void
Stokhos::AdaptivityManager::
sumInOperator(Epetra_CrsMatrix & A,const Stokhos::AdaptivityManager::Sparse3TensorHash & Cijk,int k,const Epetra_CrsMatrix & J_k) const
{
TEUCHOS_ASSERT(J_k.NumMyRows() == int(sg_basis_row_dof_.size()));
TEUCHOS_ASSERT(J_k.NumMyCols() == int(sg_basis_col_dof_.size()));
const Teuchos::Array<double> & normValues = sg_master_basis_->norm_squared();
// loop over deterministic rows
for(int localM=0;localM<J_k.NumMyRows();localM++) {
int m = J_k.GRID(localM);
// grab row basis
Teuchos::RCP<const Stokhos::ProductBasis<int,double> > rowStochBasis
= sg_basis_row_dof_[localM];
// grab row from deterministic system
int d_numEntries;
int * d_Indices;
double * d_Values;
J_k.ExtractMyRowView(localM,d_numEntries,d_Values,d_Indices);
// loop over stochastic degrees of freedom of this row
for(int rb_i=0;rb_i<rowStochBasis->size();rb_i++) {
int i = sg_master_basis_->index(rowStochBasis->term(rb_i));
double normValue = normValues[i]; // sg_master_basis->norm_squared(i);
int sg_m = getGlobalRowId(localM,rb_i);
// we wipe out old values, capacity should gurantee
// we don't allocate more often than neccessary!
std::vector<int> sg_indices;
std::vector<double> sg_values;
// sg_indices.resize(0);
// sg_values.resize(0);
// loop over each column
for(int colInd=0;colInd<d_numEntries;colInd++) {
int localN = d_Indices[colInd]; // grab local deterministic column id
// grab row basis
Teuchos::RCP<const Stokhos::ProductBasis<int,double> > colStochBasis
= sg_basis_col_dof_[localN];
// build values array
for(int cb_j=0;cb_j<colStochBasis->size();cb_j++) {
int j = sg_master_basis_->index(colStochBasis->term(cb_j));
int sg_n = getGlobalColId(localN,cb_j);
double cijk = Cijk.getValue(i,j,k);
// no reason to work it in!
if(cijk==0) continue;
if(scaleOp_)
cijk = cijk/normValue;
sg_indices.push_back(sg_n);
sg_values.push_back(cijk*d_Values[colInd]);
}
}
// add in matrix values
A.SumIntoGlobalValues(sg_m,sg_indices.size(),&sg_values[0],&sg_indices[0]);
}
}
}
// rebuild a single subblock Epetra_CrsMatrix
void rebuildSubBlock(int i,int j,const Epetra_CrsMatrix & A,const std::vector<std::pair<int,RCP<Epetra_Map> > > & subMaps,Epetra_CrsMatrix & mat)
{
// get the number of variables families
int numVarFamily = subMaps.size();
TEUCHOS_ASSERT(i>=0 && i<numVarFamily);
TEUCHOS_ASSERT(j>=0 && j<numVarFamily);
TEUCHOS_ASSERT(mat.Filled());
const Epetra_Map & gRowMap = *subMaps[i].second;
const Epetra_Map & rowMap = *Teuchos::get_extra_data<RCP<Epetra_Map> >(subMaps[i].second,"contigMap");
int colFamilyCnt = subMaps[j].first;
// compute the number of global variables
// and the row and column block offset
int numGlobalVars = 0;
int rowBlockOffset = 0;
int colBlockOffset = 0;
for(int k=0;k<numVarFamily;k++) {
numGlobalVars += subMaps[k].first;
// compute block offsets
if(k<i) rowBlockOffset += subMaps[k].first;
if(k<j) colBlockOffset += subMaps[k].first;
}
// copy all global rows to here
Epetra_Import import(gRowMap,A.RowMap());
Epetra_CrsMatrix localA(Copy,gRowMap,0);
localA.Import(A,import,Insert);
// clear out the old matrix
mat.PutScalar(0.0);
// get entry information
int numMyRows = rowMap.NumMyElements();
int maxNumEntries = A.GlobalMaxNumEntries();
// for extraction
std::vector<int> indices(maxNumEntries);
std::vector<double> values(maxNumEntries);
// for insertion
std::vector<int> colIndices(maxNumEntries);
std::vector<double> colValues(maxNumEntries);
// insert each row into subblock
// let FillComplete handle column distribution
for(int localRow=0;localRow<numMyRows;localRow++) {
int numEntries = -1;
int globalRow = gRowMap.GID(localRow);
int contigRow = rowMap.GID(localRow);
TEUCHOS_ASSERT(globalRow>=0);
TEUCHOS_ASSERT(contigRow>=0);
// extract a global row copy
int err = localA.ExtractGlobalRowCopy(globalRow, maxNumEntries, numEntries, &values[0], &indices[0]);
TEUCHOS_ASSERT(err==0);
int numOwnedCols = 0;
for(int localCol=0;localCol<numEntries;localCol++) {
int globalCol = indices[localCol];
// determinate which block this column ID is in
int block = globalCol / numGlobalVars;
bool inFamily = true;
// test the beginning of the block
inFamily &= (block*numGlobalVars+colBlockOffset <= globalCol);
inFamily &= ((block*numGlobalVars+colBlockOffset+colFamilyCnt) > globalCol);
// is this column in the variable family
if(inFamily) {
int familyOffset = globalCol-(block*numGlobalVars+colBlockOffset);
colIndices[numOwnedCols] = block*colFamilyCnt + familyOffset;
colValues[numOwnedCols] = values[localCol];
numOwnedCols++;
}
}
// insert it into the new matrix
mat.SumIntoGlobalValues(contigRow,numOwnedCols,&colValues[0],&colIndices[0]);
}
}
int main(int argc, char *argv[])
{
#ifdef EPETRA_MPI
MPI_Init(&argc,&argv);
Epetra_MpiComm Comm(MPI_COMM_WORLD);
#else
Epetra_SerialComm Comm;
#endif
Epetra_Time Time(Comm);
// Create the linear problem using the class `Trilinos_Util::CrsMatrixGallery.'
