void sample2(struct data *Afine_data, struct data *Acoarse_data,
struct data *Rmat_data, struct data *Pmat_data,
double *sol, double *rhs)
{
ML *my_ml;
struct data fsmooth, csmooth;
int fine_grid, output_level = 10, N_grids = 2, grid0 = 0, grid1 = 1;
int Nfine, Ncoarse;
Nfine = Rmat_data->from_size;
Ncoarse = Rmat_data->to_size;
fsmooth.size = Nfine;
fsmooth.ntimes = 4;
fsmooth.processor_info = Afine_data->processor_info;
csmooth.size = Ncoarse;
csmooth.processor_info = Afine_data->processor_info;
csmooth.ntimes = 1000;
fine_grid = grid1;
ML_Create (&my_ml, N_grids);
ML_Set_OutputLevel( my_ml, output_level);
ML_Init_Amatrix (my_ml, grid1, Nfine, Nfine, (void *) Afine_data);
ML_Set_Amatrix_Matvec(my_ml, grid1, mymatvec);
ML_Init_Amatrix (my_ml, grid0, Ncoarse, Ncoarse, (void *) Acoarse_data);
ML_Set_Amatrix_Matvec(my_ml, grid0, mymatvec);
ML_Init_Restrictor(my_ml, grid1, grid0, Nfine, Ncoarse,(void *)Rmat_data);
ML_Set_Restrictor_Matvec(my_ml, grid1, myrestrict);
ML_Init_Prolongator(my_ml, grid0, grid1, Ncoarse, Nfine,(void *)Pmat_data);
ML_Set_Prolongator_Matvec(my_ml, grid0, myinterp);
ML_Set_Smoother (my_ml, grid1, ML_PRESMOOTHER, (void *)&fsmooth,mysmooth, NULL);
ML_Set_Smoother (my_ml, grid0, ML_PRESMOOTHER, (void *)&csmooth,mysmooth, NULL);
ML_Gen_Solver (my_ml, 0, fine_grid, grid0 );
ML_Iterate(my_ml, sol, rhs);
ML_Destroy(&my_ml);
}
void sample3(struct data *Afine_data, struct data *Acoarse_data,
struct data *Rmat_data, struct data *Pmat_data,
double *sol, double *rhs )
{
ML *my_ml;
double *diagonal;
int i, fine_grid, output_level = 10, N_grids = 2, grid0 = 1, grid1 = 0;
int Nfine, Ncoarse;
Nfine = Rmat_data->from_size;
Ncoarse = Rmat_data->to_size;
diagonal = (double *) malloc(Nfine*sizeof(double));
for (i = 0; i < Nfine; i++) diagonal[i] = 2.;
fine_grid = grid1;
ML_Create (&my_ml, N_grids);
ML_Set_OutputLevel(my_ml, output_level);
ML_Init_Amatrix (my_ml, grid1, Nfine, Nfine, (void *) Afine_data);
ML_Set_Amatrix_Matvec(my_ml, grid1, mymatvec );
ML_Set_Amatrix_Diag (my_ml, grid1, Nfine, diagonal);
ML_Gen_Smoother_Jacobi(my_ml, grid1, ML_PRESMOOTHER, 2, ML_DEFAULT);
ML_Init_Amatrix (my_ml, grid0, Ncoarse, Ncoarse, (void *) Acoarse_data);
ML_Set_Amatrix_Matvec(my_ml, grid0, mymatvec);
ML_Set_Amatrix_Diag (my_ml, grid0, Ncoarse, diagonal);
ML_Gen_Smoother_Jacobi(my_ml, grid0, ML_PRESMOOTHER, 200, ML_DEFAULT);
ML_Init_Prolongator(my_ml, grid0, grid1, Ncoarse, Nfine, (void*)Pmat_data);
ML_Set_Prolongator_Matvec(my_ml, grid0, myinterp);
ML_Init_Restrictor(my_ml, grid1, grid0, Nfine, Ncoarse,(void *)Rmat_data);
ML_Set_Restrictor_Matvec(my_ml, grid1, myrestrict);
ML_Gen_Solver (my_ml, 0, fine_grid, grid0);
ML_free(diagonal);
ML_Iterate(my_ml, sol, rhs);
ML_Destroy(&my_ml);
}
void sample1(struct data *Afine_data, struct data *Acoarse_data,
struct data *Rmat_data, struct data *Pmat_data,
double *sol, double *rhs )
{
ML *my_ml;
int i;
int fine_grid, output_level = 10, N_grids = 2, grid0 = 0, grid1 = 1;
int Nfine, Ncoarse;
double *diagonal;
Nfine = Rmat_data->from_size;
Ncoarse = Rmat_data->to_size;
diagonal = (double *) malloc(Nfine*sizeof(double));
for (i = 0; i < Nfine; i++) diagonal[i] = 2.;
fine_grid = grid1;
ML_Create (&my_ml, N_grids);
ML_Set_OutputLevel( my_ml, output_level);
ML_Init_Amatrix (my_ml, grid1, Nfine, Nfine,(void *) Afine_data);
ML_Set_Amatrix_Getrow(my_ml, grid1, myAgetrow, my_comm, Nfine+1);
ML_Set_Amatrix_Matvec(my_ml, grid1, mymatvec);
ML_Set_Amatrix_Diag (my_ml, grid1, Nfine, diagonal);
ML_Gen_Smoother_Jacobi(my_ml, grid1, ML_PRESMOOTHER, 2, ML_DEFAULT);
ML_Init_Prolongator(my_ml, grid0, grid1, Ncoarse,Nfine,(void *)Pmat_data);
ML_Set_Prolongator_Getrow(my_ml, grid0, myPgetrow, my_comm, Ncoarse+1);
ML_Set_Prolongator_Matvec(my_ml, grid0, myinterp);
ML_Init_Restrictor(my_ml, grid1, grid0, Nfine, Ncoarse,(void *)Rmat_data);
ML_Set_Restrictor_Getrow(my_ml, grid1, myRgetrow, my_comm, Nfine+1);
ML_Set_Restrictor_Matvec(my_ml, grid1, myrestrict);
ML_Gen_AmatrixRAP(my_ml,grid1, grid0);
#ifdef SUPERLU
ML_Gen_CoarseSolverSuperLU(my_ml, grid0);
#else
ML_Gen_Smoother_Jacobi(my_ml, grid0, ML_PRESMOOTHER, 100, ML_DEFAULT);
#endif
/* ML_Gen_Smoother_Jacobi(my_ml, grid0, ML_PRESMOOTHER, 200, ML_DEFAULT); */
/* ML_Gen_Smoother_GaussSeidel(my_ml, grid0, ML_PRESMOOTHER, 200, 1.); */
ML_Gen_Solver (my_ml, 0, fine_grid, grid0);
ML_Iterate(my_ml, sol, rhs);
ML_Destroy(&my_ml);
ML_free(diagonal);
}
int main(int argc, char *argv[]){
ML *ml_object;
int i, N_grids = 3, N_levels;
double sol[5], rhs[5];
ML_Aggregate *agg_object;
int proc, nlocal, nlocal_allcolumns;
MPI_Init(&argc,&argv);
ML_Set_PrintLevel(15);
for (i = 0; i < 5; i++) sol[i] = 0.;
for (i = 0; i < 5; i++) rhs[i] = 2.;
ML_Create (&ml_object, N_grids);
proc = ml_object->comm->ML_mypid;
if (ml_object->comm->ML_nprocs != 2) {
if (proc == 0) printf("Must be run on two processors\n");
