//==============================================================================
AztecDVBR_Matrix::AztecDVBR_Matrix(fei::SharedPtr<Aztec_BlockMap> map)
: amap_(map),
Amat_(NULL),
N_update_(map->getNumLocalBlocks()),
external_(NULL),
extern_index_(NULL),
update_index_(NULL),
data_org_(NULL),
orderingUpdate_(NULL),
isLoaded_(false),
isAllocated_(false),
localNNZ_(0),
nnzPerRow_(NULL),
numRemoteBlocks_(0),
remoteInds_(NULL),
remoteBlockSizes_(NULL)
{
nnzPerRow_ = new int[N_update_];
for(int i=0; i<N_update_; i++) {
nnzPerRow_[i] = 0;
}
Amat_ = AZ_matrix_create(N_update_);
Amat_->matrix_type = AZ_VBR_MATRIX;
Amat_->matvec =
(void(*)(double*,double*,AZ_MATRIX_STRUCT*,int*))AZ_VBR_matvec_mult;
//now we can allocate and fill the rpntr array.
Amat_->rpntr = NULL;
calcRpntr();
}
void AZ_ifpack_solve(double x[], double b[], int options[], double params[],
int indx[], int bindx[], int rpntr[], int cpntr[], int bpntr[],
double val[], int data_org[], double status[], int proc_config[])
{
AZ_MATRIX *Amat;
Amat = AZ_matrix_create(data_org[AZ_N_internal]+data_org[AZ_N_border]);
options[AZ_output] = 1;
if (data_org[AZ_matrix_type] == AZ_MSR_MATRIX)
AZ_set_MSR(Amat, bindx, val, data_org, 0, NULL, AZ_LOCAL);
else if (data_org[AZ_matrix_type] == AZ_VBR_MATRIX)
AZ_set_VBR(Amat, rpntr, cpntr, bpntr, indx, bindx, val,
data_org, 0, NULL, AZ_LOCAL);
else {
fprintf(stderr,"Unknown matrix type (%d)\n",data_org[AZ_matrix_type]);
fprintf(stderr,"Matrix-free is now available via AZ_iterate()\n");
exit(1);
}
AZ_ifpack_iterate(x, b, options, params, status, proc_config, Amat);
AZ_matrix_destroy(&Amat);
}
AZ_MATRIX *user_Kn_build(struct user_partition *Node_Partition)
{
int *Kn_bindx;
double *Kn_val;
int proc_config[AZ_PROC_SIZE];
AZ_MATRIX *Kn_mat;
int *reordered_glob_nodes = NULL, *cpntr = NULL, *Kn_data_org = NULL;
int i, ii, jj, nx, gid, Nlocal_nodes, nz_ptr;
Nlocal_nodes = Node_Partition->Nlocal;
Kn_bindx = (int *) malloc((27*Nlocal_nodes+5)*sizeof(int));
Kn_val = (double *) malloc((27*Nlocal_nodes+5)*sizeof(double));
Kn_bindx[0] = Nlocal_nodes+1;
nx = (int) sqrt( ((double) Node_Partition->Nglobal) + .00001);
for (i = 0; i < Nlocal_nodes; i++) {
gid = (Node_Partition->my_global_ids)[i];
nz_ptr = Kn_bindx[i];
ii = gid%nx;
jj = (gid - ii)/nx;
if (ii != nx-1) { Kn_bindx[nz_ptr] = gid+ 1; Kn_val[nz_ptr++] = -1.;}
if (jj != nx-1) { Kn_bindx[nz_ptr] = gid+nx; Kn_val[nz_ptr++] = -1.;}
if (jj != 0) { Kn_bindx[nz_ptr] = gid-nx; Kn_val[nz_ptr++] = -1.;}
if (ii != 0) { Kn_bindx[nz_ptr] = gid- 1; Kn_val[nz_ptr++] = -1.;}
if ((ii != nx-1) && (jj != 0))
{Kn_bindx[nz_ptr] = gid-nx+1; Kn_val[nz_ptr++] = -1.;}
if ((ii != nx-1) && (jj != nx-1))
{Kn_bindx[nz_ptr] = gid+nx+1; Kn_val[nz_ptr++] = -1.;}
if ((ii != 0) && (jj != nx-1))
{Kn_bindx[nz_ptr] = gid+nx-1; Kn_val[nz_ptr++] = -1.;}
if ((ii != 0) && (jj != 0))
{Kn_bindx[nz_ptr] = gid-nx-1; Kn_val[nz_ptr++] = -1.;}
Kn_val[i] = (double) (nz_ptr - Kn_bindx[i]);
Kn_bindx[i+1] = nz_ptr;
}
AZ_set_proc_config(proc_config, COMMUNICATOR);
AZ_transform_norowreordering(proc_config,&(Node_Partition->needed_external_ids),
Kn_bindx, Kn_val, Node_Partition->my_global_ids,
&reordered_glob_nodes, &reordered_node_externs,
&Kn_data_org, Nlocal_nodes, 0, 0, 0,
&cpntr, AZ_MSR_MATRIX);
Node_Partition->Nghost = Kn_data_org[AZ_N_external];
AZ_free(reordered_glob_nodes);
/* Convert old style Aztec matrix to newer style Aztec matrix */
Kn_mat = AZ_matrix_create( Nlocal_nodes );
AZ_set_MSR(Kn_mat, Kn_bindx, Kn_val, Kn_data_org, 0, NULL, AZ_LOCAL);
return(Kn_mat);
}
AZ_MATRIX *user_Kn_build(struct user_partition *Node_Partition)
{
int *Kn_bindx;
double *Kn_val;
int proc_config[AZ_PROC_SIZE];
AZ_MATRIX *Kn_mat;
int *reordered_glob_nodes = NULL, *cpntr = NULL, *Kn_data_org = NULL;
int i, ii, jj, nx, global_id, Nlocal_nodes, nz_ptr;
Nlocal_nodes = Node_Partition->Nlocal;
Kn_bindx = (int *) malloc((27*Nlocal_nodes+5)*sizeof(int));
Kn_val = (double *) malloc((27*Nlocal_nodes+5)*sizeof(double));
Kn_bindx[0] = Nlocal_nodes+1;
nx = (int) sqrt( ((double) Node_Partition->Nglobal) + .00001);
#ifdef periodic
nx = (int) sqrt( ((double) Node_Partition->Nglobal - 2) + .00001);
Nlocal_nodes -= 2;
#endif
for (i = 0; i < Nlocal_nodes; i++) {
global_id = (Node_Partition->my_global_ids)[i];
Kn_bindx[i+1] = Kn_bindx[i] + 8;
Kn_val[i] = 8.;
inv2dnodeindex(global_id, &ii, &jj, nx);
nz_ptr = Kn_bindx[i];
Kn_bindx[nz_ptr] = southeast2d(ii,jj,nx); Kn_val[nz_ptr++] = -1.;
Kn_bindx[nz_ptr] = northwest2d(ii,jj,nx); Kn_val[nz_ptr++] = -1.;
Kn_bindx[nz_ptr] = northeast2d(ii,jj,nx); Kn_val[nz_ptr++] = -.00000001;
if (ii == 0) ii = nx-1;
else ii--;
Kn_bindx[nz_ptr] = northwest2d(ii,jj,nx); Kn_val[nz_ptr++] = -.00000001;
if (jj == 0) jj = nx-1;
else jj--;
Kn_bindx[nz_ptr] = southeast2d(ii,jj,nx); Kn_val[nz_ptr++] = -1.;
Kn_bindx[nz_ptr] = northwest2d(ii,jj,nx); Kn_val[nz_ptr++] = -1.;
