// ======================================================================
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 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);
}
int main(int argc, char *argv[])
{
int num_PDE_eqns=6, N_levels=4, nsmooth=2;
int leng, level, N_grid_pts, coarsest_level;
/* 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 = NULL;
#ifdef ML_partition
int nblocks;
int *block_list = NULL;
int k;
#endif
struct AZ_SCALING *scaling;
ML_Aggregate *ag;
double *mode, *rigid;
char filename[80];
double alpha;
int allocated = 0;
int old_prec, old_sol;
double old_tol;
/*
double *Amode, beta, biggest;
int big_ind = -1, ii;
*/
ML_Operator *Amatrix;
int *rowi_col = NULL, rowi_N, count2, ccc;
double *rowi_val = NULL;
double max_diag, min_diag, max_sum, sum;
int nBlocks, *blockIndices, Ndof;
#ifdef ML_partition
FILE *fp2;
int count;
if (argc != 2) {
printf("Usage: ml_read_elas num_processors\n");
exit(1);
}
else sscanf(argv[1],"%d",&nblocks);
#endif
#ifdef HAVE_MPI
MPI_Init(&argc,&argv);
/* get number of processors and the name of this processor */
AZ_set_proc_config(proc_config, MPI_COMM_WORLD);
#else
AZ_set_proc_config(proc_config, AZ_NOT_MPI);
#endif
/* read in the number of matrix equations */
leng = 0;
if (proc_config[AZ_node] == 0) {
# ifdef binary
fp=fopen(".data","rb");
# else
fp=fopen(".data","r");
# endif
if (fp==NULL) {
printf("couldn't open file .data\n");
exit(1);
}
# ifdef binary
fread(&leng, sizeof(int), 1, fp);
# else
fscanf(fp,"%d",&leng);
# endif
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 (proc_config[AZ_N_procs] == 1) i = AZ_linear;
else i = AZ_file;
AZ_read_update(&N_update, &update, proc_config, N_grid_pts, num_PDE_eqns,i);
AZ_read_msr_matrix(update, &val, &bindx, N_update, proc_config);
/* This code is to fix things up so that we are sure we have */
/* all block (including the ghost nodes the same size. */
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];
#ifdef SCALE_ME
ML_MSR_sym_diagonal_scaling(Amat, proc_config, &scaling_vect);
#endif
start_time = AZ_second();
options[AZ_scaling] = AZ_none;
ML_Create(&ml, N_levels);
ML_Set_PrintLevel(10);
/* set up discretization matrix and matrix vector function */
AZ_ML_Set_Amat(ml, N_levels-1, N_update, N_update, Amat, proc_config);
#ifdef ML_partition
/* this code is meant to partition the matrices so that things can be */
/* run in parallel later. */
/* It is meant to be run on only one processor. */
#ifdef MB_MODIF
fp2 = fopen(".update","w");
#else
fp2 = fopen("partition_file","w");
#endif
ML_Operator_AmalgamateAndDropWeak(&(ml->Amat[N_levels-1]), num_PDE_eqns, 0.0);
ML_Gen_Blocks_Metis(ml, N_levels-1, &nblocks, &block_list);
for (i = 0; i < nblocks; i++) {
count = 0;
for (j = 0; j < ml->Amat[N_levels-1].outvec_leng; j++) {
if (block_list[j] == i) count++;
}
fprintf(fp2," %d\n",count*num_PDE_eqns);
for (j = 0; j < ml->Amat[N_levels-1].outvec_leng; j++) {
if (block_list[j] == i) {
for (k = 0; k < num_PDE_eqns; k++) fprintf(fp2,"%d\n",j*num_PDE_eqns+k);
}
}
}
fclose(fp2);
ML_Operator_UnAmalgamateAndDropWeak(&(ml->Amat[N_levels-1]),num_PDE_eqns,0.0);
#ifdef MB_MODIF
printf(" partition file dumped in .update\n");
#endif
exit(1);
#endif
ML_Aggregate_Create( &ag );
/*
ML_Aggregate_Set_CoarsenScheme_MIS(ag);
*/
#ifdef MB_MODIF
ML_Aggregate_Set_DampingFactor(ag,1.50);
#else
ML_Aggregate_Set_DampingFactor(ag,1.5);
