatvec( HYPRE_Complex alpha,
hypre_ParCSRMatrix *A,
hypre_ParVector *x,
HYPRE_Complex beta,
hypre_ParVector *y )
{
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_vectors = hypre_VectorNumVectors(x_local);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int ierr = 0;
HYPRE_Int num_sends, i, j, jv, index, start;
HYPRE_Int vecstride = hypre_VectorVectorStride( x_local );
HYPRE_Int idxstride = hypre_VectorIndexStride( x_local );
HYPRE_Complex *x_tmp_data, **x_buf_data;
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_assert( idxstride>0 );
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;
hypre_assert( hypre_VectorNumVectors(y_local)==num_vectors );
if ( num_vectors==1 )
x_tmp = hypre_SeqVectorCreate( num_cols_offd );
else
{
hypre_assert( num_vectors>1 );
x_tmp = hypre_SeqMultiVectorCreate( num_cols_offd, num_vectors );
}
hypre_SeqVectorInitialize(x_tmp);
x_tmp_data = hypre_VectorData(x_tmp);
comm_handle = hypre_CTAlloc(hypre_ParCSRCommHandle*,num_vectors);
/*---------------------------------------------------------------------
* 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);
x_buf_data = hypre_CTAlloc( HYPRE_Complex*, num_vectors );
for ( jv=0; jv<num_vectors; ++jv )
x_buf_data[jv] = hypre_CTAlloc(HYPRE_Complex, hypre_ParCSRCommPkgSendMapStart
(comm_pkg, num_sends));
if ( num_vectors==1 )
{
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[0][index++]
= x_local_data[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)];
}
}
else
for ( jv=0; jv<num_vectors; ++jv )
{
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[jv][index++]
= x_local_data[
jv*vecstride +
idxstride*hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j) ];
}
}
hypre_assert( idxstride==1 );
/* ... The assert is because the following loop only works for 'column'
storage of a multivector. This needs to be fixed to work more generally,
at least for 'row' storage. This in turn, means either change CommPkg so
num_sends is no.zones*no.vectors (not no.zones) or, less dangerously, put
a stride in the logic of CommHandleCreate (stride either from a new arg or
a new variable inside CommPkg). Or put the num_vector iteration inside
CommHandleCreate (perhaps a new multivector variant of it).
*/
for ( jv=0; jv<num_vectors; ++jv )
{
comm_handle[jv] = hypre_ParCSRCommHandleCreate
( 1, comm_pkg, x_buf_data[jv], &(x_tmp_data[jv*num_cols_offd]) );
}
hypre_CSRMatrixMatvec( alpha, diag, x_local, beta, y_local);
for ( jv=0; jv<num_vectors; ++jv )
{
hypre_ParCSRCommHandleDestroy(comm_handle[jv]);
comm_handle[jv] = NULL;
}
hypre_TFree(comm_handle);
if (num_cols_offd) hypre_CSRMatrixMatvec( alpha, offd, x_tmp, 1.0, y_local);
hypre_SeqVectorDestroy(x_tmp);
x_tmp = NULL;
for ( jv=0; jv<num_vectors; ++jv ) hypre_TFree(x_buf_data[jv]);
hypre_TFree(x_buf_data);
return ierr;
}
/*--------------------------------------------------------------------------
* 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;
}
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_ParCSRMatrixMatvecT( HYPRE_Complex alpha,
hypre_ParCSRMatrix *A,
hypre_ParVector *x,
HYPRE_Complex beta,
hypre_ParVector *y )
{
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_Vector *y_tmp;
HYPRE_Int vecstride = hypre_VectorVectorStride( y_local );
HYPRE_Int idxstride = hypre_VectorIndexStride( y_local );
HYPRE_Complex *y_tmp_data, **y_buf_data;
HYPRE_Complex *y_local_data = hypre_VectorData(y_local);
HYPRE_Int num_rows = hypre_ParCSRMatrixGlobalNumRows(A);
HYPRE_Int num_cols = hypre_ParCSRMatrixGlobalNumCols(A);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int x_size = hypre_ParVectorGlobalSize(x);
HYPRE_Int y_size = hypre_ParVectorGlobalSize(y);
HYPRE_Int num_vectors = hypre_VectorNumVectors(y_local);
HYPRE_Int i, j, jv, index, start, num_sends;
HYPRE_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.