Kn_bindx[nz_ptr] = southwest2d(ii,jj,nx); Kn_val[nz_ptr++] = -.00000001;
if (ii == nx-1) ii = 0;
else ii++;
Kn_bindx[nz_ptr] = southeast2d(ii,jj,nx); Kn_val[nz_ptr++] = -.00000001;
}
#ifdef periodic
i = Nlocal_nodes;
Kn_bindx[i+1] = Kn_bindx[i];
Kn_val[i] = 1.;
i++;
Kn_bindx[i+1] = Kn_bindx[i];
Kn_val[i] = 1.;
Nlocal_nodes += 2;
#endif
/* Transform the global Aztec matrix into a local Aztec matrix. That is, */
/* replace global column indices by local indices and set up communication */
/* data structure 'Ke_data_org' that will be used for matvec's. */
AZ_set_proc_config(proc_config, COMMUNICATOR);
AZ_transform_norowreordering(proc_config,&(Node_Partition->needed_external_ids),
Kn_bindx, Kn_val, Node_Partition->my_global_ids,
&reordered_glob_nodes, &reordered_node_externs,
&Kn_data_org, Nlocal_nodes, 0, 0, 0,
&cpntr, AZ_MSR_MATRIX);
/* Convert old style Aztec matrix to newer style Aztec matrix */
Kn_mat = AZ_matrix_create( Nlocal_nodes );
AZ_set_MSR(Kn_mat, Kn_bindx, Kn_val, Kn_data_org, 0, NULL, AZ_LOCAL);
return(Kn_mat);
}
AZ_MATRIX *user_Ke_build(struct user_partition *Edge_Partition)
{
double dcenter, doffdiag, sigma = .0001;
int ii,jj, horv, i, nx, global_id, nz_ptr, Nlocal_edges;
/* Aztec matrix and temp variables */
int *Ke_bindx, *Ke_data_org = NULL;
double *Ke_val;
AZ_MATRIX *Ke_mat;
int proc_config[AZ_PROC_SIZE], *cpntr = NULL;
int *reordered_glob_edges = NULL, *reordered_edge_externs = NULL;
Nlocal_edges = Edge_Partition->Nlocal;
nx = (int) sqrt( ((double) Edge_Partition->Nglobal/2) + .00001);
Ke_bindx = (int *) malloc((7*Nlocal_edges+1)*sizeof(int));
Ke_val = (double *) malloc((7*Nlocal_edges+1)*sizeof(double));
Ke_bindx[0] = Nlocal_edges+1;
dcenter = 2 + 2.*sigma/((double) ( 3 * nx * nx));
doffdiag = -1 + sigma/((double) ( 6 * nx * nx));
/* Create a DMSR matrix with global column indices */
for (i = 0; i < Nlocal_edges; i++) {
global_id = (Edge_Partition->my_global_ids)[i];
Ke_val[i] = dcenter;
Ke_bindx[i+1] = Ke_bindx[i] + 6;
inv2dindex(global_id, &ii, &jj, nx, &horv);
nz_ptr = Ke_bindx[global_id];
if (horv == HORIZONTAL) {
Ke_bindx[nz_ptr] = north2d(ii,jj,nx); Ke_val[nz_ptr++] = doffdiag;
Ke_bindx[nz_ptr] = west2d(ii,jj,nx); Ke_val[nz_ptr++] = 1.;
Ke_bindx[nz_ptr] = east2d(ii,jj,nx); Ke_val[nz_ptr++] = -1.;
if (jj == 0) jj = nx-1;
else jj--;
Ke_bindx[nz_ptr] = west2d(ii,jj,nx); Ke_val[nz_ptr++] = -1.;
Ke_bindx[nz_ptr] = south2d(ii,jj,nx); Ke_val[nz_ptr++] = doffdiag;
Ke_bindx[nz_ptr] = east2d(ii,jj,nx); Ke_val[nz_ptr++] = 1.;
}
else {
Ke_bindx[nz_ptr] = north2d(ii,jj,nx); Ke_val[nz_ptr++] = -1.;
Ke_bindx[nz_ptr] = east2d(ii,jj,nx); Ke_val[nz_ptr++] = doffdiag;
Ke_bindx[nz_ptr] = south2d(ii,jj,nx); Ke_val[nz_ptr++] = 1.;
if (ii == 0) ii = nx-1;
else ii--;
Ke_bindx[nz_ptr] = west2d(ii,jj,nx); Ke_val[nz_ptr++] = doffdiag;
Ke_bindx[nz_ptr] = south2d(ii,jj,nx); Ke_val[nz_ptr++] = -1.;
Ke_bindx[nz_ptr] = north2d(ii,jj,nx); Ke_val[nz_ptr++] = 1.;
}
}
/* Transform the global Aztec matrix into a local Aztec matrix. That is, */
/* replace global column indices by local indices and set up communication */
/* data structure 'Ke_data_org' that will be used for matvec's. */
AZ_set_proc_config(proc_config, COMMUNICATOR);
AZ_transform_norowreordering(proc_config, &(Edge_Partition->needed_external_ids),
Ke_bindx, Ke_val, Edge_Partition->my_global_ids,
&reordered_glob_edges, &reordered_edge_externs,
&Ke_data_org, Nlocal_edges, 0, 0, 0,
&cpntr, AZ_MSR_MATRIX);
/* Convert old style Aztec matrix to newer style Aztec matrix */
Ke_mat = AZ_matrix_create( Nlocal_edges );
AZ_set_MSR(Ke_mat, Ke_bindx, Ke_val, Ke_data_org, 0, NULL, AZ_LOCAL);
return(Ke_mat);
}
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 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;
}
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;
}
//==============================================================================
AztecDVBR_Matrix::AztecDVBR_Matrix(const AztecDVBR_Matrix& src)
: amap_(src.amap_),
Amat_(NULL),
N_update_(src.N_update_),
external_(NULL),
extern_index_(NULL),
update_index_(NULL),
data_org_(NULL),
orderingUpdate_(NULL),
isLoaded_(src.isLoaded_),
isAllocated_(src.isAllocated_),
localNNZ_(src.localNNZ_),
nnzPerRow_(NULL),
numRemoteBlocks_(0),
remoteInds_(NULL),
remoteBlockSizes_(NULL)
{
//
//This copy constructor just takes a reference to src's amap_ (above), but
//'deep'-copies everything else. i.e., all arrays are allocated here and copied
//from src, etc.
//
//When this constructor completes, this matrix should be in the same state as
//the 'src' matrix. i.e., if src is already allocated, then this matrix will
//have its structure allocated. If src is already loaded (AZ_transform'd) then
//this matrix will have all arrays resulting from AZ_transform allocated too.
//The only thing this matrix won't get is the coefficient data from src.
