void ML_getrow_matvec(ML_Operator *matrix, double *vec, int Nvec,
double *ovec, int *Novec)
{
ML_Operator *temp, *temp2, *temp3, *temp4, *tptr;
int *cols, i;
int allocated, row_length;
if (matrix->getrow->func_ptr == NULL) {
printf("ML_getrow_matvec: empty object? \n");
exit(1);
}
temp = ML_Operator_Create(matrix->comm);
ML_Operator_Set_1Levels(temp, matrix->from, matrix->from);
ML_Operator_Set_ApplyFuncData(temp,1,Nvec,vec,Nvec,NULL,0);
ML_Operator_Set_Getrow(temp,Nvec, VECTOR_getrows);
temp->max_nz_per_row = 1;
temp->N_nonzeros = Nvec;
if (matrix->getrow->pre_comm != NULL) {
ML_exchange_rows(temp, &temp2, matrix->getrow->pre_comm);
}
else temp2 = temp;
ML_matmat_mult(matrix, temp2, &temp3);
if (matrix->getrow->post_comm != NULL)
ML_exchange_rows(temp3, &temp4, matrix->getrow->post_comm);
else temp4 = temp3;
allocated = temp4->getrow->Nrows + 1;
cols = (int *) ML_allocate(allocated*sizeof(int));
if (cols == NULL) {
printf("no space in ML_getrow_matvec()\n");
exit(1);
}
for (i = 0; i < temp4->getrow->Nrows; i++) {
ML_get_matrix_row(temp4, 1, &i, &allocated , &cols, &ovec,
&row_length, i);
if (allocated != temp4->getrow->Nrows + 1)
printf("memory problems ... we can't reallocate here\n");
}
ML_free(cols);
if ( *Novec != temp4->getrow->Nrows) {
printf("Warning: The length of ML's output vector does not agree with\n");
printf(" the user's length for the output vector (%d vs. %d).\n",
*Novec, temp4->getrow->Nrows);
printf(" indicate a problem.\n");
}
*Novec = temp4->getrow->Nrows;
if (matrix->getrow->pre_comm != NULL) {
tptr = temp2;
while ( (tptr!= NULL) && (tptr->sub_matrix != temp))
tptr = tptr->sub_matrix;
if (tptr != NULL) tptr->sub_matrix = NULL;
ML_RECUR_CSR_MSRdata_Destroy(temp2);
ML_Operator_Destroy(&temp2);
}
if (matrix->getrow->post_comm != NULL) {
tptr = temp4;
while ( (tptr!= NULL) && (tptr->sub_matrix != temp3))
tptr = tptr->sub_matrix;
if (tptr != NULL) tptr->sub_matrix = NULL;
ML_RECUR_CSR_MSRdata_Destroy(temp4);
ML_Operator_Destroy(&temp4);
}
ML_Operator_Destroy(&temp);
ML_RECUR_CSR_MSRdata_Destroy(temp3);
ML_Operator_Destroy(&temp3);
}
void ML_rap(ML_Operator *Rmat, ML_Operator *Amat,
ML_Operator *Pmat, ML_Operator *Result, int matrix_type)
{
int max_per_proc, i, j, N_input_vector;
ML_Operator *APmat, *RAPmat, *Pcomm, *RAPcomm, *APcomm, *AP2comm, *tptr;
ML_CommInfoOP *getrow_comm;
double *scales = NULL;
# ifdef ML_TIMING
double tpre,tmult,tpost,ttotal;
# endif
/* Check that N_input_vector is reasonable */
# ifdef ML_TIMING
tpre = GetClock();
ttotal = GetClock();
# endif
N_input_vector = Pmat->invec_leng;
getrow_comm = Pmat->getrow->pre_comm;
if ( getrow_comm != NULL) {
for (i = 0; i < getrow_comm->N_neighbors; i++) {
for (j = 0; j < getrow_comm->neighbors[i].N_send; j++) {
if (getrow_comm->neighbors[i].send_list[j] >= N_input_vector) {
printf("(%d) Error: N_input_vector (%d) argument to rap() is not \n", Amat->comm->ML_mypid,N_input_vector);
printf("(%d) Error: larger than %dth element (%d) sent to node %d\n", Amat->comm->ML_mypid,j+1,
getrow_comm->neighbors[i].send_list[j],
getrow_comm->neighbors[i].ML_id);
printf("(%d) Error: Amat(%d,%d) Rmat(%d,%d) Pmat(%d,%d)\n",
Amat->comm->ML_mypid,
Amat->outvec_leng,Amat->invec_leng,
