void TpetraOperatorWrapper::apply(const Tpetra::MultiVector<ST,LO,GO,NT>& X, Tpetra::MultiVector<ST,LO,GO,NT>& Y,Teuchos::ETransp mode,ST alpha, ST beta) const
{
if (!useTranspose_)
{
// allocate space for each vector
RCP<Thyra::MultiVectorBase<ST> > tX;
RCP<Thyra::MultiVectorBase<ST> > tY;
tX = Thyra::createMembers(thyraOp_->domain(),X.getNumVectors());
tY = Thyra::createMembers(thyraOp_->range(),X.getNumVectors());
Thyra::assign(tX.ptr(),0.0);
Thyra::assign(tY.ptr(),0.0);
// copy epetra X into thyra X
mapStrategy_->copyTpetraIntoThyra(X, tX.ptr());
mapStrategy_->copyTpetraIntoThyra(Y, tY.ptr()); // if this matrix isn't block square, this probably won't work!
// perform matrix vector multiplication
thyraOp_->apply(Thyra::NOTRANS,*tX,tY.ptr(),alpha,beta);
// copy thyra Y into epetra Y
mapStrategy_->copyThyraIntoTpetra(tY, Y);
}
else
{
TEUCHOS_ASSERT(false);
}
}
void OverlappingRowMatrix<MatrixType>::applyTempl(const Tpetra::MultiVector<DomainScalar,LocalOrdinal,GlobalOrdinal,Node> &X,
Tpetra::MultiVector<RangeScalar,LocalOrdinal,GlobalOrdinal,Node> &Y,
Teuchos::ETransp mode,
RangeScalar alpha,
RangeScalar beta) const
{
// Note: This isn't AztecOO compliant. But neither was Ifpack's version.
TEUCHOS_TEST_FOR_EXCEPTION(X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::OverlappingRowMatrix::apply ERROR: X.getNumVectors() != Y.getNumVectors().");
RangeScalar zero = Teuchos::ScalarTraits<RangeScalar>::zero();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const DomainScalar> > x_ptr = X.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<RangeScalar> > y_ptr = Y.get2dViewNonConst();
Y.putScalar(zero);
size_t NumVectors = Y.getNumVectors();
for (size_t i = 0 ; i < NumMyRowsA_ ; ++i) {
size_t Nnz;
// Use this class's getrow to make the below code simpler
A_->getLocalRowCopy(i,Indices_(),Values_(),Nnz);
if (mode==Teuchos::NO_TRANS){
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][i] += (RangeScalar)Values_[j] * (RangeScalar)x_ptr[k][Indices_[j]];
}
else if (mode==Teuchos::TRANS){
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][Indices_[j]] += (RangeScalar)Values_[j] * (RangeScalar)x_ptr[k][i];
}
else { //mode==Teuchos::CONJ_TRANS
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][Indices_[j]] += Teuchos::ScalarTraits<RangeScalar>::conjugate((RangeScalar)Values_[j]) * (RangeScalar)x_ptr[k][i];
}
}
for (size_t i = 0 ; i < NumMyRowsB_ ; ++i) {
size_t Nnz;
// Use this class's getrow to make the below code simpler
ExtMatrix_->getLocalRowCopy(i,Indices_(),Values_(),Nnz);
if (mode==Teuchos::NO_TRANS){
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][NumMyRowsA_+i] += (RangeScalar)Values_[j] * (RangeScalar)x_ptr[k][Indices_[j]];
}
else if (mode==Teuchos::TRANS){
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][NumMyRowsA_+Indices_[j]] += (RangeScalar)Values_[j] * (RangeScalar)x_ptr[k][i];
}
else { //mode==Teuchos::CONJ_TRANS
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][NumMyRowsA_+Indices_[j]] += Teuchos::ScalarTraits<RangeScalar>::conjugate((RangeScalar)Values_[j]) * (RangeScalar)x_ptr[k][i];
}
}
}
void
Chebyshev<MatrixType>::
applyMat (const Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& X,
Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& Y,
Teuchos::ETransp mode) const
{
TEUCHOS_TEST_FOR_EXCEPTION(X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::Chebyshev::applyMat(): X.getNumVectors() != Y.getNumVectors().");
impl_.getMatrix ()->apply (X, Y, mode);
}
void TomBlockRelaxation<MatrixType,ContainerType>::apply(
const Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& X,
Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& Y,
Teuchos::ETransp mode,
Scalar alpha,
Scalar beta) const
{
TEUCHOS_TEST_FOR_EXCEPTION(isComputed() == false, std::runtime_error,
"Ifpack2::TomBlockRelaxation::apply ERROR: isComputed() must be true prior to calling apply.");
TEUCHOS_TEST_FOR_EXCEPTION(X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::TomBlockRelaxation::apply ERROR: X.getNumVectors() != Y.getNumVectors().");
TEUCHOS_TEST_FOR_EXCEPTION(mode != Teuchos::NO_TRANS, std::runtime_error,
"Ifpack2::TomBlockRelaxation::apply ERORR: transpose modes not supported.");
Time_->start(true);
// If X and Y are pointing to the same memory location,
// we need to create an auxiliary vector, Xcopy
Teuchos::RCP< const Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node> > Xcopy;
if (X.getLocalMV().getValues() == Y.getLocalMV().getValues())
Xcopy = Teuchos::rcp( new Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>(X) );
else
Xcopy = Teuchos::rcp( &X, false );
if (ZeroStartingSolution_)
Y.putScalar(0.0);
// Flops are updated in each of the following.
