int ML_strcmp(char *input, char *string)
{
/* Similar to 'C' strcmp except this one converts everything to lower case. */
int i;
char *input_copy, *string_copy;
input_copy = (char *) ML_allocate(sizeof(char)*(strlen(input)+1));
string_copy = (char *) ML_allocate(sizeof(char)*(strlen(string)+1));
strcpy(input_copy,input);
strcpy(string_copy,string);
i = 0;
while ((input_copy[i] != '\0') && (input_copy[i] != '\n')) {
if ((input_copy[i] >= 'A') && (input_copy[i] <= 'Z'))
input_copy[i] = 'a' + input_copy[i] - 'A';
i++;
}
i = 0;
while ((string_copy[i] != '\0') && (string_copy[i] != '\n')) {
if ((string_copy[i] >= 'A') && (string_copy[i] <= 'Z'))
string_copy[i] = 'a' + string_copy[i] - 'A';
i++;
}
i = strcmp(input_copy, string_copy);
ML_free(input_copy);
ML_free(string_copy);
return(i);
}
// ================================================ ====== ==== ==== == =
// Copied from ml_agg_genP.c
static void ML_Finalize_Aux(ML_Operator *A)
{
int i;
A->getrow->func_ptr = A->aux_data->aux_func_ptr;
A->aux_data->aux_func_ptr = 0;
for (i = 0 ; i < A->aux_data->filter_size ; ++i)
ML_free((A->aux_data->filter[i]));
ML_free(A->aux_data->filter);
}
int ML_Reitzinger_Check_Hierarchy(ML *ml, ML_Operator **Tmat_array, int incr_or_decr)
{
int i,j;
int finest_level, coarsest_level;
ML_Operator *Amat, *Tmat;
double *randvec, *result, *result1;
double dnorm;
finest_level = ml->ML_finest_level;
coarsest_level = ml->ML_coarsest_level;
if (incr_or_decr == ML_INCREASING) {
if (ml->comm->ML_mypid == 0) {
printf("ML_Reitzinger_Check_Hierarchy: ML_INCREASING is not supported ");
printf(" at this time. Not checking hierarchy.\n");
}
return 1;
}
if ( ML_Get_PrintLevel() > 5 ) {
printf("ML_Reitzinger_Check_Hierarchy: Checking null space\n");
}
for (i=finest_level; i>coarsest_level; i--) {
Amat = ml->Amat+i;
Tmat = Tmat_array[i];
/* normalized random vector */
randvec = (double *) ML_allocate(Tmat->invec_leng * sizeof(double) );
ML_random_vec(randvec,Tmat->invec_leng, ml->comm);
dnorm = sqrt( ML_gdot(Tmat->invec_leng, randvec, randvec, ml->comm) );
for (j=0; j<Tmat->invec_leng; j++) randvec[j] /= dnorm;
result = (double *) ML_allocate(Amat->invec_leng * sizeof(double) );
result1 = (double *) ML_allocate(Amat->outvec_leng * sizeof(double) );
ML_Operator_Apply(Tmat, Tmat->invec_leng, randvec,
Tmat->outvec_leng, result);
ML_Operator_Apply(Amat, Amat->invec_leng, result,
Amat->outvec_leng, result1);
dnorm = sqrt( ML_gdot(Amat->outvec_leng, result1, result1, ml->comm) );
if ( (ML_Get_PrintLevel() > 5) && (ml->comm->ML_mypid == 0) ) {
printf("Level %d: for random v, ||S*T*v|| = %15.10e\n",i,dnorm);
}
ML_free(randvec);
ML_free(result);
ML_free(result1);
}
if ( (ML_Get_PrintLevel() > 5) && (ml->comm->ML_mypid == 0) ) printf("\n");
return 0;
}
int ML_Smoother_Ifpack(ML_Smoother *sm,int inlen,double x[],int outlen,
double rhs[])
{
ML_Smoother *smooth_ptr = (ML_Smoother *) sm;
void *Ifpack_Handle = smooth_ptr->smoother->data;
double* x2 = NULL,* rhs2 = NULL;
/*int i;*/
int n, kk;
int one_int = 1;
double minus_one_double = -1.0;
if (sm->init_guess == ML_NONZERO)
{
n = sm->my_level->Amat->invec_leng;
assert (n == sm->my_level->Amat->outvec_leng);
rhs2 = (double*) ML_allocate(sizeof(double) * (n + 1));
x2 = (double*) ML_allocate(sizeof(double) * (n + 1));
ML_Operator_Apply(sm->my_level->Amat, n, x, n, rhs2);
DCOPY_F77(&n, x, &one_int, x2, &one_int);
DAXPY_F77(&n, &minus_one_double, rhs, &one_int, rhs2, &one_int);
ML_Ifpack_Solve(Ifpack_Handle, x2, rhs2);
DAXPY_F77(&n, &minus_one_double, x2, &one_int, x, &one_int);
ML_free(rhs2);
ML_free(x2);
}
else
ML_Ifpack_Solve(Ifpack_Handle, x, rhs);
for (kk = 1; kk < sm->ntimes; kk++) {
n = sm->my_level->Amat->invec_leng;
assert (n == sm->my_level->Amat->outvec_leng);
rhs2 = (double*) ML_allocate(sizeof(double) * (n + 1));
x2 = (double*) ML_allocate(sizeof(double) * (n + 1));
ML_Operator_Apply(sm->my_level->Amat, n, x, n, rhs2);
DCOPY_F77(&n, x, &one_int, x2, &one_int);
DAXPY_F77(&n, &minus_one_double, rhs, &one_int, rhs2, &one_int);
ML_Ifpack_Solve(Ifpack_Handle, x2, rhs2);
DAXPY_F77(&n, &minus_one_double, x2, &one_int, x, &one_int);
ML_free(rhs2);
ML_free(x2);
}
return 0;
} /* ML_Smoother_Ifpack */
int ML_qr_fix_Destroy(void)
{
if (xCDeadNodDof == NULL)
return(0);
/* loop over all fields and release them */
if (xCDeadNodDof->xDeadNodDof) ML_free(xCDeadNodDof->xDeadNodDof);
/* free the structure itself */
ML_free(xCDeadNodDof);
xCDeadNodDof = NULL;
return(0);
}
int ML_memory_clean( char *name, int inlen )
{
int i, j, clean_flag, leng;
void *mem_ptr;
leng = inlen;
if ( inlen > 3 ) leng = 3;
if ( inlen < 0 ) leng = 0;
for ( i = 0; i < MAX_MALLOC_LOG; i++ )
{
if (malloc_leng_log[i] != -1)
{
clean_flag = 0;
for ( j = 0; j < leng; j++ )
{
if ( malloc_name_log[i][j] != name[j] )
{
clean_flag = 1;
break;
}
}
if ( clean_flag == 0 )
{
mem_ptr = (void *) malloc_addr_log[i];
ML_free( mem_ptr );
malloc_leng_log[i] = -1;
}
}
}
return 0;
}
// ======================================================================
MultiVector GetDiagonal(const Operator& A, const int offset)
{
// FIXME
if (A.GetDomainSpace() != A.GetRangeSpace())
ML_THROW("Currently only square matrices are supported", -1);
MultiVector D(A.GetDomainSpace());
D = 0.0;
ML_Operator* matrix = A.GetML_Operator();
if (matrix->getrow == NULL)
ML_THROW("getrow() not set!", -1);
int row_length;
int allocated = 128;
int* bindx = (int *) ML_allocate(allocated*sizeof(int ));
double* val = (double *) ML_allocate(allocated*sizeof(double));
for (int i = 0 ; i < matrix->getrow->Nrows; i++) {
int GlobalRow = A.GetGRID(i);
ML_get_matrix_row(matrix, 1, &i, &allocated, &bindx, &val,
&row_length, 0);
for (int j = 0; j < row_length; j++) {
D(i) = 0.0;
if (A.GetGCID(bindx[j]) == GlobalRow + offset) {
D(i) = val[j];
break;
}
}
}
ML_free(val);
ML_free(bindx);
return (D);
}
void sample1(struct data *Afine_data, struct data *Acoarse_data,
struct data *Rmat_data, struct data *Pmat_data,
double *sol, double *rhs )
{
ML *my_ml;
int i;
int fine_grid, output_level = 10, N_grids = 2, grid0 = 0, grid1 = 1;
int Nfine, Ncoarse;
double *diagonal;
Nfine = Rmat_data->from_size;
Ncoarse = Rmat_data->to_size;
diagonal = (double *) malloc(Nfine*sizeof(double));
for (i = 0; i < Nfine; i++) diagonal[i] = 2.;
fine_grid = grid1;
ML_Create (&my_ml, N_grids);
ML_Set_OutputLevel( my_ml, output_level);
ML_Init_Amatrix (my_ml, grid1, Nfine, Nfine,(void *) Afine_data);
ML_Set_Amatrix_Getrow(my_ml, grid1, myAgetrow, my_comm, Nfine+1);
ML_Set_Amatrix_Matvec(my_ml, grid1, mymatvec);
ML_Set_Amatrix_Diag (my_ml, grid1, Nfine, diagonal);
ML_Gen_Smoother_Jacobi(my_ml, grid1, ML_PRESMOOTHER, 2, ML_DEFAULT);
ML_Init_Prolongator(my_ml, grid0, grid1, Ncoarse,Nfine,(void *)Pmat_data);
ML_Set_Prolongator_Getrow(my_ml, grid0, myPgetrow, my_comm, Ncoarse+1);
ML_Set_Prolongator_Matvec(my_ml, grid0, myinterp);
ML_Init_Restrictor(my_ml, grid1, grid0, Nfine, Ncoarse,(void *)Rmat_data);
ML_Set_Restrictor_Getrow(my_ml, grid1, myRgetrow, my_comm, Nfine+1);
ML_Set_Restrictor_Matvec(my_ml, grid1, myrestrict);
ML_Gen_AmatrixRAP(my_ml,grid1, grid0);
#ifdef SUPERLU
ML_Gen_CoarseSolverSuperLU(my_ml, grid0);
#else
ML_Gen_Smoother_Jacobi(my_ml, grid0, ML_PRESMOOTHER, 100, ML_DEFAULT);
#endif
/* ML_Gen_Smoother_Jacobi(my_ml, grid0, ML_PRESMOOTHER, 200, ML_DEFAULT); */
/* ML_Gen_Smoother_GaussSeidel(my_ml, grid0, ML_PRESMOOTHER, 200, 1.); */
ML_Gen_Solver (my_ml, 0, fine_grid, grid0);
ML_Iterate(my_ml, sol, rhs);
ML_Destroy(&my_ml);
ML_free(diagonal);
}
void sample3(struct data *Afine_data, struct data *Acoarse_data,
struct data *Rmat_data, struct data *Pmat_data,
double *sol, double *rhs )
{
ML *my_ml;
double *diagonal;
int i, fine_grid, output_level = 10, N_grids = 2, grid0 = 1, grid1 = 0;
int Nfine, Ncoarse;
Nfine = Rmat_data->from_size;
Ncoarse = Rmat_data->to_size;
diagonal = (double *) malloc(Nfine*sizeof(double));
for (i = 0; i < Nfine; i++) diagonal[i] = 2.;
fine_grid = grid1;
ML_Create (&my_ml, N_grids);
ML_Set_OutputLevel(my_ml, output_level);
ML_Init_Amatrix (my_ml, grid1, Nfine, Nfine, (void *) Afine_data);
ML_Set_Amatrix_Matvec(my_ml, grid1, mymatvec );
ML_Set_Amatrix_Diag (my_ml, grid1, Nfine, diagonal);
ML_Gen_Smoother_Jacobi(my_ml, grid1, ML_PRESMOOTHER, 2, ML_DEFAULT);
ML_Init_Amatrix (my_ml, grid0, Ncoarse, Ncoarse, (void *) Acoarse_data);
ML_Set_Amatrix_Matvec(my_ml, grid0, mymatvec);
ML_Set_Amatrix_Diag (my_ml, grid0, Ncoarse, diagonal);
ML_Gen_Smoother_Jacobi(my_ml, grid0, ML_PRESMOOTHER, 200, ML_DEFAULT);
ML_Init_Prolongator(my_ml, grid0, grid1, Ncoarse, Nfine, (void*)Pmat_data);
ML_Set_Prolongator_Matvec(my_ml, grid0, myinterp);
ML_Init_Restrictor(my_ml, grid1, grid0, Nfine, Ncoarse,(void *)Rmat_data);
ML_Set_Restrictor_Matvec(my_ml, grid1, myrestrict);
ML_Gen_Solver (my_ml, 0, fine_grid, grid0);
ML_free(diagonal);
ML_Iterate(my_ml, sol, rhs);
ML_Destroy(&my_ml);
}
void ML_rap_check(ML *ml, ML_Operator *RAP, ML_Operator *R,
ML_Operator *A, ML_Operator *P, int iNvec, int oNvec)
{
int i,j;
double *vec1, *vec2, *vec3, *vec4, *vec5;
double norm1, norm2;
#ifdef DEBUG
printf("ML_rap_check begins ...\n");
#endif
if (RAP->getrow->ML_id != ML_ID_MATRIX) {
if (ml->comm->ML_mypid == 0)
printf("ML_rap_check: RAP is the wrong object (=%d). \n",
RAP->getrow->ML_id);
exit(1);
}
if (R->getrow->ML_id != ML_ID_MATRIX) {
if (ml->comm->ML_mypid == 0)
printf("ML_rap_check: R is the wrong object (=%d). \n",
RAP->getrow->ML_id);
exit(1);
}
if (P->getrow->ML_id != ML_ID_MATRIX) {
if (ml->comm->ML_mypid == 0)
printf("ML_rap_check: P is the wrong object (=%d). \n",
RAP->getrow->ML_id);
exit(1);
}
if (A->getrow->ML_id != ML_ID_MATRIX) {
if (ml->comm->ML_mypid == 0)
printf("ML_rap_check: A is the wrong object (=%d). \n",
RAP->getrow->ML_id);
exit(1);
}
/* j is the number of external variables */
j = 0;
for (i = 0; i < RAP->getrow->pre_comm->N_neighbors; i++)
j += RAP->getrow->pre_comm->neighbors[i].N_rcv;
vec1 = (double *) ML_allocate((iNvec+ 1 + j)*sizeof(double));
vec2 = (double *) ML_allocate((P->getrow->Nrows+1)*sizeof(double));
vec3 = (double *) ML_allocate((A->getrow->Nrows+1)*sizeof(double));
vec4 = (double *) ML_allocate((oNvec + 1)*sizeof(double));
vec5 = (double *) ML_allocate((oNvec + 1)*sizeof(double));
for (i = 0; i < iNvec; i++)
vec1[i] = (double) (ml->comm->ML_mypid*2301 + i*7 + 1);
j = P->getrow->Nrows;
ML_getrow_matvec(P, vec1, iNvec, vec2,&j);
i = A->getrow->Nrows;
ML_getrow_matvec(A, vec2, j, vec3,&i);
ML_getrow_matvec(R, vec3, i, vec4,&oNvec);
/* j is the number of variables sent in communication */
j = 0;
for (i = 0; i < RAP->getrow->pre_comm->N_neighbors; i++)
j += RAP->getrow->pre_comm->neighbors[i].N_send;
ML_restricted_MSR_mult( RAP, oNvec, vec1, vec5, j);
norm1 = sqrt(ML_gdot(oNvec, vec5, vec5, ml->comm));