*--------------------------------------------------------------------*/
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;
/*-----------------------------------------------------------------------
*-----------------------------------------------------------------------*/
comm_handle = hypre_CTAlloc(hypre_ParCSRCommHandle*,num_vectors);
if ( num_vectors==1 )
{
y_tmp = hypre_SeqVectorCreate(num_cols_offd);
}
else
{
y_tmp = hypre_SeqMultiVectorCreate(num_cols_offd,num_vectors);
}
hypre_SeqVectorInitialize(y_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);
y_buf_data = hypre_CTAlloc( HYPRE_Complex*, num_vectors );
for ( jv=0; jv<num_vectors; ++jv )
y_buf_data[jv] = hypre_CTAlloc(HYPRE_Complex, hypre_ParCSRCommPkgSendMapStart
(comm_pkg, num_sends));
y_tmp_data = hypre_VectorData(y_tmp);
y_local_data = hypre_VectorData(y_local);
hypre_assert( idxstride==1 ); /* only 'column' storage of multivectors
* implemented so far */
if (num_cols_offd) hypre_CSRMatrixMatvecT(alpha, offd, x_local, 0.0, y_tmp);
for ( jv=0; jv<num_vectors; ++jv )
{
/* this is where we assume multivectors are 'column' storage */
comm_handle[jv] = hypre_ParCSRCommHandleCreate
( 2, comm_pkg, &(y_tmp_data[jv*num_cols_offd]), y_buf_data[jv] );
}
hypre_CSRMatrixMatvecT(alpha, diag, x_local, beta, y_local);
for ( jv=0; jv<num_vectors; ++jv )
{
hypre_ParCSRCommHandleDestroy(comm_handle[jv]);
comm_handle[jv] = NULL;
}
hypre_TFree(comm_handle);
if ( num_vectors==1 )
{
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++)
y_local_data[hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j)]
+= y_buf_data[0][index++];
}
}
else
for ( jv=0; jv<num_vectors; ++jv )
{
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++)
y_local_data[ jv*vecstride +
idxstride*hypre_ParCSRCommPkgSendMapElmt(comm_pkg,j) ]
+= y_buf_data[jv][index++];
}
}
hypre_SeqVectorDestroy(y_tmp);
y_tmp = NULL;
for ( jv=0; jv<num_vectors; ++jv ) hypre_TFree(y_buf_data[jv]);
hypre_TFree(y_buf_data);
return ierr;
}
hypre_ParCSRBlockMatrix *
hypre_ParCSRBlockMatrixConvertFromParCSRMatrix(hypre_ParCSRMatrix *matrix,
HYPRE_Int matrix_C_block_size )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(matrix);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(matrix);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(matrix);
HYPRE_Int global_num_rows = hypre_ParCSRMatrixGlobalNumRows(matrix);
HYPRE_Int global_num_cols = hypre_ParCSRMatrixGlobalNumCols(matrix);
HYPRE_Int *row_starts = hypre_ParCSRMatrixRowStarts(matrix);
HYPRE_Int *col_starts = hypre_ParCSRMatrixColStarts(matrix);
HYPRE_Int num_cols_offd = hypre_CSRMatrixNumCols(offd);
HYPRE_Int *col_map_offd = hypre_ParCSRBlockMatrixColMapOffd(matrix);
HYPRE_Int *map_to_node=NULL, *counter=NULL, *col_in_j_map=NULL;
HYPRE_Int *matrix_C_col_map_offd = NULL;
HYPRE_Int matrix_C_num_cols_offd;
HYPRE_Int matrix_C_num_nonzeros_offd;
HYPRE_Int num_rows, num_nodes;
HYPRE_Int *offd_i = hypre_CSRMatrixI(offd);
HYPRE_Int *offd_j = hypre_CSRMatrixJ(offd);
HYPRE_Complex * offd_data = hypre_CSRMatrixData(offd);
hypre_ParCSRBlockMatrix *matrix_C;
HYPRE_Int *matrix_C_row_starts;
HYPRE_Int *matrix_C_col_starts;
hypre_CSRBlockMatrix *matrix_C_diag;
hypre_CSRBlockMatrix *matrix_C_offd;
HYPRE_Int *matrix_C_offd_i=NULL, *matrix_C_offd_j = NULL;
HYPRE_Complex *matrix_C_offd_data = NULL;
HYPRE_Int num_procs, i, j, k, k_map, count, index, start_index, pos, row;
hypre_MPI_Comm_size(comm,&num_procs);
#ifdef HYPRE_NO_GLOBAL_PARTITION
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, 2);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, 2);
for(i = 0; i < 2; i++)
{
matrix_C_row_starts[i] = row_starts[i]/matrix_C_block_size;
matrix_C_col_starts[i] = col_starts[i]/matrix_C_block_size;
}
#else
matrix_C_row_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
matrix_C_col_starts = hypre_CTAlloc(HYPRE_Int, num_procs + 1);
for(i = 0; i < num_procs + 1; i++)
{
matrix_C_row_starts[i] = row_starts[i]/matrix_C_block_size;
matrix_C_col_starts[i] = col_starts[i]/matrix_C_block_size;
}
#endif
/************* create the diagonal part ************/
matrix_C_diag = hypre_CSRBlockMatrixConvertFromCSRMatrix(diag,
matrix_C_block_size);
/******* the offd part *******************/
/* can't use the same function for the offd part - because this isn't square
and the offd j entries aren't global numbering (have to consider the offd
map) - need to look at col_map_offd first */
/* figure out the new number of offd columns (num rows is same as diag) */
num_cols_offd = hypre_CSRMatrixNumCols(offd);
num_rows = hypre_CSRMatrixNumRows(diag);
num_nodes = num_rows/matrix_C_block_size;
matrix_C_offd_i = hypre_CTAlloc(HYPRE_Int, num_nodes + 1);
matrix_C_num_cols_offd = 0;
matrix_C_offd_i[0] = 0;
matrix_C_num_nonzeros_offd = 0;
if (num_cols_offd)
{
map_to_node = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
matrix_C_num_cols_offd = 1;
map_to_node[0] = col_map_offd[0]/matrix_C_block_size;
for (i=1; i < num_cols_offd; i++)
{
map_to_node[i] = col_map_offd[i]/matrix_C_block_size;
