int main(int argc, char *argv[])
{
int num_PDE_eqns=5, N_levels=3;
/* int nsmooth=1; */
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 */
int *data_org = NULL, *update = NULL, *external = NULL;
int *update_index = NULL, *extern_index = NULL;
int *cpntr = NULL;
int *bindx = NULL, N_update, iii;
double *val = NULL;
double *xxx, *rhs;
AZ_MATRIX *Amat;
AZ_PRECOND *Pmat = NULL;
ML *ml;
FILE *fp;
int ch,i;
struct AZ_SCALING *scaling;
double solve_time, setup_time, start_time;
ML_Aggregate *ag;
int *ivec;
#ifdef VBR_VERSION
ML_Operator *B, *C, *D;
int *vbr_cnptr, *vbr_rnptr, *vbr_indx, *vbr_bindx, *vbr_bnptr, total_blk_rows;
int total_blk_cols, blk_space, nz_space;
double *vbr_val;
struct ML_CSR_MSRdata *csr_data;
#endif
#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);
#else
AZ_set_proc_config(proc_config, AZ_NOT_MPI);
#endif
#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);
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. */
AZ_read_update(&N_update, &update, proc_config, N_grid_pts, num_PDE_eqns,
AZ_linear);
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(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 );
#ifndef VBR_VERSION
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];
#else
total_blk_rows = N_update/num_PDE_eqns;
total_blk_cols = total_blk_rows;
blk_space = total_blk_rows*20;
nz_space = blk_space*num_PDE_eqns*num_PDE_eqns;
vbr_cnptr = (int *) ML_allocate(sizeof(int )*(total_blk_cols+1));
vbr_rnptr = (int *) ML_allocate(sizeof(int )*(total_blk_cols+1));
vbr_bnptr = (int *) ML_allocate(sizeof(int )*(total_blk_cols+2));
vbr_indx = (int *) ML_allocate(sizeof(int )*(blk_space+1));
vbr_bindx = (int *) ML_allocate(sizeof(int )*(blk_space+1));
vbr_val = (double *) ML_allocate(sizeof(double)*(nz_space+1));
for (i = 0; i <= total_blk_cols; i++) vbr_cnptr[i] = num_PDE_eqns;
AZ_msr2vbr(vbr_val, vbr_indx, vbr_rnptr, vbr_cnptr, vbr_bnptr,
vbr_bindx, bindx, val,
total_blk_rows, total_blk_cols, blk_space,
nz_space, -1);
data_org[AZ_N_rows] = data_org[AZ_N_internal] + data_org[AZ_N_border];
data_org[AZ_N_int_blk] = data_org[AZ_N_internal]/num_PDE_eqns;
data_org[AZ_N_bord_blk] = data_org[AZ_N_bord_blk]/num_PDE_eqns;
data_org[AZ_N_ext_blk] = data_org[AZ_N_ext_blk]/num_PDE_eqns;
data_org[AZ_matrix_type] = AZ_VBR_MATRIX;
AZ_set_VBR(Amat, vbr_rnptr, vbr_cnptr, vbr_bnptr, vbr_indx, vbr_bindx,
vbr_val, data_org, 0, NULL, AZ_LOCAL);
Amat->matrix_type = data_org[AZ_matrix_type];
#endif
start_time = AZ_second();
ML_Create(&ml, N_levels);
ML_Set_PrintLevel(3);
/* set up discretization matrix and matrix vector function */
AZ_ML_Set_Amat(ml, N_levels-1, N_update, N_update, Amat, proc_config);
ML_Aggregate_Create( &ag );
ML_Aggregate_Set_Threshold(ag,0.0);
ML_Set_SpectralNormScheme_PowerMethod(ml);
/*
To run SA:
a) set damping factor to 1 and use power method
ML_Aggregate_Set_DampingFactor(ag, 4./3.);
To run NSA:
a) set damping factor to 0
ML_Aggregate_Set_DampingFactor(ag, 0.);
To run NSR
a) set damping factor to 1 and use power method
ML_Aggregate_Set_DampingFactor(ag, 1.);
ag->Restriction_smoothagg_transpose = ML_FALSE;
ag->keep_agg_information=1;
ag->keep_P_tentative=1;
b) hack code so it calls the energy minimizing restriction
line 2973 of ml_agg_genP.c
c) turn on the NSR flag in ml_agg_energy_min.cpp
To run Emin
a) set min_eneryg = 2 and keep_agg_info = 1;
ag->minimizing_energy=2;
ag->keep_agg_information=1;
ag->cheap_minimizing_energy = 0;
ag->block_scaled_SA = 1;
*/
ag->minimizing_energy=2;
ag->keep_agg_information=1;
ag->block_scaled_SA = 1;
ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, num_PDE_eqns, NULL, N_update);
ML_Aggregate_Set_MaxCoarseSize( ag, 20);
/*
ML_Aggregate_Set_RandomOrdering( ag );
ML_Aggregate_Set_DampingFactor(ag, .1);
ag->drop_tol_for_smoothing = 1.0e-3;
ML_Aggregate_Set_Threshold(ag, 1.0e-3);
ML_Aggregate_Set_MaxCoarseSize( ag, 300);
*/
coarsest_level = ML_Gen_MultiLevelHierarchy_UsingAggregation(ml, N_levels-1, ML_DECREASING, ag);
coarsest_level = N_levels - coarsest_level;
if ( proc_config[AZ_node] == 0 )
printf("Coarse level = %d \n", coarsest_level);