for (i = 0; i < oNvec; i++) vec5[i] -= vec4[i];
norm2 = sqrt(ML_gdot(oNvec, vec5, vec5, ml->comm));
if (norm2 > norm1*1e-10) {
norm2 = sqrt(ML_gdot(oNvec, vec4, vec4, ml->comm));
if (ml->comm->ML_mypid == 0) {
printf("***************************************\n");
printf("RAP seems inaccurate:\n");
printf(" || RAP v ||_2 = %e\n\n", norm1);
printf(" || R (A (P v)) ||_2 = %e\n",norm2);
printf("***************************************\n");
}
}
ML_free(vec5);
ML_free(vec4);
ML_free(vec3);
ML_free(vec2);
ML_free(vec1);
#ifdef DEBUG
printf("ML_rap_check ends ...\n");
#endif
}
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");
ML_Operator *user_T_build(struct user_partition *Edge_Partition,
struct user_partition *Node_Partition,
ML_Operator *Kn_mat)
{
int nx, i, ii, jj, horv, Ncols, Nexterns;
int *Tmat_bindx;
double *Tmat_val;
ML_Operator *Tmat;
struct ML_CSR_MSRdata *csr_data;
ML_Comm *comm;
struct aztec_context *aztec_context;
int global_id;
int Nlocal_nodes, Nlocal_edges;
int nz_ptr;
Nlocal_nodes = Node_Partition->Nlocal;
Nlocal_edges = Edge_Partition->Nlocal;
nx = (int) sqrt( ((double) Node_Partition->Nglobal) + .00001);
#ifdef periodic
nx = (int) sqrt( ((double) Node_Partition->Nglobal - 2) + .00001);
#endif
Tmat_bindx = (int *) malloc((6*Nlocal_edges+5)*sizeof(int));
Tmat_val = (double *) malloc((6*Nlocal_edges+5)*sizeof(double));
Tmat_bindx[0] = Nlocal_edges + 1;
for (i = 0; i < Nlocal_edges; i++) {
global_id = (Edge_Partition->my_global_ids)[i];
Tmat_bindx[i+1] = Tmat_bindx[i] + 2;
#ifdef periodic
Tmat_bindx[i+1] += 1;
#endif
Tmat_val[i] = 0.0;
inv2dindex(global_id, &ii, &jj, nx, &horv);
nz_ptr = Tmat_bindx[i];
if (horv == HORIZONTAL) {
Tmat_bindx[nz_ptr] = southwest2d(ii,jj,nx); Tmat_val[nz_ptr++] = -1.;
Tmat_bindx[nz_ptr] = southeast2d(ii,jj,nx); Tmat_val[nz_ptr++] = 1.;
#ifdef periodic
Tmat_bindx[nz_ptr] = Nlocal_nodes-2; Tmat_val[nz_ptr++] = 1.;
#endif
}
else {
Tmat_bindx[nz_ptr] = northwest2d(ii,jj,nx); Tmat_val[nz_ptr++] = -1.;
Tmat_bindx[nz_ptr] = southwest2d(ii,jj,nx); Tmat_val[nz_ptr++] = 1.;
#ifdef periodic
Tmat_bindx[nz_ptr] = Nlocal_nodes - 1; Tmat_val[nz_ptr++] = 1.;
#endif
}
}
/* Convert the MSR matrix to a CSR matrix. Then use a modified Aztec */
/* routine to convert the global CSR matrix to a local ML matrix */
/* Since this routine does not compute the communication structure, */
/* we assume that it is identical to Kn's and just clone it. */
csr_data = (struct ML_CSR_MSRdata *) ML_allocate(sizeof(struct ML_CSR_MSRdata));
csr_data->columns = Tmat_bindx;
csr_data->values = Tmat_val;
ML_MSR2CSR(csr_data, Nlocal_edges, &Ncols);
aztec_context = (struct aztec_context *) Kn_mat->data;
Nexterns = (aztec_context->Amat->data_org)[AZ_N_external];
Nexterns = 0;
ML_Comm_Create( &comm);
AZ_Tmat_transform2ml(Nexterns, Node_Partition->needed_external_ids,
reordered_node_externs,
Tmat_bindx, Tmat_val, csr_data->rowptr, Nlocal_nodes,
Node_Partition->my_global_ids,
comm, Nlocal_edges, &Tmat);
ML_free(csr_data);
Tmat->data_destroy = ML_CSR_MSRdata_Destroy;
ML_CommInfoOP_Clone(&(Tmat->getrow->pre_comm), Kn_mat->getrow->pre_comm);
return(Tmat);
}
void ML_Amesos_Destroy(void *data)
{
#ifdef TFLOP
if( Level__ != -1 ) {
printf("Amesos (level %d) : Time for solve = %f (s)\n",Level__,TimeForSolve__);
if( NumSolves__ )
printf("Amesos (level %d) : avg time for solve = %f (s) ( # solve = %d)\n",Level__,TimeForSolve__/NumSolves__,NumSolves__);
else
printf("Amesos (level %d) : no solve\n",Level__);
}
#else
if( false && Level__ != -1 ) { // MS // I don't like this output any more
std::cout << std::endl;
std::cout << "Amesos (level " << Level__
<< ") : Time for solve = "
<< TimeForSolve__ << " (s)" << std::endl;
if( NumSolves__ )
std::cout << "Amesos (level " << Level__
<< ") : avg time for solve = "
<< TimeForSolve__/NumSolves__ << " (s) ( # solves = "
<< NumSolves__ << ")" << std::endl;
else
std::cout << "Amesos (level " << Level__
<< ") : no solve" << std::endl;
#ifdef ML_AMESOS_DEBUG
std::cout << "Amesos (level " << Level__
<< ") : max (over solves) ||Ax - b|| = " << setiosflags(ios::scientific) << MaxError__ << std::endl;
#endif
std::cout << std::endl;
}
#endif
Amesos_Handle_Type *Amesos_Handle = (Amesos_Handle_Type*) data;
if (Amesos_Handle->A_Base == 0) {
ML_free(Amesos_Handle);
return;
}
Amesos_BaseSolver *A_Base = (Amesos_BaseSolver *) Amesos_Handle->A_Base;
const Epetra_LinearProblem *Amesos_LinearProblem;
Amesos_LinearProblem = A_Base->GetProblem();
# ifdef ML_MPI
const Epetra_MpiComm *comm = dynamic_cast<const Epetra_MpiComm*>(&(Amesos_LinearProblem->GetOperator()->Comm()));
if (comm == 0) {
printf("ML_Amesos_Destroy: error getting MPI_Comm object\n");
exit(EXIT_FAILURE);
}
MPI_Comm subcomm = comm->GetMpiComm();
# endif
delete A_Base ;
delete Amesos_LinearProblem->GetOperator();
delete Amesos_LinearProblem ;
# ifdef ML_MPI
if (Amesos_Handle->freeMpiComm == 1) MPI_Comm_free(&subcomm);
# endif
ML_free(Amesos_Handle);
} /*ML_Amesos_Destroy()*/
void ML_interp_check(ML *ml, int coarse_level, int fine_level)
{
int ii, jj, ncoarse, nfine;
double *c_data, *f_data, coords[3], dtemp, d2, dlargest;
ML_GridFunc *coarse_funs, *fine_funs;
void *coarse_data, *fine_data;
int nfine_eqn, ncoarse_eqn, stride = 1;
/* check an interpolated linear function */
coarse_data = ml->SingleLevel[coarse_level].Grid->Grid;
fine_data = ml->SingleLevel[ fine_level].Grid->Grid;
coarse_funs = ml->SingleLevel[coarse_level].Grid->gridfcn;
fine_funs = ml->SingleLevel[ fine_level].Grid->gridfcn;
if ( (coarse_data==NULL)||(fine_data==NULL)) {
printf("ML_interp_check: grid data not found?\n");
exit(1);
}
if ( (coarse_funs==NULL)||(fine_funs==NULL)) {
printf("ML_interp_check: grid functions not found?\n");
exit(1);
}
if ( (coarse_funs->USR_grid_get_nvertices == 0) ||
( fine_funs->USR_grid_get_nvertices == 0)) {
printf("ML_interp_check: USR_grid_get_nvertices not found?\n");
exit(1);
}
ncoarse = coarse_funs->USR_grid_get_nvertices(coarse_data);
nfine = fine_funs->USR_grid_get_nvertices( fine_data);
nfine_eqn = ml->SingleLevel[coarse_level].Pmat->outvec_leng;
ncoarse_eqn = ml->SingleLevel[coarse_level].Pmat->invec_leng;
c_data = (double *) ML_allocate(ncoarse_eqn*sizeof(double));
f_data = (double *) ML_allocate(nfine_eqn*sizeof(double));
for (ii = 0; ii < ncoarse_eqn; ii++) c_data[ii] = 0.;
for (ii = 0; ii < nfine_eqn; ii++) f_data[ii] = 0.;
/* ASSUMING that for each grid point on this processor there are a */
/* set of equations and that all points at a grid point are numbered */
/* consecutively !!!!!!!!!!!!!! */
stride = nfine_eqn/nfine;
for (ii = 0 ; ii < ncoarse ; ii++) {
coarse_funs->USR_grid_get_vertex_coordinate(coarse_data,ii,coords);
for (jj = 0; jj < stride; jj++) {
c_data[ii*stride + jj] = coords[0] + 3.*coords[1] + .5;
}
}
ML_Operator_Apply(ml->SingleLevel[coarse_level].Pmat,ncoarse_eqn,c_data,
nfine_eqn,f_data);
dlargest = 0.0;
for (ii = 0 ; ii < nfine; ii++) {
fine_funs->USR_grid_get_vertex_coordinate(fine_data , ii, coords);
dtemp = coords[0] + 3.*coords[1] + .5;
d2 = ML_dabs(dtemp - f_data[ii*stride])/(ML_dabs(dtemp)+1.e-9);
/* Ray debugging
if ( d2 > 1.e-8)
printf("%d: f_data[%d] = %e %e | %e %e\n",ml->comm->ML_mypid,
ii,f_data[ii*stride],dtemp,coords[0],coords[1]);
*/
if ( d2 > dlargest) {
dlargest = d2;
}
}
ML_free(f_data);
ML_free(c_data);
}
void ML_get_row_CSR_norow_map(ML_Operator *input_matrix, int N_requested_rows,
int requested_rows[], int *allocated_space, int **columns,
double **values, int row_lengths[], int index)
{
int i, *mapper, *t1, row;
ML_Operator *next;
double *t2;
struct ML_CSR_MSRdata *matrix;
int *rowptr, *bindx, *col_ptr, itemp, j;
double *val, *val_ptr;
#ifdef DEBUG2
if (N_requested_rows != 1) {
printf("ML_get_matrix_row is currently implemented for only 1 row");
printf(" at a time.\n");
exit(1);
}
#endif
row = requested_rows[0];
#ifdef DEBUG2
if ( (row >= input_matrix->getrow->Nrows) || (row < 0) ) {
row_lengths[0] = 0;
return;
}
#endif
next = input_matrix->sub_matrix;
while ( (next != NULL) && (row < next->getrow->Nrows) ) {
input_matrix = next;
next = next->sub_matrix;
}
if (next != NULL) row -= next->getrow->Nrows;
matrix = (struct ML_CSR_MSRdata *) input_matrix->data;
rowptr = matrix->rowptr;
itemp = rowptr[row];
bindx = &(matrix->columns[itemp]);
val = &(matrix->values[itemp]);
*row_lengths = rowptr[row+1] - itemp;
if (*row_lengths+index > *allocated_space) {
*allocated_space = 2*(*allocated_space) + 1;
if (*row_lengths+index > *allocated_space)
*allocated_space = *row_lengths + 5 + index;
t1 = (int *) ML_allocate(*allocated_space*sizeof(int ));
t2 = (double *) ML_allocate(*allocated_space*sizeof(double));
if (t2 == NULL) {
printf("Not enough space to get a matrix row. A row length of \n");
printf("%d was not sufficient\n",(*allocated_space-1)/2);
fflush(stdout);
ML_avoid_unused_param( (void *) &N_requested_rows);
exit(1);
}
for (i = 0; i < index; i++) t1[i] = (*columns)[i];
for (i = 0; i < index; i++) t2[i] = (*values)[i];
ML_free(*columns); ML_free(*values);
*columns = t1;
*values = t2;
}
col_ptr = &((*columns)[index]);
val_ptr = &((*values)[index]);
for (j = 0 ; j < *row_lengths; j++) {
*col_ptr++ = *bindx++;
}
for (j = 0 ; j < *row_lengths; j++) {
*val_ptr++ = *val++;
}
if ( (input_matrix->getrow->use_loc_glob_map == ML_YES)) {
mapper = input_matrix->getrow->loc_glob_map;
for (i = 0; i < row_lengths[0]; i++)
(*columns)[i+index] = mapper[(*columns)[index+i]];
}
}
/*----------------------------------------------------------------------*
| (private) m.gee 04/05|
| set the smoother on this nonlinear level |
*----------------------------------------------------------------------*/
bool ML_NOX::ML_Nox_NonlinearLevel::Set_Smoother(ML* ml,
ML_Aggregate* ag,
int level,
int nlevel,
ML* thislevel_ml,
ML_Aggregate* thislevel_ag,
string smoothertype,
int nsmooth)
{
if (smoothertype == "SGS")
ML_Gen_Smoother_SymGaussSeidel(thislevel_ml,0,ML_BOTH,nsmooth,1.0);
else if (smoothertype == "Jacobi")
ML_Gen_Smoother_Jacobi(thislevel_ml,0,ML_BOTH,nsmooth,0.25);
else if (smoothertype == "AmesosKLU")
ML_Gen_Smoother_Amesos(thislevel_ml,0,ML_AMESOS_KLU,-1,0.0);
else if ( (smoothertype == "MLS") || (smoothertype == "Cheby") )
ML_Gen_Smoother_Cheby(thislevel_ml,0,ML_BOTH,30.,nsmooth);
else if (smoothertype == "BSGS")
{
int nblocks = 0;
int* blocks = NULL;
int* blockpde = NULL;
bool needfree = false;
// try to get nodal blocks from the VBMETIS aggregation scheme
ML_Aggregate_Get_Vblocks_CoarsenScheme_VBMETIS(ag,level,nlevel,
&nblocks,&blocks,&blockpde);
if (nblocks && blocks)
needfree=true;
else
ML_Gen_Blocks_Aggregates(ag,level,&nblocks,&blocks);
ML_Gen_Smoother_VBlockSymGaussSeidel(thislevel_ml,0,ML_BOTH,nsmooth,1.,
nblocks,blocks);
if (needfree)
{
ML_free(blocks);