if (map_to_node[i] > map_to_node[i-1]) matrix_C_num_cols_offd++;
}
matrix_C_col_map_offd = hypre_CTAlloc(HYPRE_Int, matrix_C_num_cols_offd);
col_in_j_map = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
matrix_C_col_map_offd[0] = map_to_node[0];
col_in_j_map[0] = 0;
count = 1;
j = 1;
/* fill in the col_map_off_d - these are global numbers. Then we need to
map these to j entries (these have local numbers) */
for (i=1; i < num_cols_offd; i++)
{
if (map_to_node[i] > map_to_node[i-1])
{
matrix_C_col_map_offd[count++] = map_to_node[i];
}
col_in_j_map[j++] = count - 1;
}
/* now figure the nonzeros */
matrix_C_num_nonzeros_offd = 0;
counter = hypre_CTAlloc(HYPRE_Int, matrix_C_num_cols_offd);
for (i=0; i < matrix_C_num_cols_offd; i++)
counter[i] = -1;
for (i=0; i < num_nodes; i++) /* for each block row */
{
matrix_C_offd_i[i] = matrix_C_num_nonzeros_offd;
for (j=0; j < matrix_C_block_size; j++)
{
row = i*matrix_C_block_size+j;
for (k=offd_i[row]; k < offd_i[row+1]; k++) /* go through single row */
{
k_map = col_in_j_map[offd_j[k]]; /*nodal col - see if this has
been in this block row (i)
already*/
if (counter[k_map] < i) /* not yet counted for this nodal row */
{
counter[k_map] = i;
matrix_C_num_nonzeros_offd++;
}
}
}
}
/* fill in final i entry */
matrix_C_offd_i[num_nodes] = matrix_C_num_nonzeros_offd;
}
/* create offd matrix */
matrix_C_offd = hypre_CSRBlockMatrixCreate(matrix_C_block_size, num_nodes,
matrix_C_num_cols_offd,
matrix_C_num_nonzeros_offd);
/* assign i */
hypre_CSRBlockMatrixI(matrix_C_offd) = matrix_C_offd_i;
/* create (and allocate j and data) */
if (matrix_C_num_nonzeros_offd)
{
matrix_C_offd_j = hypre_CTAlloc(HYPRE_Int, matrix_C_num_nonzeros_offd);
matrix_C_offd_data =
hypre_CTAlloc(HYPRE_Complex,
matrix_C_num_nonzeros_offd*matrix_C_block_size*
matrix_C_block_size);
hypre_CSRBlockMatrixJ(matrix_C_offd) = matrix_C_offd_j;
hypre_CSRMatrixData(matrix_C_offd) = matrix_C_offd_data;
for (i=0; i < matrix_C_num_cols_offd; i++)
counter[i] = -1;
index = 0; /*keep track of entry in matrix_C_offd_j*/
start_index = 0;
for (i=0; i < num_nodes; i++) /* for each block row */
{
for (j=0; j < matrix_C_block_size; j++) /* for each row in block */
{
row = i*matrix_C_block_size+j;
for (k=offd_i[row]; k < offd_i[row+1]; k++) /* go through single row's cols */
{
k_map = col_in_j_map[offd_j[k]]; /*nodal col for off_d */
if (counter[k_map] < start_index) /* not yet counted for this nodal row */
{
counter[k_map] = index;
matrix_C_offd_j[index] = k_map;
/*copy the data: which position (corresponds to j array) + which row + which col */
pos = (index * matrix_C_block_size * matrix_C_block_size) + (j * matrix_C_block_size) +
col_map_offd[offd_j[k]]%matrix_C_block_size;
matrix_C_offd_data[pos] = offd_data[k];
index ++;
}
else /* this col has already been listed for this row */
{
/*copy the data: which position (corresponds to j array) + which row + which col */
pos = (counter[k_map]* matrix_C_block_size * matrix_C_block_size) + (j * matrix_C_block_size) +
col_map_offd[offd_j[k]]%matrix_C_block_size;
matrix_C_offd_data[pos] = offd_data[k];
}
}
}
start_index = index; /* first index for current nodal row */
}
}
/* *********create the new matrix *************/
matrix_C = hypre_ParCSRBlockMatrixCreate(comm, matrix_C_block_size,
global_num_rows/matrix_C_block_size,
global_num_cols/matrix_C_block_size, matrix_C_row_starts,
matrix_C_col_starts, matrix_C_num_cols_offd,
hypre_CSRBlockMatrixNumNonzeros(matrix_C_diag),
matrix_C_num_nonzeros_offd);
/* use the diag and off diag matrices we have already created */
hypre_CSRBlockMatrixDestroy(hypre_ParCSRMatrixDiag(matrix_C));
hypre_ParCSRBlockMatrixDiag(matrix_C) = matrix_C_diag;
hypre_CSRBlockMatrixDestroy(hypre_ParCSRMatrixOffd(matrix_C));
hypre_ParCSRBlockMatrixOffd(matrix_C) = matrix_C_offd;
hypre_ParCSRMatrixColMapOffd(matrix_C) = matrix_C_col_map_offd;
/* *********don't bother to copy the comm_pkg *************/
hypre_ParCSRBlockMatrixCommPkg(matrix_C) = NULL;
/* CLEAN UP !!!! */
hypre_TFree(map_to_node);
hypre_TFree(col_in_j_map);
hypre_TFree(counter);
return matrix_C;
}
hypre_CSRBlockMatrix *
hypre_ParCSRBlockMatrixExtractBExt(hypre_ParCSRBlockMatrix *B,
hypre_ParCSRBlockMatrix *A, HYPRE_Int data)
{
MPI_Comm comm = hypre_ParCSRBlockMatrixComm(B);
HYPRE_Int first_col_diag = hypre_ParCSRBlockMatrixFirstColDiag(B);
HYPRE_Int *col_map_offd = hypre_ParCSRBlockMatrixColMapOffd(B);
HYPRE_Int block_size = hypre_ParCSRBlockMatrixBlockSize(B);
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRBlockMatrixCommPkg(A);
HYPRE_Int num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
HYPRE_Int *recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
HYPRE_Int *send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
HYPRE_Int *send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
hypre_ParCSRCommHandle *comm_handle;
hypre_ParCSRCommPkg *tmp_comm_pkg;
hypre_CSRBlockMatrix *diag = hypre_ParCSRBlockMatrixDiag(B);
HYPRE_Int *diag_i = hypre_CSRBlockMatrixI(diag);