/* set up smoothers */
AZ_defaults(options, params);
for (level = N_levels-1; level > coarsest_level; level--) {
/* This is the Aztec domain decomp/ilu smoother that we */
/* usually use for this problem. */
/*
options[AZ_precond] = AZ_dom_decomp;
options[AZ_subdomain_solve] = AZ_ilut;
params[AZ_ilut_fill] = 1.0;
options[AZ_reorder] = 1;
ML_Gen_SmootherAztec(ml, level, options, params,
proc_config, status, AZ_ONLY_PRECONDITIONER,
ML_PRESMOOTHER,NULL);
*/
/* 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);
parasails_thresh /= 4.;
*/
/* This is the symmetric Gauss-Seidel smoothing. 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_SymGaussSeidel(ml,level,ML_PRESMOOTHER, nsmooth,1.);
ML_Gen_Smoother_SymGaussSeidel(ml,level,ML_POSTSMOOTHER,nsmooth,1.);
*/
/* Block Gauss-Seidel with block size equal to #DOF per node. */
/* 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_BlockGaussSeidel(ml,level,ML_PRESMOOTHER,
nsmooth,0.67, num_PDE_eqns);
ML_Gen_Smoother_BlockGaussSeidel(ml,level,ML_POSTSMOOTHER,
nsmooth, 0.67, num_PDE_eqns);
*/
ML_Gen_Smoother_SymBlockGaussSeidel(ml,level,ML_POSTSMOOTHER,
1, 1.0, num_PDE_eqns);
}
ML_Gen_CoarseSolverSuperLU( ml, coarsest_level);
ML_Gen_Solver(ml, ML_MGW, N_levels-1, coarsest_level);
AZ_defaults(options, params);
options[AZ_solver] = AZ_gmres;
options[AZ_scaling] = AZ_none;
options[AZ_precond] = AZ_user_precond;
/*
options[AZ_conv] = AZ_r0;
*/
options[AZ_output] = 1;
options[AZ_max_iter] = 1500;
options[AZ_poly_ord] = 5;
options[AZ_kspace] = 130;
params[AZ_tol] = 1.0e-8;
/*
options[AZ_precond] = AZ_dom_decomp;
options[AZ_subdomain_solve] = AZ_ilut;
params[AZ_ilut_fill] = 2.0;
*/
AZ_set_ML_preconditioner(&Pmat, Amat, ml, options);
setup_time = AZ_second() - start_time;
xxx = (double *) malloc( leng*sizeof(double));
rhs=(double *)malloc(leng*sizeof(double));
for (iii = 0; iii < leng; iii++) xxx[iii] = 0.0;
/* Set rhs */
fp = fopen("AZ_capture_rhs.mat","r");
if (fp == NULL) {
if (proc_config[AZ_node] == 0) printf("taking random vector for rhs\n");
AZ_random_vector(rhs, data_org, proc_config);
AZ_reorder_vec(rhs, data_org, update_index, NULL);
}
else {
fclose(fp);
ivec =(int *)malloc((leng+1)*sizeof(int));
AZ_input_msr_matrix("AZ_capture_rhs.mat", update, &rhs, &ivec,
N_update, proc_config);
free(ivec);
AZ_reorder_vec(rhs, data_org, update_index, NULL);
}
/* Set x */
fp = fopen("AZ_capture_init_guess.mat","r");
if (fp != NULL) {
fclose(fp);
ivec =(int *)malloc((leng+1)*sizeof(int));
AZ_input_msr_matrix("AZ_capture_init_guess.mat",update, &xxx, &ivec,
N_update, proc_config);
free(ivec);
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);
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 main(int argc, char *argv[])
{
int num_PDE_eqns=3, N_levels=3, nsmooth=1;
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 */
int *data_org = NULL, *update = NULL, *external = NULL;
int *update_index = NULL, *extern_index = NULL;
int *cpntr = NULL;
int *bindx = NULL, N_update, iii;
double *val = NULL;
double *xxx, *rhs;
AZ_MATRIX *Amat;
AZ_PRECOND *Pmat = NULL;
ML *ml;
FILE *fp;
int ch,i,j, Nrigid, *garbage;
struct AZ_SCALING *scaling;
double solve_time, setup_time, start_time, *mode, *rigid;
ML_Aggregate *ag;
int nblocks, *blocks;
char filename[80];
double alpha;
int one = 1;
#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);
#else
AZ_set_proc_config(proc_config, AZ_NOT_MPI);
#endif
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. */
AZ_read_update(&N_update, &update, proc_config, N_grid_pts, num_PDE_eqns,
AZ_linear);
AZ_read_msr_matrix(update, &val, &bindx, N_update, proc_config);
AZ_transform(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();
AZ_defaults(options, params);
/*
scaling = AZ_scaling_create();
xxx = (double *) calloc( leng,sizeof(double));
rhs=(double *)calloc(leng,sizeof(double));
options[AZ_scaling] = AZ_sym_diag;
options[AZ_precond] = AZ_none;
options[AZ_max_iter] = 30;
options[AZ_keep_info] = 1;
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, NULL, scaling);
don't forget vector rescaling ...