#endif
ML_Aggregate_Set_CoarsenScheme_METIS(ag);
ML_Aggregate_Set_NodesPerAggr( ml, ag, -1, 35);
/*
ML_Aggregate_Set_Phase3AggregateCreationAggressiveness(ag, 10.001);
*/
ML_Aggregate_Set_Threshold(ag, 0.0);
ML_Aggregate_Set_MaxCoarseSize( ag, 300);
/* read in the rigid body modes */
Nrigid = 0;
/* to ensure compatibility with RBM dumping software */
if (proc_config[AZ_node] == 0) {
sprintf(filename,"rigid_body_mode%02d",Nrigid+1);
while( (fp = fopen(filename,"r")) != NULL) {
which_filename = 1;
fclose(fp);
Nrigid++;
sprintf(filename,"rigid_body_mode%02d",Nrigid+1);
}
sprintf(filename,"rigid_body_mode%d",Nrigid+1);
while( (fp = fopen(filename,"r")) != NULL) {
fclose(fp);
Nrigid++;
sprintf(filename,"rigid_body_mode%d",Nrigid+1);
}
}
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");
}
}
rhs = (double *) malloc(leng*sizeof(double));
xxx = (double *) malloc(leng*sizeof(double));
for (iii = 0; iii < leng; iii++) xxx[iii] = 0.0;
for (i = 0; i < Nrigid; i++) {
if (which_filename == 1) sprintf(filename,"rigid_body_mode%02d",i+1);
else sprintf(filename,"rigid_body_mode%d",i+1);
AZ_input_msr_matrix(filename,update,&mode,&garbage,N_update,proc_config);
AZ_reorder_vec(mode, data_org, update_index, NULL);
/* here is something to stick a rigid body mode as the initial */
/* The idea is to solve A x = 0 without smoothing with a two */
/* level method. If everything is done properly, we should */
/* converge in 2 iterations. */
/* Note: we must also zero out components of the rigid body */
/* mode that correspond to Dirichlet bcs. */
if (i == -4) {
for (iii = 0; iii < leng; iii++) xxx[iii] = mode[iii];
ccc = 0;
Amatrix = &(ml->Amat[N_levels-1]);
for (iii = 0; iii < Amatrix->outvec_leng; iii++) {
ML_get_matrix_row(Amatrix,1,&iii,&allocated,&rowi_col,&rowi_val,
&rowi_N, 0);
count2 = 0;
for (j = 0; j < rowi_N; j++) if (rowi_val[j] != 0.) count2++;
if (count2 <= 1) { xxx[iii] = 0.; ccc++; }
}
free(rowi_col); free(rowi_val);
allocated = 0; rowi_col = NULL; rowi_val = NULL;
}
/*
* Rescale matrix/rigid body modes and checking
*
AZ_sym_rescale_sl(mode, Amat->data_org, options, proc_config, scaling);
Amat->matvec(mode, rigid, Amat, proc_config);
for (j = 0; j < N_update; j++) printf("this is %d %e\n",j,rigid[j]);
*/
/* Here is some code to check that the rigid body modes are */
/* really rigid body modes. The idea is to multiply by A and */
/* then to zero out things that we "think" are boundaries. */
/* In this hardwired example, things near boundaries */
/* correspond to matrix rows that do not have 81 nonzeros. */
/*
Amode = (double *) malloc(leng*sizeof(double));
Amat->matvec(mode, Amode, Amat, proc_config);
j = 0;
biggest = 0.0;
for (ii = 0; ii < N_update; ii++) {
if ( Amat->bindx[ii+1] - Amat->bindx[ii] != 80) {
Amode[ii] = 0.; j++;
}
else {
if ( fabs(Amode[ii]) > biggest) {
biggest=fabs(Amode[ii]); big_ind = ii;
}
}
}
printf("%d entries zeroed out of %d elements\n",j,N_update);
alpha = AZ_gdot(N_update, Amode, Amode, proc_config);
beta = AZ_gdot(N_update, mode, mode, proc_config);
printf("||A r||^2 =%e, ||r||^2 = %e, ratio = %e\n",
alpha,beta,alpha/beta);
printf("the biggest is %e at row %d\n",biggest,big_ind);
free(Amode);
*/
/* orthogonalize mode with respect to previous modes. */
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_(&N_update,&alpha,&(rigid[j*N_update]), &one, mode, &one); */