ML_Destroy(&ml_object);
MPI_Finalize();
exit(1);
}
if (proc == 0) {nlocal = 2; nlocal_allcolumns = 4;}
else if (proc == 1){nlocal = 3; nlocal_allcolumns = 5;}
else {nlocal = 0; nlocal_allcolumns = 0;}
ML_Init_Amatrix (ml_object, 0, nlocal, nlocal, &proc);
ML_Set_Amatrix_Getrow(ml_object, 0, Poisson_getrow, Poisson_comm,
nlocal_allcolumns);
ML_Set_Amatrix_Matvec(ml_object, 0, Poisson_matvec);
ML_Aggregate_Create(&agg_object);
ML_Aggregate_Set_MaxCoarseSize(agg_object,1);
N_levels = ML_Gen_MGHierarchy_UsingAggregation(ml_object, 0,
ML_INCREASING, agg_object);
ML_Gen_Smoother_Jacobi(ml_object, ML_ALL_LEVELS, ML_PRESMOOTHER, 1, ML_DEFAULT);
ML_Gen_Solver (ml_object, ML_MGV, 0, N_levels-1);
ML_Iterate(ml_object, sol, rhs);
if (proc == 0) {
printf("sol(0) = %e\n",sol[1]);
fflush(stdout);
}
ML_Comm_GsumInt(ml_object->comm,1); /* just used for synchronization */
if (proc == 1) {
printf("sol(1) = %e\n",sol[0]);
printf("sol(2) = %e\n",sol[1]);
printf("sol(3) = %e\n",sol[2]);
fflush(stdout);
}
ML_Comm_GsumInt(ml_object->comm,1); /* just used for synchronization */
if (proc == 0) {
printf("sol(4) = %e\n",sol[0]);
fflush(stdout);
}
ML_Aggregate_Destroy(&agg_object);
ML_Destroy(&ml_object);
MPI_Finalize();
return 0;
}
int main(int argc, char *argv[])
{
int Nnodes=16*16; /* Total number of nodes in the problem.*/
/* 'Nnodes' must be a perfect square. */
int MaxMgLevels=6; /* Maximum number of Multigrid Levels */
int Nits_per_presmooth=1; /* # of pre & post smoothings per level */
double tolerance = 1.0e-8; /* At convergence: */
/* ||r_k||_2 < tolerance ||r_0||_2 */
int smoothPe_flag = ML_YES; /* ML_YES: smooth tentative prolongator */
/* ML_NO: don't smooth prolongator */
/***************************************************************************/
/* Select Hiptmair relaxation subsmoothers for the nodal and edge problems */
/* Choices include */
/* 1) ML_Gen_Smoother_SymGaussSeidel: this corresponds to a processor */
/* local version of symmetric Gauss-Seidel/SOR. The number of sweeps */
/* can be set via either 'edge_its' or 'nodal_its'. The damping can */
/* be set via 'edge_omega' or 'nodal_omega'. When set to ML_DDEFAULT, */
/* the damping is set to '1' on one processor. On multiple processors */
/* a lower damping value is set. This is needed to converge processor */
/* local SOR. */
/* 2) ML_Gen_Smoother_Cheby: this corresponds to polynomial relaxation. */
/* The degree of the polynomial is set via 'edge_its' or 'nodal_its'. */
/* If the degree is '-1', Marian Brezina's MLS polynomial is chosen. */
/* Otherwise, a Chebyshev polynomial is used over high frequencies */
/* [ lambda_max/alpha , lambda_max]. Lambda_max is computed. 'alpha' */
/* is hardwired in this example to correspond to twice the ratio of */
/* unknowns in the fine and coarse meshes. */
/* */
/* Using 'hiptmair_type' (see comments below) it is also possible to choose*/
/* when edge and nodal problems are relaxed within the Hiptmair smoother. */
/***************************************************************************/
void *edge_smoother=(void *) /* Edge relaxation: */
ML_Gen_Smoother_Cheby; /* ML_Gen_Smoother_Cheby */
/* ML_Gen_Smoother_SymGaussSeidel */
void *nodal_smoother=(void *) /* Nodal relaxation */
ML_Gen_Smoother_Cheby;/* ML_Gen_Smoother_Cheby */
/* ML_Gen_Smoother_SymGaussSeidel */
int edge_its = 3; /* Iterations or polynomial degree for */
int nodal_its = 3; /* edge/nodal subsmoothers. */
double nodal_omega = ML_DDEFAULT, /* SOR damping parameter for noda/edge */
edge_omega = ML_DDEFAULT; /* subsmoothers (see comments above). */
int hiptmair_type=HALF_HIPTMAIR;/* FULL_HIPTMAIR: each invokation */
/* smoothes on edges, then nodes, */
/* and then once again on edges. */
/* HALF_HIPTMAIR: each pre-invokation */
/* smoothes on edges, then nodes. */
/* Each post-invokation smoothes */
/* on nodes then edges. . */
ML_Operator *Tmat, *Tmat_trans, **Tmat_array, **Tmat_trans_array;
ML *ml_edges, *ml_nodes;
ML_Aggregate *ag;
int Nfine_edge, Ncoarse_edge, Nfine_node, Ncoarse_node, Nlevels;
int level, coarsest_level, itmp;
double edge_coarsening_rate, node_coarsening_rate, *rhs, *xxx;
void **edge_args, **nodal_args;
struct user_partition Edge_Partition = {NULL, NULL,0,0},
Node_Partition = {NULL, NULL,0,0};
struct Tmat_data Tmat_data;
int i, Ntotal;
ML_Comm *comm;
/* See Aztec User's Guide for information on these variables */
#ifdef AZTEC
AZ_MATRIX *Ke_mat, *Kn_mat;
AZ_PRECOND *Pmat = NULL;
int proc_config[AZ_PROC_SIZE], options[AZ_OPTIONS_SIZE];
double params[AZ_PARAMS_SIZE], status[AZ_STATUS_SIZE];
#endif
/* get processor information (proc id & # of procs) and set ML's printlevel. */
#ifdef ML_MPI
MPI_Init(&argc,&argv);
#endif
#ifdef AZTEC
AZ_set_proc_config(proc_config, COMMUNICATOR);
#endif