free(xxx);
free(rhs);
*/
options[AZ_scaling] = AZ_none;
ML_Create(&ml, N_levels);
/* set up discretization matrix and matrix vector function */
AZ_ML_Set_Amat(ml, N_levels-1, N_update, N_update, Amat, proc_config);
ML_Aggregate_Create( &ag );
Nrigid = 0;
if (proc_config[AZ_node] == 0) {
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));
AZ_random_vector(rhs, data_org, proc_config);
for (i = 0; i < Nrigid; i++) {
sprintf(filename,"rigid_body_mode%d",i+1);
AZ_input_msr_matrix(filename, update, &mode, &garbage,
N_update, proc_config);
/*
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]);
*/
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);
printf("alpha1 is %e\n",alpha);
}
alpha = -AZ_gdot(N_update, mode, rhs, proc_config)/AZ_gdot(N_update, mode, mode, proc_config);
printf("alpha2 is %e\n",alpha);
daxpy_(&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_(&N_update, &alpha, &(rigid[j*N_update]), &one, rhs, &one);
printf("alpha4 is %e\n",alpha);
}
for (i = 0; i < Nrigid; i++) {
alpha = -AZ_gdot(N_update, &(rigid[i*N_update]), rhs, proc_config);
printf("alpha is %e\n",alpha);
}
if (Nrigid != 0) {
ML_Aggregate_Set_NullSpace(ag, num_PDE_eqns, Nrigid, rigid, N_update);
/*
free(rigid);
*/
}
coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, N_levels-1, ML_DECREASING, ag);
coarsest_level = N_levels - coarsest_level;
/*
ML_Operator_Print(&(ml->Pmat[N_levels-2]), "Pmat");
exit(1);
*/
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--) {
j = 10;
if (level == N_levels-1) j = 10;
options[AZ_solver] = AZ_cg;
options[AZ_precond]=AZ_sym_GS; options[AZ_subdomain_solve]=AZ_icc;
/*
options[AZ_precond] = AZ_none;
*/
options[AZ_poly_ord] = 5;
ML_Gen_SmootherAztec(ml, level, options, params, proc_config, status,
j, ML_PRESMOOTHER,NULL);
ML_Gen_SmootherAztec(ml, level, options, params, proc_config, status,
j, ML_POSTSMOOTHER,NULL);
/*
ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth,1.0);
ML_Gen_Smoother_SymGaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth,1.0);
*/
/*
nblocks = ML_Aggregate_Get_AggrCount( ag, level );
ML_Aggregate_Get_AggrMap( ag, level, &blocks);
ML_Gen_Smoother_VBlockSymGaussSeidel( ml , level, ML_BOTH, nsmooth, 1.0,
nblocks, blocks);
ML_Gen_Smoother_VBlockSymGaussSeidel( ml , level, ML_POSTSMOOTHER, nsmooth, 1.0,
nblocks, blocks);
*/
/*
ML_Gen_Smoother_VBlockJacobi( ml , level, ML_PRESMOOTHER, nsmooth, .5,
nblocks, blocks);
ML_Gen_Smoother_VBlockJacobi( ml , level, ML_POSTSMOOTHER, nsmooth,.5,
nblocks, blocks);
*/
/*
ML_Gen_Smoother_GaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth);
ML_Gen_Smoother_GaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth);
*/
/*
need to change this when num_pdes is different on different levels
*/
/*
if (level == N_levels-1) {
ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth, 0.5, num_PDE_eqns);
ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth, 0.5, num_PDE_eqns);
}
else {
ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_PRESMOOTHER, nsmooth, 0.5, 2*num_PDE_eqns);
ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_POSTSMOOTHER, nsmooth, 0.5, 2*num_PDE_eqns);
}
*/
/*
*/
/*
ML_Gen_SmootherJacobi(ml , level, ML_PRESMOOTHER, nsmooth, .67);
ML_Gen_SmootherJacobi(ml , level, ML_POSTSMOOTHER, nsmooth, .67 );
*/
}
/*
ML_Gen_CoarseSolverSuperLU( ml, coarsest_level);
*/
/*
ML_Gen_SmootherSymGaussSeidel(ml , coarsest_level, ML_PRESMOOTHER, 2*nsmooth,1.);
*/
/*
ML_Gen_SmootherBlockGaussSeidel(ml , level, ML_PRESMOOTHER, 50*nsmooth, 1.0, 2*num_PDE_eqns);
*/
ML_Gen_Smoother_BlockGaussSeidel(ml , level, ML_PRESMOOTHER, 2*nsmooth, 1.0, num_PDE_eqns);
ML_Gen_Solver(ml, ML_MGV, N_levels-1, coarsest_level);
AZ_defaults(options, params);
options[AZ_solver] = AZ_GMRESR;
options[AZ_scaling] = AZ_none;
options[AZ_precond] = AZ_user_precond;
options[AZ_conv] = AZ_rhs;
options[AZ_output] = 1;
options[AZ_max_iter] = 1500;
options[AZ_poly_ord] = 5;
options[AZ_kspace] = 130;
params[AZ_tol] = 1.0e-8;
AZ_set_ML_preconditioner(&Pmat, Amat, ml, options);
setup_time = AZ_second() - start_time;
xxx = (double *) malloc( leng*sizeof(double));
/* Set rhs */
fp = fopen("AZ_capture_rhs.dat","r");
if (fp == NULL) {
if (proc_config[AZ_node] == 0) printf("taking random vector for rhs\n");
/*
AZ_random_vector(rhs, data_org, proc_config);
AZ_reorder_vec(rhs, data_org, update_index, NULL);
AZ_random_vector(xxx, data_org, proc_config);
AZ_reorder_vec(xxx, data_org, update_index, NULL);
Amat->matvec(xxx, rhs, Amat, proc_config);
*/
}
else {
ch = getc(fp);
if (ch == 'S') {
while ( (ch = getc(fp)) != '\n') ;
}
else ungetc(ch,fp);
for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++)
fscanf(fp,"%lf",&(rhs[i]));
fclose(fp);
}
for (iii = 0; iii < leng; iii++) xxx[iii] = 0.0;
/* Set x */
fp = fopen("AZ_capture_init_guess.dat","r");
if (fp != NULL) {
ch = getc(fp);
if (ch == 'S') {
while ( (ch = getc(fp)) != '\n') ;
}
else ungetc(ch,fp);
for (i = 0; i < data_org[AZ_N_internal]+data_org[AZ_N_border]; i++)
fscanf(fp,"%lf",&(xxx[i]));
fclose(fp);
options[AZ_conv] = AZ_expected_values;
}
/* 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_max_iter] = 40;
*/
AZ_iterate(xxx, rhs, options, params, status, proc_config, Amat, Pmat, scaling);
for (j = 0; j < Nrigid; j++) {
alpha = -AZ_gdot(N_update, xxx, &(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, xxx, &one);
printf("alpha5 is %e\n",alpha);
}
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);
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);
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;
}
void AZ_pbicgstab(double b[], double x[], double weight[], int options[],
double params[],int proc_config[], double status[], AZ_MATRIX *Amat,
AZ_PRECOND *precond, struct AZ_CONVERGE_STRUCT *convergence_info)
/*******************************************************************************
Vand der Vorst's (1990) variation of the Bi-Conjugate Gradient algorthm
(Sonneveld (1984,1989)) to solve the nonsymmetric matrix problem Ax = b.