ML_free(blockpde);
}
}
else if (smoothertype == "Bcheby")
{
int nblocks = 0;
int* blocks = NULL;
int* blockpde = NULL;
bool needfree = false;
// try to get nodal blocks from the VBMETIS aggregation scheme
ML_Aggregate_Get_Vblocks_CoarsenScheme_VBMETIS(ag,level,nlevel,
&nblocks,&blocks,&blockpde);
if (nblocks && blocks)
needfree=true;
else
ML_Gen_Blocks_Aggregates(ag,level,&nblocks,&blocks);
ML_Gen_Smoother_BlockDiagScaledCheby(thislevel_ml,0,ML_BOTH,30.,nsmooth,
nblocks,blocks);
if (needfree)
{
ML_free(blocks);
ML_free(blockpde);
}
}
else
{
cout << "**ERR**: ML_NOX::ML_Nox_NonlinearLevel::Setsmoother:\n"
<< "**ERR**: unknown type of smoother: " << smoothertype << "\n"
<< "**ERR**: file/line: " << __FILE__ << "/" << __LINE__ << "\n"; throw -1;
}
return true;
}
int ML_memory_free(void ** var_ptr)
{
int ndouble=sizeof(double), index, index2, *int_ptr;
char *char_ptr;
/* ------------------------------------------------------------------ */
/* Extract head and tail information and check to see if they match. */
/* If not, flag error. */
/* ------------------------------------------------------------------ */
char_ptr = (char *) (*var_ptr);
if (char_ptr != NULL)
{
int_ptr = (int *) ((ml_size_t) char_ptr - ndouble);
index = (*int_ptr) - 1;
if ( index >= 0 )
{
if (index > MAX_MALLOC_LOG)
{
if ( global_comm != NULL )
printf("%d : ML_memory_free error : header invalid(%d).\n",
global_comm->ML_mypid, index);
else
printf("ML_memory_free error : header invalid(%d).\n",index);
exit(-1);
}
int_ptr = (int *) ((ml_size_t) char_ptr + malloc_leng_log[index] -
2 * ndouble);
index2 = (*int_ptr);
if (index != index2-1)
{
if ( global_comm == NULL )
printf("ML_memory_free warning : header/tail mismatch - %d\n",
index);
else
printf("%d : ML_memory_free warning : header/tail mismatch - %d\n",
global_comm->ML_mypid, index);
printf(" (1) : header,tail indices = %d %d \n",index,index2);
printf(" (2) : %.3s length = %ld \n", malloc_name_log[index],
malloc_leng_log[index]);
}
/* ########## This check may be messed up by pointer exchanges
if ( ((ml_size_t) var_ptr) != malloc_addr_log[index])
{
printf("ML_memory_free warning : \n");
printf(" %.3s - header and log mismatch.\n",
malloc_name_log[index]);
printf("MEM LOG %d : %d \n", index, (int) malloc_addr_log[index]);
}
############## */
malloc_leng_log[index] = -1;
}
/*
else
printf("ML_memory_free : variable not found.\n");
*/
int_ptr = (int *) ((ml_size_t) char_ptr - ndouble);
ML_free(int_ptr);
}
(*var_ptr) = NULL;
return 0;
}
static PetscErrorCode MatWrapML_MPIAIJ(ML_Operator *mlmat,MatReuse reuse,Mat *newmat)
{
struct ML_CSR_MSRdata *matdata = (struct ML_CSR_MSRdata *)mlmat->data;
PetscInt *ml_cols=matdata->columns,*aj;
PetscScalar *ml_vals=matdata->values,*aa;
PetscErrorCode ierr;
PetscInt i,j,k,*gordering;
PetscInt m=mlmat->outvec_leng,n,nz_max,row;
Mat A;
PetscFunctionBegin;
if (!mlmat->getrow) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"mlmat->getrow = NULL");
n = mlmat->invec_leng;
if (m != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"m %d must equal to n %d",m,n);
if (reuse) {
A = *newmat;
for (nz_max=0,i=0; i<m; i++) nz_max = PetscMax(nz_max,ml_cols[i+1] - ml_cols[i] + 1);
} else {
PetscInt *nnzA,*nnzB,*nnz;
ierr = MatCreate(mlmat->comm->USR_comm,&A);CHKERRQ(ierr);
ierr = MatSetSizes(A,m,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr);
ierr = PetscMalloc3(m,PetscInt,&nnzA,m,PetscInt,&nnzB,m,PetscInt,&nnz);CHKERRQ(ierr);
nz_max = 0;
for (i=0; i<m; i++){
nnz[i] = ml_cols[i+1] - ml_cols[i] + 1;
if (nz_max < nnz[i]) nz_max = nnz[i];
nnzA[i] = 1; /* diag */
for (j=ml_cols[i]; j<ml_cols[i+1]; j++){
if (ml_cols[j] < m) nnzA[i]++;
}
nnzB[i] = nnz[i] - nnzA[i];
}
ierr = MatMPIAIJSetPreallocation(A,0,nnzA,0,nnzB);CHKERRQ(ierr);
ierr = PetscFree3(nnzA,nnzB,nnz);
}
/* insert mat values -- remap row and column indices */
nz_max++;
ierr = PetscMalloc2(nz_max,PetscScalar,&aa,nz_max,PetscInt,&aj);CHKERRQ(ierr);
/* create global row numbering for a ML_Operator */
ML_build_global_numbering(mlmat,&gordering,"rows");
for (i=0; i<m; i++) {
PetscInt ncols;
row = gordering[i];
k = 0;
/* diagonal entry */
aj[k] = row; aa[k++] = ml_vals[i];
/* off diagonal entries */
for (j=ml_cols[i]; j<ml_cols[i+1]; j++){
aj[k] = gordering[ml_cols[j]]; aa[k++] = ml_vals[j];
}
ncols = ml_cols[i+1] - ml_cols[i] + 1;
ierr = MatSetValues(A,1,&row,ncols,aj,aa,INSERT_VALUES);CHKERRQ(ierr);
}
ML_free(gordering);
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
*newmat = A;
ierr = PetscFree2(aa,aj);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int ML_Aggregate_CoarsenUser(ML_Aggregate *ml_ag, ML_Operator *Amatrix,
ML_Operator **Pmatrix, ML_Comm *comm)
{
unsigned int nbytes, length;
int i, j, k, Nrows, exp_Nrows;
int diff_level;
int aggr_count, index, mypid, num_PDE_eqns;
int *aggr_index = NULL, nullspace_dim;
int Ncoarse, count;
int *new_ia = NULL, *new_ja = NULL, new_Nrows;
int exp_Ncoarse;
int *aggr_cnt_array = NULL;
int level, index3, max_agg_size;
int **rows_in_aggs = NULL, lwork, info;
double *new_val = NULL, epsilon;
double *nullspace_vect = NULL, *qr_tmp = NULL;
double *tmp_vect = NULL, *work = NULL, *new_null = NULL;
ML_SuperNode *aggr_head = NULL, *aggr_curr, *supernode;
struct ML_CSR_MSRdata *csr_data;
int total_nz = 0;
char str[80];
int * graph_decomposition = NULL;
ML_Aggregate_Viz_Stats * aggr_viz_and_stats;
ML_Aggregate_Viz_Stats * grid_info;
int Nprocs;
char * unamalg_bdry = NULL;
char* label;
int N_dimensions;
double* x_coord = NULL;
double* y_coord = NULL;
double* z_coord = NULL;
/* ------------------- execution begins --------------------------------- */
label = ML_GetUserLabel();
sprintf(str, "%s (level %d) :", label, ml_ag->cur_level);
/* ============================================================= */
/* get the machine information and matrix references */
/* ============================================================= */
mypid = comm->ML_mypid;
Nprocs = comm->ML_nprocs;
epsilon = ml_ag->threshold;
num_PDE_eqns = ml_ag->num_PDE_eqns;
nullspace_dim = ml_ag->nullspace_dim;
nullspace_vect = ml_ag->nullspace_vect;
Nrows = Amatrix->outvec_leng;
if (mypid == 0 && 5 < ML_Get_PrintLevel()) {
printf("%s num PDE eqns = %d\n",
str,
num_PDE_eqns);
}
/* ============================================================= */
/* check the system size versus null dimension size */
/* ============================================================= */
if ( Nrows % num_PDE_eqns != 0 )
{
printf("ML_Aggregate_CoarsenUser ERROR : Nrows must be multiples");
printf(" of num_PDE_eqns.\n");
exit(EXIT_FAILURE);
}
diff_level = ml_ag->max_levels - ml_ag->cur_level - 1;
if ( diff_level > 0 ) num_PDE_eqns = nullspace_dim; /* ## 12/20/99 */
/* ============================================================= */
/* set up the threshold for weight-based coarsening */
/* ============================================================= */
diff_level = ml_ag->begin_level - ml_ag->cur_level;
if (diff_level == 0)
ml_ag->curr_threshold = ml_ag->threshold;
epsilon = ml_ag->curr_threshold;
ml_ag->curr_threshold *= 0.5;
if (mypid == 0 && 7 < ML_Get_PrintLevel())
printf("%s current eps = %e\n", str, epsilon);
epsilon = epsilon * epsilon;
ML_Operator_AmalgamateAndDropWeak(Amatrix, num_PDE_eqns, epsilon);
Nrows /= num_PDE_eqns;
exp_Nrows = Nrows;
/* ********************************************************************** */
/* allocate memory for aggr_index, which will contain the decomposition */
/* ********************************************************************** */
nbytes = (Nrows*num_PDE_eqns) * sizeof(int);
if ( nbytes > 0 ) {
ML_memory_alloc((void**) &aggr_index, nbytes, "ACJ");
if( aggr_index == NULL ) {
fprintf( stderr,
"*ML*ERR* not enough memory for %d bytes\n"
"*ML*ERR* (file %s, line %d)\n",
nbytes,
__FILE__,
__LINE__ );
exit( EXIT_FAILURE );
}
}
else aggr_index = NULL;
for( i=0 ; i<Nrows*num_PDE_eqns ; i++ ) aggr_index[i] = -1;
unamalg_bdry = (char *) ML_allocate( sizeof(char) * (Nrows+1) );
if( unamalg_bdry == NULL ) {
fprintf( stderr,
"*ML*ERR* on proc %d, not enough space for %d bytes\n"
"*ML*ERR* (file %s, line %d)\n",
mypid,
(int)sizeof(char) * Nrows,
__FILE__,
__LINE__ );
exit( EXIT_FAILURE );
}
N_dimensions = ml_ag->N_dimensions;
grid_info = (ML_Aggregate_Viz_Stats*) Amatrix->to->Grid->Grid;
x_coord = grid_info->x;
if (N_dimensions > 1 && x_coord)
y_coord = grid_info->y;
else
y_coord = 0;
if (N_dimensions > 2 && x_coord)
z_coord = grid_info->z;
else
z_coord = 0;
aggr_count = ML_GetUserPartitions(Amatrix,unamalg_bdry,
epsilon,
x_coord,y_coord,z_coord,
aggr_index,&total_nz);
#ifdef ML_MPI
MPI_Allreduce( &Nrows, &i, 1, MPI_INT, MPI_SUM, Amatrix->comm->USR_comm );
MPI_Allreduce( &aggr_count, &j, 1, MPI_INT, MPI_SUM, Amatrix->comm->USR_comm );
#else
i = Nrows;
j = aggr_count;
#endif
if( mypid == 0 && 7 < ML_Get_PrintLevel() ) {
printf("%s Using %d (block) aggregates (globally)\n",
str,
j );
printf("%s # (block) aggre/ # (block) rows = %8.5f %% ( = %d / %d)\n",
str,
100.0*j/i,
j, i);
}
j = ML_gsum_int( aggr_count, comm );
if (mypid == 0 && 7 < ML_Get_PrintLevel()) {
printf("%s %d (block) aggregates (globally)\n",
str, j );
}
/* ********************************************************************** */
/* I allocate room to copy aggr_index and pass this value to the user, */
/* who will be able to analyze and visualize this after the construction */
/* of the levels. This way, the only price we have to pay for stats and */
/* viz is essentially a little bit of memory. */
/* this memory will be cleaned with the object ML_Aggregate ml_ag. */
/* I set the pointers using the ML_Aggregate_Info structure. This is */
/* allocated using ML_Aggregate_Info_Setup(ml,MaxNumLevels) */
/* ********************************************************************** */
if (Amatrix->to->Grid->Grid != NULL) {
graph_decomposition = (int *)ML_allocate(sizeof(int)*(Nrows+1));
if( graph_decomposition == NULL ) {
fprintf( stderr,
"*ML*ERR* Not enough memory for %d bytes\n"
"*ML*ERR* (file %s, line %d)\n",
(int)sizeof(int)*Nrows,
__FILE__,
__LINE__ );
exit( EXIT_FAILURE );
}
for( i=0 ; i<Nrows ; i++ ) graph_decomposition[i] = aggr_index[i];
aggr_viz_and_stats = (ML_Aggregate_Viz_Stats *) (Amatrix->to->Grid->Grid);
aggr_viz_and_stats->graph_decomposition = graph_decomposition;
aggr_viz_and_stats->Nlocal = Nrows;
aggr_viz_and_stats->Naggregates = aggr_count;
aggr_viz_and_stats->local_or_global = ML_LOCAL_INDICES;
aggr_viz_and_stats->is_filled = ML_YES;
aggr_viz_and_stats->Amatrix = Amatrix;
}
/* ********************************************************************** */
/* take the decomposition as created by METIS and form the aggregates */
/* ********************************************************************** */
