//==============================================================================
void AztecDVBR_Matrix::loadComplete() {
//
// This is where we call the Aztec function AZ_transform, which calculates
// communication parameters and re-orders the equations for use as a
// global distributed matrix.
//
MPI_Comm thisComm = amap_->getCommunicator();
// Sync processors.
MPI_Barrier(thisComm);
#ifndef FEI_SER
int thisProc = 0;
MPI_Comm_rank(thisComm, &thisProc);
#endif
AZ_transform(amap_->getProcConfig(), &external_, Amat_->bindx, Amat_->val,
amap_->getBlockUpdate(), &update_index_, &extern_index_, &data_org_,
N_update_, Amat_->indx, Amat_->bpntr, Amat_->rpntr,
&(Amat_->cpntr), AZ_VBR_MATRIX);
data_org_[AZ_internal_use] = 1;
Amat_->data_org = data_org_;
//On return from AZ_transform, the array update_index_ contains a mapping
//to the local re-ordering of the indices of the update_ array. Now we will
//fill the orderingUpdate array with the reverse of that mapping. i.e., a
//record of how to get back to the original ordering of the update indices.
orderingUpdate_ = new int[N_update_];
for(int ii=0; ii<N_update_; ii++)
orderingUpdate_[update_index_[ii]] = ii;
// Sync processors.
#ifndef FEI_SER
MPI_Barrier(thisComm);
#endif
setLoaded(true);
return;
}
void init_matrix_vector_structures(int proc_config[], int *update_index[], int
*update[], int *data_org[], int *external[],
int *extern_index[], int input_option, double
*val[], int *bindx[], int *indx[], int
*bpntr[], int *rpntr[], int *cpntr[])
/*
* Read in the points to be updated on this processor, create the global
* distributed form of the application matrix, and then convert it to a
* local distributed form for AZTEC kernels. Along the way, initialize the
* following quantities:
* update_index[], update[], data_org[], a[], bindx[], bpntr[], cpntr[],
* rpntr[], indx[], external[], extern_index[].
*
* Author: Ray Tuminaro, Div 1422, SNL
* Date: 3/15/95
*
* Parameters
*
* proc_config == On input, processor information:
* proc_config[AZ_node] = name of this processor
* proc_config[AZ_N_procs] = # of processors used
* update == On output, list of pts to be updated on this node
* val,bindx == On output, local distributed form of arrays
* holding matrix values
* external == On output, list of external vector elements
* update_index == On output, ordering of update and external
* extern_index == locally on this processor. For example
* 'update_index[i]' gives the index location
* of the block which has the global index
* 'update[i]'.
* data_org == On output, indicates how the data is set out 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.
* input_option == Indicates how update[] will be initialized.
* = 0, linear decomposition
* = 1, points read from file 'update'.
* = 2, box decomposition
* See AZ_read_update() comments for more details.
*
* The default finite difference MSR problem corresponds to a setting up
* a series of uncoupled 3D Poisson equations on a cube.
* To solve other problems, the call 'add_row_3D(...)' in
* 'create_msr_matrix()' can be changed to 'add_row_5pt()' or
* 'add_row_9pt()'.
*/
{
int N_update; /* Number of pts updated on this processor */
int MSRorVBR;
int chunks;
int blk_size, num_blk_cols,num_blk_rows,size,kk, convert_to_vbr = 0;
double *val2;
int *bindx2;
MSRorVBR = AZ_MSR_MATRIX;
if (application == 1) MSRorVBR = AZ_VBR_MATRIX;
chunks = num_PDE_eqns;
if (MSRorVBR == AZ_VBR_MATRIX) chunks = 1;
/* 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, chunks,
input_option);
/* create the matrix: each processor creates only the */
/* rows appearing in update[] ... however this row is */
/* created as if it were on a serial machine (i.e. using */
/* the global column numbers) */
if (application == 1)
create_vbr_matrix(*update, val, indx, N_update, rpntr, bpntr, bindx);
else {
*indx = NULL; *bpntr = NULL; *rpntr = NULL; *cpntr = NULL;
if (application == 0) create_msr_matrix(*update, val, bindx, N_update);
if (application == 2) create_fe_matrix(*update, proc_config[AZ_node],
bindx, val, N_update);
if (application == 3) {
AZ_read_msr_matrix(*update, val, bindx, N_update, proc_config);
}
}
/* convert matrix to a distributed parallel matrix */
AZ_transform(proc_config, external, *bindx, *val, *update, update_index,
extern_index, data_org, N_update, *indx, *bpntr, *rpntr, cpntr,
MSRorVBR);
if ( (convert_to_vbr == 1) && (application == 3) ) {
if (proc_config[AZ_node] == 0 ) {
printf("enter the block size\n");
scanf("%d",&blk_size);
}
AZ_broadcast((char *) &blk_size, sizeof(int), proc_config, AZ_PACK);
AZ_broadcast((char *) NULL , 0 , proc_config, AZ_SEND);
if ( N_update%blk_size != 0 ) {
(void) fprintf(stderr," The block size must be a multiple of the number of rows per processor.\n");
exit(-1);
}
num_blk_rows = N_update/blk_size;
num_blk_cols = ( (*data_org)[AZ_N_external] + N_update)/blk_size;
*cpntr = (int *) AZ_allocate( (num_blk_cols+2)*sizeof(int));
*rpntr = (int *) AZ_allocate( (num_blk_cols+2)*sizeof(int));
*bpntr = (int *) AZ_allocate( (num_blk_cols+2)*sizeof(int));
size = 20*(num_blk_cols+2);
*indx = (int *) AZ_allocate(size*sizeof(int));
bindx2 = *bindx;
val2 = *val;
*bindx = (int *) AZ_allocate(size*sizeof(int));
*val = (double *) AZ_allocate(size*blk_size*blk_size*sizeof(double));
for (kk = 0 ; kk < num_blk_cols ; kk++ ) (*cpntr)[kk] = blk_size;
AZ_msr2vbr(*val,*indx,*rpntr,*cpntr,*bpntr,*bindx,bindx2,val2,
num_blk_rows,num_blk_cols,size,size*blk_size*blk_size,blk_size);
MSRorVBR = AZ_VBR_MATRIX;
N_update /= blk_size;
num_PDE_eqns = blk_size;
for (kk = 0 ; kk < N_update ; kk++ )
(*update)[kk] = (*update)[blk_size*kk]/blk_size;
for (kk = 0 ; kk < (*data_org)[AZ_N_external] ; kk++ )
(*external)[kk] = (*external)[blk_size*kk]/blk_size;
(*data_org)[AZ_matrix_type] = AZ_VBR_MATRIX;
(*data_org)[AZ_N_int_blk ] /= blk_size;
(*data_org)[AZ_N_bord_blk] /= blk_size;
(*data_org)[AZ_N_ext_blk ] /= blk_size;
AZ_free(bindx2); AZ_free(val2);
}
} /* init_matrix_vector_structures */
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[])
{
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[])
{
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 ;
}
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);
}