Author: John N. Shadid, SNL, 1421
=======
Return code: void
============
Parameter list:
===============
b: Right hand side of linear system.
x: On input, contains the initial guess. On output contains the
solution to the linear system.
weight: Vector of weights for convergence norm #4.
options: Determines specific solution method and other parameters.
params: Drop tolerance and convergence tolerance info.
proc_config: Machine configuration. proc_config[AZ_node] is the node
number. proc_config[AZ_N_procs] is the number of processors.
status: On output, indicates termination status:
0: terminated normally.
-1: maximum number of iterations taken without achieving
convergence.
-2: Breakdown. The algorithm can not proceed due to
numerical difficulties (usually a divide by zero).
-3: Internal residual differs from the computed residual due
to a significant loss of precision.
Amat: Structure used to represent the matrix (see file az_aztec.h
and Aztec User's Guide).
precond: Structure used to represent the preconditionner
(see file az_aztec.h and Aztec User's Guide).
*******************************************************************************/
{
/* local variables */
register int i;
int N, NN, one = 1, iter=1, r_avail = AZ_TRUE, j;
int precond_flag, print_freq, proc;
int brkdown_will_occur = AZ_FALSE;
double alpha = 1.0, beta, true_scaled_r=0.0;
double *v, *r, *rtilda, *p, *phat, *s, *shat;
double omega = 1.0, dot_vec[2], tmp[2], init_time = 0.0;
double rhonm1 = 1.0, rhon, sigma, brkdown_tol = DBL_EPSILON;
double scaled_r_norm= -1.0, actual_residual = -1.0, rec_residual= -1.0;
double dtemp;
int *data_org, str_leng, first_time = AZ_TRUE;
char label[64],suffix[32], prefix[64];
/**************************** execution begins ******************************/
sprintf(suffix," in cgstab%d",options[AZ_recursion_level]);
/* set string that will be used */
/* for manage_memory label */
/* set prefix for printing */
str_leng = 0;
for (i = 0; i < 16; i++) prefix[str_leng++] = ' ';
for (i = 0 ; i < options[AZ_recursion_level]; i++ ) {
prefix[str_leng++] = ' '; prefix[str_leng++] = ' '; prefix[str_leng++] = ' ';
prefix[str_leng++] = ' '; prefix[str_leng++] = ' ';
}
prefix[str_leng] = '\0';
data_org = Amat->data_org;
/* pull needed values out of parameter arrays */
N = data_org[AZ_N_internal] + data_org[AZ_N_border];
precond_flag = options[AZ_precond];
proc = proc_config[AZ_node];
print_freq = options[AZ_print_freq];
/* Initialize some values in convergence info struct */
convergence_info->print_info = print_freq;
convergence_info->iteration = 0;
convergence_info->sol_updated = 1; /* BiCGStab always updates solution */
convergence_info->epsilon = params[AZ_tol]; /* Test against this */
/* allocate memory for required vectors */
NN = N + data_org[AZ_N_external];
if (NN == 0) NN++; /* make sure everybody allocates something*/
NN = NN + (NN%2); /* make sure things are aligned for the */
/* assembly coded matvec() on the Intel. */
sprintf(label,"phat%s",suffix);
phat = (double *) AZ_manage_memory(7*NN*sizeof(double), AZ_ALLOC,
AZ_SYS+az_iterate_id, label,&j);
p = &(phat[1*NN]);
shat = &(phat[2*NN]); /* NOTE: phat and shat must be aligned */
/* so that the assembly dgemv */
/* works on the paragon. */
s = &(phat[3*NN]);
r = &(phat[4*NN]);
rtilda = &(phat[5*NN]);
v = &(phat[6*NN]);
AZ_compute_residual(b, x, r, proc_config, Amat);
/* v, p <- 0 */
for (i = 0; i < N; i++) v[i] = p[i] = 0.0;
/* set rtilda */
if (options[AZ_aux_vec] == AZ_resid)
DCOPY_F77(&N, r, &one, rtilda, &one);
else
AZ_random_vector(rtilda, data_org, proc_config);
/*
* Compute a few global scalars:
* 1) ||r|| corresponding to options[AZ_conv]
* 2) scaled ||r|| corresponding to options[AZ_conv]
* 3) rho = <rtilda, r>
*/
AZ_compute_global_scalars(Amat, x, b, r,