total_nz = ML_Comm_GsumInt( comm, total_nz);
i = ML_Comm_GsumInt( comm, Nrows);
if ( mypid == 0 && 7 < ML_Get_PrintLevel())
printf("%s Total (block) nnz = %d ( = %5.2f/(block)row)\n",
str,
total_nz,1.0*total_nz/i);
if ( ml_ag->operator_complexity == 0.0 ) {
ml_ag->fine_complexity = total_nz;
ml_ag->operator_complexity = total_nz;
}
else ml_ag->operator_complexity += total_nz;
/* fix aggr_index for num_PDE_eqns > 1 */
for (i = Nrows - 1; i >= 0; i-- ) {
for (j = num_PDE_eqns-1; j >= 0; j--) {
aggr_index[i*num_PDE_eqns+j] = aggr_index[i];
}
}
if ( mypid == 0 && 8 < ML_Get_PrintLevel())
{
printf("Calling ML_Operator_UnAmalgamateAndDropWeak\n");
fflush(stdout);
}
ML_Operator_UnAmalgamateAndDropWeak(Amatrix, num_PDE_eqns, epsilon);
Nrows *= num_PDE_eqns;
exp_Nrows *= num_PDE_eqns;
/* count the size of each aggregate */
aggr_cnt_array = (int *) ML_allocate(sizeof(int)*(aggr_count+1));
for (i = 0; i < aggr_count ; i++) aggr_cnt_array[i] = 0;
for (i = 0; i < exp_Nrows; i++) {
if (aggr_index[i] >= 0) {
if( aggr_index[i] >= aggr_count ) {
fprintf( stderr,
"*ML*WRN* on process %d, something weird happened...\n"
"*ML*WRN* node %d belong to aggregate %d (#aggr = %d)\n"
"*ML*WRN* (file %s, line %d)\n",
comm->ML_mypid,
i,
aggr_index[i],
aggr_count,
__FILE__,
__LINE__ );
} else {
aggr_cnt_array[aggr_index[i]]++;
}
}
}
/* ============================================================= */
/* Form tentative prolongator */
/* ============================================================= */
Ncoarse = aggr_count;
/* ============================================================= */
/* check and copy aggr_index */
/* ------------------------------------------------------------- */
level = ml_ag->cur_level;
nbytes = (Nrows+1) * sizeof( int );
ML_memory_alloc((void**) &(ml_ag->aggr_info[level]), nbytes, "AGl");
count = aggr_count;
for ( i = 0; i < Nrows; i+=num_PDE_eqns )
{
if ( aggr_index[i] >= 0 )
{
for ( j = 0; j < num_PDE_eqns; j++ )
ml_ag->aggr_info[level][i+j] = aggr_index[i];
if (aggr_index[i] >= count) count = aggr_index[i] + 1;
}
/*else
*{
* printf("%d : CoarsenMIS error : aggr_index[%d] < 0\n",
* mypid,i);
* exit(1);
*}*/
}
ml_ag->aggr_count[level] = count; /* for relaxing boundary points */
/* ============================================================= */
/* set up the new operator */
/* ------------------------------------------------------------- */
new_Nrows = Nrows;
exp_Ncoarse = Nrows;
for ( i = 0; i < new_Nrows; i++ )
{
if ( aggr_index[i] >= exp_Ncoarse )
{
printf("*ML*WRN* index out of bound %d = %d(%d)\n",
i, aggr_index[i],
exp_Ncoarse);
}
}
nbytes = ( new_Nrows+1 ) * sizeof(int);
ML_memory_alloc((void**)&(new_ia), nbytes, "AIA");
nbytes = ( new_Nrows+1) * nullspace_dim * sizeof(int);
ML_memory_alloc((void**)&(new_ja), nbytes, "AJA");
nbytes = ( new_Nrows+1) * nullspace_dim * sizeof(double);
ML_memory_alloc((void**)&(new_val), nbytes, "AVA");
for ( i = 0; i < new_Nrows*nullspace_dim; i++ ) new_val[i] = 0.0;
/* ------------------------------------------------------------- */
/* set up the space for storing the new null space */
/* ------------------------------------------------------------- */
nbytes = (Ncoarse+1) * nullspace_dim * nullspace_dim * sizeof(double);
ML_memory_alloc((void**)&(new_null),nbytes,"AGr");
if( new_null == NULL ) {
fprintf( stderr,
"*ML*ERR* on process %d, not enough memory for %d bytes\n"
"*ML*ERR* (file %s, line %d)\n",
mypid,
nbytes,
__FILE__,
__LINE__ );
exit( EXIT_FAILURE );
}
for (i = 0; i < Ncoarse*nullspace_dim*nullspace_dim; i++)
new_null[i] = 0.0;
/* ------------------------------------------------------------- */
/* initialize the row pointer for the CSR prolongation operator */
/* (each row will have at most nullspace_dim nonzero entries) */
/* ------------------------------------------------------------- */
for (i = 0; i <= Nrows; i++) new_ia[i] = i * nullspace_dim;
/* trying this when a Dirichlet row is taken out */
j = 0;
new_ia[0] = 0;
for (i = 0; i < Nrows; i++) {
if (aggr_index[i] != -1) j += nullspace_dim;
new_ia[i+1] = j;
}
/* ------------------------------------------------------------- */
/* generate an array to store which aggregate has which rows.Then*/
/* loop through the rows of A checking which aggregate each row */
/* is in, and adding it to the appropriate spot in rows_in_aggs */
/* ------------------------------------------------------------- */
ML_memory_alloc((void**)&rows_in_aggs,aggr_count*sizeof(int*),"MLs");
for (i = 0; i < aggr_count; i++) {
nbytes = aggr_cnt_array[i]+1;
rows_in_aggs[i] = (int *) ML_allocate(nbytes*sizeof(int));
aggr_cnt_array[i] = 0;
if (rows_in_aggs[i] == NULL) {
printf("*ML*ERR* couldn't allocate memory in CoarsenMETIS\n");
exit(1);
}
}
for (i = 0; i < exp_Nrows; i+=num_PDE_eqns) {
if ( aggr_index[i] >= 0 && aggr_index[i] < aggr_count)
{
for (j = 0; j < num_PDE_eqns; j++)
{
index = aggr_cnt_array[aggr_index[i]]++;
rows_in_aggs[aggr_index[i]][index] = i + j;
}
}
}
/* ------------------------------------------------------------- */
/* allocate work arrays for QR factorization */
/* work and lwork are needed for lapack's QR routine. These */
/* settings seemed easiest since I don't quite understand */
/* what they do, but may want to do something better here later */
/* ------------------------------------------------------------- */
max_agg_size = 0;
for (i = 0; i < aggr_count; i++)
{
if (aggr_cnt_array[i] > max_agg_size) max_agg_size = aggr_cnt_array[i];
}
nbytes = max_agg_size * nullspace_dim * sizeof(double);
ML_memory_alloc((void**)&qr_tmp, nbytes, "AGu");
nbytes = nullspace_dim * sizeof(double);
ML_memory_alloc((void**)&tmp_vect, nbytes, "AGv");
lwork = nullspace_dim;
nbytes = nullspace_dim * sizeof(double);
ML_memory_alloc((void**)&work, nbytes, "AGw");
/* ------------------------------------------------------------- */
/* perform block QR decomposition */
/* ------------------------------------------------------------- */
for (i = 0; i < aggr_count; i++)
{
/* ---------------------------------------------------------- */
/* set up the matrix we want to decompose into Q and R: */
/* ---------------------------------------------------------- */
length = aggr_cnt_array[i];
if (nullspace_vect == NULL)
{
for (j = 0; j < (int) length; j++)
{
index = rows_in_aggs[i][j];
for (k = 0; k < nullspace_dim; k++)
{
if ( unamalg_bdry[index/num_PDE_eqns] == 'T')
qr_tmp[k*length+j] = 0.;
else
{
if (index % num_PDE_eqns == k) qr_tmp[k*length+j] = 1.0;
else qr_tmp[k*length+j] = 0.0;
}
}
}
}
else
{
for (k = 0; k < nullspace_dim; k++)
{
for (j = 0; j < (int) length; j++)
{
index = rows_in_aggs[i][j];
if ( unamalg_bdry[index/num_PDE_eqns] == 'T')
qr_tmp[k*length+j] = 0.;
else {
if (index < Nrows) {
qr_tmp[k*length+j] = nullspace_vect[k*Nrows+index];
}
else {
fprintf( stderr,
"*ML*ERR* in QR\n"
"*ML*ERR* (file %s, line %d)\n",
__FILE__,
__LINE__ );
exit( EXIT_FAILURE );
}
}
}
}
}
/* ---------------------------------------------------------- */
/* now calculate QR using an LAPACK routine */
/* ---------------------------------------------------------- */
if (aggr_cnt_array[i] >= nullspace_dim) {
DGEQRF_F77(&(aggr_cnt_array[i]), &nullspace_dim, qr_tmp,
&(aggr_cnt_array[i]), tmp_vect, work, &lwork, &info);
if (info != 0)
pr_error("ErrOr in CoarsenMIS : dgeqrf returned a non-zero %d %d\n",
aggr_cnt_array[i],i);
if (work[0] > lwork)
{
lwork=(int) work[0];
ML_memory_free((void**) &work);
ML_memory_alloc((void**) &work, sizeof(double)*lwork, "AGx");
}
else lwork=(int) work[0];
/* ---------------------------------------------------------- */
/* the upper triangle of qr_tmp is now R, so copy that into */
/* the new nullspace */
/* ---------------------------------------------------------- */
for (j = 0; j < nullspace_dim; j++)
for (k = j; k < nullspace_dim; k++)
new_null[i*nullspace_dim+j+k*Ncoarse*nullspace_dim] =
qr_tmp[j+aggr_cnt_array[i]*k];
/* ---------------------------------------------------------- */
/* to get this block of P, need to run qr_tmp through another */
/* LAPACK function: */
/* ---------------------------------------------------------- */
if ( aggr_cnt_array[i] < nullspace_dim ){
printf("Error in dorgqr on %d row (dims are %d, %d)\n",i,aggr_cnt_array[i],
nullspace_dim);
printf("ERROR : performing QR on a MxN matrix where M<N.\n");
}
DORGQR_F77(&(aggr_cnt_array[i]), &nullspace_dim, &nullspace_dim,
qr_tmp, &(aggr_cnt_array[i]), tmp_vect, work, &lwork, &info);
if (info != 0) {
printf("Error in dorgqr on %d row (dims are %d, %d)\n",i,aggr_cnt_array[i],
nullspace_dim);
pr_error("Error in CoarsenMIS: dorgqr returned a non-zero\n");
}
if (work[0] > lwork)
{
lwork=(int) work[0];
ML_memory_free((void**) &work);
ML_memory_alloc((void**) &work, sizeof(double)*lwork, "AGy");
}
else lwork=(int) work[0];
/* ---------------------------------------------------------- */
/* now copy Q over into the appropriate part of P: */
/* The rows of P get calculated out of order, so I assume the */
/* Q is totally dense and use what I know of how big each Q */
/* will be to determine where in ia, ja, etc each nonzero in */
/* Q belongs. If I did not assume this, I would have to keep */
/* all of P in memory in order to determine where each entry */
/* should go */
/* ---------------------------------------------------------- */
for (j = 0; j < aggr_cnt_array[i]; j++)
{
index = rows_in_aggs[i][j];
if ( index < Nrows )
{
index3 = new_ia[index];
for (k = 0; k < nullspace_dim; k++)
{
new_ja [index3+k] = i * nullspace_dim + k;
new_val[index3+k] = qr_tmp[ k*aggr_cnt_array[i]+j];
}
}
else
{
fprintf( stderr,
"*ML*ERR* in QR: index out of bounds (%d - %d)\n",
index,
Nrows );
}
}
}
else {
/* We have a small aggregate such that the QR factorization can not */
/* be performed. Instead let us copy the null space from the fine */
/* into the coarse grid nullspace and put the identity for the */
/* prolongator???? */
for (j = 0; j < nullspace_dim; j++)
for (k = 0; k < nullspace_dim; k++)
new_null[i*nullspace_dim+j+k*Ncoarse*nullspace_dim] =
qr_tmp[j+aggr_cnt_array[i]*k];
for (j = 0; j < aggr_cnt_array[i]; j++) {
index = rows_in_aggs[i][j];
index3 = new_ia[index];
for (k = 0; k < nullspace_dim; k++) {
new_ja [index3+k] = i * nullspace_dim + k;
if (k == j) new_val[index3+k] = 1.;
else new_val[index3+k] = 0.;
}
}
}
}
ML_Aggregate_Set_NullSpace(ml_ag, num_PDE_eqns, nullspace_dim,
new_null, Ncoarse*nullspace_dim);
ML_memory_free( (void **) &new_null);
/* ------------------------------------------------------------- */
/* set up the csr_data data structure */
/* ------------------------------------------------------------- */
ML_memory_alloc((void**) &csr_data, sizeof(struct ML_CSR_MSRdata),"CSR");
csr_data->rowptr = new_ia;