ML_Set_PrintLevel(10); /* set ML's output level: 0 gives least output */
/* Set the # of global nodes/edges and partition both the edges and the */
/* nodes over the processors. NOTE: I believe we assume that if an edge */
/* is assigned to a processor at least one of its nodes must be also */
/* assigned to that processor. */
Node_Partition.Nglobal = Nnodes;
Edge_Partition.Nglobal = Node_Partition.Nglobal*2;
Node_Partition.type = NODE;
Edge_Partition.type = EDGE;
#define perxodic
#ifdef periodic
Node_Partition.Nglobal += 2;
#endif
partition_edges(&Edge_Partition);
partition_nodes(&Node_Partition);
xxx = (double *) ML_allocate((Edge_Partition.Nlocal+100)*sizeof(double));
rhs = (double *) ML_allocate((Edge_Partition.Nlocal+100)*sizeof(double));
for (i = 0; i < Edge_Partition.Nlocal + 100; i++) xxx[i] = -1.;
for (i = 0; i < Edge_Partition.Nlocal; i++) xxx[i] = (double)
Edge_Partition.my_global_ids[i];
update_ghost_edges(xxx, (void *) &Edge_Partition);
/* Create an empty multigrid hierarchy and set the 'MaxMGLevels-1'th */
/* level discretization within this hierarchy to the ML matrix */
/* representing Ke (Maxwell edge discretization). */
ML_Create(&ml_edges, MaxMgLevels);
#ifdef AZTEC
/* Build Ke as an Aztec matrix. Use built-in function AZ_ML_Set_Amat() */
/* to convert to an ML matrix and put in hierarchy. */
Ke_mat = user_Ke_build(&Edge_Partition);
AZ_ML_Set_Amat(ml_edges, MaxMgLevels-1, Edge_Partition.Nlocal,
Edge_Partition.Nlocal, Ke_mat, proc_config);
#else
/* Build Ke directly as an ML matrix. */
ML_Init_Amatrix (ml_edges, MaxMgLevels-1, Edge_Partition.Nlocal,
Edge_Partition.Nlocal, &Edge_Partition);
Ntotal = Edge_Partition.Nlocal;
if (Edge_Partition.nprocs == 2) Ntotal += Edge_Partition.Nghost;
ML_Set_Amatrix_Getrow(ml_edges, MaxMgLevels-1, Ke_getrow, update_ghost_edges, Ntotal);
ML_Set_Amatrix_Matvec(ml_edges, MaxMgLevels-1, Ke_matvec);
#endif
/* Build an Aztec matrix representing an auxiliary nodal PDE problem. */
/* This should be a variable coefficient Poisson problem (with unknowns*/
/* at the nodes). The coefficients should be chosen to reflect the */
/* conductivity of the original edge problems. */
/* Create an empty multigrid hierarchy. Convert the Aztec matrix to an */
/* ML matrix and put it in the 'MaxMGLevels-1' level of the hierarchy. */
/* Note it is possible to multiply T'*T for get this matrix though this*/
/* will not incorporate material properties. */
ML_Create(&ml_nodes, MaxMgLevels);
#ifdef AZTEC
Kn_mat = user_Kn_build( &Node_Partition);
AZ_ML_Set_Amat(ml_nodes, MaxMgLevels-1, Node_Partition.Nlocal,
Node_Partition.Nlocal, Kn_mat, proc_config);
#else
ML_Init_Amatrix (ml_nodes, MaxMgLevels-1 , Node_Partition.Nlocal,
Node_Partition.Nlocal, &Node_Partition);
Ntotal = Node_Partition.Nlocal;
if (Node_Partition.nprocs == 2) Ntotal += Node_Partition.Nghost;
ML_Set_Amatrix_Getrow(ml_nodes, MaxMgLevels-1, Kn_getrow, update_ghost_nodes, Ntotal);
#endif
/* Build an ML matrix representing the null space of the PDE problem. */
/* This should be a discrete gradient (nodes to edges). */
#ifdef AZTEC
Tmat = user_T_build (&Edge_Partition, &Node_Partition,
&(ml_nodes->Amat[MaxMgLevels-1]));
#else
Tmat = ML_Operator_Create(ml_nodes->comm);
Tmat_data.edge = &Edge_Partition;
Tmat_data.node = &Node_Partition;
Tmat_data.Kn = &(ml_nodes->Amat[MaxMgLevels-1]);
ML_Operator_Set_ApplyFuncData( Tmat, Node_Partition.Nlocal,
Edge_Partition.Nlocal, ML_EMPTY, (void *) &Tmat_data,
Edge_Partition.Nlocal, NULL, 0);
ML_Operator_Set_Getrow( Tmat, ML_INTERNAL, Edge_Partition.Nlocal,Tmat_getrow);
ML_Operator_Set_ApplyFunc(Tmat, ML_INTERNAL, Tmat_matvec);
ML_Comm_Create( &comm);
ML_CommInfoOP_Generate( &(Tmat->getrow->pre_comm), update_ghost_nodes,
&Node_Partition,comm, Tmat->invec_leng,
Node_Partition.Nghost);
#endif
/********************************************************************/
/* Set some ML parameters. */
/*------------------------------------------------------------------*/
ML_Set_ResidualOutputFrequency(ml_edges, 1);
ML_Set_Tolerance(ml_edges, 1.0e-8);
ML_Aggregate_Create( &ag );
ML_Aggregate_Set_CoarsenScheme_Uncoupled(ag);
ML_Aggregate_Set_DampingFactor(ag, 0.0); /* must use 0 for maxwell */
ML_Aggregate_Set_MaxCoarseSize(ag, 30);
ML_Aggregate_Set_Threshold(ag, 0.0);
/********************************************************************/
/* Set up Tmat_trans */
/*------------------------------------------------------------------*/
Tmat_trans = ML_Operator_Create(ml_edges->comm);
ML_Operator_Transpose_byrow(Tmat, Tmat_trans);
Nlevels=ML_Gen_MGHierarchy_UsingReitzinger(ml_edges, &ml_nodes,MaxMgLevels-1,
ML_DECREASING,ag,Tmat,Tmat_trans,
&Tmat_array,&Tmat_trans_array,
smoothPe_flag, 1.5);
/* Set the Hiptmair subsmoothers */
if (nodal_smoother == (void *) ML_Gen_Smoother_SymGaussSeidel) {
nodal_args = ML_Smoother_Arglist_Create(2);
