shared_ptr<Epetra_CrsMatrix> sparseCholesky(const Epetra_CrsMatrix &mat) {
// Note: we assume the matrix mat is symmetric and positive-definite
size_t size = mat.NumGlobalCols();
if (mat.NumGlobalRows() != size)
throw std::invalid_argument("sparseCholesky(): matrix must be square");
int *rowOffsets = 0;
int *colIndices = 0;
double *values = 0;
mat.ExtractCrsDataPointers(rowOffsets, colIndices, values);
Epetra_SerialComm comm;
Epetra_LocalMap rowMap(static_cast<int>(size), 0 /* index_base */, comm);
Epetra_LocalMap columnMap(static_cast<int>(size), 0 /* index_base */, comm);
shared_ptr<Epetra_CrsMatrix> result = boost::make_shared<Epetra_CrsMatrix>(
Copy, rowMap, columnMap, mat.GlobalMaxNumEntries());
arma::Mat<double> localMat;
arma::Mat<double> localCholesky;
std::vector<bool> processed(size, false);
for (size_t r = 0; r < size; ++r) {
if (processed[r])
continue;
int localSize = rowOffsets[r + 1] - rowOffsets[r];
localMat.set_size(localSize, localSize);
localMat.fill(0.);
localCholesky.set_size(localSize, localSize);
for (int s = 0; s < localSize; ++s) {
int row = colIndices[rowOffsets[r] + s];
for (int c = 0; c < localSize; ++c) {
int col = colIndices[rowOffsets[row] + c];
if (col != colIndices[rowOffsets[r] + c])
throw std::invalid_argument("sparseCholesky(): matrix is not "
"block-diagonal");
localMat(s, c) = values[rowOffsets[row] + c];
}
}
assert(arma::norm(localMat - localMat.t(), "fro") <
1e-12 * arma::norm(localMat, "fro"));
localCholesky = arma::chol(localMat); // localCholesky: U
for (int s = 0; s < localSize; ++s) {
int row = colIndices[rowOffsets[r] + s];
processed[row] = true;
#ifndef NDEBUG
int errorCode =
#endif
result->InsertGlobalValues(row, s + 1 /* number of values */,
localCholesky.colptr(s),
colIndices + rowOffsets[r]);
assert(errorCode == 0);
}
}
result->FillComplete(columnMap, rowMap);
return result;
}
shared_ptr<const Epetra_CrsMatrix> IndexPermutation::permutationMatrix() const
{
const int size = m_permutedIndices.size();
Epetra_SerialComm comm; // To be replaced once we begin to use MPI
Epetra_LocalMap rowMap(size, 0 /* index_base */, comm);
Epetra_LocalMap colMap(size, 0 /* index_base */, comm);
shared_ptr<Epetra_CrsMatrix> result = boost::make_shared<Epetra_CrsMatrix>(
Copy, rowMap, colMap, 1 /* one entry per row */);
const double ONE = 1.;
for (int i = 0; i < size; ++i)
result->InsertGlobalValues(m_permutedIndices[i], 1 /* one entry */,
&ONE, &i);
result->FillComplete();
return result;
}
void CrsMatrixWrapper<ST>::nullifyRowsAndCols(
const Teuchos::ArrayView<const real_t>& rowMask,
const Teuchos::ArrayView<const real_t>& colView,
ST mdv)
{
const_TrilinosMap_ptr rowMap(mat.getRowMap());
RCP<VectorType<real_t> > lclCol = rcp(new VectorType<real_t>(rowMap,
colView, colView.size(), 1));
RCP<VectorType<real_t> > gblCol = rcp(new VectorType<real_t>(
mat.getColMap(), 1));
const ImportType importer(rowMap, mat.getColMap());
gblCol->doImport(*lclCol, importer, Tpetra::INSERT);
Teuchos::ArrayRCP<const real_t> colMask(gblCol->getData(0));
const ST zero = Teuchos::ScalarTraits<ST>::zero();
resumeFill();
// Can't use OpenMP here as replaceLocalValues() is not thread-safe.