csr_data->columns = new_ja;
csr_data->values = new_val;
ML_Operator_Set_ApplyFuncData( *Pmatrix, nullspace_dim*Ncoarse, Nrows,
csr_data, Nrows, NULL, 0);
(*Pmatrix)->data_destroy = ML_CSR_MSR_ML_memorydata_Destroy;
(*Pmatrix)->getrow->pre_comm = ML_CommInfoOP_Create();
(*Pmatrix)->max_nz_per_row = 1;
ML_Operator_Set_Getrow((*Pmatrix), Nrows, CSR_getrow);
ML_Operator_Set_ApplyFunc((*Pmatrix), CSR_matvec);
(*Pmatrix)->max_nz_per_row = 1;
/* this must be set so that the hierarchy generation does not abort early
in adaptive SA */
(*Pmatrix)->num_PDEs = nullspace_dim;
/* ------------------------------------------------------------- */
/* clean up */
/* ------------------------------------------------------------- */
ML_free(unamalg_bdry);
ML_memory_free((void**)&aggr_index);
ML_free(aggr_cnt_array);
for (i = 0; i < aggr_count; i++) ML_free(rows_in_aggs[i]);
ML_memory_free((void**)&rows_in_aggs);
ML_memory_free((void**)&qr_tmp);
ML_memory_free((void**)&tmp_vect);
ML_memory_free((void**)&work);
aggr_curr = aggr_head;
while ( aggr_curr != NULL )
{
supernode = aggr_curr;
aggr_curr = aggr_curr->next;
if ( supernode->length > 0 ) ML_free( supernode->list );
ML_free( supernode );
}
return Ncoarse*nullspace_dim;
} /* ML_Aggregate_CoarsenUser */
int main(int argc, char *argv[])
{
int num_PDE_eqns=1, N_levels=3, nsmooth=2;
int leng, level, N_grid_pts, coarsest_level;
int leng1,leng2;
/* See Aztec User's Guide for more information on the */
/* variables that follow. */
int proc_config[AZ_PROC_SIZE], options[AZ_OPTIONS_SIZE];
double params[AZ_PARAMS_SIZE], status[AZ_STATUS_SIZE];
/* data structure for matrix corresponding to the fine grid */
double *val = NULL, *xxx, *rhs, solve_time, setup_time, start_time;
AZ_MATRIX *Amat;
AZ_PRECOND *Pmat = NULL;
ML *ml;
FILE *fp;
int i, j, Nrigid, *garbage, nblocks=0, *blocks = NULL, *block_pde=NULL;
struct AZ_SCALING *scaling;
ML_Aggregate *ag;
double *mode, *rigid=NULL, alpha;
char filename[80];
int one = 1;
int proc,nprocs;
char pathfilename[100];
#ifdef ML_MPI
MPI_Init(&argc,&argv);
/* get number of processors and the name of this processor */
AZ_set_proc_config(proc_config, MPI_COMM_WORLD);
proc = proc_config[AZ_node];
nprocs = proc_config[AZ_N_procs];
#else
AZ_set_proc_config(proc_config, AZ_NOT_MPI);
proc = 0;
nprocs = 1;
#endif
if (proc_config[AZ_node] == 0) {
sprintf(pathfilename,"%s/inputfile",argv[1]);
ML_Reader_ReadInput(pathfilename, &context);
}
else context = (struct reader_context *) ML_allocate(sizeof(struct reader_context));
AZ_broadcast((char *) context, sizeof(struct reader_context), proc_config,
AZ_PACK);
AZ_broadcast((char *) NULL , 0 , proc_config, AZ_SEND);
N_levels = context->N_levels;
printf("N_levels %d\n",N_levels);
nsmooth = context->nsmooth;
num_PDE_eqns = context->N_dofPerNode;
printf("num_PDE_eqns %d\n",num_PDE_eqns);
ML_Set_PrintLevel(context->output_level);
/* read in the number of matrix equations */
leng = 0;
if (proc_config[AZ_node] == 0) {
sprintf(pathfilename,"%s/data_matrix.txt",argv[1]);
fp=fopen(pathfilename,"r");
if (fp==NULL) {
printf("**ERR** couldn't open file data_matrix.txt\n");
exit(1);
}
fscanf(fp,"%d",&leng);
fclose(fp);
}
leng = AZ_gsum_int(leng, proc_config);
N_grid_pts=leng/num_PDE_eqns;
/* initialize the list of global indices. NOTE: the list of global */
/* indices must be in ascending order so that subsequent calls to */
/* AZ_find_index() will function properly. */
#if 0
if (proc_config[AZ_N_procs] == 1) i = AZ_linear;
else i = AZ_file;
#endif
i = AZ_linear;
/* cannot use AZ_input_update for variable blocks (forgot why, but debugged through it)*/
/* make a linear distribution of the matrix */
/* if the linear distribution does not align with the blocks, */
/* this is corrected in ML_AZ_Reader_ReadVariableBlocks */
leng1 = leng/nprocs;
leng2 = leng-leng1*nprocs;
if (proc >= leng2)
{
leng2 += (proc*leng1);
}
else
{
leng1++;
leng2 = proc*leng1;
}
N_update = leng1;
update = (int*)AZ_allocate((N_update+1)*sizeof(int));
if (update==NULL)
{
(void) fprintf (stderr, "Not enough space to allocate 'update'\n");
fflush(stderr); exit(EXIT_FAILURE);
}
for (i=0; i<N_update; i++) update[i] = i+leng2;
#if 0 /* debug */
printf("proc %d N_update %d\n",proc_config[AZ_node],N_update);
fflush(stdout);
#endif
sprintf(pathfilename,"%s/data_vblocks.txt",argv[1]);
ML_AZ_Reader_ReadVariableBlocks(pathfilename,&nblocks,&blocks,&block_pde,
&N_update,&update,proc_config);
#if 0 /* debug */
printf("proc %d N_update %d\n",proc_config[AZ_node],N_update);
fflush(stdout);
#endif
sprintf(pathfilename,"%s/data_matrix.txt",argv[1]);
AZ_input_msr_matrix(pathfilename,update, &val, &bindx, N_update, proc_config);
/* This code is to fix things up so that we are sure we have */
/* all blocks (including the ghost nodes) the same size. */
/* not sure, whether this is a good idea with variable blocks */
/* the examples inpufiles (see top of this file) don't need it */
/* anyway */
/*
AZ_block_MSR(&bindx, &val, N_update, num_PDE_eqns, update);
*/
AZ_transform_norowreordering(proc_config, &external, bindx, val, update, &update_index,
&extern_index, &data_org, N_update, 0, 0, 0, &cpntr,
AZ_MSR_MATRIX);
Amat = AZ_matrix_create( leng );
AZ_set_MSR(Amat, bindx, val, data_org, 0, NULL, AZ_LOCAL);
Amat->matrix_type = data_org[AZ_matrix_type];
data_org[AZ_N_rows] = data_org[AZ_N_internal] + data_org[AZ_N_border];
start_time = AZ_second();
options[AZ_scaling] = AZ_none;
ML_Create(&ml, N_levels);
/* set up discretization matrix and matrix vector function */
AZ_ML_Set_Amat(ml, 0, N_update, N_update, Amat, proc_config);
ML_Set_ResidualOutputFrequency(ml, context->output);
ML_Set_Tolerance(ml, context->tol);
ML_Aggregate_Create( &ag );
if (ML_strcmp(context->agg_coarsen_scheme,"Mis") == 0) {
ML_Aggregate_Set_CoarsenScheme_MIS(ag);
}
else if (ML_strcmp(context->agg_coarsen_scheme,"Uncoupled") == 0) {
ML_Aggregate_Set_CoarsenScheme_Uncoupled(ag);
}
else if (ML_strcmp(context->agg_coarsen_scheme,"Coupled") == 0) {
ML_Aggregate_Set_CoarsenScheme_Coupled(ag);
}
else if (ML_strcmp(context->agg_coarsen_scheme,"Metis") == 0) {
ML_Aggregate_Set_CoarsenScheme_METIS(ag);
for (i=0; i<N_levels; i++)
ML_Aggregate_Set_NodesPerAggr(ml,ag,i,9);
}
else if (ML_strcmp(context->agg_coarsen_scheme,"VBMetis") == 0) {
/* when no blocks read, use standard metis assuming constant block sizes */
if (!blocks)
ML_Aggregate_Set_CoarsenScheme_METIS(ag);
else {
ML_Aggregate_Set_CoarsenScheme_VBMETIS(ag);
ML_Aggregate_Set_Vblocks_CoarsenScheme_VBMETIS(ag,0,N_levels,nblocks,
blocks,block_pde,N_update);
}
for (i=0; i<N_levels; i++)
ML_Aggregate_Set_NodesPerAggr(ml,ag,i,9);
}
else {
printf("**ERR** ML: Unknown aggregation scheme %s\n",context->agg_coarsen_scheme);
exit(-1);
}
ML_Aggregate_Set_DampingFactor(ag, context->agg_damping);
ML_Aggregate_Set_MaxCoarseSize( ag, context->maxcoarsesize);
ML_Aggregate_Set_Threshold(ag, context->agg_thresh);
if (ML_strcmp(context->agg_spectral_norm,"Calc") == 0) {
ML_Set_SpectralNormScheme_Calc(ml);
}
else if (ML_strcmp(context->agg_spectral_norm,"Anorm") == 0) {
ML_Set_SpectralNormScheme_Anorm(ml);
}
else {
printf("**WRN** ML: Unknown spectral norm scheme %s\n",context->agg_spectral_norm);
}
/* read in the rigid body modes */
Nrigid = 0;
if (proc_config[AZ_node] == 0) {
sprintf(filename,"data_nullsp%d.txt",Nrigid);
sprintf(pathfilename,"%s/%s",argv[1],filename);
while( (fp = fopen(pathfilename,"r")) != NULL) {
fclose(fp);
Nrigid++;
sprintf(filename,"data_nullsp%d.txt",Nrigid);
sprintf(pathfilename,"%s/%s",argv[1],filename);
}
}
Nrigid = AZ_gsum_int(Nrigid,proc_config);
if (Nrigid != 0) {
rigid = (double *) ML_allocate( sizeof(double)*Nrigid*(N_update+1) );
if (rigid == NULL) {
printf("Error: Not enough space for rigid body modes\n");
}
}
/* Set rhs */
sprintf(pathfilename,"%s/data_rhs.txt",argv[1]);
fp = fopen(pathfilename,"r");
if (fp == NULL) {
rhs=(double *)ML_allocate(leng*sizeof(double));
if (proc_config[AZ_node] == 0) printf("taking linear vector for rhs\n");
for (i = 0; i < N_update; i++) rhs[i] = (double) update[i];
}
else {
fclose(fp);
if (proc_config[AZ_node] == 0) printf("reading rhs from a file\n");
AZ_input_msr_matrix(pathfilename, update, &rhs, &garbage, N_update,
proc_config);
}
AZ_reorder_vec(rhs, data_org, update_index, NULL);
for (i = 0; i < Nrigid; i++) {
sprintf(filename,"data_nullsp%d.txt",i);
sprintf(pathfilename,"%s/%s",argv[1],filename);
AZ_input_msr_matrix(pathfilename, update, &mode, &garbage, N_update,
proc_config);
AZ_reorder_vec(mode, data_org, update_index, NULL);
#if 0 /* test the given rigid body mode, output-vector should be ~0 */
Amat->matvec(mode, rigid, Amat, proc_config);
for (j = 0; j < N_update; j++) printf("this is %d %e\n",j,rigid[j]);
#endif
for (j = 0; j < i; j++) {
alpha = -AZ_gdot(N_update, mode, &(rigid[j*N_update]), proc_config)/
AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]),
proc_config);
DAXPY_F77(&N_update, &alpha, &(rigid[j*N_update]), &one, mode, &one);
}
/* rhs orthogonalization */
alpha = -AZ_gdot(N_update, mode, rhs, proc_config)/
AZ_gdot(N_update, mode, mode, proc_config);
DAXPY_F77(&N_update, &alpha, mode, &one, rhs, &one);
for (j = 0; j < N_update; j++) rigid[i*N_update+j] = mode[j];
free(mode);
free(garbage);
}
for (j = 0; j < Nrigid; j++) {
alpha = -AZ_gdot(N_update, rhs, &(rigid[j*N_update]), proc_config)/
AZ_gdot(N_update, &(rigid[j*N_update]), &(rigid[j*N_update]),
proc_config);
DAXPY_F77(&N_update, &alpha, &(rigid[j*N_update]), &one, rhs, &one);
}
#if 0 /* for testing the default nullsp */
ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, 6, NULL, N_update);
#else
if (Nrigid != 0) {
ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, Nrigid, rigid, N_update);
}
#endif
if (rigid) ML_free(rigid);
ag->keep_agg_information = 1;
coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, 0,
ML_INCREASING, ag);
coarsest_level--;
if ( proc_config[AZ_node] == 0 )
printf("Coarse level = %d \n", coarsest_level);
#if 0
/* set up smoothers */
if (!blocks)
blocks = (int *) ML_allocate(sizeof(int)*N_update);
#endif
for (level = 0; level < coarsest_level; level++) {
num_PDE_eqns = ml->Amat[level].num_PDEs;
/* Sparse approximate inverse smoother that acutally does both */
/* pre and post smoothing. */
if (ML_strcmp(context->smoother,"Parasails") == 0) {
ML_Gen_Smoother_ParaSails(ml , level, ML_PRESMOOTHER, nsmooth,
parasails_sym, parasails_thresh,
parasails_nlevels, parasails_filter,
(int) parasails_loadbal, parasails_factorized);
}
/* This is the symmetric Gauss-Seidel smoothing that we usually use. */
/* In parallel, it is not a true Gauss-Seidel in that each processor */
/* does a Gauss-Seidel on its local submatrix independent of the */
/* other processors. */