// Various matrix examples are supported; please refer to the
// Trilinos tutorial for more details.
// create Aztec stuff
int proc_config[AZ_PROC_SIZE], options[AZ_OPTIONS_SIZE];
#ifdef ML_MPI
/* get number of processors and the name of this processor */
AZ_set_proc_config(proc_config, MPI_COMM_WORLD);
int proc = proc_config[AZ_node];
int nprocs = proc_config[AZ_N_procs];
#else
AZ_set_proc_config(proc_config, AZ_NOT_MPI);
int proc = 0;
int nprocs = 1;
#endif
// read in the matrix size
FILE *fp = fopen("ExampleMatrices/cantilever2D/data_matrix.txt","r");
int leng;
fscanf(fp,"%d",&leng);
int num_PDE_eqns=2;
int N_grid_pts = leng/num_PDE_eqns;
// make a linear distribution of the matrix respecting the blocks size
int leng1 = leng/nprocs;
int leng2 = leng-leng1*nprocs;
if (proc >= leng2)
{
leng2 += (proc*leng1);
}
else
{
leng1++;
leng2 = proc*leng1;
}
int N_update = leng1;
int* update = new int[N_update+1];
int i;
double *val=NULL;
int *bindx=NULL;
for (i=0; i<N_update; i++) update[i] = i+leng2;
// create the Epetra_CrSMatrix
Epetra_Map* StandardMap = new Epetra_Map(leng,N_update,update,0,Comm);
Epetra_CrsMatrix* A = new Epetra_CrsMatrix(Copy,*StandardMap,1);
AZ_input_msr_matrix("ExampleMatrices/cantilever2D/data_matrix.txt",
update, &val, &bindx, N_update, proc_config);
for (i=0; i<leng; i++)
{
int row = update[i];
A->SumIntoGlobalValues(row,1,&(val[i]),&row);
A->SumIntoGlobalValues(row,bindx[i+1]-bindx[i],&(val[bindx[i]]),&(bindx[bindx[i]]));
}
A->TransformToLocal();
// create solution and right-hand side (MultiVectors are fine as well)
Epetra_Vector* LHS = new Epetra_Vector(A->OperatorDomainMap());
Epetra_Vector* RHS = new Epetra_Vector(A->OperatorRangeMap());
LHS->Random();
RHS->Random();
// build the epetra linear problem
Epetra_LinearProblem Problem(A, LHS, RHS);
// Construct a solver object for this problem
AztecOO solver(Problem);
// =========================== begin of ML part ===========================
// create a parameter list for ML options
ParameterList MLList;
// set defaults for classic smoothed aggregation
ML_Epetra::SetDefaults("SA",MLList);
MLList.set("aggregation: damping factor", 0.0);
// number of relaxation sweeps
MLList.set("adaptive: max sweeps", 10);
// number of additional null space vectors to compute
MLList.set("adaptive: num vectors",2);
#if 1
ML_Epetra::MultiLevelPreconditioner* MLPrec =
new ML_Epetra::MultiLevelPreconditioner(dynamic_cast<Epetra_RowMatrix&>(*A), MLList, false);
// need to allocate and fill the null space (also the
// default one, as in this case). This vector is no longer
// needed after a call to ComputeAdaptivePreconditioner().
int NullSpaceSize = 2;
vector<double> NullSpace((NullSpaceSize*A->NumMyRows()));
for (i = 0 ; i < A->NumMyRows() ; ++i)
{
NullSpace[i] = 1.0;
++i;
NullSpace[i] = 0.0;
}
for (i = A->NumMyRows() ; i < 2*A->NumMyRows() ; ++i)
{
NullSpace[i] = 0.0;
++i;
NullSpace[i] = 1.0;
}
MLPrec->ComputeAdaptivePreconditioner(NullSpaceSize,&NullSpace[0]);
#else
ML_Epetra::MultiLevelPreconditioner* MLPrec =
new ML_Epetra::MultiLevelPreconditioner(dynamic_cast<Epetra_RowMatrix&>(*A), MLList);
#endif
// tell AztecOO to use this preconditioner, then solve
solver.SetPrecOperator(MLPrec);
// =========================== end of ML part =============================
solver.SetAztecOption(AZ_solver, AZ_gmres);
solver.SetAztecOption(AZ_output, 32);
// solve with 500 iterations and 1e-12 tolerance
solver.Iterate(1550, 1e-5);
delete MLPrec;
// compute the real residual
double residual, diff;
if( Comm.MyPID()==0 ) {
cout << "||b-Ax||_2 = " << residual << endl;
cout << "||x_exact - x||_2 = " << diff << endl;
cout << "Total Time = " << Time.ElapsedTime() << endl;
}
#ifdef EPETRA_MPI
MPI_Finalize();
#endif
return(0);
}