ML_Smoother_Arglist_Set(nodal_args, 0, &nodal_its);
ML_Smoother_Arglist_Set(nodal_args, 1, &nodal_omega);
}
if (edge_smoother == (void *) ML_Gen_Smoother_SymGaussSeidel) {
edge_args = ML_Smoother_Arglist_Create(2);
ML_Smoother_Arglist_Set(edge_args, 0, &edge_its);
ML_Smoother_Arglist_Set(edge_args, 1, &edge_omega);
}
if (nodal_smoother == (void *) ML_Gen_Smoother_Cheby) {
nodal_args = ML_Smoother_Arglist_Create(2);
ML_Smoother_Arglist_Set(nodal_args, 0, &nodal_its);
Nfine_node = Tmat_array[MaxMgLevels-1]->invec_leng;
Nfine_node = ML_gsum_int(Nfine_node, ml_edges->comm);
}
if (edge_smoother == (void *) ML_Gen_Smoother_Cheby) {
edge_args = ML_Smoother_Arglist_Create(2);
ML_Smoother_Arglist_Set(edge_args, 0, &edge_its);
Nfine_edge = Tmat_array[MaxMgLevels-1]->outvec_leng;
Nfine_edge = ML_gsum_int(Nfine_edge, ml_edges->comm);
}
/****************************************************
* Set up smoothers for all levels but the coarsest. *
****************************************************/
coarsest_level = MaxMgLevels - Nlevels;
for (level = MaxMgLevels-1; level > coarsest_level; level--)
{
if (edge_smoother == (void *) ML_Gen_Smoother_Cheby) {
Ncoarse_edge = Tmat_array[level-1]->outvec_leng;
Ncoarse_edge = ML_gsum_int(Ncoarse_edge, ml_edges->comm);
edge_coarsening_rate = 2.*((double) Nfine_edge)/ ((double) Ncoarse_edge);
ML_Smoother_Arglist_Set(edge_args, 1, &edge_coarsening_rate);
Nfine_edge = Ncoarse_edge;
}
if (nodal_smoother == (void *) ML_Gen_Smoother_Cheby) {
Ncoarse_node = Tmat_array[level-1]->invec_leng;
Ncoarse_node = ML_gsum_int(Ncoarse_node, ml_edges->comm);
node_coarsening_rate = 2.*((double) Nfine_node)/ ((double) Ncoarse_node);
ML_Smoother_Arglist_Set(nodal_args, 1, &node_coarsening_rate);
Nfine_node = Ncoarse_node;
}
ML_Gen_Smoother_Hiptmair(ml_edges, level, ML_BOTH, Nits_per_presmooth,
Tmat_array, Tmat_trans_array, NULL, edge_smoother,
edge_args, nodal_smoother,nodal_args, hiptmair_type);
}
/*******************************************
* Set up coarsest level smoother
*******************************************/
if (edge_smoother == (void *) ML_Gen_Smoother_Cheby) {
edge_coarsening_rate = (double) Nfine_edge;
ML_Smoother_Arglist_Set(edge_args, 1, &edge_coarsening_rate);
}
if (nodal_smoother == (void *) ML_Gen_Smoother_Cheby) {
node_coarsening_rate = (double) Nfine_node;
ML_Smoother_Arglist_Set(nodal_args,1,&node_coarsening_rate);
}
ML_Gen_CoarseSolverSuperLU( ml_edges, coarsest_level);
/* Must be called before invoking the preconditioner */
ML_Gen_Solver(ml_edges, ML_MGV, MaxMgLevels-1, coarsest_level);
/* Set the initial guess and the right hand side. Invoke solver */
xxx = (double *) ML_allocate(Edge_Partition.Nlocal*sizeof(double));
ML_random_vec(xxx, Edge_Partition.Nlocal, ml_edges->comm);
rhs = (double *) ML_allocate(Edge_Partition.Nlocal*sizeof(double));
ML_random_vec(rhs, Edge_Partition.Nlocal, ml_edges->comm);
#ifdef AZTEC
/* Choose the Aztec solver and criteria. Also tell Aztec that */
/* ML will be supplying the preconditioner. */
AZ_defaults(options, params);
options[AZ_solver] = AZ_fixed_pt;
options[AZ_solver] = AZ_gmres;
options[AZ_kspace] = 80;
params[AZ_tol] = tolerance;
AZ_set_ML_preconditioner(&Pmat, Ke_mat, ml_edges, options);
options[AZ_conv] = AZ_noscaled;
AZ_iterate(xxx, rhs, options, params, status, proc_config, Ke_mat, Pmat, NULL);
#else
ML_Iterate(ml_edges, xxx, rhs);
#endif
/* clean up. */
ML_Smoother_Arglist_Delete(&nodal_args);
ML_Smoother_Arglist_Delete(&edge_args);
ML_Aggregate_Destroy(&ag);
ML_Destroy(&ml_edges);
ML_Destroy(&ml_nodes);
#ifdef AZTEC
AZ_free((void *) Ke_mat->data_org);
AZ_free((void *) Ke_mat->val);
AZ_free((void *) Ke_mat->bindx);
if (Ke_mat != NULL) AZ_matrix_destroy(&Ke_mat);
if (Pmat != NULL) AZ_precond_destroy(&Pmat);
if (Kn_mat != NULL) AZ_matrix_destroy(&Kn_mat);
#endif
free(xxx);
free(rhs);
ML_Operator_Destroy(&Tmat);
ML_Operator_Destroy(&Tmat_trans);
ML_MGHierarchy_ReitzingerDestroy(MaxMgLevels-2, &Tmat_array, &Tmat_trans_array);
#ifdef ML_MPI
MPI_Finalize();
#endif
return 0;
}
PetscErrorCode PCSetUp_ML(PC pc)
{
PetscErrorCode ierr;
PetscMPIInt size;
FineGridCtx *PetscMLdata;
ML *ml_object;
ML_Aggregate *agg_object;
ML_Operator *mlmat;
PetscInt nlocal_allcols,Nlevels,mllevel,level,level1,m,fine_level,bs;
Mat A,Aloc;
GridCtx *gridctx;
PC_MG *mg = (PC_MG*)pc->data;
PC_ML *pc_ml = (PC_ML*)mg->innerctx;
PetscBool isSeq, isMPI;
KSP smoother;
PC subpc;
PetscInt mesh_level, old_mesh_level;
PetscFunctionBegin;
A = pc->pmat;
ierr = MPI_Comm_size(((PetscObject)A)->comm,&size);CHKERRQ(ierr);
if (pc->setupcalled) {
if (pc->flag == SAME_NONZERO_PATTERN && pc_ml->reuse_interpolation) {
/*
Reuse interpolaton instead of recomputing aggregates and updating the whole hierarchy. This is less expensive for
multiple solves in which the matrix is not changing too quickly.