switch (PrecType_) {
case Ifpack2::JACOBI:
ApplyInverseJacobi(*Xcopy,Y);
break;
case Ifpack2::GS:
ApplyInverseGS(*Xcopy,Y);
break;
case Ifpack2::SGS:
ApplyInverseSGS(*Xcopy,Y);
break;
default:
throw std::runtime_error("Ifpack2::TomBlockRelaxation::apply internal logic error.");
}
++NumApply_;
Time_->stop();
ApplyTime_ += Time_->totalElapsedTime();
}
void ReorderFilter<MatrixType>::permuteOriginalToReorderedTempl(const Tpetra::MultiVector<DomainScalar,local_ordinal_type,global_ordinal_type,node_type> &originalX,
Tpetra::MultiVector<RangeScalar,local_ordinal_type,global_ordinal_type,node_type> &reorderedY) const
{
TEUCHOS_TEST_FOR_EXCEPTION(originalX.getNumVectors() != reorderedY.getNumVectors(), std::runtime_error,
"Ifpack2::ReorderFilter::permuteOriginalToReordered ERROR: X.getNumVectors() != Y.getNumVectors().");
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const DomainScalar> > x_ptr = originalX.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<RangeScalar> > y_ptr = reorderedY.get2dViewNonConst();
for(size_t k=0; k < originalX.getNumVectors(); k++)
for(local_ordinal_type i=0; (size_t)i< originalX.getLocalLength(); i++)
y_ptr[k][perm_[i]] = (RangeScalar)x_ptr[k][i];
}
void ReorderFilter<MatrixType>::permuteReorderedToOriginalTempl(const Tpetra::MultiVector<DomainScalar,LocalOrdinal,GlobalOrdinal,Node> &reorderedX,
Tpetra::MultiVector<RangeScalar,LocalOrdinal,GlobalOrdinal,Node> &originalY) const
{
TEUCHOS_TEST_FOR_EXCEPTION(reorderedX.getNumVectors() != originalY.getNumVectors(), std::runtime_error,
"Ifpack2::ReorderFilter::permuteReorderedToOriginal ERROR: X.getNumVectors() != Y.getNumVectors().");
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const DomainScalar> > x_ptr = reorderedX.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<RangeScalar> > y_ptr = originalY.get2dViewNonConst();
for(size_t k=0; k < reorderedX.getNumVectors(); k++)
for(LocalOrdinal i=0; (size_t)i< reorderedX.getLocalLength(); i++)
y_ptr[k][reverseperm_[i]] = (RangeScalar)x_ptr[k][i];
}
void Hiptmair<MatrixType>::
applyHiptmairSmoother(const Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& X,
Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& Y) const
{
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcpFromRef;
typedef Tpetra::MultiVector<scalar_type, local_ordinal_type,
global_ordinal_type, node_type> MV;
const scalar_type ZERO = STS::zero ();
const scalar_type ONE = STS::one ();
RCP<MV> res1 = rcp (new MV (A_->getRowMap (), X.getNumVectors ()));
RCP<MV> vec1 = rcp (new MV (A_->getRowMap (), X.getNumVectors ()));
RCP<MV> res2 = rcp (new MV (PtAP_->getRowMap (), X.getNumVectors ()));
RCP<MV> vec2 = rcp (new MV (PtAP_->getRowMap (), X.getNumVectors ()));
if (preOrPost_ == "pre" || preOrPost_ == "both") {
// apply initial relaxation to primary space
A_->apply (Y, *res1);
res1->update (ONE, X, -ONE);
vec1->putScalar (ZERO);
ifpack2_prec1_->apply (*res1, *vec1);
Y.update (ONE, *vec1, ONE);
}
// project to auxiliary space and smooth
A_->apply (Y, *res1);
res1->update (ONE, X, -ONE);
P_->apply (*res1, *res2, Teuchos::TRANS);
vec2->putScalar (ZERO);
ifpack2_prec2_->apply (*res2, *vec2);
P_->apply (*vec2, *vec1, Teuchos::NO_TRANS);
Y.update (ONE,*vec1,ONE);
if (preOrPost_ == "post" || preOrPost_ == "both") {
// smooth again on primary space
A_->apply (Y, *res1);
res1->update (ONE, X, -ONE);
vec1->putScalar (ZERO);
ifpack2_prec1_->apply (*res1, *vec1);
Y.update (ONE, *vec1, ONE);
}
}
void BorderedOperator<Scalar, LocalOrdinal, GlobalOrdinal, Node >::apply(
const Tpetra::MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node >& X,
Tpetra::MultiVector<Scalar, LocalOrdinal, GlobalOrdinal, Node >& Y,
Teuchos::ETransp mode,
Scalar coefAx,
Scalar coefY ) const
{
//bool opHasTrans = A_->hasTransposeApply();
//TEUCHOS_TEST_FOR_EXCEPTION( mode && !opHasTrans, std::runtime_error,
//"Ifpack2::BorderedOperator::apply() ERROR: The operator does not implement transpose.");
TEUCHOS_TEST_FOR_EXCEPTION(X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::BorderedOperator::apply() ERROR: X.getNumVectors() != Y.getNumVectors().");
A_->apply(X, Y, mode, coefAx, coefY );
}
void ReorderFilter<MatrixType>::
apply (const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> &X,
Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> &Y,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
typedef Teuchos::ScalarTraits<scalar_type> STS;
// Note: This isn't AztecOO compliant. But neither was Ifpack's version.