HYPRE_Int *diag_j = hypre_CSRBlockMatrixJ(diag);
HYPRE_Complex *diag_data = hypre_CSRBlockMatrixData(diag);
hypre_CSRBlockMatrix *offd = hypre_ParCSRBlockMatrixOffd(B);
HYPRE_Int *offd_i = hypre_CSRBlockMatrixI(offd);
HYPRE_Int *offd_j = hypre_CSRBlockMatrixJ(offd);
HYPRE_Complex *offd_data = hypre_CSRBlockMatrixData(offd);
HYPRE_Int *B_int_i;
HYPRE_Int *B_int_j;
HYPRE_Complex *B_int_data;
HYPRE_Int num_cols_B, num_nonzeros;
HYPRE_Int num_rows_B_ext;
HYPRE_Int num_procs, my_id;
hypre_CSRBlockMatrix *B_ext;
HYPRE_Int *B_ext_i;
HYPRE_Int *B_ext_j;
HYPRE_Complex *B_ext_data;
HYPRE_Int *jdata_recv_vec_starts;
HYPRE_Int *jdata_send_map_starts;
HYPRE_Int i, j, k, l, counter, bnnz;
HYPRE_Int start_index;
HYPRE_Int j_cnt, jrow;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
bnnz = block_size * block_size;
num_cols_B = hypre_ParCSRMatrixGlobalNumCols(B);
num_rows_B_ext = recv_vec_starts[num_recvs];
B_int_i = hypre_CTAlloc(HYPRE_Int, send_map_starts[num_sends]+1);
B_ext_i = hypre_CTAlloc(HYPRE_Int, num_rows_B_ext+1);
/*--------------------------------------------------------------------------
* generate B_int_i through adding number of row-elements of offd and diag
* for corresponding rows. B_int_i[j+1] contains the number of elements of
* a row j (which is determined through send_map_elmts)
*--------------------------------------------------------------------------*/
B_int_i[0] = 0;
j_cnt = 0;
num_nonzeros = 0;
for (i=0; i < num_sends; i++)
{
for (j = send_map_starts[i]; j < send_map_starts[i+1]; j++)
{
jrow = send_map_elmts[j];
B_int_i[++j_cnt] = offd_i[jrow+1] - offd_i[jrow]
+ diag_i[jrow+1] - diag_i[jrow];
num_nonzeros += B_int_i[j_cnt];
}
}
/*--------------------------------------------------------------------------
* initialize communication
*--------------------------------------------------------------------------*/
comm_handle = hypre_ParCSRCommHandleCreate(11,comm_pkg,
&B_int_i[1],&B_ext_i[1]);
B_int_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros);
if (data) B_int_data = hypre_CTAlloc(HYPRE_Complex, num_nonzeros*bnnz);
jdata_send_map_starts = hypre_CTAlloc(HYPRE_Int, num_sends+1);
jdata_recv_vec_starts = hypre_CTAlloc(HYPRE_Int, num_recvs+1);
start_index = B_int_i[0];
jdata_send_map_starts[0] = start_index;
counter = 0;
for (i=0; i < num_sends; i++)
{
num_nonzeros = counter;
for (j = send_map_starts[i]; j < send_map_starts[i+1]; j++)
{
jrow = send_map_elmts[j];
for (k=diag_i[jrow]; k < diag_i[jrow+1]; k++)
{
B_int_j[counter] = diag_j[k]+first_col_diag;
if (data) {
for(l = 0; l < bnnz; l++)
B_int_data[counter*bnnz+ l] = diag_data[k*bnnz+ l];
}
counter++;
}
for (k=offd_i[jrow]; k < offd_i[jrow+1]; k++)
{
B_int_j[counter] = col_map_offd[offd_j[k]];
if (data) {
for(l = 0; l < bnnz; l++)
B_int_data[counter*bnnz+ l] =
offd_data[k*bnnz+ l];
}
counter++;
}
}
num_nonzeros = counter - num_nonzeros;
start_index += num_nonzeros;
jdata_send_map_starts[i+1] = start_index;
}
tmp_comm_pkg = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(tmp_comm_pkg) = comm;
hypre_ParCSRCommPkgNumSends(tmp_comm_pkg) = num_sends;
hypre_ParCSRCommPkgNumRecvs(tmp_comm_pkg) = num_recvs;
hypre_ParCSRCommPkgSendProcs(tmp_comm_pkg) = hypre_ParCSRCommPkgSendProcs(comm_pkg);
hypre_ParCSRCommPkgRecvProcs(tmp_comm_pkg) = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
hypre_ParCSRCommPkgSendMapStarts(tmp_comm_pkg) = jdata_send_map_starts;
hypre_ParCSRCommHandleDestroy(comm_handle);
comm_handle = NULL;
/*--------------------------------------------------------------------------
* after communication exchange B_ext_i[j+1] contains the number of elements
* of a row j !
* evaluate B_ext_i and compute num_nonzeros for B_ext
*--------------------------------------------------------------------------*/
for (i=0; i < num_recvs; i++)
for (j = recv_vec_starts[i]; j < recv_vec_starts[i+1]; j++)
B_ext_i[j+1] += B_ext_i[j];
num_nonzeros = B_ext_i[num_rows_B_ext];
B_ext = hypre_CSRBlockMatrixCreate(block_size, num_rows_B_ext, num_cols_B,
num_nonzeros);
B_ext_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros);
if (data) B_ext_data = hypre_CTAlloc(HYPRE_Complex, num_nonzeros*bnnz);
for (i=0; i < num_recvs; i++)
{
start_index = B_ext_i[recv_vec_starts[i]];
num_nonzeros = B_ext_i[recv_vec_starts[i+1]]-start_index;
jdata_recv_vec_starts[i+1] = B_ext_i[recv_vec_starts[i+1]];
}
hypre_ParCSRCommPkgRecvVecStarts(tmp_comm_pkg) = jdata_recv_vec_starts;
comm_handle = hypre_ParCSRCommHandleCreate(11,tmp_comm_pkg,B_int_j,B_ext_j);
hypre_ParCSRCommHandleDestroy(comm_handle);
comm_handle = NULL;
if (data)
{
comm_handle = hypre_ParCSRBlockCommHandleCreate(1, bnnz,tmp_comm_pkg,
B_int_data, B_ext_data);
hypre_ParCSRBlockCommHandleDestroy(comm_handle);
comm_handle = NULL;
}
hypre_CSRBlockMatrixI(B_ext) = B_ext_i;
hypre_CSRBlockMatrixJ(B_ext) = B_ext_j;
if (data) hypre_CSRBlockMatrixData(B_ext) = B_ext_data;
hypre_TFree(B_int_i);
hypre_TFree(B_int_j);
if (data) hypre_TFree(B_int_data);
hypre_TFree(jdata_send_map_starts);
hypre_TFree(jdata_recv_vec_starts);
hypre_TFree(tmp_comm_pkg);