*/
ml_object = pc_ml->ml_object;
gridctx = pc_ml->gridctx;
Nlevels = pc_ml->Nlevels;
fine_level = Nlevels - 1;
gridctx[fine_level].A = A;
ierr = PetscObjectTypeCompare((PetscObject) A, MATSEQAIJ, &isSeq);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject) A, MATMPIAIJ, &isMPI);CHKERRQ(ierr);
if (isMPI){
ierr = MatConvert_MPIAIJ_ML(A,PETSC_NULL,MAT_INITIAL_MATRIX,&Aloc);CHKERRQ(ierr);
} else if (isSeq) {
Aloc = A;
ierr = PetscObjectReference((PetscObject)Aloc);CHKERRQ(ierr);
} else SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.",((PetscObject)A)->type_name);
ierr = MatGetSize(Aloc,&m,&nlocal_allcols);CHKERRQ(ierr);
PetscMLdata = pc_ml->PetscMLdata;
ierr = MatDestroy(&PetscMLdata->Aloc);CHKERRQ(ierr);
PetscMLdata->A = A;
PetscMLdata->Aloc = Aloc;
ML_Init_Amatrix(ml_object,0,m,m,PetscMLdata);
ML_Set_Amatrix_Matvec(ml_object,0,PetscML_matvec);
mesh_level = ml_object->ML_finest_level;
while (ml_object->SingleLevel[mesh_level].Rmat->to) {
old_mesh_level = mesh_level;
mesh_level = ml_object->SingleLevel[mesh_level].Rmat->to->levelnum;
/* clean and regenerate A */
mlmat = &(ml_object->Amat[mesh_level]);
ML_Operator_Clean(mlmat);
ML_Operator_Init(mlmat,ml_object->comm);
ML_Gen_AmatrixRAP(ml_object, old_mesh_level, mesh_level);
}
level = fine_level - 1;
if (size == 1) { /* convert ML P, R and A into seqaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_REUSE_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
} else { /* convert ML P and R into shell format, ML A into mpiaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_MPIAIJ(mlmat,MAT_REUSE_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
}
for (level=0; level<fine_level; level++) {
if (level > 0){
ierr = PCMGSetResidual(pc,level,PCMGDefaultResidual,gridctx[level].A);CHKERRQ(ierr);
}
ierr = KSPSetOperators(gridctx[level].ksp,gridctx[level].A,gridctx[level].A,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
}
ierr = PCMGSetResidual(pc,fine_level,PCMGDefaultResidual,gridctx[fine_level].A);CHKERRQ(ierr);
ierr = KSPSetOperators(gridctx[fine_level].ksp,gridctx[level].A,gridctx[fine_level].A,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
ierr = PCSetUp_MG(pc);CHKERRQ(ierr);
PetscFunctionReturn(0);
} else {
/* since ML can change the size of vectors/matrices at any level we must destroy everything */
ierr = PCReset_ML(pc);CHKERRQ(ierr);
ierr = PCReset_MG(pc);CHKERRQ(ierr);
}
}
/* setup special features of PCML */
/*--------------------------------*/
/* covert A to Aloc to be used by ML at fine grid */
pc_ml->size = size;
ierr = PetscObjectTypeCompare((PetscObject) A, MATSEQAIJ, &isSeq);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject) A, MATMPIAIJ, &isMPI);CHKERRQ(ierr);
if (isMPI){
ierr = MatConvert_MPIAIJ_ML(A,PETSC_NULL,MAT_INITIAL_MATRIX,&Aloc);CHKERRQ(ierr);
} else if (isSeq) {
Aloc = A;
ierr = PetscObjectReference((PetscObject)Aloc);CHKERRQ(ierr);
} else SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.",((PetscObject)A)->type_name);
/* create and initialize struct 'PetscMLdata' */
ierr = PetscNewLog(pc,FineGridCtx,&PetscMLdata);CHKERRQ(ierr);
pc_ml->PetscMLdata = PetscMLdata;
ierr = PetscMalloc((Aloc->cmap->n+1)*sizeof(PetscScalar),&PetscMLdata->pwork);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_SELF,&PetscMLdata->x);CHKERRQ(ierr);
ierr = VecSetSizes(PetscMLdata->x,Aloc->cmap->n,Aloc->cmap->n);CHKERRQ(ierr);
ierr = VecSetType(PetscMLdata->x,VECSEQ);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_SELF,&PetscMLdata->y);CHKERRQ(ierr);
ierr = VecSetSizes(PetscMLdata->y,A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(PetscMLdata->y,VECSEQ);CHKERRQ(ierr);
PetscMLdata->A = A;
PetscMLdata->Aloc = Aloc;
/* create ML discretization matrix at fine grid */
/* ML requires input of fine-grid matrix. It determines nlevels. */
ierr = MatGetSize(Aloc,&m,&nlocal_allcols);CHKERRQ(ierr);
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
ML_Create(&ml_object,pc_ml->MaxNlevels);
ML_Comm_Set_UsrComm(ml_object->comm,((PetscObject)A)->comm);
pc_ml->ml_object = ml_object;
ML_Init_Amatrix(ml_object,0,m,m,PetscMLdata);