// Note: The localized maps mean the matvec is trivial (and has no import)
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::ReorderFilter::apply: X.getNumVectors() != Y.getNumVectors().");
const scalar_type zero = STS::zero ();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const scalar_type> > x_ptr = X.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<scalar_type> > y_ptr = Y.get2dViewNonConst();
Y.putScalar (zero);
const size_t NumVectors = Y.getNumVectors ();
for (size_t i = 0; i < A_->getNodeNumRows (); ++i) {
size_t Nnz;
// Use this class's getrow to make the below code simpler
getLocalRowCopy (i, Indices_ (), Values_ (), Nnz);
if (mode == Teuchos::NO_TRANS) {
for (size_t j = 0; j < Nnz; ++j) {
for (size_t k = 0; k < NumVectors; ++k) {
y_ptr[k][i] += Values_[j] * x_ptr[k][Indices_[j]];
}
}
}
else if (mode == Teuchos::TRANS) {
for (size_t j = 0; j < Nnz; ++j) {
for (size_t k = 0; k < NumVectors; ++k) {
y_ptr[k][Indices_[j]] += Values_[j] * x_ptr[k][i];
}
}
}
else { //mode==Teuchos::CONJ_TRANS
for (size_t j = 0; j < Nnz; ++j) {
for (size_t k = 0; k < NumVectors; ++k) {
y_ptr[k][Indices_[j]] += STS::conjugate(Values_[j]) * x_ptr[k][i];
}
}
}
}
}
void TomBlockRelaxation<MatrixType,ContainerType>::DoJacobi(const Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>& X, Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>& Y) const
{
size_t NumVectors = X.getNumVectors();
Scalar one=Teuchos::ScalarTraits<Scalar>::one();
// Note: Flop counts copied naively from Ifpack.
if (OverlapLevel_ == 0) {
// Non-overlapping Jacobi
for (LocalOrdinal i = 0 ; i < NumLocalBlocks_ ; i++) {
// may happen that a partition is empty
if (Containers_[i]->getNumRows() == 0) continue;
Containers_[i]->apply(X,Y,Teuchos::NO_TRANS,DampingFactor_,one);
ApplyFlops_ += NumVectors * 2 * NumGlobalRows_;
}
}
else {
// Overlapping Jacobi
for (LocalOrdinal i = 0 ; i < NumLocalBlocks_ ; i++) {
// may happen that a partition is empty
if (Containers_[i]->getNumRows() == 0) continue;
Containers_[i]->weightedApply(X,Y,*W_,Teuchos::NO_TRANS,DampingFactor_,one);
// NOTE: do not count (for simplicity) the flops due to overlapping rows
ApplyFlops_ += NumVectors * 4 * NumGlobalRows_;
}
}
}
void SingletonFilter<MatrixType>::CreateReducedRHSTempl(const Tpetra::MultiVector<DomainScalar,LocalOrdinal,GlobalOrdinal,Node>& LHS,
const Tpetra::MultiVector<RangeScalar,LocalOrdinal,GlobalOrdinal,Node>& RHS,
Tpetra::MultiVector<RangeScalar,LocalOrdinal,GlobalOrdinal,Node>& ReducedRHS)
{
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const RangeScalar > > RHS_ptr = RHS.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const DomainScalar> > LHS_ptr = LHS.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<RangeScalar> > ReducedRHS_ptr = ReducedRHS.get2dViewNonConst();
size_t NumVectors = LHS.getNumVectors();
for (size_t i = 0 ; i < NumRows_ ; ++i)
for (size_t k = 0 ; k < NumVectors ; ++k)
ReducedRHS_ptr[k][i] = RHS_ptr[k][InvReorder_[i]];
for (size_t i = 0 ; i < NumRows_ ; ++i) {
LocalOrdinal ii = InvReorder_[i];
size_t Nnz;
A_->getLocalRowCopy(ii,Indices_(),Values_(),Nnz);
for (size_t j = 0 ; j < Nnz ; ++j) {
if (Reorder_[Indices_[j]] == -1) {
for (size_t k = 0 ; k < NumVectors ; ++k)
ReducedRHS_ptr[k][i] -= (RangeScalar)Values_[j] * (RangeScalar)LHS_ptr[k][Indices_[j]];
}
}
}
}
void TomBlockRelaxation<MatrixType,ContainerType>::applyMat(
const Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& X,
Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& Y,
Teuchos::ETransp mode) const
{
TEUCHOS_TEST_FOR_EXCEPTION(isComputed() == false, std::runtime_error,
"Ifpack2::TomBlockRelaxation::applyMat() ERROR: isComputed() must be true prior to calling applyMat().");
TEUCHOS_TEST_FOR_EXCEPTION(X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::TomBlockRelaxation::applyMat() ERROR: X.getNumVectors() != Y.getNumVectors().");
A_->apply(X, Y, mode);
}
void
Chebyshev<MatrixType>::
applyMat (const Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& X,
Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& Y,
Teuchos::ETransp mode) const
{
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors () != Y.getNumVectors (), std::invalid_argument,
"Ifpack2::Chebyshev::applyMat: X.getNumVectors() != Y.getNumVectors().");
Teuchos::RCP<const row_matrix_type> A = impl_.getMatrix ();
TEUCHOS_TEST_FOR_EXCEPTION(
A.is_null (), std::runtime_error, "Ifpack2::Chebyshev::applyMat: The input "
"matrix A is null. Please call setMatrix() with a nonnull input matrix "
"before calling this method.");
A->apply (X, Y, mode);
}
void
Chebyshev<MatrixType>::
apply (const Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& X,
Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& Y,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
{
Teuchos::TimeMonitor timeMon (*Time_);
// compute() calls initialize() if it hasn't already been called.