}
#ifndef MB_MODIF
printf(" after mb %e %e %e\n",mode[0],mode[1],mode[2]);
#endif
for (j = 0; j < N_update; j++) rigid[i*N_update+j] = mode[j];
free(mode);
free(garbage); garbage = NULL;
}
if (Nrigid != 0) {
ML_Aggregate_Set_BlockDiagScaling(ag);
ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, Nrigid, rigid, N_update);
free(rigid);
}
#ifdef SCALE_ME
ML_Aggregate_Scale_NullSpace(ag, scaling_vect, N_update);
#endif
coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, N_levels-1,
ML_DECREASING, ag);
AZ_defaults(options, params);
coarsest_level = N_levels - coarsest_level;
if ( proc_config[AZ_node] == 0 )
printf("Coarse level = %d \n", coarsest_level);
/* set up smoothers */
for (level = N_levels-1; level > coarsest_level; level--) {
/*
ML_Gen_Smoother_BlockGaussSeidel(ml, level,ML_BOTH, 1, 1., num_PDE_eqns);
*/
/* Sparse approximate inverse smoother that acutally does both */
/* pre and post smoothing. */
/*
ML_Gen_Smoother_ParaSails(ml , level, ML_PRESMOOTHER, nsmooth,
parasails_sym, parasails_thresh,
parasails_nlevels, parasails_filter,
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. */
/* ML_Gen_Smoother_Cheby(ml, level, ML_BOTH, 30., nsmooth); */
Ndof = ml->Amat[level].invec_leng;
ML_Gen_Blocks_Aggregates(ag, level, &nBlocks, &blockIndices);
ML_Gen_Smoother_BlockDiagScaledCheby(ml, level, ML_BOTH, 30.,nsmooth,
nBlocks, blockIndices);
/*
ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_BOTH, nsmooth,1.);
*/
/* 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/num_PDE_eqns;
blocks = (int *) ML_allocate(sizeof(int)*N_update);
for (i =0; i < ml->Amat[level].invec_leng; i++)
blocks[i] = i/num_PDE_eqns;
ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml , level, ML_PRESMOOTHER,
nsmooth, 1., nblocks, blocks);
ML_Gen_Smoother_VBlockSymGaussSeidelSequential(ml, level, ML_POSTSMOOTHER,
nsmooth, 1., nblocks, blocks);
free(blocks);
*/
/* Block Jacobi Smoothing */
/*
nblocks = ml->Amat[level].invec_leng/num_PDE_eqns;
blocks = (int *) ML_allocate(sizeof(int)*N_update);
for (i =0; i < ml->Amat[level].invec_leng; i++)
blocks[i] = i/num_PDE_eqns;
ML_Gen_Smoother_VBlockJacobi(ml , level, ML_BOTH, nsmooth,
ML_ONE_STEP_CG, nblocks, blocks);
free(blocks);
*/
/* Jacobi Smoothing */
/*
ML_Gen_Smoother_Jacobi(ml , level, ML_PRESMOOTHER, nsmooth, ML_ONE_STEP_CG);
ML_Gen_Smoother_Jacobi(ml , level, ML_POSTSMOOTHER, nsmooth,ML_ONE_STEP_CG);
*/
/* This does a block Gauss-Seidel (not true GS in parallel) */
/* where each processor has 'nblocks' blocks. */
/*
nblocks = 250;
ML_Gen_Blocks_Metis(ml, level, &nblocks, &blocks);
ML_Gen_Smoother_VBlockJacobi(ml , level, ML_BOTH, nsmooth,ML_ONE_STEP_CG,
nblocks, blocks);
free(blocks);
*/
num_PDE_eqns = 6;
}
/* Choose coarse grid solver: mls, superlu, symGS, or Aztec */
/*
ML_Gen_Smoother_Cheby(ml, coarsest_level, ML_BOTH, 30., nsmooth);
ML_Gen_CoarseSolverSuperLU( ml, coarsest_level);
*/
/*
ML_Gen_Smoother_SymGaussSeidel(ml , coarsest_level, ML_BOTH, nsmooth,1.);
*/
old_prec = options[AZ_precond];
old_sol = options[AZ_solver];
old_tol = params[AZ_tol];
params[AZ_tol] = 1.0e-9;
params[AZ_tol] = 1.0e-5;
options[AZ_precond] = AZ_Jacobi;
options[AZ_solver] = AZ_cg;
options[AZ_poly_ord] = 1;
options[AZ_conv] = AZ_r0;
options[AZ_orth_kvecs] = AZ_TRUE;
j = AZ_gsum_int(ml->Amat[coarsest_level].outvec_leng, proc_config);
options[AZ_keep_kvecs] = j - 6;