//
nnzPerRow_ = new int[N_update_];
int i;
for(i=0; i<N_update_; i++) {
nnzPerRow_[i] = src.nnzPerRow_[i];
}
Amat_ = AZ_matrix_create(N_update_);
Amat_->matrix_type = AZ_VBR_MATRIX;
Amat_->matvec =
(void(*)(double*,double*,AZ_MATRIX_STRUCT*,int*))AZ_VBR_matvec_mult;
Amat_->rpntr = NULL;
calcRpntr();
if (isAllocated_) {
Amat_->bpntr = new int[N_update_+1];
for(i=0; i<N_update_+1; i++) Amat_->bpntr[i] = src.Amat_->bpntr[i];
int totalNNZBlks = Amat_->bpntr[N_update_];
Amat_->bindx = new int[totalNNZBlks];
for(i=0; i<totalNNZBlks; i++) Amat_->bindx[i] = src.Amat_->bindx[i];
Amat_->indx = new int[totalNNZBlks+1];
for(i=0; i<totalNNZBlks+1; i++) Amat_->indx[i] = src.Amat_->indx[i];
Amat_->val = new double[localNNZ_];
for(i=0; i<localNNZ_; i++) Amat_->val[i] = 0.0;
}
if (isLoaded_) {
int dataOrgLength = src.data_org_[AZ_total_send] + AZ_send_list;
data_org_ = (int*) AZ_allocate(dataOrgLength * sizeof(int));
for(i=0; i<dataOrgLength; i++) data_org_[i] = src.data_org_[i];
Amat_->data_org = data_org_;
int extLength = src.data_org_[AZ_N_ext_blk];
external_ = (int*) AZ_allocate(extLength * sizeof(int));
extern_index_ = (int*) AZ_allocate(extLength * sizeof(int));
for(i=0; i<extLength; i++) {
external_[i] = src.external_[i];
extern_index_[i] = src.extern_index_[i];
}
update_index_ = (int*) AZ_allocate(N_update_ * sizeof(int));
orderingUpdate_ = new int[N_update_];
for(i=0; i<N_update_; i++) {
update_index_[i] = src.update_index_[i];
orderingUpdate_[i] = src.orderingUpdate_[i];
}
int cpntrLength = N_update_ + src.data_org_[AZ_N_ext_blk] + 1;
Amat_->cpntr = (int*) AZ_allocate(cpntrLength * sizeof(int));
for(i=0; i<cpntrLength; i++) Amat_->cpntr[i] = src.Amat_->cpntr[i];
}
}
int test_azoo_scaling(Epetra_CrsMatrix& A,
Epetra_Vector& x,
Epetra_Vector& b,
bool verbose)
{
Epetra_Vector vec1(x);
Epetra_Vector vec2(x);
Epetra_Vector diag(x);
Epetra_Vector vec3(x);
Epetra_Vector vec4(x);
Epetra_Vector rhs(x);
Epetra_Vector soln_none(x);
Epetra_Vector soln_jacobi(x);
Epetra_Vector soln_rowsum(x);
Epetra_Vector soln_symdiag(x);
vec1.PutScalar(1.0);
A.Multiply(false, vec1, vec2);
A.ExtractDiagonalCopy(diag);
double* diag_vals = NULL;
diag.ExtractView(&diag_vals);
int* options = new int[AZ_OPTIONS_SIZE];
double* params = new double[AZ_PARAMS_SIZE];
AZ_defaults(options, params);
options[AZ_output] = verbose ? 1 : AZ_none;
options[AZ_scaling] = AZ_Jacobi;
AztecOO::MatrixData mdata(&A);
AZ_MATRIX* Amat = AZ_matrix_create(vec1.Map().NumMyElements());
AZ_set_MATFREE(Amat, (void*)(&mdata), Epetra_Aztec_matvec);
AZ_SCALING* scaling = AZ_scaling_create();
double* xvals = NULL, *bvals = NULL;
x.ExtractView(&xvals);
b.ExtractView(&bvals);
int err = AztecOO_scale_epetra(AZ_SCALE_MAT_RHS_SOL, Amat,
options, bvals, xvals, NULL, scaling);
if (err != 0) {
if (verbose) {
cout << "AztecOO_scale_epetra returned err="<<err<<endl;
}
return(err);
}
A.Multiply(false, vec1, vec3);
vec4.Multiply(1.0, diag, vec3, 0.0);
double vec2nrm, vec4nrm;
vec2.Norm2(&vec2nrm);
vec4.Norm2(&vec4nrm);
if (fabs(vec2nrm - vec4nrm) > 1.e-6) {
return(-1);
}
//now call the scaling function again, just to allow for
//testing memory-leak issues.
err = AztecOO_scale_epetra(AZ_SCALE_MAT_RHS_SOL, Amat,
options, bvals, xvals, NULL, scaling);
if (err != 0) {
if (verbose) {
cout << "AztecOO_scale_epetra returned err="<<err<<endl;
}
return(err);
}
AztecOO_scale_epetra(AZ_DESTROY_SCALING_DATA, Amat, options,
bvals, xvals, NULL, scaling);
x.PutScalar(1.0);
Epetra_CrsMatrix* Atmp = create_and_fill_crs_matrix(A.RowMap());
Atmp->Multiply(false, x, rhs);
x.PutScalar(0.0);
AztecOO azoo(&A, &x, &b);
azoo.SetAztecOption(AZ_scaling, AZ_Jacobi);
if (verbose) {
azoo.SetAztecOption(AZ_output, 1);
}
else {
azoo.SetAztecOption(AZ_output, AZ_none);
}
azoo.Iterate(100, 1.e-6);
delete Atmp;
Epetra_CrsMatrix* Atmp1 = create_and_fill_crs_matrix(A.RowMap());
x.PutScalar(1.0);
Atmp1->Multiply(false, x, rhs);
soln_rowsum.PutScalar(0.0);
AztecOO azoo1(Atmp1, &soln_rowsum, &rhs);
azoo1.SetAztecOption(AZ_scaling, AZ_row_sum);
azoo1.Iterate(100, 1.e-8);
delete Atmp1;
Epetra_CrsMatrix* Atmp2 = create_and_fill_crs_matrix(A.RowMap());
x.PutScalar(1.0);
Atmp2->Multiply(false, x, rhs);
soln_symdiag.PutScalar(0.0);
AztecOO azoo2(Atmp2, &soln_symdiag, &rhs);
azoo2.SetAztecOption(AZ_scaling, AZ_sym_diag);
azoo2.Iterate(100, 1.e-8);
delete Atmp2;
Epetra_CrsMatrix* Atmp3 = create_and_fill_crs_matrix(A.RowMap());
x.PutScalar(1.0);
Atmp3->Multiply(false, x, rhs);
soln_none.PutScalar(0.0);
AztecOO azoo3(Atmp3, &soln_none, &rhs);
azoo3.SetAztecOption(AZ_scaling, AZ_none);
azoo3.Iterate(100, 1.e-8);
delete Atmp3;
Epetra_CrsMatrix* Atmp4 = create_and_fill_crs_matrix(A.RowMap());
x.PutScalar(1.0);
Atmp4->Multiply(false, x, rhs);
soln_jacobi.PutScalar(0.0);
AztecOO azoo4(Atmp4, &soln_jacobi, &rhs);
azoo4.SetAztecOption(AZ_scaling, AZ_Jacobi);
azoo4.Iterate(100, 1.e-8);
delete Atmp4;
//at this point, soln_none, soln_jacobi, soln_rowsum and soln_symdiag
//should be the same or at least close to the same, since the
//matrix used in the solution has well-behaved coefficients.
//form vec1 = soln_none - soln_rowsum
vec1.PutScalar(0.0);
vec1.Update(1.0, soln_none, 0.0);
vec1.Update(-1.0, soln_rowsum, 1.0);
double norm_check1= 0.0;
vec1.Norm2(&norm_check1);
//form vec2 = soln_none - soln_symdiag
vec2.PutScalar(0.0);
vec2.Update(1.0, soln_none, 0.0);
vec2.Update(-1.0, soln_symdiag, 1.0);
double norm_check2= 0.0;
vec2.Norm2(&norm_check2);
//form vec3 = soln_none - soln_jacobi
vec3.PutScalar(0.0);
vec3.Update(1.0, soln_none, 0.0);
vec3.Update(-1.0, soln_jacobi, 1.0);
double norm_check3= 0.0;
vec3.Norm2(&norm_check3);
if (std::abs(norm_check1) > 1.e-6) {
if (verbose) {
cerr << "AZ_row_sum scaling produced bad soln"
<< endl;
}
return(-1);
}
if (std::abs(norm_check2) > 1.e-6) {
if (verbose) {
cerr << "AZ_sym_diag scaling produced bad soln"
<< endl;
}
return(-1);
}
if (std::abs(norm_check3) > 1.e-6) {
if (verbose) {
cerr << "AZ_Jacobi scaling produced bad soln"
<< endl;
}
return(-1);
}
options[AZ_pre_calc] = AZ_reuse;
err = AztecOO_scale_epetra(AZ_SCALE_MAT_RHS_SOL, Amat,
options, bvals, xvals, NULL, scaling);
if (err == 0) {
if (verbose) {
cerr << "AztecOO_scale_epetra failed to return err when"
<< " asked to reuse non-existent scaling data."<<endl;
}
return(-1);
}
options[AZ_keep_info] = 1;
options[AZ_pre_calc] = AZ_calc;
err = AztecOO_scale_epetra(AZ_SCALE_MAT_RHS_SOL, Amat,
options, bvals, xvals, NULL, scaling);
if (err != 0) {
if (verbose) {
cerr << "AztecOO_scale_epetra returned err=="<<err<<endl;
}
return(err);
}
options[AZ_keep_info] = 0;
options[AZ_pre_calc] = AZ_reuse;
err = AztecOO_scale_epetra(AZ_SCALE_MAT_RHS_SOL, Amat,
options, bvals, xvals, NULL, scaling);
if (err != 0) {
if (verbose) {
cerr << "AztecOO_scale_epetra returned err=="<<err
<<" when asked to reuse scaling data"<<endl;
}
return(err);
}
options[AZ_pre_calc] = AZ_calc;
err = AztecOO_scale_epetra(AZ_DESTROY_SCALING_DATA, Amat,
options, bvals, xvals, NULL, scaling);
if (err != 0) {
if (verbose) {
std::cerr << "AztecOO_scale_epetra returned err=="<<err
<< " when asked to destroy scaling data."<<std::endl;
}
return(err);
}
AZ_matrix_destroy(&Amat);
delete [] options;
delete [] params;
AZ_scaling_destroy(&scaling);
AZ_manage_memory(0, AZ_CLEAR_ALL, 0, 0, 0);
return(0);
}
int create_and_transform_simple_matrix(int matrix_type,
int N,
double diag_term,
int* proc_config,
AZ_MATRIX*& Amat,
int*& external,
int*& update_index,
int*& external_index)
{
//We're going to create a very simple tri-diagonal matrix with diag_term
//on the diagonal, and -1.0 on the off-diagonals.