return B_ext;
}
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;
}
}
/*
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_CSRMatrix *
hypre_ParCSRMatrixExtractAExt( hypre_ParCSRMatrix *A,
HYPRE_Int data,
HYPRE_Int ** pA_ext_row_map )
{
/* Note that A's role as the first factor in A*A^T is used only
through ...CommPkgT(A), which basically says which rows of A
(columns of A^T) are needed. In all the other places where A
serves as an input, it is through its role as A^T, the matrix
whose data needs to be passed between processors. */
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
HYPRE_Int first_col_diag = hypre_ParCSRMatrixFirstColDiag(A);
HYPRE_Int first_row_index = hypre_ParCSRMatrixFirstRowIndex(A);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkgT(A);
/* ... CommPkgT(A) should identify all rows of A^T needed for A*A^T (that is
* generally a bigger set than ...CommPkg(A), the rows of B needed for A*B) */
HYPRE_Int num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
HYPRE_Int *recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
HYPRE_Int num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
HYPRE_Int *send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
HYPRE_Int *send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
hypre_CSRMatrix *diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Int *diag_i = hypre_CSRMatrixI(diag);
HYPRE_Int *diag_j = hypre_CSRMatrixJ(diag);
HYPRE_Complex *diag_data = hypre_CSRMatrixData(diag);
hypre_CSRMatrix *offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *offd_i = hypre_CSRMatrixI(offd);
HYPRE_Int *offd_j = hypre_CSRMatrixJ(offd);
HYPRE_Complex *offd_data = hypre_CSRMatrixData(offd);
HYPRE_Int num_cols_A, num_nonzeros;
HYPRE_Int num_rows_A_ext;
hypre_CSRMatrix *A_ext;
HYPRE_Int *A_ext_i;
HYPRE_Int *A_ext_j;
HYPRE_Complex *A_ext_data;
num_cols_A = hypre_ParCSRMatrixGlobalNumCols(A);
num_rows_A_ext = recv_vec_starts[num_recvs];
hypre_ParCSRMatrixExtractBExt_Arrays
( &A_ext_i, &A_ext_j, &A_ext_data, pA_ext_row_map,
&num_nonzeros,
data, 1, comm, comm_pkg,
num_cols_A, num_recvs, num_sends,
first_col_diag, first_row_index,
recv_vec_starts, send_map_starts, send_map_elmts,
diag_i, diag_j, offd_i, offd_j, col_map_offd,
diag_data, offd_data
);
A_ext = hypre_CSRMatrixCreate(num_rows_A_ext,num_cols_A,num_nonzeros);
hypre_CSRMatrixI(A_ext) = A_ext_i;
hypre_CSRMatrixJ(A_ext) = A_ext_j;
if (data) hypre_CSRMatrixData(A_ext) = A_ext_data;
return A_ext;
}
hypre_ParCSRMatrix *hypre_ParCSRAAt( hypre_ParCSRMatrix *A )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Complex *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Complex *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *A_col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *A_ext_row_map;
HYPRE_Int *row_starts_A = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int num_rows_diag_A = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_cols_offd_A = hypre_CSRMatrixNumCols(A_offd);
hypre_ParCSRMatrix *C;
HYPRE_Int *col_map_offd_C;
hypre_CSRMatrix *C_diag;
HYPRE_Complex *C_diag_data;
HYPRE_Int *C_diag_i;
HYPRE_Int *C_diag_j;
hypre_CSRMatrix *C_offd;
HYPRE_Complex *C_offd_data=NULL;
HYPRE_Int *C_offd_i=NULL;
HYPRE_Int *C_offd_j=NULL;
HYPRE_Int *new_C_offd_j;
HYPRE_Int C_diag_size;
HYPRE_Int C_offd_size;
HYPRE_Int last_col_diag_C;
HYPRE_Int num_cols_offd_C;
hypre_CSRMatrix *A_ext;
HYPRE_Complex *A_ext_data;
HYPRE_Int *A_ext_i;
HYPRE_Int *A_ext_j;
HYPRE_Int num_rows_A_ext=0;
HYPRE_Int first_row_index_A = hypre_ParCSRMatrixFirstRowIndex(A);
HYPRE_Int first_col_diag_A = hypre_ParCSRMatrixFirstColDiag(A);
HYPRE_Int *B_marker;
HYPRE_Int i;
HYPRE_Int i1, i2, i3;
HYPRE_Int jj2, jj3;
HYPRE_Int jj_count_diag, jj_count_offd;
HYPRE_Int jj_row_begin_diag, jj_row_begin_offd;
HYPRE_Int start_indexing = 0; /* start indexing for C_data at 0 */
HYPRE_Int count;
HYPRE_Int n_rows_A, n_cols_A;
HYPRE_Complex a_entry;
HYPRE_Complex a_b_product;
HYPRE_Complex zero = 0.0;
n_rows_A = hypre_ParCSRMatrixGlobalNumRows(A);
n_cols_A = hypre_ParCSRMatrixGlobalNumCols(A);
if (n_cols_A != n_rows_A)
{
hypre_printf(" Error! Incompatible matrix dimensions!\n");
return NULL;
}
/*-----------------------------------------------------------------------
* Extract A_ext, i.e. portion of A that is stored on neighbor procs
* and needed locally for A^T in the matrix matrix product A*A^T
*-----------------------------------------------------------------------*/
if (num_rows_diag_A != n_rows_A)
{
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings
*--------------------------------------------------------------------*/
if (!hypre_ParCSRMatrixCommPkg(A))
{
hypre_MatTCommPkgCreate(A);
}
A_ext = hypre_ParCSRMatrixExtractAExt( A, 1, &A_ext_row_map );
A_ext_data = hypre_CSRMatrixData(A_ext);
A_ext_i = hypre_CSRMatrixI(A_ext);
A_ext_j = hypre_CSRMatrixJ(A_ext);
num_rows_A_ext = hypre_CSRMatrixNumRows(A_ext);
}
/*-----------------------------------------------------------------------
* Allocate marker array.