// Thus, we only need to check isComputed().
TEUCHOS_TEST_FOR_EXCEPTION(! isComputed(), std::runtime_error,
"Ifpack2::Chebyshev::apply(): You must call the compute() method before "
"you may call apply().");
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors() != Y.getNumVectors(),
std::runtime_error,
"Ifpack2::Chebyshev::apply(): X and Y must have the same number of "
"columns. X.getNumVectors() = " << X.getNumVectors() << " != "
<< "Y.getNumVectors() = " << Y.getNumVectors() << ".");
#ifdef HAVE_TEUCHOS_DEBUG
{
// The relation 'isSameAs' is transitive. It's also a collective,
// so we don't have to do a "shared" test for exception (i.e., a
// global reduction on the test value).
TEUCHOS_TEST_FOR_EXCEPTION(
! X.getMap ()->isSameAs (*getDomainMap ()),
std::runtime_error,
"Ifpack2::Chebyshev: The domain Map of the matrix must be the same as "
"the Map of the input vector(s) X.");
TEUCHOS_TEST_FOR_EXCEPTION(
! Y.getMap ()->isSameAs (*getRangeMap ()),
std::runtime_error,
"Ifpack2::Chebyshev: The range Map of the matrix must be the same as "
"the Map of the output vector(s) Y.");
}
#endif // HAVE_TEUCHOS_DEBUG
applyImpl (X, Y, mode, alpha, beta);
}
++NumApply_;
ApplyTime_ += Time_->totalElapsedTime ();
}
void SingletonFilter<MatrixType>::UpdateLHSTempl(const Tpetra::MultiVector<DomainScalar,LocalOrdinal,GlobalOrdinal,Node>& ReducedLHS,
Tpetra::MultiVector<RangeScalar,LocalOrdinal,GlobalOrdinal,Node>& LHS)
{
Teuchos::ArrayRCP<Teuchos::ArrayRCP<RangeScalar> > LHS_ptr = LHS.get2dViewNonConst();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const DomainScalar> > ReducedLHS_ptr = ReducedLHS.get2dView();
for (size_t i = 0 ; i < NumRows_ ; ++i)
for (size_t k = 0 ; k < LHS.getNumVectors() ; ++k)
LHS_ptr[k][InvReorder_[i]] = (RangeScalar)ReducedLHS_ptr[k][i];
}
void SingletonFilter<MatrixType>::apply(const Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node> &X,
Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node> &Y,
Teuchos::ETransp mode,
Scalar alpha,
Scalar beta) const
{
// Note: This isn't AztecOO compliant. But neither was Ifpack's version.
TEUCHOS_TEST_FOR_EXCEPTION(X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::SingletonFilter::apply ERROR: X.getNumVectors() != Y.getNumVectors().");
Scalar zero = Teuchos::ScalarTraits<Scalar>::zero();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const Scalar> > x_ptr = X.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<Scalar> > y_ptr = Y.get2dViewNonConst();
Y.putScalar(zero);
size_t NumVectors = Y.getNumVectors();
for (size_t i = 0 ; i < NumRows_ ; ++i) {
size_t Nnz;
// Use this class's getrow to make the below code simpler
getLocalRowCopy(i,Indices_(),Values_(),Nnz);
if (mode==Teuchos::NO_TRANS){
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][i] += Values_[j] * x_ptr[k][Indices_[j]];
}
else if (mode==Teuchos::TRANS){
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][Indices_[j]] += Values_[j] * x_ptr[k][i];
}
else { //mode==Teuchos::CONJ_TRANS
for (size_t j = 0 ; j < Nnz ; ++j)
for (size_t k = 0 ; k < NumVectors ; ++k)
y_ptr[k][Indices_[j]] += Teuchos::ScalarTraits<Scalar>::conjugate(Values_[j]) * x_ptr[k][i];
}
}
}
void RILUK<MatrixType>::generateXY(Teuchos::ETransp mode,
const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& Xin,
const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& Yin,
Teuchos::RCP<const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> >& Xout,
Teuchos::RCP<Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> >& Yout) const {
// Generate an X and Y suitable for performing Solve() and Multiply() methods
TEUCHOS_TEST_FOR_EXCEPTION(Xin.getNumVectors()!=Yin.getNumVectors(), std::runtime_error,
"Ifpack2::RILUK::GenerateXY ERROR: X and Y not the same size");
//cout << "Xin = " << Xin << endl;
Xout = Teuchos::rcp( (const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> *) &Xin, false );
Yout = Teuchos::rcp( (Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> *) &Yin, false );
if (!isOverlapped_) return; // Nothing more to do
if (isOverlapped_) {
// Make sure the number of vectors in the multivector is the same as before.