Amat = AZ_matrix_create(N);
int* update = new int[N];
int i;
int numprocs = proc_config[AZ_N_procs];
int first_eqn = proc_config[AZ_node]*N;
int adjustment = 0;
for(i=0; i<N; ++i) {
update[i] = first_eqn+i;
if (update[i] == 0 || update[i] == numprocs*N-1) ++adjustment;
}
int* data_org;
//global row 0 and global-N-1 (N*numprocs-1) will have 2 nonzeros in the first
//and last rows, and there will be 3 nonzeros in all other rows.
//If you are brave enough to attempt to modify any of the following code,
//bear in mind that the number of nonzeros per row (3) is hard-coded in
//a few places.
int nnz = 3*N - adjustment;
double* val = new double[nnz+2];
int* bindx = new int[nnz+2];
int* indx = NULL;
int* rpntr = NULL;
int* cpntr = NULL;
int* bpntr = NULL;
int offs = N+1;
for(i=0; i<N; ++i) {
val[i] = diag_term;
bindx[i] = offs;
int num_off_diagonals = 2;
if (update[i]==0 || update[i]==numprocs*N-1) num_off_diagonals = 1;
offs += num_off_diagonals;
}
bindx[N] = offs;
for(i=0; i<N; ++i) {
int global_row = update[i];
if (global_row > 0) {
int ks = bindx[i];
val[ks] = -1.0;
bindx[ks] = global_row-1;
}
if (global_row < numprocs*N-1) {
int ke = bindx[i+1]-1;
val[ke] = -1.0;
bindx[ke] = global_row+1;
}
}
if (matrix_type == AZ_VBR_MATRIX) {
//AZ_transform allocates cpntr
rpntr = new int[N+1];
bpntr = new int[N+1];
indx = new int[nnz+2];
offs = 0;
for(i=0; i<N; ++i) {
rpntr[i] = i;
bpntr[i] = offs;
if (update[i]==0 || update[i]==numprocs*N-1) offs += 2;
else offs += 3;
}
rpntr[N] = N;
bpntr[N] = offs;
for(i=0; i<N; ++i) {
int global_col = update[i] - 1;
if (update[i]==0) ++global_col;
for(int j=bpntr[i]; j<=bpntr[i+1]-1; ++j) {
if (global_col == update[i]) val[j] = diag_term;
else val[j] = -1.0;
bindx[j] = global_col++;
}
}
for(i=0; i<nnz+2; ++i) {
indx[i] = i;
}
}
AZ_transform(proc_config, &external, bindx, val, update, &update_index,
&external_index, &data_org, N, indx, bpntr, rpntr,
&cpntr, matrix_type);
if (matrix_type == AZ_MSR_MATRIX) {
AZ_set_MSR(Amat, bindx, val, data_org, N, update, AZ_LOCAL);
}
else {
AZ_set_VBR(Amat, rpntr, cpntr, bpntr, indx, bindx, val, data_org,
N, update, AZ_LOCAL);
}
Amat->must_free_data_org = 1;
return(0);
}
int main(int argc, char *argv[])
{
int proc_config[AZ_PROC_SIZE];/* Processor information. */
int options[AZ_OPTIONS_SIZE]; /* Array used to select solver options. */
double params[AZ_PARAMS_SIZE]; /* User selected solver paramters. */
double status[AZ_STATUS_SIZE]; /* Information returned from AZ_solve(). */
int *bindx_real; /* index and values arrays for MSR matrices */
double *val_real, *val_imag;
int * update; /* List of global eqs owned by the processor */
double *x_real, *b_real; /* initial guess/solution, RHS */
double *x_imag, *b_imag;
unsigned int N_local; /* Number of equations on this node */
double residual; /* Used for computing residual */
double *xx_real, *xx_imag, *xx; /* Known exact solution */
int myPID, nprocs;
AZ_MATRIX *Amat_real; /* Real matrix structure */
AZ_MATRIX *Amat; /* Komplex matrix to be solved. */
AZ_PRECOND *Prec; /* Komplex preconditioner */
double *x, *b; /* Komplex Initial guess and RHS */
int i;
/******************************/
/* First executable statement */
/******************************/
#ifdef AZTEC_MPI
MPI_Init(&argc,&argv);
#endif
/* Get number of processors and the name of this processor */
#ifdef AZTEC_MPI
AZ_set_proc_config(proc_config,MPI_COMM_WORLD);
#else
AZ_set_proc_config(proc_config,0);
#endif
nprocs = proc_config[AZ_N_procs];
myPID = proc_config[AZ_node];
printf("proc %d of %d is alive\n",myPID, nprocs);
/* Define two real diagonal matrices. Will use as real and imaginary parts */
/* Get the number of local equations from the command line */
if (argc!=2)
{
if (myPID==0) printf("Usage: %s number_of_local_equations\n",argv[0]);
exit(1);
}
N_local = atoi(argv[1]);
const unsigned int N_local_max = 1000000;
if (N_local > N_local_max) {
if (myPID==0)
printf("No more than %d local equation allowed\n", N_local_max);
exit(1);
}
/* Need N_local+1 elements for val/bindx arrays */
val_real = malloc((N_local+1)*sizeof(double));
val_imag = malloc((N_local+1)*sizeof(double));
/* bindx_imag is not needed since real/imag have same pattern */
bindx_real = malloc((N_local+1)*sizeof(int));
update = malloc((N_local+1)*sizeof(int)); /* Malloc equation update list */
b_real = malloc((N_local+1)*sizeof(double)); /* Malloc x and b arrays */
b_imag = malloc((N_local+1)*sizeof(double));
x_real = malloc((N_local+1)*sizeof(double));
x_imag = malloc((N_local+1)*sizeof(double));
xx_real = malloc((N_local+1)*sizeof(double));
xx_imag = malloc((N_local+1)*sizeof(double));
for (i=0; i<N_local; i++)
{
val_real[i] = 10 + i/(N_local/10); /* Some very fake diagonals */
val_imag[i] = 10 - i/(N_local/10); /* Should take exactly 20 GMRES steps */
x_real[i] = 0.0; /* Zero initial guess */
x_imag[i] = 0.0;
xx_real[i] = 1.0; /* Let exact solution = 1 */
xx_imag[i] = 0.0;
/* Generate RHS to match exact solution */
b_real[i] = val_real[i]*xx_real[i] - val_imag[i]*xx_imag[i];
b_imag[i] = val_imag[i]*xx_real[i] + val_real[i]*xx_imag[i];
/* All bindx[i] have same value since no off-diag terms */
bindx_real[i] = N_local + 1;
/* each processor owns equations
myPID*N_local through myPID*N_local + N_local - 1 */
update[i] = myPID*N_local + i;
}
bindx_real[N_local] = N_local+1; /* Need this last index */
/* Register Aztec Matrix for Real Part, only imaginary values are needed*/
Amat_real = AZ_matrix_create(N_local);
AZ_set_MSR(Amat_real, bindx_real, val_real, NULL, N_local, update, AZ_GLOBAL);
/* initialize AZTEC options */
AZ_defaults(options, params);
options[AZ_solver] = AZ_gmres; /* Use CG with no preconditioning */
options[AZ_precond] = AZ_none;
options[AZ_kspace] = 21;
options[AZ_max_iter] = 21;
params[AZ_tol] = 1.e-14;
/**************************************************************/
/* Construct linear system. Form depends on input parameters */
/**************************************************************/
/**************************************************************/
/* Method 1: Construct A, x, and b in one call. */
/* Useful if using A,x,b only one time. Equivalent to Method 2*/
/**************************************************************/
AZK_create_linsys_ri2k (x_real, x_imag, b_real, b_imag,
options, params, proc_config,
Amat_real, val_imag, &x, &b, &Amat);
/**************************************************************/
/* Method 2: Construct A, x, and b in separate calls. */
/* Useful for having more control over the construction. */
/* Note that the matrix must be constructed first. */
/**************************************************************/
/* Uncomment these three calls and comment out the above call
AZK_create_matrix_ri2k (options, params, proc_config,
Amat_real, val_imag, &Amat);
AZK_create_vector_ri2k(options, params, proc_config, Amat,
x_real, x_imag, &x);
AZK_create_vector_ri2k(options, params, proc_config, Amat,
b_real, b_imag, &b);
*/