*-----------------------------------------------------------------------*/
B_marker = hypre_CTAlloc(HYPRE_Int, num_rows_diag_A+num_rows_A_ext );
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
for ( i1=0; i1<num_rows_diag_A+num_rows_A_ext; ++i1 )
{
B_marker[i1] = -1;
}
hypre_ParAat_RowSizes(
&C_diag_i, &C_offd_i, B_marker,
A_diag_i, A_diag_j,
A_offd_i, A_offd_j, A_col_map_offd,
A_ext_i, A_ext_j, A_ext_row_map,
&C_diag_size, &C_offd_size,
num_rows_diag_A, num_cols_offd_A,
num_rows_A_ext,
first_col_diag_A, first_row_index_A
);
#if 0
/* debugging output: */
hypre_printf("A_ext_row_map (%i):",num_rows_A_ext);
for ( i1=0; i1<num_rows_A_ext; ++i1 ) hypre_printf(" %i",A_ext_row_map[i1] );
hypre_printf("\nC_diag_i (%i):",C_diag_size);
for ( i1=0; i1<=num_rows_diag_A; ++i1 ) hypre_printf(" %i",C_diag_i[i1] );
hypre_printf("\nC_offd_i (%i):",C_offd_size);
for ( i1=0; i1<=num_rows_diag_A; ++i1 ) hypre_printf(" %i",C_offd_i[i1] );
hypre_printf("\n");
#endif
/*-----------------------------------------------------------------------
* Allocate C_diag_data and C_diag_j arrays.
* Allocate C_offd_data and C_offd_j arrays.
*-----------------------------------------------------------------------*/
last_col_diag_C = first_row_index_A + num_rows_diag_A - 1;
C_diag_data = hypre_CTAlloc(HYPRE_Complex, C_diag_size);
C_diag_j = hypre_CTAlloc(HYPRE_Int, C_diag_size);
if (C_offd_size)
{
C_offd_data = hypre_CTAlloc(HYPRE_Complex, C_offd_size);
C_offd_j = hypre_CTAlloc(HYPRE_Int, C_offd_size);
}
/*-----------------------------------------------------------------------
* Second Pass: Fill in C_diag_data and C_diag_j.
* Second Pass: Fill in C_offd_data and C_offd_j.
*-----------------------------------------------------------------------*/
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
jj_count_diag = start_indexing;
jj_count_offd = start_indexing;
for ( i1=0; i1<num_rows_diag_A+num_rows_A_ext; ++i1 )
{
B_marker[i1] = -1;
}
/*-----------------------------------------------------------------------
* Loop over interior c-points.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_A; i1++)
{
/*--------------------------------------------------------------------
* Create diagonal entry, C_{i1,i1}
*--------------------------------------------------------------------*/
B_marker[i1] = jj_count_diag;
jj_row_begin_diag = jj_count_diag;
jj_row_begin_offd = jj_count_offd;
C_diag_data[jj_count_diag] = zero;
C_diag_j[jj_count_diag] = i1;
jj_count_diag++;
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_offd.
*-----------------------------------------------------------------*/
/* There are 3 CSRMatrix or CSRBooleanMatrix objects here:
ext*ext, ext*diag, and ext*offd belong to another processor.
diag*offd and offd*diag don't count - never share a column by definition.
So we have to do 4 cases:
diag*ext, offd*ext, diag*diag, and offd*offd.
*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
a_entry = A_diag_data[jj2];
/* diag*ext */
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of (A_ext)^T
* That is, rows i3 having a column i2 of A_ext.
* For now, for each row i3 of A_ext we crudely check _all_
* columns to see whether one matches i2.
* For each entry (i2,i3) of (A_ext)^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3) . This contributes to both the diag and offd
* blocks of C.
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_A_ext; i3++ ) {
for ( jj3=A_ext_i[i3]; jj3<A_ext_i[i3+1]; jj3++ ) {
if ( A_ext_j[jj3]==i2+first_col_diag_A ) {
/* row i3, column i2 of A_ext; or,
row i2, column i3 of (A_ext)^T */
a_b_product = a_entry * A_ext_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if ( A_ext_row_map[i3] < first_row_index_A ||
A_ext_row_map[i3] > last_col_diag_C ) { /* offd */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_offd) {
B_marker[i3+num_rows_diag_A] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3;
jj_count_offd++;
}
else
C_offd_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
else { /* diag */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_diag) {
B_marker[i3+num_rows_diag_A] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3-first_col_diag_A;
jj_count_diag++;
}
else
C_diag_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
}
}
}
}
if (num_cols_offd_A)
{
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
a_entry = A_offd_data[jj2];
/* offd * ext */
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of (A_ext)^T
* That is, rows i3 having a column i2 of A_ext.
* For now, for each row i3 of A_ext we crudely check _all_
* columns to see whether one matches i2.
* For each entry (i2,i3) of (A_ext)^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3) . This contributes to both the diag and offd
* blocks of C.
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_A_ext; i3++ ) {
for ( jj3=A_ext_i[i3]; jj3<A_ext_i[i3+1]; jj3++ ) {
if ( A_ext_j[jj3]==A_col_map_offd[i2] ) {
/* row i3, column i2 of A_ext; or,
row i2, column i3 of (A_ext)^T */
a_b_product = a_entry * A_ext_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if ( A_ext_row_map[i3] < first_row_index_A ||
A_ext_row_map[i3] > last_col_diag_C ) { /* offd */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_offd) {
B_marker[i3+num_rows_diag_A] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3;
jj_count_offd++;
}
else
C_offd_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
else { /* diag */
if (B_marker[i3+num_rows_diag_A] < jj_row_begin_diag) {
B_marker[i3+num_rows_diag_A] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3-first_row_index_A;
jj_count_diag++;
}
else
C_diag_data[B_marker[i3+num_rows_diag_A]] += a_b_product;
}
}
}
}
}
}
/* diag * diag */
/*-----------------------------------------------------------------
* Loop over entries (columns) i2 in row i1 of A_diag.
* For each such column we will find the contributions of the
* corresponding rows i2 of A^T to C=A*A^T . Now we only look
* at the local part of A^T - with columns (rows of A) living
* on this processor.