if (OverlapX_!=Teuchos::null) {
if (OverlapX_->getNumVectors()!=Xin.getNumVectors()) {
OverlapX_ = Teuchos::null;
OverlapY_ = Teuchos::null;
}
}
if (OverlapX_==Teuchos::null) { // Need to allocate space for overlap X and Y
OverlapX_ = Teuchos::rcp( new Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>(U_->getColMap(), Xout->getNumVectors()) );
OverlapY_ = Teuchos::rcp( new Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>(L_->getRowMap(), Yout->getNumVectors()) );
}
if (mode == Teuchos::NO_TRANS) {
OverlapX_->doImport(*Xout,*U_->getGraph()->getImporter(), Tpetra::INSERT); // Import X values for solve
}
else {
OverlapX_->doImport(*Xout,*L_->getGraph()->getExporter(), Tpetra::INSERT); // Import X values for solve
}
Xout = OverlapX_;
Yout = OverlapY_; // Set pointers for Xout and Yout to point to overlap space
//cout << "OverlapX_ = " << *OverlapX_ << endl;
}
}
void
Chebyshev<MatrixType>::
apply (const Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& X,
Tpetra::MultiVector<scalar_type, local_ordinal_type, global_ordinal_type, node_type>& Y,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
const std::string timerName ("Ifpack2::Chebyshev::apply");
Teuchos::RCP<Teuchos::Time> timer = Teuchos::TimeMonitor::lookupCounter (timerName);
if (timer.is_null ()) {
timer = Teuchos::TimeMonitor::getNewCounter (timerName);
}
// Start timing here.
{
Teuchos::TimeMonitor timeMon (*timer);
// compute() calls initialize() if it hasn't already been called.
// Thus, we only need to check isComputed().
TEUCHOS_TEST_FOR_EXCEPTION(
! isComputed (), std::runtime_error,
"Ifpack2::Chebyshev::apply(): You must call the compute() method before "
"you may call apply().");
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors () != Y.getNumVectors (), std::runtime_error,
"Ifpack2::Chebyshev::apply(): X and Y must have the same number of "
"columns. X.getNumVectors() = " << X.getNumVectors() << " != "
<< "Y.getNumVectors() = " << Y.getNumVectors() << ".");
applyImpl (X, Y, mode, alpha, beta);
}
++NumApply_;
// timer->totalElapsedTime() returns the total time over all timer
// calls. Thus, we use = instead of +=.
ApplyTime_ = timer->totalElapsedTime ();
}
void SingletonFilter<MatrixType>::SolveSingletonsTempl(const Tpetra::MultiVector<DomainScalar,LocalOrdinal,GlobalOrdinal,Node>& RHS,
Tpetra::MultiVector<RangeScalar,LocalOrdinal,GlobalOrdinal,Node>& LHS)
{
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const DomainScalar> > RHS_ptr = RHS.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<RangeScalar> > LHS_ptr = LHS.get2dViewNonConst();
for (size_t i = 0 ; i < NumSingletons_ ; ++i) {
LocalOrdinal ii = SingletonIndex_[i];
// get the diagonal value for the singleton
size_t Nnz;
A_->getLocalRowCopy(ii,Indices_(),Values_(),Nnz);
for (size_t j = 0 ; j < Nnz ; ++j) {
if (Indices_[j] == ii) {
for (size_t k = 0 ; k < LHS.getNumVectors() ; ++k)
LHS_ptr[k][ii] = (RangeScalar)RHS_ptr[k][ii] / (RangeScalar)Values_[j];
}
}
}
}
void
apply(
const Tpetra::MultiVector<double,int,int> & X,
Tpetra::MultiVector<double,int,int> & Y,
Teuchos::ETransp mode = Teuchos::NO_TRANS,
double alpha = Teuchos::ScalarTraits<double>::one(),
double beta = Teuchos::ScalarTraits<double>::zero()
) const
{
for (size_t k = 0; k < Y.getNumVectors(); k++) {
const auto x_data = X.getData(k);
const auto x0_data = x0_.getData();
auto y_data = Y.getDataNonConst(k);
for (size_t i = 0; i < y_data.size(); i++) {
y_data[i] = 2 * x0_data[i] * x_data[i];
}
}
return;
}
void IdentitySolver<MatrixType>::
apply (const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& X,
Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& Y,
Teuchos::ETransp /*mode*/,
scalar_type alpha,
scalar_type beta) const
{
using Teuchos::RCP;
typedef Teuchos::ScalarTraits<scalar_type> STS;
typedef Tpetra::MultiVector<scalar_type, local_ordinal_type,
global_ordinal_type, node_type> MV;
TEUCHOS_TEST_FOR_EXCEPTION(
! isComputed (), std::runtime_error,
"Ifpack2::IdentitySolver::apply: If compute() has not yet been called, "
"or if you have changed the matrix via setMatrix(), "
"you must call compute() before you may call this method.");
// "Identity solver" does what it says: it's the identity operator.
// We have to Export if the domain and range Maps are not the same.
// Otherwise, this operator would be a permutation, not the identity.
if (export_.is_null ()) {
Y.update (alpha, X, beta);
}
else {
if (alpha == STS::one () && beta == STS::zero ()) { // the common case
Y.doExport (X, *export_, Tpetra::REPLACE);
}
else {
// We know that the domain and range Maps are compatible. First
// bring X into the range Map via Export. Then compute in place
// in Y.