/**************************************************************/
/* Build exact solution vector. */
/* Check residual of init guess and exact solution */
/**************************************************************/
AZK_create_vector_ri2k(options, params, proc_config, Amat,
xx_real, xx_imag, &xx);
residual = AZK_residual_norm(x, b, options, params, proc_config, Amat);
if (proc_config[AZ_node]==0)
printf("\n\n\nNorm of residual using initial guess = %12.4g\n",residual);
residual = AZK_residual_norm(xx, b, options, params, proc_config, Amat);
AZK_destroy_vector(options, params, proc_config, Amat, &xx);
if (proc_config[AZ_node]==0)
printf("\n\n\nNorm of residual using exact solution = %12.4g\n",residual);
/**************************************************************/
/* Create preconditioner */
/**************************************************************/
AZK_create_precon(options, params, proc_config, x, b, Amat, &Prec);
/**************************************************************/
/* Solve linear system using Aztec. */
/**************************************************************/
AZ_iterate(x, b, options, params, status, proc_config, Amat, Prec, NULL);
/**************************************************************/
/* Extract solution. */
/**************************************************************/
AZK_extract_solution_k2ri(options, params, proc_config, Amat, Prec, x,
x_real, x_imag);
/**************************************************************/
/* Destroy Preconditioner. */
/**************************************************************/
AZK_destroy_precon (options, params, proc_config, Amat, &Prec);
/**************************************************************/
/* Destroy linear system. */
/**************************************************************/
AZK_destroy_linsys (options, params, proc_config, &x, &b, &Amat);
if (proc_config[AZ_node]==0)
{
printf("True residual norm squared = %22.16g\n",status[AZ_r]);
printf("True scaled res norm squared = %22.16g\n",status[AZ_scaled_r]);
printf("Computed res norm squared = %22.16g\n",status[AZ_rec_r]);
}
/* Print comparison between known exact and computed solution */
{double sum = 0.0;
for (i=0; i<N_local; i++) sum += fabs(x_real[i]-xx_real[i]);
for (i=0; i<N_local; i++) sum += fabs(x_imag[i]-xx_imag[i]);
printf("Processor %d: Difference between exact and computed solution = %12.4g\n",
proc_config[AZ_node],sum);
}
/* Free memory allocated */
free((void *) val_real );
free((void *) bindx_real );
free((void *) val_imag );
free((void *) update );
free((void *) b_real );
free((void *) b_imag );
free((void *) x_real );
free((void *) x_imag );
free((void *) xx_real );
free((void *) xx_imag );
AZ_matrix_destroy(&Amat_real);
#ifdef AZTEC_MPI
MPI_Finalize();
#endif
return 0 ;
}
int main(int argc, char *argv[])
{
/* See Aztec User's Guide for the variables that follow: */
int proc_config[AZ_PROC_SIZE];/* Processor information. */
int N_update; /* # of unknowns updated on this node */
int *update; /* vector elements updated on this node */
int *data_orgA; /* Array to specify data layout */
int *externalA; /* vector elements needed by this node. */
int *update_indexA; /* ordering of update[] and external[] */
int *extern_indexA; /* locally on this processor. */
int *bindxA; /* Sparse matrix to be solved is stored */
double *valA; /* in these MSR arrays. */
AZ_MATRIX *mat_curl_edge; /* curl operator matrix */
int *data_orgB; /* Array to specify data layout */
int *externalB; /* vector elements needed by this node. */
int *update_indexB; /* ordering of update[] and external[] */
int *extern_indexB; /* locally on this processor. */
int *bindxB; /* Sparse matrix to be solved is stored */
double *valB; /* in these MSR arrays. */
AZ_MATRIX *mat_curl_face; /* curl operator matrix */
int *bc_indx;
int n_bc;
double *efield;
double *bfield;
double *epsilon;
double *tmp_vec;
double *tmp_vec2;
int i, nrow, x, y, z;
int k, t;
long startTime, endTime;
int myrank;
int vec_len;
/* get number of processors and the name of this processor */
#ifdef AZ_MPI
MPI_Init(&argc,&argv);
AZ_set_proc_config(proc_config, MPI_COMM_WORLD);
MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
#else
myrank = 0;
AZ_set_proc_config(proc_config, AZ_NOT_MPI);
#endif
nrow = ncomp * nx * ny * nz; /* overll number of matrix rows */
// Define partitioning: matrix rows (ascending order) owned by this node
// Here it is done automatically, but it can also be specified by hand
AZ_read_update(&N_update, &update, proc_config, nrow, 1, AZ_linear);
// In the following we set up the matrix for the edge centered curl operator
// All the steps are described in detail in the AZTEC manual.
// first: allocate space for the first matrix.
bindxA = (int *) malloc((N_update*MAX_NZ_ROW+1)*sizeof(int));
valA = (double *) malloc((N_update*MAX_NZ_ROW+1)*sizeof(double));
if (valA == NULL) perror("Error: Not enough space to create matrix");
// Initialize the index for the first off diagonal element
bindxA[0] = N_update+1;
// Create the matrix row by row. Each processor creates only rows appearing
// in update[] (using global col. numbers).
for (i = 0; i < N_update; i++)
create_curl_matrix_row_edge(update[i], i, valA, bindxA);
// convert matrix to a local distributed matrix
AZ_transform(proc_config, &externalA, bindxA, valA, update, &update_indexA,
&extern_indexA, &data_orgA, N_update, NULL, NULL, NULL, NULL,
AZ_MSR_MATRIX);
// convert the matrix arrays into a matrix structure, used in the
// matrix vector multiplication
mat_curl_edge = AZ_matrix_create(data_orgA[AZ_N_internal] + data_orgA[AZ_N_border]);
AZ_set_MSR(mat_curl_edge, bindxA, valA, data_orgA, 0, NULL, AZ_LOCAL);
// at this point the edge centered curl matrix is completed.
// In the following we set up the matrix for the face centered curl operator
// All the steps are described in detail in the AZTEC manual.
// first: allocate space for the first matrix.
bindxB = (int *) malloc((N_update*MAX_NZ_ROW+1)*sizeof(int));
valB = (double *) malloc((N_update*MAX_NZ_ROW+1)*sizeof(double));
if (valB == NULL) perror("Error: Not enough space to create matrix");
// Initialize the index for the first off diagonal element
bindxB[0] = N_update+1;
// Create the matrix row by row. Each processor creates only rows appearing
// in update[] (using global col. numbers).
for (i = 0; i < N_update; i++)
create_curl_matrix_row_face(update[i], i, valB, bindxB);
// convert matrix to a local distributed matrix
AZ_transform(proc_config, &externalB, bindxB, valB, update, &update_indexB,
&extern_indexB, &data_orgB,
N_update, NULL, NULL, NULL, NULL,
AZ_MSR_MATRIX);
// convert the matrix arrays into a matrix structure, used in the
// matrix vector multiplication
mat_curl_face = AZ_matrix_create(data_orgB[AZ_N_internal] + data_orgB[AZ_N_border]);
AZ_set_MSR(mat_curl_face, bindxB, valB, data_orgB, 0, NULL, AZ_LOCAL);
// at this point the face centered curl matrix is completed.