weight, &rec_residual, &scaled_r_norm, options,
data_org, proc_config,&r_avail,r,rtilda, &rhon,
convergence_info);
true_scaled_r = scaled_r_norm;
if ((options[AZ_output] != AZ_none) &&
(options[AZ_output] != AZ_last) &&
(options[AZ_output] != AZ_warnings) &&
(options[AZ_output] != AZ_summary) &&
(options[AZ_conv]!=AZTECOO_conv_test) && (proc == 0))
(void) AZ_printf_out("%siter: 0 residual = %e\n",prefix,scaled_r_norm);
for (iter = 1; iter <= options[AZ_max_iter] && !(convergence_info->converged)
&& !(convergence_info->isnan); iter++) {
if (brkdown_will_occur) {
AZ_scale_true_residual( x, b, v,
weight, &actual_residual, &true_scaled_r, options,
data_org, proc_config, Amat, convergence_info);
AZ_terminate_status_print(AZ_breakdown, iter, status, rec_residual,
params, true_scaled_r, actual_residual, options,
proc_config);
return;
}
beta = (rhon/rhonm1) * (alpha/omega);
if (fabs(rhon) < brkdown_tol) { /* possible problem */
if (AZ_breakdown_f(N, r, rtilda, rhon, proc_config))
brkdown_will_occur = AZ_TRUE;
else
brkdown_tol = 0.1 * fabs(rhon);
}
rhonm1 = rhon;
/* p = r + beta*(p - omega*v) */
/* phat = M^-1 p */
/* v = A phat */
dtemp = beta * omega;
for (i = 0; i < N; i++) p[i] = r[i] + beta * p[i] - dtemp * v[i];
DCOPY_F77(&N, p, &one, phat, &one);
if (iter==1) init_time = AZ_second();
if (precond_flag)
precond->prec_function(phat,options,proc_config,params,Amat,precond);
if (iter==1) status[AZ_first_precond] = AZ_second() - init_time;
Amat->matvec(phat, v, Amat, proc_config);
sigma = AZ_gdot(N, rtilda, v, proc_config);
if (fabs(sigma) < brkdown_tol) { /* possible problem */
if (AZ_breakdown_f(N, rtilda, v, sigma, proc_config)) {
/* break down */
AZ_scale_true_residual( x, b, v,
weight, &actual_residual, &true_scaled_r,
options, data_org,proc_config, Amat,
convergence_info);
AZ_terminate_status_print(AZ_breakdown, iter, status, rec_residual,
params, true_scaled_r, actual_residual,
options, proc_config);
return;
}
else brkdown_tol = 0.1 * fabs(sigma);
}
alpha = rhon / sigma;
/* s = r - alpha*v */
/* shat = M^-1 s */
/* r = A shat (r is a tmp here for t ) */
for (i = 0; i < N; i++) s[i] = r[i] - alpha * v[i];
DCOPY_F77(&N, s, &one, shat, &one);
if (precond_flag)
precond->prec_function(shat,options,proc_config,params,Amat,precond);
Amat->matvec(shat, r, Amat, proc_config);
/* omega = (t,s)/(t,t) with r = t */
dot_vec[0] = DDOT_F77(&N, r, &one, s, &one);
dot_vec[1] = DDOT_F77(&N, r, &one, r, &one);
AZ_gdot_vec(2, dot_vec, tmp, proc_config);
if (fabs(dot_vec[1]) < DBL_MIN) {
omega = 0.0;
brkdown_will_occur = AZ_TRUE;
}
else omega = dot_vec[0] / dot_vec[1];
/* x = x + alpha*phat + omega*shat */
/* r = s - omega*r */
DAXPY_F77(&N, &alpha, phat, &one, x, &one);
DAXPY_F77(&N, &omega, shat, &one, x, &one);
for (i = 0; i < N; i++) r[i] = s[i] - omega * r[i];
/*
* Compute a few global scalars:
* 1) ||r|| corresponding to options[AZ_conv]
* 2) scaled ||r|| corresponding to options[AZ_conv]
* 3) rho = <rtilda, r>
*/
AZ_compute_global_scalars(Amat, x, b, r,
weight, &rec_residual, &scaled_r_norm, options,
data_org, proc_config, &r_avail, r, rtilda, &rhon,
convergence_info);
if ( (iter%print_freq == 0) && proc == 0)
(void) AZ_printf_out("%siter: %4d residual = %e\n",prefix,iter,
scaled_r_norm);
/* convergence tests */
if (options[AZ_check_update_size] & convergence_info->converged) {
dtemp = alpha/omega;
DAXPY_F77(&N, &dtemp, phat, &one, shat, &one);
convergence_info->converged = AZ_compare_update_vs_soln(N, -1.,omega, shat, x,
params[AZ_update_reduction],
options[AZ_output], proc_config, &first_time);
}
if (convergence_info->converged) {
AZ_scale_true_residual(x, b, v,
weight, &actual_residual, &true_scaled_r, options,
data_org, proc_config, Amat, convergence_info);
/*
* Note: epsilon and params[AZ_tol] may not be equal due to a previous
* call to AZ_get_new_eps().