MV X_tmp (Y.getMap (), Y.getNumVectors ());
X_tmp.doExport (X, *export_, Tpetra::REPLACE);
Y.update (alpha, X_tmp, beta);
}
}
++numApply_;
}
void TomBlockRelaxation<MatrixType,ContainerType>::ApplyInverseJacobi(
const Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>& X,
Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>& Y) const
{
size_t NumVectors = X.getNumVectors();
Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node> AY( Y.getMap(),NumVectors );
// Initial matvec not needed
int starting_iteration=0;
if(ZeroStartingSolution_) {
DoJacobi(X,Y);
starting_iteration=1;
}
for (int j = starting_iteration; j < NumSweeps_ ; j++) {
applyMat(Y,AY);
AY.update(1.0,X,-1.0);
DoJacobi(AY,Y);
// Flops for matrix apply & update
ApplyFlops_ += NumVectors * (2 * NumGlobalNonzeros_ + 2 * NumGlobalRows_);
}
}
void PrecPrecomputed::applyTempl(
const Tpetra::MultiVector<DomainScalar, Ordinal, Ordinal, Node>& X,
Tpetra::MultiVector<RangeScalar, Ordinal, Ordinal, Node>& Z,
Teuchos::ETransp mode,
RangeScalar alpha,
RangeScalar beta) const
{
using Teuchos::RCP;
using Teuchos::rcp;
typedef Tpetra::MultiVector<DomainScalar, Ordinal, Ordinal, Node> MV;
Teuchos::Time timer ("ILUT::apply");
{ // Timer scope for timing apply()
Teuchos::TimeMonitor timeMon (timer, true);
TEUCHOS_TEST_FOR_EXCEPTION(
! isComputed (), std::runtime_error,
"Ifpack2::ILUT::apply: You must call compute() to compute the incomplete "
"factorization, before calling apply().");
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors() != Z.getNumVectors(), std::runtime_error,
"Ifpack2::ILUT::apply: X and Y must have the same number of columns. "
"X has " << X.getNumVectors () << " columns, but Y has "
<< Z.getNumVectors () << " columns.");
TEUCHOS_TEST_FOR_EXCEPTION(
beta != STS::zero (), std::logic_error,
"Ifpack2::ILUT::apply: This method does not currently work when beta != 0.");
// If X and Y are pointing to the same memory location,
// we need to create an auxiliary vector, Xcopy
RCP<const MV> Xcopy;
if (X.getLocalMV ().getValues () == Z.getLocalMV ().getValues ()) {
Xcopy = rcp (new MV (X));
}
else {
Xcopy = rcpFromRef (X);
//ZACH always taken
}
//if (mode == Teuchos::NO_TRANS) { // Solve L U Y = X
/*
L_->template localSolve<RangeScalar,DomainScalar> (*Xcopy, Y, Teuchos::NO_TRANS);
U_->template localSolve<RangeScalar,RangeScalar> (Y, Y, Teuchos::NO_TRANS);
*/
//dump(X, "##X");
//dump(*Xcopy, "##Xcopy");
//ZACH the problem is here: This actually needs to be a global triangular solve
L_->template localSolve<RangeScalar,DomainScalar> (*Xcopy, *Y_, Teuchos::NO_TRANS);
//dump(*Y_, "##Y");
U_->template localSolve<RangeScalar,RangeScalar> (*Y_, Z, Teuchos::NO_TRANS);
//U_->template localSolve<RangeScalar,RangeScalar> (*Xcopy, Z, Teuchos::NO_TRANS);
//dump(Z, "##Z");
//U_->template localSolve<RangeScalar,DomainScalar> (*Xcopy, Y, Teuchos::NO_TRANS);
//U_->template localSolve<RangeScalar,RangeScalar> (Y, Y, Teuchos::TRANS);
//}
//else { // Solve U^* L^* Y = X
/*
U_->template localSolve<RangeScalar,DomainScalar> (*Xcopy, Y, mode);
L_->template localSolve<RangeScalar,RangeScalar> (Y, Y, mode);
*/
//}
if (alpha != STS::one ()) {
Z.scale (alpha);
}
}
++NumApply_;
ApplyTime_ += timer.totalElapsedTime ();
}
void Hiptmair<MatrixType>::
apply (const Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& X,
Tpetra::MultiVector<typename MatrixType::scalar_type,
typename MatrixType::local_ordinal_type,
typename MatrixType::global_ordinal_type,
typename MatrixType::node_type>& Y,
Teuchos::ETransp mode,
typename MatrixType::scalar_type alpha,
typename MatrixType::scalar_type beta) const
{
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcpFromRef;
typedef Tpetra::MultiVector<scalar_type, local_ordinal_type,
global_ordinal_type, node_type> MV;
TEUCHOS_TEST_FOR_EXCEPTION(
! isComputed (), std::runtime_error,
"Ifpack2::Hiptmair::apply: You must call compute() before you may call apply().");
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors () != Y.getNumVectors (), std::invalid_argument,
"Ifpack2::Hiptmair::apply: The MultiVector inputs X and Y do not have the "
"same number of columns. X.getNumVectors() = " << X.getNumVectors ()
<< " != Y.getNumVectors() = " << Y.getNumVectors () << ".");
// Catch unimplemented cases: alpha != 1, beta != 0, mode != NO_TRANS.
TEUCHOS_TEST_FOR_EXCEPTION(
alpha != STS::one (), std::logic_error,
"Ifpack2::Hiptmair::apply: alpha != 1 has not been implemented.");
TEUCHOS_TEST_FOR_EXCEPTION(
beta != STS::zero (), std::logic_error,
"Ifpack2::Hiptmair::apply: zero != 0 has not been implemented.");
TEUCHOS_TEST_FOR_EXCEPTION(
mode != Teuchos::NO_TRANS, std::logic_error,
"Ifpack2::Hiptmair::apply: mode != Teuchos::NO_TRANS has not been implemented.");
Teuchos::Time timer ("apply");
{ // The body of code to time
Teuchos::TimeMonitor timeMon (timer);
// If X and Y are pointing to the same memory location,
// we need to create an auxiliary vector, Xcopy
RCP<const MV> Xcopy;
{
auto X_lcl_host = X.template getLocalView<Kokkos::HostSpace> ();
auto Y_lcl_host = Y.template getLocalView<Kokkos::HostSpace> ();
if (X_lcl_host.ptr_on_device () == Y_lcl_host.ptr_on_device ()) {
Xcopy = rcp (new MV (X, Teuchos::Copy));
} else {
Xcopy = rcpFromRef (X);
}
}
RCP<MV> Ycopy = rcpFromRef (Y);
if (ZeroStartingSolution_) {
Ycopy->putScalar (STS::zero ());
}
// apply Hiptmair Smoothing
applyHiptmairSmoother (*Xcopy, *Ycopy);
}
++NumApply_;
ApplyTime_ += timer.totalElapsedTime ();
}
size_t findUniqueGids(
Tpetra::MultiVector<gno_t, lno_t, gno_t> &keys,
Tpetra::Vector<gno_t, lno_t, gno_t> &gids
)
{
// Input: Tpetra MultiVector of keys; key length = numVectors()
// May contain duplicate keys within a processor.