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
a_entry = A_diag_data[jj2];
/*-----------------------------------------------------------
* Loop over entries (columns) i3 in row i2 of A^T
* That is, rows i3 having a column i2 of A (local part).
* For now, for each row i3 of A we crudely check _all_
* columns to see whether one matches i2.
* This i3-loop is for the diagonal block of A.
* It contributes to the diagonal block of C.
* For each entry (i2,i3) of A^T, add A(i1,i2)*A(i3,i2)
* to C(i1,i3)
*-----------------------------------------------------------*/
for ( i3=0; i3<num_rows_diag_A; i3++ ) {
for ( jj3=A_diag_i[i3]; jj3<A_diag_i[i3+1]; jj3++ ) {
if ( A_diag_j[jj3]==i2 ) {
/* row i3, column i2 of A; or,
row i2, column i3 of A^T */
a_b_product = a_entry * A_diag_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, mark it and increment
* counter.
*--------------------------------------------------------*/
if (B_marker[i3] < jj_row_begin_diag)
{
B_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[B_marker[i3]] += a_b_product;
}
}
}
} /* end of i3 loop */
} /* end of third i2 loop */
/* offd * offd */
/*-----------------------------------------------------------
* Loop over offd columns i2 of A in A*A^T. Then
* loop over offd entries (columns) i3 in row i2 of A^T
* That is, rows i3 having a column i2 of A (local part).
* For now, for each row i3 of A we crudely check _all_
* columns to see whether one matches i2.
* This i3-loop is for the off-diagonal block of A.
* It contributes to the diag block of C.
* For each entry (i2,i3) of A^T, add A*A^T to C
*-----------------------------------------------------------*/
if (num_cols_offd_A) {
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
a_entry = A_offd_data[jj2];
for ( i3=0; i3<num_rows_diag_A; i3++ ) {
/* ... note that num_rows_diag_A == num_rows_offd_A */
for ( jj3=A_offd_i[i3]; jj3<A_offd_i[i3+1]; jj3++ ) {
if ( A_offd_j[jj3]==i2 ) {
/* row i3, column i2 of A; or,
row i2, column i3 of A^T */
a_b_product = a_entry * A_offd_data[jj3];
/*--------------------------------------------------------
* Check B_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution
*--------------------------------------------------------*/
if (B_marker[i3] < jj_row_begin_diag)
{
B_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[B_marker[i3]] += a_b_product;
}
}
}
} /* end of last i3 loop */
} /* end of if (num_cols_offd_A) */
} /* end of fourth and last i2 loop */
#if 0 /* debugging printout */
hypre_printf("end of i1 loop: i1=%i jj_count_diag=%i\n", i1, jj_count_diag );
hypre_printf(" C_diag_j=");
for ( jj3=0; jj3<jj_count_diag; ++jj3) hypre_printf("%i ",C_diag_j[jj3]);
hypre_printf(" C_diag_data=");
for ( jj3=0; jj3<jj_count_diag; ++jj3) hypre_printf("%f ",C_diag_data[jj3]);
hypre_printf("\n");
hypre_printf(" C_offd_j=");
for ( jj3=0; jj3<jj_count_offd; ++jj3) hypre_printf("%i ",C_offd_j[jj3]);
hypre_printf(" C_offd_data=");
for ( jj3=0; jj3<jj_count_offd; ++jj3) hypre_printf("%f ",C_offd_data[jj3]);
hypre_printf("\n");
hypre_printf( " B_marker =" );
for ( it=0; it<num_rows_diag_A+num_rows_A_ext; ++it )
hypre_printf(" %i", B_marker[it] );
hypre_printf( "\n" );
#endif
} /* end of i1 loop */
/*-----------------------------------------------------------------------
* Delete 0-columns in C_offd, i.e. generate col_map_offd and reset
* C_offd_j. Note that (with the indexing we have coming into this
* block) col_map_offd_C[i3]==A_ext_row_map[i3].
*-----------------------------------------------------------------------*/
for ( i=0; i<num_rows_diag_A+num_rows_A_ext; ++i )
B_marker[i] = -1;
for ( i=0; i<C_offd_size; i++ )
B_marker[ C_offd_j[i] ] = -2;
count = 0;
for (i=0; i < num_rows_diag_A + num_rows_A_ext; i++) {
if (B_marker[i] == -2) {
B_marker[i] = count;
count++;
}
}
num_cols_offd_C = count;
if (num_cols_offd_C) {
col_map_offd_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_C);
new_C_offd_j = hypre_CTAlloc(HYPRE_Int,C_offd_size);
/* ... a bit big, but num_cols_offd_C is too small. It might be worth
computing the correct size, which is sum( no. columns in row i, over all rows i )
*/
for (i=0; i < C_offd_size; i++) {
new_C_offd_j[i] = B_marker[C_offd_j[i]];
col_map_offd_C[ new_C_offd_j[i] ] = A_ext_row_map[ C_offd_j[i] ];
}
hypre_TFree(C_offd_j);
C_offd_j = new_C_offd_j;
}
/*----------------------------------------------------------------
* Create C
*----------------------------------------------------------------*/
C = hypre_ParCSRMatrixCreate(comm, n_rows_A, n_rows_A, row_starts_A,
row_starts_A, num_cols_offd_C,
C_diag_size, C_offd_size);
/* Note that C does not own the partitionings */
hypre_ParCSRMatrixSetRowStartsOwner(C,0);
hypre_ParCSRMatrixSetColStartsOwner(C,0);
C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrixData(C_diag) = C_diag_data;
hypre_CSRMatrixI(C_diag) = C_diag_i;
hypre_CSRMatrixJ(C_diag) = C_diag_j;
if (num_cols_offd_C)
{
C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrixData(C_offd) = C_offd_data;
hypre_CSRMatrixI(C_offd) = C_offd_i;
hypre_CSRMatrixJ(C_offd) = C_offd_j;
hypre_ParCSRMatrixOffd(C) = C_offd;
hypre_ParCSRMatrixColMapOffd(C) = col_map_offd_C;
}
else
hypre_TFree(C_offd_i);
/*-----------------------------------------------------------------------
* Free B_ext and marker array.