// allocate memory for the fields and a temporary vector
vec_len = N_update + data_orgA[AZ_N_external];
efield = (double *) malloc(vec_len*sizeof(double));
bfield = (double *) malloc(vec_len*sizeof(double));
epsilon = (double *) malloc(vec_len*sizeof(double));
tmp_vec = (double *) malloc(vec_len*sizeof(double));
tmp_vec2 = (double *) malloc(vec_len*sizeof(double));
// setup the boundary condition. We will get an arry that tells us
// which positions need to be updated and where the results needs
// to be stored in the E field.
setup_bc(update, update_indexB, N_update, &bc_indx, &n_bc);
// initialize the field vectors
for(k = 0; k < vec_len; k++){
efield[k] = 0.;
bfield[k] = 0.;
epsilon[k] = 1.;
tmp_vec[k] = 0.;
}
// initialize the dielectric structure. Ugly hard-coded stuff,
// needs to be cleaned out...
for(y=45; y<55; y++){
for(x = y; x<100; x++)
epsilon[compZ + pos_to_row(x, y, 0)] = 0.95;
}
// reorder the dielectric vector in order to align with the B field
AZ_reorder_vec(epsilon, data_orgA, update_indexA, NULL);
printf("Begin iteration \n");
// just some timing ...
startTime = currentTimeMillis();
// *******************
// begin of the time stepping loop
// *******************
for( t = 0; t < nsteps; t++){
// first we do the e field update
// convert the B field to the H field
for(k = 0 ; k < vec_len; k++)
bfield[k] *= epsilon[k];
// setup the initial condition
for( k = 0; k < n_bc; k++){
x = bc_indx[4*k];
y = bc_indx[4*k+1];
z = bc_indx[4*k+2];
efield[bc_indx[4*k+3]] =
sin((double) y * 5. * 3.14159 / (double) ny) *
sin(omega * dt * (double) (t + 1));
}
// E field update:
// tmp_vec = Curl_Op * bfield
// efield = efield + c^2 * dt * tmp_vec
AZ_MSR_matvec_mult( bfield, tmp_vec, mat_curl_edge, proc_config);
// reorder the result in tmp_vec so that it aligns with the
// decomposition of the E field
AZ_invorder_vec(tmp_vec, data_orgA, update_indexA, NULL, tmp_vec2);
AZ_reorder_vec(tmp_vec2, data_orgB, update_indexB, NULL);
// update the efield
for(k = 0 ; k < N_update; k++)
efield[k] = efield[k] + c2 * tmp_vec2[k] * dt;
// bfield update :
// tmp_vec = DualCurl_Op * efield
// bfield = bfield - tmp_vec * dt
AZ_MSR_matvec_mult( efield, tmp_vec, mat_curl_face, proc_config);
// reorder the result so that it fits the decomposition of the bfield
AZ_invorder_vec(tmp_vec, data_orgB, update_indexB, NULL, tmp_vec2);
AZ_reorder_vec(tmp_vec2, data_orgA, update_indexA, NULL);
// update the b field
for(k = 0; k < N_update; k++)
bfield[k] = bfield[k] - tmp_vec2[k] * dt;
if(myrank == 0)
printf("Taking step %d at time %g\n", t,
(double) (currentTimeMillis() - startTime) / 1000.);
}
// ******************
// end of timestepping loop
// *****************
endTime = currentTimeMillis();
printf("After iteration: %g\n", (double)(endTime - startTime) / 1000. );
#if 1
system("rm efield.txt bfield.txt");
// dump filed data: efield
AZ_invorder_vec(efield, data_orgB, update_indexB, NULL, tmp_vec);
write_file("efield.txt", tmp_vec, N_update);
// dump filed data: bfield
AZ_invorder_vec(bfield, data_orgA, update_indexA, NULL, tmp_vec);
write_file("bfield.txt", tmp_vec, N_update);
#endif
/* Free allocated memory */
AZ_matrix_destroy( &mat_curl_edge);
free((void *) update); free((void *) update_indexA);
free((void *) externalA); free((void *) extern_indexA);
free((void *) bindxA); free((void *) valA); free((void *) data_orgA);
AZ_matrix_destroy( &mat_curl_face);
free((void *) externalB); free((void *) extern_indexB);
free((void *) bindxB); free((void *) valB); free((void *) data_orgB);
free((void *) efield); free((void *) bfield);
free((void *) tmp_vec); free((void *) tmp_vec2);
#ifdef AZ_MPI
MPI_Finalize();
#endif
return(1);
}
int main(int argc, char *argv[])
{
char global[]="global";
char local[]="local";
int proc_config[AZ_PROC_SIZE];/* Processor information. */
int options[AZ_OPTIONS_SIZE]; /* Array used to select solver options. */
double params[AZ_PARAMS_SIZE]; /* User selected solver paramters. */
int *data_org;
/* Array to specify data layout */
double status[AZ_STATUS_SIZE]; /* Information returned from AZ_solve(). */
int *update; /* vector elements updated on this node. */
int *external;
/* vector elements needed by this node. */
int *update_index;
/* ordering of update[] and external[] */
int *extern_index;
/* locally on this processor. */
int *indx; /* MSR format of real and imag parts */
int *bindx;
int *bpntr;
int *rpntr;
int *cpntr;
AZ_MATRIX *Amat;
AZ_PRECOND *Prec;
double *val;
double *x, *b, *xexact, *xsolve;
int n_nonzeros, n_blk_nonzeros;
int N_update; /* # of block unknowns updated on this node */
int N_local;
/* Number scalar equations on this node */
int N_global, N_blk_global; /* Total number of equations */
int N_external, N_blk_eqns;
double *val_msr;
int *bindx_msr;
double norm, d ;
int matrix_type;
int has_global_indices, option;
int i, j, m, mp ;
int ione = 1;
#ifdef TEST_SINGULAR
double * xnull; /* will contain difference of given exact solution and computed solution*/
double * Axnull; /* Product of A time xnull */
double norm_Axnull;
#endif
#ifdef AZTEC_MPI
double MPI_Wtime(void) ;
#endif
double time ;
#ifdef AZTEC_MPI
MPI_Init(&argc,&argv);
#endif
/* get number of processors and the name of this processor */
#ifdef AZTEC_MPI
AZ_set_proc_config(proc_config,MPI_COMM_WORLD);
#else
AZ_set_proc_config(proc_config,0);
#endif
printf("proc %d of %d is alive\n",
proc_config[AZ_node],proc_config[AZ_N_procs]) ;
#ifdef AZTEC_MPI
MPI_Barrier(MPI_COMM_WORLD) ;
#endif
#ifdef VBRMATRIX
if(argc != 3)
perror("error: enter name of data and partition file on command line") ;
#else
if(argc != 2) perror("error: enter name of data file on command line") ;
#endif
/* Set exact solution to NULL */
xexact = NULL;
/* Read matrix file and distribute among processors.