// May contain duplicate keys across processors.
// Input: Empty Tpetra Vector with same map for holding the results
// Output: Filled gids vector, containing unique global numbers for
// each unique key. Global numbers are in range [0,#UniqueKeys).
size_t num_keys = keys.getLocalLength();
size_t num_entries = keys.getNumVectors();
#ifdef HAVE_ZOLTAN2_MPI
MPI_Comm mpicomm = Teuchos::getRawMpiComm(*(keys.getMap()->getComm()));
#else
// Zoltan's siMPI will be used here
{
int flag;
MPI_Initialized(&flag);
if (!flag) {
int narg = 0;
char **argv = NULL;
MPI_Init(&narg, &argv);
}
}
MPI_Comm mpicomm = MPI_COMM_WORLD; // Will get MPI_COMM_WORLD from siMPI
#endif
int num_gid = sizeof(gno_t)/sizeof(ZOLTAN_ID_TYPE) * num_entries;
int num_user = sizeof(gno_t);
// Buffer the keys for Zoltan_DD
Teuchos::ArrayRCP<const gno_t> *tmpKeyVecs =
new Teuchos::ArrayRCP<const gno_t>[num_entries];
for (size_t v = 0; v < num_entries; v++) tmpKeyVecs[v] = keys.getData(v);
ZOLTAN_ID_PTR ddkeys = new ZOLTAN_ID_TYPE[num_gid * num_keys];
size_t idx = 0;
for (size_t i = 0; i < num_keys; i++) {
for (size_t v = 0; v < num_entries; v++) {
ZOLTAN_ID_PTR ddkey = &(ddkeys[idx]);
TPL_Traits<ZOLTAN_ID_PTR,gno_t>::ASSIGN(ddkey, tmpKeyVecs[v][i]);
idx += TPL_Traits<ZOLTAN_ID_PTR,gno_t>::NUM_ID;
}
}
delete [] tmpKeyVecs;
// Allocate memory for the result
char *ddnewgids = new char[num_user * num_keys];
// Compute the new GIDs
size_t nUnique = findUniqueGidsCommon<gno_t>(num_keys, num_gid,
ddkeys, ddnewgids, mpicomm);
// Copy the result into the output vector
gno_t *result = (gno_t *)ddnewgids;
for (size_t i = 0; i < num_keys; i++)
gids.replaceLocalValue(i, result[i]);
// Clean up
delete [] ddkeys;
delete [] ddnewgids;
return nUnique;
}
void DenseContainer<MatrixType,LocalScalarType>::
weightedApply (const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& X,
Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& Y,
const Tpetra::Vector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& D,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
using Teuchos::ArrayRCP;
using Teuchos::ArrayView;
using Teuchos::Range1D;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcp_const_cast;
using std::cerr;
using std::endl;
typedef Teuchos::ScalarTraits<scalar_type> STS;
// The local operator template parameter might have a different
// Scalar type than MatrixType. This means that we might have to
// convert X and Y to the Tpetra::MultiVector specialization that
// the local operator wants. This class' X_ and Y_ internal fields
// are of the right type for the local operator, so we can use those
// as targets.
// Tpetra::MultiVector specialization corresponding to the global operator.
typedef Tpetra::MultiVector<scalar_type,
local_ordinal_type, global_ordinal_type, node_type> global_mv_type;
// Tpetra::Vector specialization corresponding to the local
// operator. We will use this for the subset permutation of the
// diagonal scaling D.
typedef Tpetra::Vector<local_scalar_type, local_ordinal_type,
global_ordinal_type, node_type> local_vec_type;
const char prefix[] = "Ifpack2::DenseContainer::weightedApply: ";
TEUCHOS_TEST_FOR_EXCEPTION(
! IsComputed_, std::runtime_error, prefix << "You must have called the "
"compute() method before you may call this method. You may call "
"weightedApply() as many times as you want after calling compute() once, "
"but you must have called compute() at least once first.");
const size_t numVecs = X.getNumVectors ();
TEUCHOS_TEST_FOR_EXCEPTION(
numVecs != Y.getNumVectors (), std::runtime_error,
prefix << "X and Y have different numbers of vectors (columns). X has "
<< X.getNumVectors () << ", but Y has " << X.getNumVectors () << ".");
if (numVecs == 0) {
return; // done! nothing to do
}
// The local operator works on a permuted subset of the local parts
// of X and Y. The subset and permutation are defined by the index
// array returned by getLocalRows(). If the permutation is trivial
// and the subset is exactly equal to the local indices, then we
// could use the local parts of X and Y exactly, without needing to
// permute. Otherwise, we have to use temporary storage to permute
// X and Y. For now, we always use temporary storage.