else if (ML_strcmp(context->smoother,"GaussSeidel") == 0) {
ML_Gen_Smoother_GaussSeidel(ml , level, ML_BOTH, nsmooth,1.);
}
else if (ML_strcmp(context->smoother,"SymGaussSeidel") == 0) {
ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_BOTH, nsmooth,1.);
}
else if (ML_strcmp(context->smoother,"Poly") == 0) {
ML_Gen_Smoother_Cheby(ml, level, ML_BOTH, 30., nsmooth);
}
else if (ML_strcmp(context->smoother,"BlockGaussSeidel") == 0) {
ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_BOTH, nsmooth,1.,
num_PDE_eqns);
}
else if (ML_strcmp(context->smoother,"VBSymGaussSeidel") == 0) {
if (blocks) ML_free(blocks);
if (block_pde) ML_free(block_pde);
blocks = NULL;
block_pde = NULL;
nblocks = 0;
ML_Aggregate_Get_Vblocks_CoarsenScheme_VBMETIS(ag,level,N_levels,&nblocks,
&blocks,&block_pde);
if (blocks==NULL) ML_Gen_Blocks_Aggregates(ag, level, &nblocks, &blocks);
ML_Gen_Smoother_VBlockSymGaussSeidel(ml , level, ML_BOTH, nsmooth,1.,
nblocks, blocks);
}
/* This is a true Gauss Seidel in parallel. This seems to work for */
/* elasticity problems. However, I don't believe that this is very */
/* efficient in parallel. */
/*
nblocks = ml->Amat[level].invec_leng;
for (i =0; i < nblocks; i++) blocks[i] = i;
ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml , level, ML_PRESMOOTHER,
nsmooth, 1., nblocks, blocks);
ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml, level, ML_POSTSMOOTHER,
nsmooth, 1., nblocks, blocks);
*/
/* Jacobi Smoothing */
else if (ML_strcmp(context->smoother,"Jacobi") == 0) {
ML_Gen_Smoother_Jacobi(ml , level, ML_PRESMOOTHER, nsmooth,.4);
ML_Gen_Smoother_Jacobi(ml , level, ML_POSTSMOOTHER, nsmooth,.4);
}
/* This does a block Gauss-Seidel (not true GS in parallel) */
/* where each processor has 'nblocks' blocks. */
/* */
else if (ML_strcmp(context->smoother,"Metis") == 0) {
if (blocks) ML_free(blocks);
if (block_pde) ML_free(block_pde);
nblocks = 250;
ML_Gen_Blocks_Metis(ml, level, &nblocks, &blocks);
ML_Gen_Smoother_VBlockSymGaussSeidel(ml , level, ML_BOTH, nsmooth,1.,
nblocks, blocks);
}
else {
printf("unknown smoother %s\n",context->smoother);
exit(1);
}
}
/* set coarse level solver */
nsmooth = context->coarse_its;
/* Sparse approximate inverse smoother that acutally does both */
/* pre and post smoothing. */
if (ML_strcmp(context->coarse_solve,"Parasails") == 0) {
ML_Gen_Smoother_ParaSails(ml , coarsest_level, ML_PRESMOOTHER, nsmooth,
parasails_sym, parasails_thresh,
parasails_nlevels, parasails_filter,
(int) parasails_loadbal, parasails_factorized);
}
else if (ML_strcmp(context->coarse_solve,"GaussSeidel") == 0) {
ML_Gen_Smoother_GaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1.);
}
else if (ML_strcmp(context->coarse_solve,"Poly") == 0) {
ML_Gen_Smoother_Cheby(ml, coarsest_level, ML_BOTH, 30., nsmooth);
}
else if (ML_strcmp(context->coarse_solve,"SymGaussSeidel") == 0) {
ML_Gen_Smoother_SymGaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1.);
}
else if (ML_strcmp(context->coarse_solve,"BlockGaussSeidel") == 0) {
ML_Gen_Smoother_BlockGaussSeidel(ml, coarsest_level, ML_BOTH, nsmooth,1.,
num_PDE_eqns);
}
else if (ML_strcmp(context->coarse_solve,"Aggregate") == 0) {
if (blocks) ML_free(blocks);
if (block_pde) ML_free(block_pde);
ML_Gen_Blocks_Aggregates(ag, coarsest_level, &nblocks, &blocks);
ML_Gen_Smoother_VBlockSymGaussSeidel(ml , coarsest_level, ML_BOTH,
nsmooth,1., nblocks, blocks);
}
else if (ML_strcmp(context->coarse_solve,"Jacobi") == 0) {
ML_Gen_Smoother_Jacobi(ml , coarsest_level, ML_BOTH, nsmooth,.5);
}
else if (ML_strcmp(context->coarse_solve,"Metis") == 0) {
if (blocks) ML_free(blocks);
if (block_pde) ML_free(block_pde);
nblocks = 250;
ML_Gen_Blocks_Metis(ml, coarsest_level, &nblocks, &blocks);
ML_Gen_Smoother_VBlockSymGaussSeidel(ml , coarsest_level, ML_BOTH,
nsmooth,1., nblocks, blocks);
}
else if (ML_strcmp(context->coarse_solve,"SuperLU") == 0) {
ML_Gen_CoarseSolverSuperLU( ml, coarsest_level);
}
else if (ML_strcmp(context->coarse_solve,"Amesos") == 0) {
ML_Gen_Smoother_Amesos(ml,coarsest_level,ML_AMESOS_KLU,-1, 0.0);
}
else {
printf("unknown coarse grid solver %s\n",context->coarse_solve);
exit(1);
}
ML_Gen_Solver(ml, ML_MGV, 0, coarsest_level);
AZ_defaults(options, params);
if (ML_strcmp(context->krylov,"Cg") == 0) {
options[AZ_solver] = AZ_cg;
}
else if (ML_strcmp(context->krylov,"Bicgstab") == 0) {
options[AZ_solver] = AZ_bicgstab;
}
else if (ML_strcmp(context->krylov,"Tfqmr") == 0) {
options[AZ_solver] = AZ_tfqmr;
}
else if (ML_strcmp(context->krylov,"Gmres") == 0) {
options[AZ_solver] = AZ_gmres;
}
else {
printf("unknown krylov method %s\n",context->krylov);
}
if (blocks) ML_free(blocks);
if (block_pde) ML_free(block_pde);
options[AZ_scaling] = AZ_none;
options[AZ_precond] = AZ_user_precond;
options[AZ_conv] = AZ_r0;
options[AZ_output] = 1;
options[AZ_max_iter] = context->max_outer_its;
options[AZ_poly_ord] = 5;
options[AZ_kspace] = 130;
params[AZ_tol] = context->tol;
options[AZ_output] = context->output;
ML_free(context);
AZ_set_ML_preconditioner(&Pmat, Amat, ml, options);
setup_time = AZ_second() - start_time;
xxx = (double *) malloc( leng*sizeof(double));
for (iii = 0; iii < leng; iii++) xxx[iii] = 0.0;
/* Set x */
/*
there is no initguess supplied with these examples for the moment....
*/
fp = fopen("initguessfile","r");
if (fp != NULL) {
fclose(fp);
if (proc_config[AZ_node]== 0) printf("reading initial guess from file\n");
AZ_input_msr_matrix("data_initguess.txt", update, &xxx, &garbage, N_update,
proc_config);
options[AZ_conv] = AZ_expected_values;
}
else if (proc_config[AZ_node]== 0) printf("taking 0 initial guess \n");
AZ_reorder_vec(xxx, data_org, update_index, NULL);
/* if Dirichlet BC ... put the answer in */
for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++) {
if ( (val[i] > .99999999) && (val[i] < 1.0000001))
xxx[i] = rhs[i];
}
fp = fopen("AZ_no_multilevel.dat","r");
scaling = AZ_scaling_create();
start_time = AZ_second();
if (fp != NULL) {
fclose(fp);
options[AZ_precond] = AZ_none;
options[AZ_scaling] = AZ_sym_diag;
options[AZ_ignore_scaling] = AZ_TRUE;
options[AZ_keep_info] = 1;
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling);
/*
options[AZ_pre_calc] = AZ_reuse;
options[AZ_conv] = AZ_expected_values;
if (proc_config[AZ_node] == 0)
printf("\n-------- Second solve with improved convergence test -----\n");
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling);
if (proc_config[AZ_node] == 0)
printf("\n-------- Third solve with improved convergence test -----\n");
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling);
*/
}
else {
options[AZ_keep_info] = 1;
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling);
options[AZ_pre_calc] = AZ_reuse;
options[AZ_conv] = AZ_expected_values;
/*
if (proc_config[AZ_node] == 0)
printf("\n-------- Second solve with improved convergence test -----\n");
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling);
if (proc_config[AZ_node] == 0)
printf("\n-------- Third solve with improved convergence test -----\n");
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling);
*/
}
solve_time = AZ_second() - start_time;
if (proc_config[AZ_node] == 0)
printf("Solve time = %e, MG Setup time = %e\n", solve_time, setup_time);
if (proc_config[AZ_node] == 0)
printf("Printing out a few entries of the solution ...\n");
for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++)
if (update[j] == 7) {printf("solution(gid = %d) = %10.4e\n",
update[j],xxx[update_index[j]]); fflush(stdout);}
j = AZ_gsum_int(7, proc_config); /* sync processors */
for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++)
if (update[j] == 23) {printf("solution(gid = %d) = %10.4e\n",
update[j],xxx[update_index[j]]); fflush(stdout);}
j = AZ_gsum_int(7, proc_config); /* sync processors */
for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++)
if (update[j] == 47) {printf("solution(gid = %d) = %10.4e\n",
update[j],xxx[update_index[j]]); fflush(stdout);}
j = AZ_gsum_int(7, proc_config); /* sync processors */
for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++)
if (update[j] == 101) {printf("solution(gid = %d) = %10.4e\n",
update[j],xxx[update_index[j]]); fflush(stdout);}
j = AZ_gsum_int(7, proc_config); /* sync processors */
for (j=0;j<Amat->data_org[AZ_N_internal]+ Amat->data_org[AZ_N_border];j++)
if (update[j] == 171) {printf("solution(gid = %d) = %10.4e\n",
update[j],xxx[update_index[j]]); fflush(stdout);}
ML_Aggregate_Destroy(&ag);
ML_Destroy(&ml);
AZ_free((void *) Amat->data_org);
AZ_free((void *) Amat->val);
AZ_free((void *) Amat->bindx);
AZ_free((void *) update);
AZ_free((void *) external);
AZ_free((void *) extern_index);
AZ_free((void *) update_index);
AZ_scaling_destroy(&scaling);
if (Amat != NULL) AZ_matrix_destroy(&Amat);
if (Pmat != NULL) AZ_precond_destroy(&Pmat);
free(xxx);
free(rhs);
#ifdef ML_MPI
MPI_Finalize();
#endif
return 0;
}
// ======================================================================
int ML_Operator_Add2(ML_Operator *A, ML_Operator *B, ML_Operator *C,
int matrix_type, double scalarA, double scalarB)
{
int A_allocated = 0, *A_bindx = NULL, B_allocated = 0, *B_bindx = NULL;
double *A_val = NULL, *B_val = NULL, *hashed_vals;
int i, A_length, B_length, *hashed_inds;
int max_nz_per_row = 0, min_nz_per_row=1e6, j;
int hash_val, index_length;
int *columns, *rowptr, nz_ptr, hash_used, global_col;
double *values;
struct ML_CSR_MSRdata *temp;
int *A_gids, *B_gids;
int max_per_proc;
#ifdef ML_WITH_EPETRA
int count;
#endif
if (A->getrow == NULL)
pr_error("ML_Operator_Add: A does not have a getrow function.\n");
if (B->getrow == NULL)
pr_error("ML_Operator_Add: B does not have a getrow function.\n");
if (A->getrow->Nrows != B->getrow->Nrows) {
printf("ML_Operator_Add: Can not add, two matrices do not have the same");
printf(" number of rows %d vs %d",A->getrow->Nrows,B->getrow->Nrows);
exit(1);
}
if (A->invec_leng != B->invec_leng) {
printf("ML_Operator_Add: Can not add, two matrices do not have the same");
printf(" number of columns %d vs %d",A->getrow->Nrows,B->getrow->Nrows);
exit(1);
}
/* let's just count some things */
index_length = A->invec_leng + 1;
if (A->getrow->pre_comm != NULL) {
ML_CommInfoOP_Compute_TotalRcvLength(A->getrow->pre_comm);
index_length += A->getrow->pre_comm->total_rcv_length;
}
if (B->getrow->pre_comm != NULL) {
ML_CommInfoOP_Compute_TotalRcvLength(B->getrow->pre_comm);
index_length += B->getrow->pre_comm->total_rcv_length;
}
ML_create_unique_col_id(A->invec_leng, &A_gids, A->getrow->pre_comm,
&max_per_proc,A->comm);
ML_create_unique_col_id(B->invec_leng, &B_gids, B->getrow->pre_comm,
&max_per_proc,B->comm);
hashed_inds = (int *) ML_allocate(sizeof(int)*index_length);
hashed_vals = (double *) ML_allocate(sizeof(double)*index_length);