Rmat->outvec_leng,Rmat->invec_leng,
Pmat->outvec_leng,Pmat->invec_leng);
fflush(stdout);
exit(1);
}
}
}
}
ML_create_unique_col_id(N_input_vector, &(Pmat->getrow->loc_glob_map),
getrow_comm, &max_per_proc, Pmat->comm);
Pmat->getrow->use_loc_glob_map = ML_YES;
if (Amat->getrow->pre_comm != NULL)
ML_exchange_rows( Pmat, &Pcomm, Amat->getrow->pre_comm);
else Pcomm = Pmat;
#ifdef DEBUG
if ( Pmat->comm->ML_mypid == 0 )
printf("ML_rap : A * P begins...\n");
#endif
# ifdef ML_TIMING
tpre = GetClock() - tpre;
tmult = GetClock();
# endif
ML_matmat_mult(Amat, Pcomm , &APmat);
# ifdef ML_TIMING
tmult = GetClock() - tmult;
tpost = GetClock();
# endif
#ifdef DEBUG
if ( Pmat->comm->ML_mypid == 0 )
printf("ML_rap : A * P ends.\n");
#endif
ML_free(Pmat->getrow->loc_glob_map); Pmat->getrow->loc_glob_map = NULL;
Pmat->getrow->use_loc_glob_map = ML_NO;
if (Amat->getrow->pre_comm != NULL) {
tptr = Pcomm;
while ( (tptr!= NULL) && (tptr->sub_matrix != Pmat))
tptr = tptr->sub_matrix;
if (tptr != NULL) tptr->sub_matrix = NULL;
ML_RECUR_CSR_MSRdata_Destroy(Pcomm);
ML_Operator_Destroy(&Pcomm);
}
if (Amat->getrow->post_comm != NULL) {
ML_exchange_rows(APmat, &APcomm, Amat->getrow->post_comm);
}
else APcomm = APmat;
/* Take into account any scaling in Amat */
if (Rmat->from != NULL)
ML_DVector_GetDataPtr(Rmat->from->Amat_Normalization,&scales);
if (scales != NULL)
ML_Scale_CSR(APcomm, scales, 0);
if (Rmat->getrow->pre_comm != NULL)
ML_exchange_rows( APcomm, &AP2comm, Rmat->getrow->pre_comm);
else AP2comm = APcomm;
# ifdef ML_TIMING
tpost = GetClock() - tpost;
if ( Pmat->comm->ML_mypid == 0 && ML_Get_PrintLevel() > 5) {
int level=-1;
if (Amat->from != NULL)
level = Amat->from->levelnum-1;
printf("Timing summary (in seconds) for product RAP on level %d\n", level);
printf(" (level %d) RAP right: pre-multiply communication time = %3.2e\n", level, tpre);
printf(" (level %d) RAP right: multiply time = %3.2e\n", level, tmult);
printf(" (level %d) RAP right: post-multiply communication time = %3.2e\n", level, tpost);
}
# endif
#ifdef DEBUG
if ( Pmat->comm->ML_mypid == 0 )
printf("ML_rap : R * AP begins...\n");
#endif
# ifdef ML_TIMING
tmult = GetClock();
# endif
ML_matmat_mult(Rmat,AP2comm, &RAPmat);
#ifdef DEBUG
if ( Pmat->comm->ML_mypid == 0 )
printf("ML_rap : R * AP ends.\n");
#endif
ML_RECUR_CSR_MSRdata_Destroy(AP2comm);
ML_Operator_Destroy(&AP2comm);
# ifdef ML_TIMING
tmult = GetClock()-tmult;
tpost = GetClock();
# endif
if (Rmat->getrow->post_comm != NULL)
ML_exchange_rows( RAPmat, &RAPcomm, Rmat->getrow->post_comm);
else RAPcomm = RAPmat;
scales = NULL;
if (Rmat->to != NULL)
ML_DVector_GetDataPtr(Rmat->to->Amat_Normalization,&scales);
if (scales != NULL)
ML_Scale_CSR(RAPcomm, scales, 1);
RAPcomm->num_PDEs = Amat->num_PDEs;
RAPcomm->num_rigid = Amat->num_rigid;
if (matrix_type == ML_MSR_MATRIX)
ML_back_to_local(RAPcomm, Result, max_per_proc);
else if (matrix_type == ML_CSR_MATRIX)
ML_back_to_csrlocal(RAPcomm, Result, max_per_proc);
else if (matrix_type == ML_EpetraCRS_MATRIX)
#ifdef ML_WITH_EPETRA
ML_back_to_epetraCrs(RAPcomm, Result, Rmat, Pmat);
#else
pr_error("ML_RAP: ML_EpetraCRS_MATRIX requires epetra to be compiled in.\n");
#endif
else pr_error("ML_RAP: Unknown matrix type\n");