Returns with this processor's set of rows */
#ifdef VBRMATRIX
read_hb(argv[1], proc_config, &N_global, &n_nonzeros,
&val_msr, &bindx_msr, &x, &b, &xexact);
create_vbr(argv[2], proc_config, &N_global, &N_blk_global,
&n_nonzeros, &n_blk_nonzeros, &N_update, &update,
bindx_msr, val_msr, &val, &indx,
&rpntr, &cpntr, &bpntr, &bindx);
if(proc_config[AZ_node] == 0)
{
free ((void *) val_msr);
free ((void *) bindx_msr);
free ((void *) cpntr);
}
matrix_type = AZ_VBR_MATRIX;
#ifdef AZTEC_MPI
MPI_Barrier(MPI_COMM_WORLD) ;
#endif
distrib_vbr_matrix( proc_config, N_global, N_blk_global,
&n_nonzeros, &n_blk_nonzeros,
&N_update, &update,
&val, &indx, &rpntr, &cpntr, &bpntr, &bindx,
&x, &b, &xexact);
#else
read_hb(argv[1], proc_config, &N_global, &n_nonzeros,
&val, &bindx, &x, &b, &xexact);
#ifdef AZTEC_MPI
MPI_Barrier(MPI_COMM_WORLD) ;
#endif
distrib_msr_matrix(proc_config, N_global, &n_nonzeros, &N_update,
&update, &val, &bindx, &x, &b, &xexact);
#ifdef DEBUG
for (i = 0; i<N_update; i++)
if (val[i] == 0.0 ) printf("Zero diagonal at row %d\n",i);
#endif
matrix_type = AZ_MSR_MATRIX;
#endif
/* convert matrix to a local distributed matrix */
cpntr = NULL;
AZ_transform(proc_config, &external, bindx, val, update,
&update_index, &extern_index, &data_org,
N_update, indx, bpntr, rpntr, &cpntr,
matrix_type);
printf("Processor %d: Completed AZ_transform\n",proc_config[AZ_node]) ;
has_global_indices = 0;
option = AZ_LOCAL;
#ifdef VBRMATRIX
N_local = rpntr[N_update];
#else
N_local = N_update;
#endif
Amat = AZ_matrix_create(N_local);
#ifdef VBRMATRIX
AZ_set_VBR(Amat, rpntr, cpntr, bpntr, indx, bindx, val, data_org,
N_update, update, option);
#else
AZ_set_MSR(Amat, bindx, val, data_org, N_update, update, option);
#endif
printf("proc %d Completed AZ_create_matrix\n",proc_config[AZ_node]) ;
#ifdef AZTEC_MPI
MPI_Barrier(MPI_COMM_WORLD) ;
#endif
/* initialize AZTEC options */
AZ_defaults(options, params);
options[AZ_solver] = AZ_gmres;
options[AZ_precond] = AZ_sym_GS;
options[AZ_poly_ord] = 1;
options[AZ_graph_fill] = 1;
params[AZ_rthresh] = 0.0E-7;
params[AZ_athresh] = 0.0E-7;
options[AZ_overlap] = 1;
/*
params[AZ_ilut_fill] = 2.0;
params[AZ_drop] = 0.01;
options[AZ_overlap] = 0;
options[AZ_reorder] = 0;
params[AZ_rthresh] = 1.0E-1;
params[AZ_athresh] = 1.0E-1;
options[AZ_precond] = AZ_dom_decomp ;
options[AZ_subdomain_solve] = AZ_bilu_ifp;
options[AZ_reorder] = 0;
options[AZ_graph_fill] = 0;
params[AZ_rthresh] = 1.0E-7;
params[AZ_athresh] = 1.0E-7;
options[AZ_poly_ord] = 1;
options[AZ_precond] = AZ_Jacobi;
params[AZ_omega] = 1.0;
options[AZ_precond] = AZ_none ;
options[AZ_poly_ord] = 1;
options[AZ_precond] = AZ_Jacobi ;
options[AZ_scaling] = AZ_sym_row_sum ;
options[AZ_scaling] = AZ_sym_diag;
options[AZ_conv] = AZ_noscaled;
options[AZ_scaling] = AZ_Jacobi ;
options[AZ_precond] = AZ_dom_decomp ;
options[AZ_subdomain_solve] = AZ_icc ;
options[AZ_subdomain_solve] = AZ_ilut ;
params[AZ_omega] = 1.2;
params[AZ_ilut_fill] = 2.0;
params[AZ_drop] = 0.01;
options[AZ_reorder] = 0;
options[AZ_overlap] = 0;
options[AZ_type_overlap] = AZ_symmetric;
options[AZ_precond] = AZ_dom_decomp ;
options[AZ_subdomain_solve] = AZ_bilu ;
options[AZ_graph_fill] = 0;
options[AZ_overlap] = 0;
options[AZ_precond] = AZ_dom_decomp ;
options[AZ_subdomain_solve] = AZ_bilu_ifp ;
options[AZ_graph_fill] = 0;
options[AZ_overlap] = 0;
params[AZ_rthresh] = 1.0E-3;
params[AZ_athresh] = 1.0E-3;
options[AZ_poly_ord] = 1;
options[AZ_precond] = AZ_Jacobi ; */
options[AZ_kspace] = 600 ;
options[AZ_max_iter] = 600 ;
params[AZ_tol] = 1.0e-14;
#ifdef BGMRES
options[AZ_gmres_blocksize] = 3;
options[AZ_gmres_num_rhs] = 1;
#endif
#ifdef DEBUG
if (proc_config[AZ_N_procs]==1)
write_vec("rhs.dat", N_local, b);
#endif
/* xsolve is a little longer vector needed to account for external
entries. Make it and copy x (initial guess) into it.
*/
if (has_global_indices)
{
N_external = 0;
}
else
{
N_external = data_org[AZ_N_external];
}
xsolve = (double *) calloc(N_local + N_external,
sizeof(double)) ;
for (i=0; i<N_local; i++) xsolve[i] = x[i];
/* Reorder rhs and xsolve to match matrix ordering from AZ_transform */
if (!has_global_indices)
{
AZ_reorder_vec(b, data_org, update_index, rpntr) ;
AZ_reorder_vec(xsolve, data_org, update_index, rpntr) ;
}
#ifdef VBRMATRIX
AZ_check_vbr(N_update, data_org[AZ_N_ext_blk], AZ_LOCAL,
bindx, bpntr, cpntr, rpntr, proc_config);
#else
AZ_check_msr(bindx, N_update, N_external, AZ_LOCAL, proc_config);
#endif
printf("Processor %d of %d N_local = %d N_external = %d NNZ = %d\n",
proc_config[AZ_node],proc_config[AZ_N_procs],N_local,N_external,
n_nonzeros);
/* solve the system of equations using b as the right hand side */
Prec = AZ_precond_create(Amat,AZ_precondition, NULL);
AZ_iterate(xsolve, b, options, params, status, proc_config,
Amat, Prec, NULL);
/*AZ_ifpack_iterate(xsolve, b, options, params, status, proc_config,
Amat);*/
if (proc_config[AZ_node]==0)
{
printf("True residual norm = %22.16g\n",status[AZ_r]);
printf("True scaled res = %22.16g\n",status[AZ_scaled_r]);
printf("Computed res norm = %22.16g\n",status[AZ_rec_r]);
}
#ifdef TEST_SINGULAR
xnull = (double *) calloc(N_local + N_external, sizeof(double)) ;
Axnull = (double *) calloc(N_local + N_external, sizeof(double)) ;
for (i=0; i<N_local; i++) xnull[i] = xexact[i];
if (!has_global_indices) AZ_reorder_vec(xnull, data_org, update_index, rpntr);
for (i=0; i<N_local; i++) xnull[i] -= xsolve[i]; /* fill with nullerence */
Amat->matvec(xnull, Axnull, Amat, proc_config);
norm_Axnull = AZ_gvector_norm(N_local, 2, Axnull, proc_config);
if (proc_config[AZ_node]==0) printf("Norm of A(xexact-xsolve) = %12.4g\n",norm_Axnull);
free((void *) xnull);
free((void *) Axnull);
#endif
/* Get solution back into original ordering */
if (!has_global_indices) {
AZ_invorder_vec(xsolve, data_org, update_index, rpntr, x);
free((void *) xsolve);
}
else {
free((void *) x);
x = xsolve;
}
#ifdef DEBUG
if (proc_config[AZ_N_procs]==1)
write_vec("solution.dat", N_local, x);
#endif
if (xexact != NULL)
{
double sum = 0.0;
double largest = 0.0;
for (i=0; i<N_local; i++) sum += fabs(x[i]-xexact[i]);
printf("Processor %d: Difference between exact and computed solution = %12.4g\n",
proc_config[AZ_node],sum);
for (i=0; i<N_local; i++) largest = AZ_MAX(largest,fabs(xexact[i]));
printf("Processor %d: Difference divided by max abs value of exact = %12.4g\n",
proc_config[AZ_node],sum/largest);
}
free((void *) val);
free((void *) bindx);
#ifdef VBRMATRIX
free((void *) rpntr);
free((void *) bpntr);
free((void *) indx);
#endif
free((void *) b);
free((void *) x);
if (xexact!=NULL) free((void *) xexact);
AZ_free((void *) update);
AZ_free((void *) update_index);
AZ_free((void *) external);
AZ_free((void *) extern_index);
AZ_free((void *) data_org);
if (cpntr!=NULL) AZ_free((void *) cpntr);
AZ_precond_destroy(&Prec);
AZ_matrix_destroy(&Amat);
#ifdef AZTEC_MPI
MPI_Finalize() ;
#endif
/* end main
*/
return 0 ;
}
void init_guess_and_rhs(int update_index[], int update[], double *x[],double
*ax[],int data_org[], double val[], int indx[], int
bindx[], int rpntr[], int cpntr[], int bpntr[], int
proc_config[])
/*
* Set the initial guess and the right hand side where the right hand side
* is obtained by doing a matrix-vector multiplication.