for (i = 0; i < index_length; i++) hashed_inds[i] = -1;
for (i = 0; i < index_length; i++) hashed_vals[i] = 0.;
nz_ptr = 0;
for (i = 0 ; i < A->getrow->Nrows; i++) {
hash_used = 0;
ML_get_matrix_row(A, 1, &i, &A_allocated, &A_bindx, &A_val,
&A_length, 0);
for (j = 0; j < A_length; j++) {
global_col = A_gids[A_bindx[j]];
ML_hash_it(global_col, hashed_inds, index_length,&hash_used,&hash_val);
hashed_inds[hash_val] = global_col;
hashed_vals[hash_val] += scalarA * A_val[j];
A_bindx[j] = hash_val;
}
ML_get_matrix_row(B, 1, &i, &B_allocated, &B_bindx, &B_val,
&B_length, 0);
for (j = 0; j < B_length; j++) {
global_col = B_gids[B_bindx[j]];
ML_hash_it(global_col, hashed_inds, index_length,&hash_used, &hash_val);
hashed_inds[hash_val] = global_col;
hashed_vals[hash_val] += scalarB*B_val[j];
B_bindx[j] = hash_val;
}
for (j = 0; j < A_length; j++) {
nz_ptr++;
hashed_inds[A_bindx[j]] = -1;
hashed_vals[A_bindx[j]] = 0.;
}
for (j = 0; j < B_length; j++) {
if (hashed_inds[B_bindx[j]] != -1) {
nz_ptr++;
hashed_inds[B_bindx[j]] = -1;
hashed_vals[B_bindx[j]] = 0.;
}
}
}
nz_ptr++;
columns = 0;
values = 0;
rowptr = (int *) ML_allocate(sizeof(int)*(A->outvec_leng+1));
if (matrix_type == ML_CSR_MATRIX) {
columns= (int *) ML_allocate(sizeof(int)*nz_ptr);
values = (double *) ML_allocate(sizeof(double)*nz_ptr);
}
#ifdef ML_WITH_EPETRA
else if (matrix_type == ML_EpetraCRS_MATRIX) {
columns= (int *) ML_allocate(sizeof(int)*(index_length+1));
values = (double *) ML_allocate(sizeof(double)*(index_length+1));
}
#endif
else {
pr_error("ML_Operator_Add: Unknown matrix type\n");
}
nz_ptr = 0;
rowptr[0] = 0;
for (i = 0 ; i < A->getrow->Nrows; i++) {
hash_used = 0;
ML_get_matrix_row(A, 1, &i, &A_allocated, &A_bindx, &A_val,
&A_length, 0);
for (j = 0; j < A_length; j++) {
global_col = A_gids[A_bindx[j]];
ML_hash_it(global_col, hashed_inds, index_length,&hash_used, &hash_val);
hashed_inds[hash_val] = global_col;
hashed_vals[hash_val] += scalarA * A_val[j];
A_bindx[j] = hash_val;
}
ML_get_matrix_row(B, 1, &i, &B_allocated, &B_bindx, &B_val,
&B_length, 0);
for (j = 0; j < B_length; j++) {
global_col = B_gids[B_bindx[j]];
ML_hash_it(global_col, hashed_inds, index_length,&hash_used, &hash_val);
hashed_inds[hash_val] = global_col;
hashed_vals[hash_val] += scalarB*B_val[j];
B_bindx[j] = hash_val;
}
#ifdef ML_WITH_EPETRA
if (matrix_type == ML_EpetraCRS_MATRIX) {
for (j = 0; j < A_length; j++) {
columns[j] = hashed_inds[A_bindx[j]];
values[j] = hashed_vals[A_bindx[j]];
nz_ptr++;
hashed_inds[A_bindx[j]] = -1;
hashed_vals[A_bindx[j]] = 0.;
}
count = A_length;
for (j = 0; j < B_length; j++) {
if (hashed_inds[B_bindx[j]] != -1) {
columns[count] = hashed_inds[B_bindx[j]];
values[count++] = hashed_vals[B_bindx[j]];
nz_ptr++;
hashed_inds[B_bindx[j]] = -1;
hashed_vals[B_bindx[j]] = 0.;
}
}
ML_Epetra_CRSinsert(C,i,columns,values,count);
}
else {
#endif
for (j = 0; j < A_length; j++) {
columns[nz_ptr] = hashed_inds[A_bindx[j]];
values[nz_ptr] = hashed_vals[A_bindx[j]];
nz_ptr++;
hashed_inds[A_bindx[j]] = -1;
hashed_vals[A_bindx[j]] = 0.;
}
for (j = 0; j < B_length; j++) {
if (hashed_inds[B_bindx[j]] != -1) {
columns[nz_ptr] = hashed_inds[B_bindx[j]];
values[nz_ptr] = hashed_vals[B_bindx[j]];
nz_ptr++;
hashed_inds[B_bindx[j]] = -1;
hashed_vals[B_bindx[j]] = 0.;
}
}
#ifdef ML_WITH_EPETRA
}
#endif
rowptr[i+1] = nz_ptr;
j = rowptr[i+1] - rowptr[i];
if (j > max_nz_per_row)
max_nz_per_row = j;
if (j < min_nz_per_row && j>0)
min_nz_per_row = j;
}
if (matrix_type == ML_CSR_MATRIX) {
temp = (struct ML_CSR_MSRdata *) ML_allocate(sizeof(struct ML_CSR_MSRdata));
if (temp == NULL) pr_error("ML_Operator_Add: no space for temp\n");
temp->columns = columns;
temp->values = values;
temp->rowptr = rowptr;
ML_Operator_Set_ApplyFuncData(C, B->invec_leng, A->outvec_leng,
temp,A->outvec_leng, NULL,0);
ML_Operator_Set_Getrow(C, A->outvec_leng, CSR_getrow);
ML_Operator_Set_ApplyFunc (C, CSR_matvec);
ML_globalcsr2localcsr(C, max_per_proc);
C->data_destroy = ML_CSR_MSRdata_Destroy;
C->max_nz_per_row = max_nz_per_row;
C->min_nz_per_row = min_nz_per_row;
C->N_nonzeros = nz_ptr;
}
#ifdef ML_WITH_EPETRA
else {
ML_free(rowptr);
ML_free(columns);
ML_free(values);
}
#endif
ML_free(A_gids);
ML_free(B_gids);
ML_free(hashed_vals);
ML_free(hashed_inds);
ML_free(A_val);
ML_free(A_bindx);
ML_free(B_val);
ML_free(B_bindx);
return 1;
}
void ML_get_matrix_row(ML_Operator *input_matrix, int N_requested_rows,
int requested_rows[], int *allocated_space, int **columns,
double **values, int row_lengths[], int index)
{
int i, *mapper, *t1, row;
ML_Operator *next;
double *t2;
void *data;
int (*getfunction)(void *,int,int*,int,int*,double*,int*);
#ifdef DEBUG2
if (N_requested_rows != 1) {
printf("ML_get_matrix_row is currently implemented for only 1 row");
printf(" at a time.\n");
exit(1);
}
#endif
row = requested_rows[0];
#ifdef DEBUG2
if ( (row >= input_matrix->getrow->Nrows) || (row < 0) ) {
row_lengths[0] = 0;
return;
}
#endif
if (input_matrix->getrow->row_map != NULL) {
if (input_matrix->getrow->row_map[row] != -1)
row = input_matrix->getrow->row_map[row];
else {
row_lengths[0] = 0;
ML_avoid_unused_param( (void *) &N_requested_rows);
return;}
}
next = input_matrix->sub_matrix;
while ( (next != NULL) && (row < next->getrow->Nrows) ) {
input_matrix = next;
next = next->sub_matrix;
}
if (next != NULL) row -= next->getrow->Nrows;
data = (void *) input_matrix;
getfunction = (int (*)(void *,int,int*,int,int*,double*,int*))
input_matrix->getrow->func_ptr;
while(getfunction(data,1,&row,*allocated_space-index,
&((*columns)[index]), &((*values)[index]), row_lengths) == 0) {
*allocated_space = 2*(*allocated_space) + 1;
t1 = (int *) ML_allocate(*allocated_space*sizeof(int ));
if (t1 == NULL) {
printf("Not enough space to get a matrix row. A row length of \n");
printf("%d Was not sufficient\n",(*allocated_space-1)/2);
fflush(stdout);
exit(1);
}
else {
for (i = 0; i < index; i++) t1[i] = (*columns)[i];
if (*columns != NULL) ML_free(*columns);
*columns = t1;
}
t2 = (double *) ML_allocate(*allocated_space*sizeof(double));
if (t2 == NULL) {
printf("Not enough space to get a matrix row. A row length of \n");
printf("%d Was not sufficient\n",(*allocated_space-1)/2);
fflush(stdout);
exit(1);
}
for (i = 0; i < index; i++) t2[i] = (*values)[i];
if (*values != NULL) ML_free(*values);
*values = t2;
}
if ( (input_matrix->getrow->use_loc_glob_map == ML_YES)) {
mapper = input_matrix->getrow->loc_glob_map;
for (i = 0; i < row_lengths[0]; i++)
(*columns)[i+index] = mapper[(*columns)[index+i]];
}
}
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);
}
char *ML_allocate(unsigned int isize) {
/*
* Allocate memory and record the event by placing an entry in the
* ml_widget_head list. Also recored the size of this entry as well.
*
* Note: we actually allocate more memory than is requested (7 doubles more).
* This additional memory is used to record the 'size' and to mark the
* memory with a header and trailer which we can later check to see if
* they were overwritten.
*
*/
char *ptr, *header_start, *header_end;
struct ml_widget *ml_widget;
int *size_ptr, i, size;
double *dptr;
size = (int) isize;
size = size + 7*sizeof(double);
ml_widget = (struct ml_widget *) malloc(sizeof(struct ml_widget));
/* THIS MALLOC NEEDS TO STAY A */
/* MALLOC AND NOT AN ML_ALLOCATE */
if (ml_widget == NULL) return(NULL);
ptr = (char *) malloc(size); /* THIS MALLOC NEEDS TO STAY A */
/* MALLOC AND NOT AN ML_ALLOCATE */
if (ptr == NULL) {
ML_free(ml_widget); /* THIS FREE() NEEDS TO STAY A */
/* FREE AND NOT AN ML_FREE */
return(NULL);
}
ml_allo_count++;
/* put trash in the space to make sure nobody is expecting zeros */
for (i = 0 ; i < size/sizeof(char) ; i++ )
ptr[i] = 'f';
/* record the entry */
ml_widget->order = ml_allo_count;
if (size == -7*sizeof(double) ) {
printf("allocating 0 space %u (%d)\n",ptr,size);
i = 0;
size = 1/i;
ml_widget = NULL;
}
ml_widget->size = size - 7*sizeof(double);
ml_widget->next = ml_widget_head;
ml_widget_head = ml_widget;
ml_widget->address = ptr;
size_ptr = (int *) ptr;
size_ptr[0] = size - 7*sizeof(double);
dptr = (double *) ptr;
/* mark the header */
header_start = (char *) &(dptr[1]);
for (i = 0 ; i < 3*sizeof(double)/sizeof(char) ; i++ )
header_start[i] = 'x';
/* mark the trailer */
header_end = &(ptr[ (size/sizeof(char)) - 1]);
header_start = (char *) &(dptr[4]);
header_start = & (header_start[(size-7*sizeof(double))/sizeof(char)]);
while (header_start <= header_end) {
*header_start = 'x';
header_start++;
}
return( (char *) &(dptr[4]) );
}
char *ML_myrealloc(void *vptr, unsigned int new_size) {
struct ml_widget *current, *prev;
int i, *iptr, size, *new_size_ptr;
char *header_start, *header_end, *ptr;
char *data1, *data2, *new_ptr, *new_header_start, *new_header_end;
int newmsize, smaller;
double *dptr, *new_dptr;
ptr = (char *) vptr;
if (ptr == NULL) {
printf("Trying to realloc a NULL ptr\n");
exit(1);
}
else {
current = ml_widget_head;
prev = NULL;
data1 = ptr;
dptr = (double *) ptr;
--dptr;
--dptr;
--dptr;
--dptr;
ptr = (char *) dptr;
while (current != NULL) {
if (current->address == ptr) break;
else { prev = current; current = current->next; }
}
if (current == NULL) {
printf("the pointer %u was not found and thus can not be realloc.\n",
ptr);
exit(1);
}
else {
/* check to see if the header is corrupted */
iptr = (int *) ptr;
header_start = (char *) &(dptr[1]);
for (i = 0 ; i < 3*sizeof(double)/sizeof(char) ; i++ ) {
if (header_start[i] != 'x') {
printf("realloc header is corrupted for %u (%d,%d)\n",ptr,
current->size,current->order);
size = 0;
size = 1/size;
}
/* DO WE CHECK THE TRAILER ???? */
}
size = iptr[0];
newmsize = new_size + 7*sizeof(double);
new_ptr = (char *) malloc(newmsize); /* THIS MALLOC NEEDS TO STAY A */
/* MALLOC AND NOT AN ML_ALLOCATE */
if (new_ptr == NULL) return(NULL);
new_size_ptr = (int *) new_ptr;
new_size_ptr[0] = new_size;
new_dptr = (double *) new_ptr;
data2 = (char *) &(new_dptr[4]);
new_header_start = (char *) &(new_dptr[1]);
for (i = 0 ; i < 3*sizeof(double)/sizeof(char) ; i++ )
new_header_start[i] = 'x';
new_header_end = &(new_ptr[ (newmsize/sizeof(char)) - 1]);
new_header_start = (char *) &(new_dptr[4]);
new_header_start= &(new_header_start[new_size/sizeof(char)]);
while (new_header_start <= new_header_end) {
*new_header_start = 'x';
new_header_start++;
}
smaller = current->size;
if (smaller > new_size ) smaller = new_size;
for (i = 0 ; i < smaller; i++) data2[i] = data1[i];
ML_free(dptr);
current->size = new_size;
current->address = (char *) new_dptr;
return( (char *) &(new_dptr[4]));
}
}
}
// ================================================ ====== ==== ==== == =
//! Build pi operator described by Bochev, Siefert, Tuminaro, Xu and Zhu (2007).