//#pragma omp parallel for
for (LO lclrow = 0; lclrow < mat.getNodeNumRows(); lclrow++) {
Teuchos::ArrayView<const LO> indices;
Teuchos::ArrayView<const ST> values;
std::vector<GO> cols;
std::vector<ST> vals;
mat.getLocalRowView(lclrow, indices, values);
GO row = rowMap->getGlobalElement(lclrow);
for (size_t c = 0; c < indices.size(); c++) {
const LO lclcol = indices[c];
const GO col = mat.getColMap()->getGlobalElement(lclcol);
if (rowMask[lclrow] > 0. || colMask[lclcol] > 0.) {
cols.push_back(lclcol);
vals.push_back(row==col ? mdv : zero);
}
}
if (cols.size() > 0)
mat.replaceLocalValues(lclrow, cols, vals);
}
fillComplete(true);
}
std::unique_ptr<DiscreteBoundaryOperator<ResultType>>
ElementaryLocalOperator<BasisFunctionType, ResultType>::
assembleWeakFormInSparseMode(LocalAssembler &assembler,
const AssemblyOptions &options) const {
#ifdef WITH_TRILINOS
if (boost::is_complex<BasisFunctionType>::value)
throw std::runtime_error(
"ElementaryLocalOperator::assembleWeakFormInSparseMode(): "
"sparse-mode assembly of identity operators for "
"complex-valued basis functions is not supported yet");
const Space<BasisFunctionType> &testSpace = *this->dualToRange();
const Space<BasisFunctionType> &trialSpace = *this->domain();
// Fill local submatrices
const GridView &view = testSpace.gridView();
const size_t elementCount = view.entityCount(0);
std::vector<int> elementIndices(elementCount);
for (size_t i = 0; i < elementCount; ++i)
elementIndices[i] = i;
std::vector<arma::Mat<ResultType>> localResult;
assembler.evaluateLocalWeakForms(elementIndices, localResult);
// Global DOF indices corresponding to local DOFs on elements
std::vector<std::vector<GlobalDofIndex>> testGdofs(elementCount);
std::vector<std::vector<GlobalDofIndex>> trialGdofs(elementCount);
std::vector<std::vector<BasisFunctionType>> testLdofWeights(elementCount);
std::vector<std::vector<BasisFunctionType>> trialLdofWeights(elementCount);
gatherGlobalDofs(testSpace, trialSpace, testGdofs, trialGdofs,
testLdofWeights, trialLdofWeights);
// Multiply matrix entries by DOF weights
for (size_t e = 0; e < elementCount; ++e)
for (size_t trialDof = 0; trialDof < trialGdofs[e].size(); ++trialDof)
for (size_t testDof = 0; testDof < testGdofs[e].size(); ++testDof)
localResult[e](testDof, trialDof) *=
conj(testLdofWeights[e][testDof]) * trialLdofWeights[e][trialDof];
// Estimate number of entries in each row
// This will be useful when we begin to use MPI
// // Get global DOF indices for which this process is responsible
// const int testGlobalDofCount = testSpace.globalDofCount();
// Epetra_Map rowMap(testGlobalDofCount, 0 /* index-base */, comm);
// std::vector<int> myTestGlobalDofs(rowMap.MyGlobalElements(),
// rowMap.MyGlobalElements() +
// rowMap.NumMyElements());
// const int myTestGlobalDofCount = myTestGlobalDofs.size();
const int testGlobalDofCount = testSpace.globalDofCount();
const int trialGlobalDofCount = trialSpace.globalDofCount();
arma::Col<int> nonzeroEntryCountEstimates(testGlobalDofCount);
nonzeroEntryCountEstimates.fill(0);
// Upper estimate for the number of global trial DOFs coupled to a given
// global test DOF: sum of the local trial DOF counts for each element that
// contributes to the global test DOF in question
for (size_t e = 0; e < elementCount; ++e)