*
* Author: Ray Tuminaro, Div 1422, SNL
* Date : 3/15/95
*
* Parameters
*
* update_index == On input, ordering of update and external
* locally on this processor. For example
* 'update_index[i]' gives the index location
* of the block which has the global index
* 'update[i]'.
* update == On input, list of pts to be updated on this node
* data_org == On input, indicates how data is set on this node.
* For example, data_org[] contains information on
* how many unknowns are internal, external and
* border unknowns as well as which points need
* to be communicated. See User's Guide for more
* details.
* val, indx, == On input, holds matrix nonzeros. See User's Guide
* bindx, rpntr, for more details.
* cpntr, bpntr
* x == On output, 'x' is allocated and set to all zeros.
* ax == On output, 'ax' is allocated and is set to the
* result of a matrix-vector product.
*/
{
int i,j;
int temp,num;
double sum = 0.0;
AZ_MATRIX *Amat;
temp = data_org[AZ_N_int_blk] + data_org[AZ_N_bord_blk];
num = data_org[AZ_N_internal] + data_org[AZ_N_border];
/* allocate vectors */
i = num + data_org[AZ_N_external];
*x = (double *) AZ_allocate((i+1)*sizeof(double));
*ax = (double *) AZ_allocate((i+1)*sizeof(double));
if (*ax == NULL) {
(void) fprintf(stderr, "Not enough space in init_guess_and_rhs() for ax\n");
exit(1);
}
for (j = 0 ; j < i ; j++ ) (*x)[j] = 0.0;
for (j = 0 ; j < i ; j++ ) (*ax)[j] = 0.0;
/* initialize 'x' to a function which will be used in matrix-vector product */
if (data_org[AZ_matrix_type] == AZ_VBR_MATRIX) {
for (i = 0; i < temp; i++) {
for (j = rpntr[i]; j < rpntr[i+1]; j++) {
(*x)[j] = (double) (update[i]) + (double)(j-rpntr[i]) /
(double)(num_PDE_eqns);
}
}
}
else {
for (i = 0; i < temp; i++) {
(*x)[i] = (double) (update[i]) / (double) (num_PDE_eqns);
}
}
/* Reorder 'x' so that it conforms to the transformed matrix */
AZ_reorder_vec(*x,data_org,update_index,rpntr);
if (application == 2) {
/* take out the constant vector. Used for the */
/* finite element problem because it is singular */
sum = AZ_gsum_double(sum, proc_config);
i = AZ_gsum_int(num, proc_config);
if (i != 0) sum = sum / ((double) i);
for (i = 0; i < num; i++) (*x)[i] -= sum;
}
Amat = AZ_matrix_create(num);
if (data_org[AZ_matrix_type] == AZ_MSR_MATRIX)
AZ_set_MSR(Amat, bindx, val, data_org, 0, NULL, AZ_LOCAL);
else if (data_org[AZ_matrix_type] == AZ_VBR_MATRIX)
AZ_set_VBR(Amat, rpntr,cpntr, bpntr, indx,bindx, val, data_org, 0, NULL,AZ_LOCAL);
Amat->matvec(*x, *ax, Amat, proc_config);
AZ_matrix_destroy( &Amat );
for (i = 0; i < num; i++) (*x)[i] = 0.0;
} /* init_guess_and_rhs */
AZ_MATRIX *user_Ke_build(struct user_partition *Edge_Partition)
{
double dcenter, doff, sigma = .0001;
int ii,jj, horv, i, nx, global_id, nz_ptr, Nlocal_edges;
/* Aztec matrix and temp variables */
int *Ke_bindx, *Ke_data_org = NULL;
double *Ke_val;
AZ_MATRIX *Ke_mat;
int proc_config[AZ_PROC_SIZE], *cpntr = NULL;
int *reordered_glob_edges = NULL, *reordered_edge_externs = NULL;
Nlocal_edges = Edge_Partition->Nlocal;
nx = (int) sqrt( ((double) Edge_Partition->Nglobal/2) + .00001);
Ke_bindx = (int *) malloc((7*Nlocal_edges+1)*sizeof(int));
Ke_val = (double *) malloc((7*Nlocal_edges+1)*sizeof(double));
Ke_bindx[0] = Nlocal_edges+1;
dcenter = 2 + 2.*sigma/((double) ( 3 * nx * nx));
doff = -1 + sigma/((double) ( 6 * nx * nx));
for (i = 0; i < Nlocal_edges; i++) {
global_id = (Edge_Partition->my_global_ids)[i];
invindex(global_id, &ii, &jj, nx, &horv);
nz_ptr = Ke_bindx[i];
Ke_val[i] = dcenter;
if (horv == HORIZONTAL) {
if (jj != 0) {
Ke_bindx[nz_ptr] = north(ii,jj,nx); Ke_val[nz_ptr++] = doff;
Ke_bindx[nz_ptr] = east(ii,jj,nx); Ke_val[nz_ptr++] = -1.;
if (ii != 0) {Ke_bindx[nz_ptr]=west(ii,jj,nx); Ke_val[nz_ptr++]= 1.;}
jj--;
}
else {
Ke_val[i] = 1. + 2.*sigma/((double) ( 3 * nx * nx));
jj = nx-1;
}
Ke_bindx[nz_ptr] = east(ii,jj,nx); Ke_val[nz_ptr++] = 1.;
if (ii != 0){ Ke_bindx[nz_ptr]=west(ii,jj,nx); Ke_val[nz_ptr++]=-1.;}
if (jj != 0){ Ke_bindx[nz_ptr]=south(ii,jj,nx); Ke_val[nz_ptr++]=doff;}
}
else {
if (ii != 0) {
Ke_bindx[nz_ptr] = north(ii,jj,nx); Ke_val[nz_ptr++] = -1.;
Ke_bindx[nz_ptr] = east(ii,jj,nx); Ke_val[nz_ptr++] = doff;
if (jj != 0) {Ke_bindx[nz_ptr]=south(ii,jj,nx); Ke_val[nz_ptr++]=1.;}
ii--;
}
else {
Ke_val[i] = 1 + 2.*sigma/((double) ( 3 * nx * nx));
ii = nx-1;
}
Ke_bindx[nz_ptr] = north(ii,jj,nx); Ke_val[nz_ptr++] = 1.;
if (ii != 0) {Ke_bindx[nz_ptr]=west(ii,jj,nx); Ke_val[nz_ptr++]=doff;}
if (jj != 0) {Ke_bindx[nz_ptr]=south(ii,jj,nx); Ke_val[nz_ptr++]=-1.;}
}
Ke_bindx[i+1] = nz_ptr;
}
AZ_set_proc_config(proc_config, COMMUNICATOR);
AZ_transform_norowreordering(proc_config, &(Edge_Partition->needed_external_ids),
Ke_bindx, Ke_val, Edge_Partition->my_global_ids,
&reordered_glob_edges, &reordered_edge_externs,
&Ke_data_org, Nlocal_edges, 0, 0, 0,
&cpntr, AZ_MSR_MATRIX);
AZ_free(reordered_glob_edges);
AZ_free(reordered_edge_externs);
Edge_Partition->Nghost = Ke_data_org[AZ_N_external];
Ke_mat = AZ_matrix_create( Nlocal_edges );
AZ_set_MSR(Ke_mat, Ke_bindx, Ke_val, Ke_data_org, 0, NULL, AZ_LOCAL);
return(Ke_mat);
}