ML_Set_Amatrix_Getrow(ml_object,0,PetscML_getrow,PetscML_comm,nlocal_allcols);
ML_Set_Amatrix_Matvec(ml_object,0,PetscML_matvec);
ML_Set_Symmetrize(ml_object,pc_ml->Symmetrize ? ML_YES : ML_NO);
/* aggregation */
ML_Aggregate_Create(&agg_object);
pc_ml->agg_object = agg_object;
{
MatNullSpace mnull;
ierr = MatGetNearNullSpace(A,&mnull);CHKERRQ(ierr);
if (pc_ml->nulltype == PCML_NULLSPACE_AUTO) {
if (mnull) pc_ml->nulltype = PCML_NULLSPACE_USER;
else if (bs > 1) pc_ml->nulltype = PCML_NULLSPACE_BLOCK;
else pc_ml->nulltype = PCML_NULLSPACE_SCALAR;
}
switch (pc_ml->nulltype) {
case PCML_NULLSPACE_USER: {
PetscScalar *nullvec;
const PetscScalar *v;
PetscBool has_const;
PetscInt i,j,mlocal,nvec,M;
const Vec *vecs;
if (!mnull) SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_USER,"Must provide explicit null space using MatSetNearNullSpace() to use user-specified null space");
ierr = MatGetSize(A,&M,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetLocalSize(Aloc,&mlocal,PETSC_NULL);CHKERRQ(ierr);
ierr = MatNullSpaceGetVecs(mnull,&has_const,&nvec,&vecs);CHKERRQ(ierr);
ierr = PetscMalloc((nvec+!!has_const)*mlocal*sizeof *nullvec,&nullvec);CHKERRQ(ierr);
if (has_const) for (i=0; i<mlocal; i++) nullvec[i] = 1.0/M;
for (i=0; i<nvec; i++) {
ierr = VecGetArrayRead(vecs[i],&v);CHKERRQ(ierr);
for (j=0; j<mlocal; j++) nullvec[(i+!!has_const)*mlocal + j] = v[j];
ierr = VecRestoreArrayRead(vecs[i],&v);CHKERRQ(ierr);
}
ierr = ML_Aggregate_Set_NullSpace(agg_object,bs,nvec+!!has_const,nullvec,mlocal);CHKERRQ(ierr);
ierr = PetscFree(nullvec);CHKERRQ(ierr);
} break;
case PCML_NULLSPACE_BLOCK:
ierr = ML_Aggregate_Set_NullSpace(agg_object,bs,bs,0,0);CHKERRQ(ierr);
break;
case PCML_NULLSPACE_SCALAR:
break;
default: SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_SUP,"Unknown null space type");
}
}
ML_Aggregate_Set_MaxCoarseSize(agg_object,pc_ml->MaxCoarseSize);
/* set options */
switch (pc_ml->CoarsenScheme) {
case 1:
ML_Aggregate_Set_CoarsenScheme_Coupled(agg_object);break;
case 2:
ML_Aggregate_Set_CoarsenScheme_MIS(agg_object);break;
case 3:
ML_Aggregate_Set_CoarsenScheme_METIS(agg_object);break;
}
ML_Aggregate_Set_Threshold(agg_object,pc_ml->Threshold);
ML_Aggregate_Set_DampingFactor(agg_object,pc_ml->DampingFactor);
if (pc_ml->SpectralNormScheme_Anorm){
ML_Set_SpectralNormScheme_Anorm(ml_object);
}
agg_object->keep_agg_information = (int)pc_ml->KeepAggInfo;
agg_object->keep_P_tentative = (int)pc_ml->Reusable;
agg_object->block_scaled_SA = (int)pc_ml->BlockScaling;
agg_object->minimizing_energy = (int)pc_ml->EnergyMinimization;
agg_object->minimizing_energy_droptol = (double)pc_ml->EnergyMinimizationDropTol;
agg_object->cheap_minimizing_energy = (int)pc_ml->EnergyMinimizationCheap;
if (pc_ml->OldHierarchy) {
Nlevels = ML_Gen_MGHierarchy_UsingAggregation(ml_object,0,ML_INCREASING,agg_object);
} else {
Nlevels = ML_Gen_MultiLevelHierarchy_UsingAggregation(ml_object,0,ML_INCREASING,agg_object);
}
if (Nlevels<=0) SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_OUTOFRANGE,"Nlevels %d must > 0",Nlevels);
pc_ml->Nlevels = Nlevels;
fine_level = Nlevels - 1;
ierr = PCMGSetLevels(pc,Nlevels,PETSC_NULL);CHKERRQ(ierr);
/* set default smoothers */
for (level=1; level<=fine_level; level++){
if (size == 1){
ierr = PCMGGetSmoother(pc,level,&smoother);CHKERRQ(ierr);
ierr = KSPSetType(smoother,KSPRICHARDSON);CHKERRQ(ierr);
ierr = KSPGetPC(smoother,&subpc);CHKERRQ(ierr);
ierr = PCSetType(subpc,PCSOR);CHKERRQ(ierr);
} else {
ierr = PCMGGetSmoother(pc,level,&smoother);CHKERRQ(ierr);
ierr = KSPSetType(smoother,KSPRICHARDSON);CHKERRQ(ierr);
ierr = KSPGetPC(smoother,&subpc);CHKERRQ(ierr);
ierr = PCSetType(subpc,PCSOR);CHKERRQ(ierr);
}
}
ierr = PetscObjectOptionsBegin((PetscObject)pc);CHKERRQ(ierr);
ierr = PCSetFromOptions_MG(pc);CHKERRQ(ierr); /* should be called in PCSetFromOptions_ML(), but cannot be called prior to PCMGSetLevels() */
ierr = PetscOptionsEnd();CHKERRQ(ierr);
ierr = PetscMalloc(Nlevels*sizeof(GridCtx),&gridctx);CHKERRQ(ierr);
pc_ml->gridctx = gridctx;
/* wrap ML matrices by PETSc shell matrices at coarsened grids.