*/
if (!(convergence_info->converged) && options[AZ_conv]!=AZTECOO_conv_test) {
if (AZ_get_new_eps(&convergence_info->epsilon, scaled_r_norm, true_scaled_r,
options, proc_config) == AZ_QUIT) {
/*
* Computed residual has converged, actual residual has not converged,
* AZ_get_new_eps() has decided that it is time to quit.
*/
AZ_terminate_status_print(AZ_loss, iter, status, rec_residual, params,
true_scaled_r, actual_residual, options,
proc_config);
return;
}
}
}
}
iter--;
if ( (iter%print_freq != 0) && (proc == 0) && (options[AZ_output] != AZ_none)
&& (options[AZ_output] != AZ_warnings) &&
(options[AZ_conv]!=AZTECOO_conv_test))
(void) AZ_printf_out("%siter: %4d residual = %e\n", prefix,iter,
scaled_r_norm);
/* check if we exceeded maximum number of iterations */
if (convergence_info->converged) {
i = AZ_normal;
scaled_r_norm = true_scaled_r;
}
else if (convergence_info->isnan) i = AZ_breakdown;
else
i = AZ_maxits;
AZ_terminate_status_print(i, iter, status, rec_residual, params,
scaled_r_norm, actual_residual, options,
proc_config);
} /* bicgstab */
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;
}
void AZ_pqmrs(double b[], double x[], double weight[], int options[],
double params[], int proc_config[], double status[], AZ_MATRIX *Amat,
AZ_PRECOND *precond, struct AZ_CONVERGE_STRUCT *convergence_info)
/*******************************************************************************
Freund's transpose free QMR routine to solve the nonsymmetric matrix problem
Ax = b. NOTE: this routine differs from Freund's paper in that we compute
ubar (= M^-1 u ) and qbar (= M^-1 q) instead of u and q defined in Freund's
paper.
IMPORTANT NOTE: While an estimate of the 2-norm of the qmr residual is
available (see comment below), the actual qmr residual is not normally
computed as part of the qmr algorithm. Thus, if the user uses a convergence
condition (see AZ_compute_global_scalars()) that is based on the 2-norm of the
residual there is no need to compute the residual (i.e. r_avail = AZ_FALSE).
However, if another norm of r is requested, AZ_compute_global_scalars() will
set r_avail = AZ_TRUE and the algorithm will compute the residual.
Author: John N. Shadid, SNL, 1421
=======
Return code: void
============
Parameter list:
===============
b: Right hand side of linear system.
x: On input, contains the initial guess. On output contains the
solution to the linear system.
weight: Vector of weights for convergence norm #4.
options: Determines specific solution method and other parameters.
params: Drop tolerance and convergence tolerance info.
proc_config: Machine configuration. proc_config[AZ_node] is the node
number. proc_config[AZ_N_procs] is the number of processors.
status: On output, indicates termination status:
0: terminated normally.
-1: maximum number of iterations taken without achieving
convergence.
-2: Breakdown. The algorithm can not proceed due to
numerical difficulties (usually a divide by zero).
-3: Internal residual differs from the computed residual due
to a significant loss of precision.
Amat: Structure used to represent the matrix (see az_aztec.h
and Aztec User's Guide).
Oprecond: Structure used to represent the preconditionner
(see file az_aztec.h and Aztec User's Guide).
*******************************************************************************/
{
/* local variables */
register int i;
int N, NN, converged, one = 1, iter= 1,r_avail = AZ_FALSE, j;
int precond_flag, print_freq, proc;
int brkdown_will_occur = AZ_FALSE;
double alpha, beta = 0.0, true_scaled_r=0.0;
double *ubar, *v, *r_cgs, *rtilda, *Aubar, *qbar, *Aqbar, *d, *Ad = NULL;
double rhonm1, rhon, est_residual, actual_residual = -1.0;
double scaled_r_norm, sigma, epsilon, brkdown_tol = DBL_EPSILON;
double omega, c, norm_r_n_cgs, norm_r_nm1_cgs;
double tau_m, nu_m, eta_m, init_time = 0.0;
double tau_mm1, nu_mm1 = 0.0, eta_mm1 = 0.0, doubleone = 1.0;
register double dtemp;
double W_norm = 0.0;
int offset = 0;
int *data_org, str_leng, first_time = AZ_TRUE;
char label[64],suffix[32], prefix[64];
/**************************** execution begins ******************************/
sprintf(suffix," in qmrcgs%d",options[AZ_recursion_level]);
/* set string that will be used */
/* for manage_memory label */
/* set prefix for printing */
str_leng = 0;
for (i = 0; i < 16; i++) prefix[str_leng++] = ' ';
for (i = 0 ; i < options[AZ_recursion_level]; i++ ) {
prefix[str_leng++] = ' '; prefix[str_leng++] = ' '; prefix[str_leng++] = ' ';
prefix[str_leng++] = ' '; prefix[str_leng++] = ' ';
}
prefix[str_leng] = '\0';
data_org = Amat->data_org;
/* pull needed values out of parameter arrays */
N = data_org[AZ_N_internal] + data_org[AZ_N_border];
precond_flag = options[AZ_precond];
epsilon = params[AZ_tol];
proc = proc_config[AZ_node];
print_freq = options[AZ_print_freq];
/* allocate memory for required vectors */
NN = N + data_org[AZ_N_external];
if (NN == 0) NN++; /* make sure everyone allocates something */
NN = NN + (NN%2);
/* make sure things are aligned on double words for paragon */
sprintf(label,"ubar%s",suffix);
ubar = (double *) AZ_manage_memory(8*NN*sizeof(double),
AZ_ALLOC,AZ_SYS,label,&j);
v = &(ubar[1*NN]);
Aubar = &(ubar[2*NN]);
d = &(ubar[3*NN]);
qbar = &(ubar[4*NN]);
rtilda = &(ubar[5*NN]);
Aqbar = &(ubar[6*NN]);
r_cgs = &(ubar[7*NN]);
AZ_compute_residual(b, x, r_cgs, proc_config, Amat);
/* d, qbar, Aqbar, v = 0 */
for (i = 0; i < N; i++)
d[i] = qbar[i] = Aqbar[i] = v[i] = 0.0;
/* set rtilda */
if (options[AZ_aux_vec] == AZ_resid) dcopy_(&N, r_cgs, &one, rtilda, &one);
else AZ_random_vector(rtilda, data_org, proc_config);
/*
* Compute a few global scalars:
* 1) ||r_cgs|| corresponding to options[AZ_conv]
* 2) scaled ||r_cgs|| corresponding to options[AZ_conv]
* 3) rhon = <rtilda, r_cgs>
* Note: step 1) is performed if r_avail = AZ_TRUE on entry or
* AZ_FIRST_TIME is passed in. Otherwise, ||r_cgs|| is taken as
* est_residual.