for (size_t testLdof = 0; testLdof < testGdofs[e].size(); ++testLdof) {
int testGdof = testGdofs[e][testLdof];
if (testGdof >= 0)
nonzeroEntryCountEstimates(testGdof) += trialGdofs[e].size();
}
Epetra_SerialComm comm; // To be replaced once we begin to use MPI
Epetra_LocalMap rowMap(testGlobalDofCount, 0 /* index_base */, comm);
Epetra_LocalMap colMap(trialGlobalDofCount, 0 /* index_base */, comm);
shared_ptr<Epetra_FECrsMatrix> result =
boost::make_shared<Epetra_FECrsMatrix>(
Copy, rowMap, colMap, nonzeroEntryCountEstimates.memptr());
// TODO: make each process responsible for a subset of elements
// Find maximum number of local dofs per element
size_t maxLdofCount = 0;
for (size_t e = 0; e < elementCount; ++e)
maxLdofCount =
std::max(maxLdofCount, testGdofs[e].size() * trialGdofs[e].size());
// Initialise sparse matrix with zeros at required positions
arma::Col<double> zeros(maxLdofCount);
zeros.fill(0.);
for (size_t e = 0; e < elementCount; ++e)
result->InsertGlobalValues(testGdofs[e].size(), &testGdofs[e][0],
trialGdofs[e].size(), &trialGdofs[e][0],
zeros.memptr());
// Add contributions from individual elements
for (size_t e = 0; e < elementCount; ++e)
epetraSumIntoGlobalValues(*result, testGdofs[e], trialGdofs[e],
localResult[e]);
result->GlobalAssemble();
// If assembly mode is equal to ACA and we have AHMED,
// construct the block cluster tree. Otherwise leave it uninitialized.
typedef ClusterConstructionHelper<BasisFunctionType> CCH;
typedef AhmedDofWrapper<CoordinateType> AhmedDofType;
typedef ExtendedBemCluster<AhmedDofType> AhmedBemCluster;
typedef bbxbemblcluster<AhmedDofType, AhmedDofType> AhmedBemBlcluster;
shared_ptr<AhmedBemBlcluster> blockCluster;
shared_ptr<IndexPermutation> test_o2pPermutation, test_p2oPermutation;
shared_ptr<IndexPermutation> trial_o2pPermutation, trial_p2oPermutation;
#ifdef WITH_AHMED
if (options.assemblyMode() == AssemblyOptions::ACA) {
const AcaOptions &acaOptions = options.acaOptions();
bool indexWithGlobalDofs = acaOptions.mode != AcaOptions::HYBRID_ASSEMBLY;
typedef ClusterConstructionHelper<BasisFunctionType> CCH;
shared_ptr<AhmedBemCluster> testClusterTree;
CCH::constructBemCluster(testSpace, indexWithGlobalDofs, acaOptions,
testClusterTree, test_o2pPermutation,
test_p2oPermutation);
// TODO: construct a hermitian H-matrix if possible
shared_ptr<AhmedBemCluster> trialClusterTree;
CCH::constructBemCluster(trialSpace, indexWithGlobalDofs, acaOptions,
trialClusterTree, trial_o2pPermutation,
trial_p2oPermutation);
unsigned int blockCount = 0;
bool useStrongAdmissibilityCondition = !indexWithGlobalDofs;
blockCluster.reset(CCH::constructBemBlockCluster(
acaOptions, false /* hermitian */, *testClusterTree, *trialClusterTree,
useStrongAdmissibilityCondition, blockCount).release());
}
#endif
// Create and return a discrete operator represented by the matrix that
// has just been calculated
return std::unique_ptr<DiscreteBoundaryOperator<ResultType>>(
new DiscreteSparseBoundaryOperator<ResultType>(
result, this->symmetry(), NO_TRANSPOSE, blockCluster,
trial_o2pPermutation, test_o2pPermutation));
#else // WITH_TRILINOS
throw std::runtime_error(
"ElementaryLocalOperator::assembleWeakFormInSparseMode(): "
"To enable assembly in sparse mode, recompile BEM++ "
"with the symbol WITH_TRILINOS defined.");
#endif
}