Level 0 is the finest grid for ML, but coarsest for PETSc! */
gridctx[fine_level].A = A;
level = fine_level - 1;
if (size == 1){ /* convert ML P, R and A into seqaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Pmat[mllevel]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].P);CHKERRQ(ierr);
mlmat = &(ml_object->Rmat[mllevel-1]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].R);CHKERRQ(ierr);
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
} else { /* convert ML P and R into shell format, ML A into mpiaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Pmat[mllevel]);
ierr = MatWrapML_SHELL(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].P);CHKERRQ(ierr);
mlmat = &(ml_object->Rmat[mllevel-1]);
ierr = MatWrapML_SHELL(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].R);CHKERRQ(ierr);
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_MPIAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
}
/* create vectors and ksp at all levels */
for (level=0; level<fine_level; level++){
level1 = level + 1;
ierr = VecCreate(((PetscObject)gridctx[level].A)->comm,&gridctx[level].x);CHKERRQ(ierr);
ierr = VecSetSizes(gridctx[level].x,gridctx[level].A->cmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(gridctx[level].x,VECMPI);CHKERRQ(ierr);
ierr = PCMGSetX(pc,level,gridctx[level].x);CHKERRQ(ierr);
ierr = VecCreate(((PetscObject)gridctx[level].A)->comm,&gridctx[level].b);CHKERRQ(ierr);
ierr = VecSetSizes(gridctx[level].b,gridctx[level].A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(gridctx[level].b,VECMPI);CHKERRQ(ierr);
ierr = PCMGSetRhs(pc,level,gridctx[level].b);CHKERRQ(ierr);
ierr = VecCreate(((PetscObject)gridctx[level1].A)->comm,&gridctx[level1].r);CHKERRQ(ierr);
ierr = VecSetSizes(gridctx[level1].r,gridctx[level1].A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(gridctx[level1].r,VECMPI);CHKERRQ(ierr);
ierr = PCMGSetR(pc,level1,gridctx[level1].r);CHKERRQ(ierr);
if (level == 0){
ierr = PCMGGetCoarseSolve(pc,&gridctx[level].ksp);CHKERRQ(ierr);
} else {
ierr = PCMGGetSmoother(pc,level,&gridctx[level].ksp);CHKERRQ(ierr);
}
}
ierr = PCMGGetSmoother(pc,fine_level,&gridctx[fine_level].ksp);CHKERRQ(ierr);
/* create coarse level and the interpolation between the levels */
for (level=0; level<fine_level; level++){
level1 = level + 1;
ierr = PCMGSetInterpolation(pc,level1,gridctx[level].P);CHKERRQ(ierr);
ierr = PCMGSetRestriction(pc,level1,gridctx[level].R);CHKERRQ(ierr);
if (level > 0){
ierr = PCMGSetResidual(pc,level,PCMGDefaultResidual,gridctx[level].A);CHKERRQ(ierr);
}
ierr = KSPSetOperators(gridctx[level].ksp,gridctx[level].A,gridctx[level].A,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
}
ierr = PCMGSetResidual(pc,fine_level,PCMGDefaultResidual,gridctx[fine_level].A);CHKERRQ(ierr);
ierr = KSPSetOperators(gridctx[fine_level].ksp,gridctx[level].A,gridctx[fine_level].A,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
/* setupcalled is set to 0 so that MG is setup from scratch */
pc->setupcalled = 0;
ierr = PCSetUp_MG(pc);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
HYPRE_Int HYPRE_ParCSRMLSetup( HYPRE_Solver solver, HYPRE_ParCSRMatrix A,
HYPRE_ParVector b, HYPRE_ParVector x )
{
HYPRE_Int i, my_id, nprocs, coarsest_level, level, sweeps, nlevels;
HYPRE_Int *row_partition, localEqns, length;
HYPRE_Int Nblocks, *blockList;
double wght;
MH_Context *context;
MH_Matrix *mh_mat;
/* -------------------------------------------------------- */
/* fetch the ML pointer */
/* -------------------------------------------------------- */
MH_Link *link = (MH_Link *) solver;
ML *ml = link->ml_ptr;
nlevels = link->nlevels;
/* -------------------------------------------------------- */
/* set up the parallel environment */
/* -------------------------------------------------------- */
hypre_MPI_Comm_rank(link->comm, &my_id);
hypre_MPI_Comm_size(link->comm, &nprocs);
/* -------------------------------------------------------- */
/* fetch the matrix row partition information and put it */
/* into the matrix data object (for matvec and getrow) */
/* -------------------------------------------------------- */
HYPRE_ParCSRMatrixGetRowPartitioning( A, &row_partition );
localEqns = row_partition[my_id+1] - row_partition[my_id];
context = (MH_Context *) malloc(sizeof(MH_Context));
link->contxt = context;
context->comm = link->comm;
context->globalEqns = row_partition[nprocs];
context->partition = (HYPRE_Int *) malloc(sizeof(HYPRE_Int)*(nprocs+1));
for (i=0; i<=nprocs; i++) context->partition[i] = row_partition[i];
hypre_TFree( row_partition );
mh_mat = ( MH_Matrix * ) malloc( sizeof( MH_Matrix) );
context->Amat = mh_mat;
HYPRE_ParCSRMLConstructMHMatrix(A,mh_mat,link->comm,
context->partition,context);
/* -------------------------------------------------------- */
/* set up the ML communicator information */
/* -------------------------------------------------------- */
ML_Set_Comm_Communicator(ml, link->comm);
ML_Set_Comm_MyRank(ml, my_id);
ML_Set_Comm_Nprocs(ml, nprocs);
ML_Set_Comm_Send(ml, MH_Send);
ML_Set_Comm_Recv(ml, MH_Irecv);
ML_Set_Comm_Wait(ml, MH_Wait);
/* -------------------------------------------------------- */