options[AZ_max_iter] = options[AZ_keep_kvecs];
ML_Gen_SmootherAztec(ml, coarsest_level, options, params,
proc_config, status, options[AZ_keep_kvecs], ML_PRESMOOTHER, NULL);
options[AZ_conv] = AZ_noscaled;
options[AZ_keep_kvecs] = 0;
options[AZ_orth_kvecs] = 0;
options[AZ_precond] = old_prec;
options[AZ_solver] = old_sol;
params[AZ_tol] = old_tol;
/* */
#ifdef RST_MODIF
ML_Gen_Solver(ml, ML_MGV, N_levels-1, coarsest_level);
#else
#ifdef MB_MODIF
ML_Gen_Solver(ml, ML_SAAMG, N_levels-1, coarsest_level);
#else
ML_Gen_Solver(ml, ML_MGFULLV, N_levels-1, coarsest_level);
#endif
#endif
options[AZ_solver] = AZ_GMRESR;
options[AZ_solver] = AZ_cg;
options[AZ_scaling] = AZ_none;
options[AZ_precond] = AZ_user_precond;
options[AZ_conv] = AZ_r0;
options[AZ_conv] = AZ_noscaled;
options[AZ_output] = 1;
options[AZ_max_iter] = 500;
options[AZ_poly_ord] = 5;
options[AZ_kspace] = 40;
params[AZ_tol] = 4.8e-6;
AZ_set_ML_preconditioner(&Pmat, Amat, ml, options);
setup_time = AZ_second() - start_time;
/* Set rhs */
fp = fopen("AZ_capture_rhs.dat","r");
if (fp == NULL) {
AZ_random_vector(rhs, data_org, proc_config);
if (proc_config[AZ_node] == 0) printf("taking random vector for rhs\n");
for (i = 0; i < -N_update; i++) {
rhs[i] = (double) update[i]; rhs[i] = 7.;
}
}
else {
if (proc_config[AZ_node]== 0) printf("reading rhs guess from file\n");
AZ_input_msr_matrix("AZ_capture_rhs.dat", update, &rhs, &garbage,
N_update, proc_config);
free(garbage);
}
AZ_reorder_vec(rhs, data_org, update_index, NULL);
printf("changing rhs by multiplying with A\n");
Amat->matvec(rhs, xxx, Amat, proc_config);
for (i = 0; i < N_update; i++) rhs[i] = xxx[i];
fp = fopen("AZ_capture_init_guess.dat","r");
if (fp != NULL) {
fclose(fp);
if (proc_config[AZ_node]== 0) printf("reading initial guess from file\n");
AZ_input_msr_matrix("AZ_capture_init_guess.dat", update, &xxx, &garbage,
N_update, proc_config);
free(garbage);
xxx = (double *) realloc(xxx, sizeof(double)*(
Amat->data_org[AZ_N_internal]+
Amat->data_org[AZ_N_border] +
Amat->data_org[AZ_N_external]));
}
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;
options[AZ_conv] = AZ_noscaled;
options[AZ_conv] = AZ_r0;
params[AZ_tol] = 1.0e-7;
/* ML_Iterate(ml, xxx, rhs); */
alpha = sqrt(AZ_gdot(N_update, xxx, xxx, proc_config));
printf("init guess = %e\n",alpha);
alpha = sqrt(AZ_gdot(N_update, rhs, rhs, proc_config));
printf("rhs = %e\n",alpha);
#ifdef SCALE_ME
ML_MSR_scalerhs(rhs, scaling_vect, data_org[AZ_N_internal] +
data_org[AZ_N_border]);
ML_MSR_scalesol(xxx, scaling_vect, data_org[AZ_N_internal] +
data_org[AZ_N_border]);
#endif
max_diag = 0.;
min_diag = 1.e30;
max_sum = 0.;
for (i = 0; i < N_update; i++) {
if (Amat->val[i] < 0.) printf("woops negative diagonal A(%d,%d) = %e\n",
i,i,Amat->val[i]);
if (Amat->val[i] > max_diag) max_diag = Amat->val[i];
if (Amat->val[i] < min_diag) min_diag = Amat->val[i];
sum = fabs(Amat->val[i]);
for (j = Amat->bindx[i]; j < Amat->bindx[i+1]; j++) {
sum += fabs(Amat->val[j]);
}
if (sum > max_sum) max_sum = sum;
}
printf("Largest diagonal = %e, min diag = %e large abs row sum = %e\n",
max_diag, min_diag, max_sum);
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 HAVE_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;
}
// ================================================ ====== ==== ==== == =
// 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);
}