*/
AZ_compute_global_scalars(Amat, x, b, r_cgs,
weight, &est_residual, &scaled_r_norm, options,
data_org, proc_config, &r_avail, r_cgs, rtilda,
&rhon, convergence_info);
true_scaled_r = scaled_r_norm;
if ((options[AZ_output] != AZ_none) && (options[AZ_output] != AZ_last) &&
(options[AZ_output] != AZ_warnings) && (proc == 0))
(void) fprintf(stdout, "%siter: 0 residual = %e\n",prefix,scaled_r_norm);
norm_r_nm1_cgs = est_residual;
tau_mm1 = norm_r_nm1_cgs;
rhonm1 = rhon;
/* Set up aux-vector if we need to compute the qmr residual */
if (r_avail) {
sprintf(label,"Ad%s",suffix);
Ad = (double *) AZ_manage_memory(NN*sizeof(double),AZ_ALLOC,
AZ_SYS, label, &j);
for (i = 0; i < N; i++) Ad[i] = 0.0;
}
converged = scaled_r_norm < epsilon;
for (iter = 1; iter <= options[AZ_max_iter] && !converged; iter++) {
if (fabs(rhon) < brkdown_tol) { /* possible breakdown problem */
if (AZ_breakdown_f(N, r_cgs, rtilda, rhon, proc_config))
brkdown_will_occur = AZ_TRUE;
else
brkdown_tol = 0.1 * fabs(rhon);
}
/* ubar = M^-1 r_cgs + beta*qbar */
/* Aubar = A ubar */
/* v = A ubar + beta ( A qbar + beta pnm1 ) */
/* = Aubar + beta ( Aqbar + beta v) */
dcopy_(&N, r_cgs, &one, ubar, &one);
if (iter==1) init_time = AZ_second();
if (precond_flag)
precond->prec_function(ubar,options,proc_config,params,Amat,precond);
if (iter==1) status[AZ_first_precond] = AZ_second() - init_time;
for (i = 0; i < N; i++) ubar[i] = ubar[i] + beta * qbar[i];
Amat->matvec(ubar, Aubar, Amat, proc_config);
daxpy_(&N, &beta, v, &one, Aqbar, &one);
for (i = 0; i < N; i++) v[i] = Aubar[i] + beta * Aqbar[i];
sigma = AZ_gdot(N, rtilda, v, proc_config);
if (fabs(sigma) < brkdown_tol) { /* possible problem */
if (AZ_breakdown_f(N, rtilda, v, sigma, proc_config)) {
/* break down */
AZ_scale_true_residual(x, b, v,
weight, &actual_residual, &true_scaled_r,
options, data_org, proc_config, Amat,
convergence_info);
AZ_terminate_status_print(AZ_breakdown, iter, status, est_residual,
params, true_scaled_r, actual_residual,
options, proc_config);
return;
}
else brkdown_tol = 0.1 * fabs(sigma);
}
alpha = rhon / sigma;
/* qbar = ubar - alpha* M^-1 v */
/* Aqbar = A qbar */
/* r_cgs = r_cgs - alpha (A ubar + A qbar) */
/* = r_cgs - alpha (Aubar + Aqbar) */
dcopy_(&N, v, &one, qbar, &one);
if (precond_flag)
precond->prec_function(qbar,options,proc_config,params,Amat,precond);
for (i = 0; i < N; i++) qbar[i] = ubar[i] - alpha * qbar[i];
Amat->matvec(qbar, Aqbar, Amat, proc_config);
for (i = 0; i < N; i++) r_cgs[i] = r_cgs[i] - alpha*(Aubar[i] + Aqbar[i]);
/* QMRS scaling and iterates weights 5.11 */
norm_r_n_cgs = sqrt(AZ_gdot(N, r_cgs, r_cgs, proc_config));
/* m is odd in Freund's paper */
omega = sqrt(norm_r_nm1_cgs * norm_r_n_cgs);
nu_m = omega / tau_mm1;
c = 1.0 / sqrt(1.0 + nu_m * nu_m);
tau_m = tau_mm1 * nu_m * c;
eta_m = c * c * alpha;
if (brkdown_will_occur) {
AZ_scale_true_residual(x, b, v,
weight, &actual_residual, &true_scaled_r, options,
data_org, proc_config, Amat, convergence_info);
AZ_terminate_status_print(AZ_breakdown, iter, status, est_residual,
params, true_scaled_r, actual_residual, options,
proc_config);
return;
}
dtemp = nu_mm1 *nu_mm1 * eta_mm1 / alpha;
for (i = 0; i < N; i++) d[i] = ubar[i] + dtemp * d[i];
daxpy_(&N, &eta_m, d, &one, x, &one); /* x = x - eta_m d */
if (r_avail) {
for (i = 0; i < N; i++) Ad[i] = Aubar[i] + dtemp * Ad[i];
}
/* save some values */
eta_mm1 = eta_m; tau_mm1 = tau_m;
nu_mm1 = nu_m; norm_r_nm1_cgs = norm_r_n_cgs;
/* m is even in Freund's paper */
omega = norm_r_n_cgs;
if (tau_mm1 == 0.0) nu_m = 0.0;