/* set up the ML matrix information */
/* -------------------------------------------------------- */
ML_Init_Amatrix(ml, nlevels-1, localEqns, localEqns, (void *) context);
ML_Set_Amatrix_Matvec(ml, nlevels-1, MH_MatVec);
length = localEqns;
for (i=0; i<mh_mat->recvProcCnt; i++ ) length += mh_mat->recvLeng[i];
ML_Set_Amatrix_Getrow(ml, nlevels-1, MH_GetRow, MH_ExchBdry, length);
/* -------------------------------------------------------- */
/* create an aggregate context */
/* -------------------------------------------------------- */
ML_Aggregate_Create(&(link->ml_ag));
link->ml_ag->max_levels = link->nlevels;
ML_Aggregate_Set_Threshold( link->ml_ag, link->ag_threshold );
/* -------------------------------------------------------- */
/* perform aggregation */
/* -------------------------------------------------------- */
coarsest_level = ML_Gen_MGHierarchy_UsingAggregation(ml, nlevels-1,
ML_DECREASING, link->ml_ag);
if ( my_id == 0 )
hypre_printf("ML : number of levels = %d\n", coarsest_level);
coarsest_level = nlevels - coarsest_level;
/* -------------------------------------------------------- */
/* set up smoother and coarse solver */
/* -------------------------------------------------------- */
for (level = nlevels-1; level > coarsest_level; level--)
{
sweeps = link->pre_sweeps;
wght = link->jacobi_wt;
switch ( link->pre )
{
case 0 :
ML_Gen_SmootherJacobi(ml, level, ML_PRESMOOTHER, sweeps, wght);
break;
case 1 :
ML_Gen_SmootherGaussSeidel(ml, level, ML_PRESMOOTHER, sweeps);
break;
case 2 :
ML_Gen_SmootherSymGaussSeidel(ml,level,ML_PRESMOOTHER,sweeps,1.0);
break;
case 3 :
Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level );
ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList );
ML_Gen_SmootherVBlockGaussSeidel(ml,level,ML_PRESMOOTHER,
sweeps, Nblocks, blockList);
break;
case 4 :
Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level );
ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList );
ML_Gen_SmootherVBlockJacobi(ml,level,ML_PRESMOOTHER,
sweeps, wght, Nblocks, blockList);
break;
}
sweeps = link->post_sweeps;
switch ( link->post )
{
case 0 :
ML_Gen_SmootherJacobi(ml, level, ML_POSTSMOOTHER, sweeps, wght);
break;
case 1 :
ML_Gen_SmootherGaussSeidel(ml, level, ML_POSTSMOOTHER, sweeps);
break;
case 2 :
ML_Gen_SmootherSymGaussSeidel(ml,level,ML_POSTSMOOTHER,sweeps,1.0);
break;
case 3 :
Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level );
ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList );
ML_Gen_SmootherVBlockGaussSeidel(ml,level,ML_POSTSMOOTHER,
sweeps, Nblocks, blockList);
break;
case 4 :
Nblocks = ML_Aggregate_Get_AggrCount( link->ml_ag, level );
ML_Aggregate_Get_AggrMap( link->ml_ag, level, &blockList );
ML_Gen_SmootherVBlockJacobi(ml,level,ML_POSTSMOOTHER,
sweeps, wght, Nblocks, blockList);
break;
}
}
ML_Gen_CoarseSolverSuperLU(ml, coarsest_level);
//ML_Gen_SmootherGaussSeidel(ml, coarsest_level, ML_PRESMOOTHER, 100);
ML_Gen_Solver(ml, ML_MGV, nlevels-1, coarsest_level);
return 0;
}
int main(int argc, char *argv[]){
ML *ml_object;
int i, N_grids = 3, N_levels;
double sol[129], rhs[129];
ML_Aggregate *agg_object;
ML_Operator *data;
ML_Krylov *kdata;
#ifdef ML_MPI
MPI_Init(&argc,&argv);
#endif
for (i = 0; i < 129; i++) sol[i] = 0.;
for (i = 0; i < 129; i++) rhs[i] = 2.;
ML_Create (&ml_object, N_grids);
ML_Init_Amatrix (ml_object, 0, 129, 129, NULL);
ML_Set_Amatrix_Getrow(ml_object, 0, Poisson_getrow, NULL, 129);
ML_Set_Amatrix_Matvec(ml_object, 0, Poisson_matvec);
ML_Set_PrintLevel(10);
ML_Aggregate_Create(&agg_object);
ML_Aggregate_Set_MaxCoarseSize(agg_object,1);
N_levels = ML_Gen_MGHierarchy_UsingAggregation(ml_object, 0,
ML_INCREASING, agg_object);
/******** Begin code to set a Jacobi smoother ******
ML_Gen_Smoother_Jacobi(ml_object, ML_ALL_LEVELS, ML_PRESMOOTHER, 1, ML_DEFAULT);
******** End code to set a Jacobi smoother ******/
/******** Begin code to set a user-defined smoother ******/
ML_Get_Amatrix(ml_object, 0, &data);
ML_Set_Smoother(ml_object, 0, ML_BOTH, data, user_smoothing,"mine");
ML_Get_Amatrix(ml_object, 1, &data);
ML_Set_Smoother(ml_object, 1, ML_BOTH, data, user_smoothing,"mine");
ML_Get_Amatrix(ml_object, 2, &data);
ML_Set_Smoother(ml_object, 2, ML_BOTH, data, user_smoothing,"mine");
ML_Gen_Solver (ml_object, ML_MGV, 0, N_levels-1);
/* This example uses an internal CG solver within ML */
/* ML has limited Krylov methods support. It is intended */
/* that ML be used with another package that supplies */
/* more sophisticated Krylov solver options (such as those */
/* found in the Trilinos or Aztec packages. */
kdata = ML_Krylov_Create(ml_object->comm);
ML_Krylov_Set_PrintFreq( kdata, 1 );
ML_Krylov_Set_Method(kdata, ML_CG);
ML_Krylov_Set_Amatrix(kdata, &(ml_object->Amat[0]));
ML_Krylov_Set_PreconFunc(kdata, ML_MGVSolve_Wrapper);
ML_Krylov_Set_Precon(kdata, ml_object);
ML_Krylov_Set_Tolerance(kdata, 1.e-7);
ML_Krylov_Solve(kdata, 129, rhs, sol);
ML_Krylov_Destroy( &kdata );
ML_Aggregate_Destroy(&agg_object);
ML_Destroy(&ml_object);
/******** End code to set a user-defined smoother ******/
printf("answer is %e %e %e %e %e\n",sol[0],sol[1],sol[2],sol[3],sol[4]);
#ifdef ML_MPI
MPI_Finalize();
#endif
exit(EXIT_SUCCESS);
}