*-----------------------------------------------------------------------*/
if (num_cols_offd_A)
{
hypre_CSRMatrixDestroy(A_ext);
A_ext = NULL;
}
hypre_TFree(B_marker);
if ( num_rows_diag_A != n_rows_A )
hypre_TFree(A_ext_row_map);
return C;
}
HYPRE_Int
main( HYPRE_Int argc,
char *argv[] )
{
hypre_CSRMatrix *matrix;
hypre_CSRMatrix *matrix1;
hypre_ParCSRMatrix *par_matrix;
hypre_Vector *x_local;
hypre_Vector *y_local;
hypre_Vector *y2_local;
hypre_ParVector *x;
hypre_ParVector *x2;
hypre_ParVector *y;
hypre_ParVector *y2;
HYPRE_Int vecstride_x, idxstride_x, vecstride_y, idxstride_y;
HYPRE_Int num_procs, my_id;
HYPRE_Int local_size;
HYPRE_Int num_vectors;
HYPRE_Int global_num_rows, global_num_cols;
HYPRE_Int first_index;
HYPRE_Int i, j, ierr=0;
double *data, *data2;
HYPRE_Int *row_starts, *col_starts;
char file_name[80];
/* Initialize MPI */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &num_procs);
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &my_id);
hypre_printf(" my_id: %d num_procs: %d\n", my_id, num_procs);
if (my_id == 0)
{
matrix = hypre_CSRMatrixRead("input");
hypre_printf(" read input\n");
}
row_starts = NULL;
col_starts = NULL;
par_matrix = hypre_CSRMatrixToParCSRMatrix(hypre_MPI_COMM_WORLD, matrix,
row_starts, col_starts);
hypre_printf(" converted\n");
matrix1 = hypre_ParCSRMatrixToCSRMatrixAll(par_matrix);
hypre_sprintf(file_name,"matrix1.%d",my_id);
if (matrix1) hypre_CSRMatrixPrint(matrix1, file_name);
hypre_ParCSRMatrixPrint(par_matrix,"matrix");
hypre_ParCSRMatrixPrintIJ(par_matrix,0,0,"matrixIJ");
par_matrix = hypre_ParCSRMatrixRead(hypre_MPI_COMM_WORLD,"matrix");
global_num_cols = hypre_ParCSRMatrixGlobalNumCols(par_matrix);
hypre_printf(" global_num_cols %d\n", global_num_cols);
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(par_matrix);
col_starts = hypre_ParCSRMatrixColStarts(par_matrix);
first_index = col_starts[my_id];
local_size = col_starts[my_id+1] - first_index;
num_vectors = 3;
x = hypre_ParMultiVectorCreate( hypre_MPI_COMM_WORLD, global_num_cols,
col_starts, num_vectors );
hypre_ParVectorSetPartitioningOwner(x,0);
hypre_ParVectorInitialize(x);
x_local = hypre_ParVectorLocalVector(x);
data = hypre_VectorData(x_local);
vecstride_x = hypre_VectorVectorStride(x_local);
idxstride_x = hypre_VectorIndexStride(x_local);
for ( j=0; j<num_vectors; ++j )
for (i=0; i < local_size; i++)
data[i*idxstride_x + j*vecstride_x] = first_index+i+1 + 100*j;
x2 = hypre_ParMultiVectorCreate( hypre_MPI_COMM_WORLD, global_num_cols,
col_starts, num_vectors );
hypre_ParVectorSetPartitioningOwner(x2,0);
hypre_ParVectorInitialize(x2);
hypre_ParVectorSetConstantValues(x2,2.0);
row_starts = hypre_ParCSRMatrixRowStarts(par_matrix);
first_index = row_starts[my_id];
local_size = row_starts[my_id+1] - first_index;
y = hypre_ParMultiVectorCreate( hypre_MPI_COMM_WORLD, global_num_rows,
row_starts, num_vectors );
hypre_ParVectorSetPartitioningOwner(y,0);
hypre_ParVectorInitialize(y);
y_local = hypre_ParVectorLocalVector(y);
y2 = hypre_ParMultiVectorCreate( hypre_MPI_COMM_WORLD, global_num_rows,
row_starts, num_vectors );
hypre_ParVectorSetPartitioningOwner(y2,0);
hypre_ParVectorInitialize(y2);
y2_local = hypre_ParVectorLocalVector(y2);
data2 = hypre_VectorData(y2_local);
vecstride_y = hypre_VectorVectorStride(y2_local);
idxstride_y = hypre_VectorIndexStride(y2_local);
for ( j=0; j<num_vectors; ++j )
for (i=0; i < local_size; i++)
data2[i*idxstride_y+j*vecstride_y] = first_index+i+1 + 100*j;
hypre_ParVectorSetConstantValues(y,1.0);
hypre_printf(" initialized vectors, first_index=%i\n", first_index);
hypre_ParVectorPrint(x, "vectorx");
hypre_ParVectorPrint(y, "vectory");
hypre_MatvecCommPkgCreate(par_matrix);
hypre_ParCSRMatrixMatvec ( 1.0, par_matrix, x, 1.0, y);
hypre_printf(" did matvec\n");
hypre_ParVectorPrint(y, "result");
ierr = hypre_ParCSRMatrixMatvecT ( 1.0, par_matrix, y2, 1.0, x2);
hypre_printf(" did matvecT %d\n", ierr);
hypre_ParVectorPrint(x2, "transp");
hypre_ParCSRMatrixDestroy(par_matrix);
hypre_ParVectorDestroy(x);
hypre_ParVectorDestroy(x2);
hypre_ParVectorDestroy(y);
hypre_ParVectorDestroy(y2);
if (my_id == 0) hypre_CSRMatrixDestroy(matrix);
if (matrix1) hypre_CSRMatrixDestroy(matrix1);
/* Finalize MPI */
hypre_MPI_Finalize();
return 0;
}