int ML_Epetra::FaceMatrixFreePreconditioner::BuildNullspace(Epetra_MultiVector *& nullspace){
int Nf=FaceRangeMap_->NumMyElements();
/* Pull the coordinates from Teuchos */
double * xcoord=List_.get("x-coordinates",(double*)0);
double * ycoord=List_.get("y-coordinates",(double*)0);
double * zcoord=List_.get("z-coordinates",(double*)0);
dim=(xcoord!=0) + (ycoord!=0) + (zcoord!=0);
// Sanity check
if(dim!=3){if(!Comm_->MyPID()) printf("ERROR: FaceMatrixFreePreconditioner only works in 3D"); ML_CHK_ERR(-1);}
// Build the (unimported) coordinate multivector
double **d_coords=new double* [dim];
d_coords[0]=xcoord; d_coords[1]=ycoord;
if(dim==3) d_coords[2]=zcoord;
Epetra_MultiVector n_coords_domain(View,*NodeDomainMap_,d_coords,dim);
Epetra_MultiVector *n_coords;
// Import coordinate info
if(FaceNode_Matrix_->Importer()){
n_coords=new Epetra_MultiVector(FaceNode_Matrix_->ColMap(),dim);
n_coords->PutScalar(0.0);
n_coords->Import(n_coords_domain,*FaceNode_Matrix_->Importer(),Add);
}
else n_coords=&n_coords_domain;
// Sanity HAQ - Only works on Hexes
if(FaceNode_Matrix_->GlobalMaxNumEntries()!=4)
{if(!Comm_->MyPID()) printf("ERROR: FaceMatrixFreePreconditioner only works on Hexes"); ML_CHK_ERR(-2);}
// Allocate vector
nullspace=new Epetra_MultiVector(*FaceDomainMap_,3);
// Fill matrix - NTS this will *NOT* do periodic BCs correctly.
double *a=new double[dim];
double *b=new double[dim];
double *c=new double[dim];
for(int i=0;i<Nf;i++){
int Ni,*indices;
double *values;
FaceNode_Matrix_->ExtractMyRowView(i,Ni,values,indices);
if(Ni != 4){
printf("ERROR: Face %d has only %d nodes\n",i,Ni);
ML_free(a);
ML_free(b);
ML_free(c);
ML_CHK_ERR(-1);
}
a[0] = (*n_coords)[0][indices[1]] - (*n_coords)[0][indices[0]];
a[1] = (*n_coords)[1][indices[1]] - (*n_coords)[1][indices[0]];
a[2] = (*n_coords)[2][indices[1]] - (*n_coords)[2][indices[0]];
b[0] = (*n_coords)[0][indices[2]] - (*n_coords)[0][indices[0]];
b[1] = (*n_coords)[1][indices[2]] - (*n_coords)[1][indices[0]];
b[2] = (*n_coords)[2][indices[2]] - (*n_coords)[2][indices[0]];
cross_product(a,b,c);
// HAQ - Hardwiring for hexes
// HAQ - Absolute value, presuming all hexes are actually pointed the same way. This is a HAQ!!!
(*nullspace)[0][i]=ABS(c[0])/6.0;
(*nullspace)[1][i]=ABS(c[1])/6.0;
(*nullspace)[2][i]=ABS(c[2])/6.0;
}
/* Cleanup */
if(FaceNode_Matrix_->Importer()) delete n_coords;
delete [] a; delete [] b; delete [] c;
delete [] d_coords;
return 0;
}
ML_Operator *user_T_build(struct user_partition *Edge_Partition,
struct user_partition *Node_Partition,
ML_Operator *Kn_mat, ML_Comm *comm)
{
int nx, i, ii, jj, horv, Ncols, Nexterns;
int *Tmat_bindx;
double *Tmat_val;
ML_Operator *Tmat;
struct ML_CSR_MSRdata *csr_data;
struct aztec_context *aztec_context;
int global_id;
int Nlocal_nodes, Nlocal_edges;
int nz_ptr;
Nlocal_nodes = Node_Partition->Nlocal;
Nlocal_edges = Edge_Partition->Nlocal;
nx = (int) sqrt( ((double) Node_Partition->Nglobal) + .00001);
Tmat_bindx = (int *) malloc((3*Nlocal_edges+1)*sizeof(int));
Tmat_val = (double *) malloc((3*Nlocal_edges+1)*sizeof(double));
Tmat_bindx[0] = Nlocal_edges + 1;
for (i = 0; i < Nlocal_edges; i++) {
global_id = (Edge_Partition->my_global_ids)[i];
Tmat_val[i] = 0.0;
invindex(global_id, &ii, &jj, nx, &horv);
nz_ptr = Tmat_bindx[i];
ii--;
if (horv == HORIZONTAL) {
if(ii != -1) {
Tmat_bindx[nz_ptr] = southwest(ii,jj,nx); Tmat_val[nz_ptr++] = -1.;
}
Tmat_bindx[nz_ptr] = southeast(ii,jj,nx); Tmat_val[nz_ptr++] = 1.;
}
else {
if (ii == -1) ii = nx-1;
Tmat_bindx[nz_ptr] = northwest(ii,jj,nx); Tmat_val[nz_ptr++] = -1.;
if (jj != 0) {
Tmat_bindx[nz_ptr] = southwest(ii,jj,nx); Tmat_val[nz_ptr++] = 1.;}
}
Tmat_bindx[i+1] = nz_ptr;
}
csr_data = (struct ML_CSR_MSRdata *) ML_allocate(sizeof(struct ML_CSR_MSRdata));
csr_data->columns = Tmat_bindx;
csr_data->values = Tmat_val;
ML_MSR2CSR(csr_data, Nlocal_edges, &Ncols);
aztec_context = (struct aztec_context *) Kn_mat->data;
Nexterns = (aztec_context->Amat->data_org)[AZ_N_external];
AZ_Tmat_transform2ml(Nexterns, Node_Partition->needed_external_ids,
reordered_node_externs,
Tmat_bindx, Tmat_val, csr_data->rowptr, Nlocal_nodes,
Node_Partition->my_global_ids,
comm, Nlocal_edges, &Tmat);
ML_free(csr_data);
Tmat->data_destroy = ML_CSR_MSRdata_Destroy;
ML_CommInfoOP_Clone(&(Tmat->getrow->pre_comm), Kn_mat->getrow->pre_comm);
return(Tmat);
}
// ================================================ ====== ==== ==== == =
// Copied from ml_agg_genP.c
static void ML_Init_Aux(ML_Operator* A, Teuchos::ParameterList &List) {
int i, j, n, count, num_PDEs, BlockRow, BlockCol;
double threshold;
int* columns;
double* values;
int allocated, entries = 0;
int N_dimensions;
int DiagID;
double DiagValue;
int** filter;
double dist;
double *LaplacianDiag;
int Nghost;
// Boundary exchange the coords
double *x_coord=0, *y_coord=0, *z_coord=0;
RefMaxwell_SetupCoordinates(A,List,x_coord,y_coord,z_coord);
int dim=(x_coord!=0) + (y_coord!=0) + (z_coord!=0);
/* Sanity Checks */
if(dim == 0 || ((!x_coord && (y_coord || z_coord)) || (x_coord && !y_coord && z_coord))){
std::cerr<<"Error: Coordinates not defined. This is necessary for aux aggregation (found "<<dim<<" coordinates).\n";
exit(-1);
}
num_PDEs = A->num_PDEs;
N_dimensions = dim;
threshold = A->aux_data->threshold;
ML_Operator_AmalgamateAndDropWeak(A, num_PDEs, 0.0);
n = A->invec_leng;
Nghost = ML_CommInfoOP_Compute_TotalRcvLength(A->getrow->pre_comm);
LaplacianDiag = (double *) ML_allocate((A->getrow->Nrows+Nghost+1)*
sizeof(double));
filter = (int**) ML_allocate(sizeof(int*) * n);
allocated = 128;
columns = (int *) ML_allocate(allocated * sizeof(int));
values = (double *) ML_allocate(allocated * sizeof(double));
for (i = 0 ; i < n ; ++i) {
BlockRow = i;
DiagID = -1;
DiagValue = 0.0;
ML_get_matrix_row(A,1,&i,&allocated,&columns,&values, &entries,0);
for (j = 0; j < entries; j++) {
BlockCol = columns[j];
if (BlockRow != BlockCol) {
dist = 0.0;
switch (N_dimensions) {
case 3:
dist += (z_coord[BlockRow] - z_coord[BlockCol]) * (z_coord[BlockRow] - z_coord[BlockCol]);
case 2:
dist += (y_coord[BlockRow] - y_coord[BlockCol]) * (y_coord[BlockRow] - y_coord[BlockCol]);
case 1:
dist += (x_coord[BlockRow] - x_coord[BlockCol]) * (x_coord[BlockRow] - x_coord[BlockCol]);
}
if (dist == 0.0) {
printf("node %d = %e ", i, x_coord[BlockRow]);
if (N_dimensions > 1) printf(" %e ", y_coord[BlockRow]);
if (N_dimensions > 2) printf(" %e ", z_coord[BlockRow]);
printf("\n");
printf("node %d = %e ", j, x_coord[BlockCol]);
if (N_dimensions > 1) printf(" %e ", y_coord[BlockCol]);
if (N_dimensions > 2) printf(" %e ", z_coord[BlockCol]);
printf("\n");
printf("Operator has inlen = %d and outlen = %d\n",
A->invec_leng, A->outvec_leng);
}
dist = 1.0 / dist;
DiagValue += dist;
}
else if (columns[j] == i) {
DiagID = j;
}
}
if (DiagID == -1) {
fprintf(stderr, "ERROR: matrix has no diagonal!\n"
"ERROR: (file %s, line %d)\n",
__FILE__, __LINE__);
exit(EXIT_FAILURE);
}
LaplacianDiag[BlockRow] = DiagValue;
}
if ( A->getrow->pre_comm != NULL )
ML_exchange_bdry(LaplacianDiag,A->getrow->pre_comm,A->getrow->Nrows,
A->comm, ML_OVERWRITE,NULL);
for (i = 0 ; i < n ; ++i) {
BlockRow = i;
ML_get_matrix_row(A,1,&i,&allocated,&columns,&values, &entries,0);
for (j = 0; j < entries; j++) {
BlockCol = columns[j];
if (BlockRow != BlockCol) {
dist = 0.0;
switch (N_dimensions) {
case 3:
dist += (z_coord[BlockRow] - z_coord[BlockCol]) * (z_coord[BlockRow] - z_coord[BlockCol]);
case 2:
dist += (y_coord[BlockRow] - y_coord[BlockCol]) * (y_coord[BlockRow] - y_coord[BlockCol]);
case 1:
dist += (x_coord[BlockRow] - x_coord[BlockCol]) * (x_coord[BlockRow] - x_coord[BlockCol]);
}
dist = 1.0 / dist;
values[j] = dist;
}
}
count = 0;
for (j = 0 ; j < entries ; ++j) {
if ( (i != columns[j]) &&
(values[j]*values[j] <
LaplacianDiag[BlockRow]*LaplacianDiag[columns[j]]*threshold*threshold)){
columns[count++] = columns[j];
}
}
/* insert the rows */
filter[BlockRow] = (int*) ML_allocate(sizeof(int) * (count + 1));
filter[BlockRow][0] = count;
for (j = 0 ; j < count ; ++j)
filter[BlockRow][j + 1] = columns[j];
}
ML_free(columns);
ML_free(values);
ML_free(LaplacianDiag);
ML_Operator_UnAmalgamateAndDropWeak(A, num_PDEs, 0.0);
A->aux_data->aux_func_ptr = A->getrow->func_ptr;
A->getrow->func_ptr = ML_Aux_Getrow;
A->aux_data->filter = filter;
A->aux_data->filter_size = n;
// Cleanup
ML_free(x_coord);
ML_free(y_coord);
ML_free(z_coord);
}