else nu_m = omega / tau_mm1;
c = 1.0 / sqrt(1.0 + nu_m * nu_m);
tau_m = tau_mm1 * nu_m * c;
if (options[AZ_check_update_size]) {
eta_m = eta_m/(c*c*alpha);
for (i = 0; i < N; i++) ubar[i] = eta_m*d[i];
}
eta_m = c * c * alpha;
dtemp = nu_mm1 * nu_mm1 * eta_mm1 / alpha;
for (i = 0; i < N; i++) d[i] = qbar[i] + dtemp * d[i];
daxpy_(&N, &eta_m, d, &one, x, &one); /* x = x - eta_m d */
if (r_avail) {
for (i = 0; i < N; i++) Ad[i] = Aqbar[i] + dtemp * Ad[i];
}
/* save some values */
eta_mm1 = eta_m; tau_mm1 = tau_m;
nu_mm1 = nu_m; norm_r_nm1_cgs = norm_r_n_cgs;
rhonm1 = rhon;
if (r_avail) {
for (i = 0; i < N; i++) Aubar[i] = r_cgs[i] - (eta_m - alpha) * Ad[i];
/* Note: Aubar temporarily holds qmr residual */
}
else {
/*
* We want to estimate the 2-norm of the qmr residual. Freund gives the
* bound ||r|| <= tau_m * sqrt(2*iter+1). We use this bound until we get
* close to the solution. At that point we compute the real residual norm
* and use this to estimate the norm of ||W|| in Freund's paper.
*/
dtemp = sqrt((double) (2 * iter + 1));
if ((scaled_r_norm < epsilon * dtemp) && !offset) {
AZ_scale_true_residual(x, b,
Aubar, weight, &actual_residual, &true_scaled_r,
options, data_org, proc_config, Amat,
convergence_info);
if (tau_m != 0.0) W_norm = actual_residual / tau_m;
if (W_norm < 1.0) W_norm = 1.0;
offset = 2 * iter + 1;
est_residual = actual_residual;
}
else est_residual = sqrt((double)(2 * iter + 1 - offset) +
W_norm * W_norm) * tau_m;
}
/*
* Compute a few global scalars:
* 1) ||r|| corresponding to options[AZ_conv]
* 2) scaled ||r|| corresponding to options[AZ_conv]
* 3) rhon = <rtilda, r_cgs>
* Note: step 1) is performed if r_avail = AZ_TRUE or AZ_FIRST_TIME
* is passed in. Otherwise, ||r|| is taken as est_residual.
*/
AZ_compute_global_scalars(Amat, x, b,
Aubar, weight, &est_residual, &scaled_r_norm,
options, data_org, proc_config, &r_avail, rtilda,
r_cgs, &rhon, convergence_info);
if ( (iter%print_freq == 0) && proc == 0 )
(void) fprintf(stdout, "%siter: %4d residual = %e\n",prefix,iter,
scaled_r_norm);
/* convergence tests */
converged = scaled_r_norm < epsilon;
if (options[AZ_check_update_size] & converged) {
daxpy_(&N, &doubleone , d, &one, ubar, &one);
converged = AZ_compare_update_vs_soln(N, -1.,eta_m, ubar, x,
params[AZ_update_reduction],
options[AZ_output], proc_config, &first_time);
}
if (converged) {
AZ_scale_true_residual(x, b, Aubar,
weight, &actual_residual, &true_scaled_r, options,
data_org, proc_config, Amat,convergence_info);
converged = true_scaled_r < params[AZ_tol];
/*
* Note: epsilon and params[AZ_tol] may not be equal due to a previous
* call to AZ_get_new_eps().
*/
if (!converged &&
(AZ_get_new_eps(&epsilon, scaled_r_norm, true_scaled_r,
proc_config) == AZ_QUIT)) {
/*
* Computed residual has converged, actual residual has not converged,
* AZ_get_new_eps() has decided that it is time to quit.
*/
AZ_terminate_status_print(AZ_loss, iter, status, est_residual, params,
true_scaled_r, actual_residual, options,
proc_config);
return;
}
}
beta = rhon / rhonm1;
}
iter--;
if ( (iter%print_freq != 0) && (proc == 0) && (options[AZ_output] != AZ_none)
&& (options[AZ_output] != AZ_warnings))
(void) fprintf(stdout, "%siter: %4d residual = %e\n",prefix,iter,
scaled_r_norm);
/* check if we exceeded maximum number of iterations */
if (converged) {
i = AZ_normal;
scaled_r_norm = true_scaled_r;
}
else
i = AZ_maxits;
AZ_terminate_status_print(i, iter, status, est_residual, params,
scaled_r_norm, actual_residual, options,
proc_config);
} /* pqmrs */