//
// Create temporary permuted versions of the input and output.
// (Re)allocate X_ and/or Y_ only if necessary. We'll use them to
// store the permuted versions of X resp. Y. Note that X_local has
// the domain Map of the operator, which may be a permuted subset of
// the local Map corresponding to X.getMap(). Similarly, Y_local
// has the range Map of the operator, which may be a permuted subset
// of the local Map corresponding to Y.getMap(). numRows_ here
// gives the number of rows in the row Map of the local operator.
//
// FIXME (mfh 20 Aug 2013) There might be an implicit assumption
// here that the row Map and the range Map of that operator are
// the same.
//
// FIXME (mfh 20 Aug 2013) This "local permutation" functionality
// really belongs in Tpetra.
if (X_.is_null ()) {
X_ = rcp (new local_mv_type (localMap_, numVecs));
}
RCP<local_mv_type> X_local = X_;
TEUCHOS_TEST_FOR_EXCEPTION(
X_local->getLocalLength () != numRows_, std::logic_error,
"Ifpack2::DenseContainer::weightedApply: "
"X_local has length " << X_local->getLocalLength () << ", which does "
"not match numRows_ = " << numRows_ << ". Please report this bug to "
"the Ifpack2 developers.");
ArrayView<const local_ordinal_type> localRows = this->getLocalRows ();
Details::MultiVectorLocalGatherScatter<global_mv_type, local_mv_type> mvgs;
mvgs.gather (*X_local, X, localRows);
// We must gather the output multivector Y even on input to
// applyImpl(), since the local operator might use it as an initial
// guess for a linear solve. We have no way of knowing whether it
// does or does not.
if (Y_.is_null ()) {
Y_ = rcp (new local_mv_type (localMap_, numVecs));
}
RCP<local_mv_type> Y_local = Y_;
TEUCHOS_TEST_FOR_EXCEPTION(
Y_local->getLocalLength () != numRows_, std::logic_error,
"Ifpack2::DenseContainer::weightedApply: "
"Y_local has length " << X_local->getLocalLength () << ", which does "
"not match numRows_ = " << numRows_ << ". Please report this bug to "
"the Ifpack2 developers.");
mvgs.gather (*Y_local, Y, localRows);
// Apply the diagonal scaling D to the input X. It's our choice
// whether the result has the original input Map of X, or the
// permuted subset Map of X_local. If the latter, we also need to
// gather D into the permuted subset Map. We choose the latter, to
// save memory and computation. Thus, we do the following:
//
// 1. Gather D into a temporary vector D_local.
// 2. Create a temporary X_scaled to hold diag(D_local) * X_local.
// 3. Compute X_scaled := diag(D_loca) * X_local.
local_vec_type D_local (localMap_);
TEUCHOS_TEST_FOR_EXCEPTION(
D_local.getLocalLength () != numRows_, std::logic_error,
"Ifpack2::DenseContainer::weightedApply: "
"D_local has length " << D_local.getLocalLength () << ", which does "
"not match numRows_ = " << numRows_ << ". Please report this bug to "
"the Ifpack2 developers.");
mvgs.gather (D_local, D, localRows);
local_mv_type X_scaled (localMap_, numVecs);
X_scaled.elementWiseMultiply (STS::one (), D_local, *X_local, STS::zero ());
// Y_temp will hold the result of M^{-1}*X_scaled. If beta == 0, we
// can write the result of Inverse_->apply() directly to Y_local, so
// Y_temp may alias Y_local. Otherwise, if beta != 0, we need
// temporary storage for M^{-1}*X_scaled, so Y_temp must be
// different than Y_local.
RCP<local_mv_type> Y_temp;
if (beta == STS::zero ()) {
Y_temp = Y_local;
} else {
Y_temp = rcp (new local_mv_type (localMap_, numVecs));
}
// Apply the local operator: Y_temp := M^{-1} * X_scaled
this->applyImpl (X_scaled, *Y_temp, mode, STS::one (), STS::zero ());
// Y_local := beta * Y_local + alpha * diag(D_local) * Y_temp.
//
// Note that we still use the permuted subset scaling D_local here,
// because Y_temp has the same permuted subset Map. That's good, in
// fact, because it's a subset: less data to read and multiply.
Y_local->elementWiseMultiply (alpha, D_local, *Y_temp, beta);
// Copy the permuted subset output vector Y_local into the original
// output multivector Y.
mvgs.scatter (Y, *Y_local, localRows);
}
void DenseContainer<MatrixType, LocalScalarType>::
apply (const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& X,
Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type>& Y,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
using Teuchos::ArrayView;
using Teuchos::as;
using Teuchos::RCP;
using Teuchos::rcp;
// The local operator might have a different Scalar type than
// MatrixType. This means that we might have to convert X and Y to
// the Tpetra::MultiVector specialization that the local operator
// wants. This class' X_ and Y_ internal fields are of the right
// type for the local operator, so we can use those as targets.
// Tpetra::MultiVector specialization corresponding to the global operator.
typedef Tpetra::MultiVector<scalar_type,
local_ordinal_type, global_ordinal_type, node_type> global_mv_type;
const char prefix[] = "Ifpack2::DenseContainer::weightedApply: ";
TEUCHOS_TEST_FOR_EXCEPTION(
! IsComputed_, std::runtime_error, prefix << "You must have called the "
"compute() method before you may call this method. You may call "
"apply() as many times as you want after calling compute() once, "
"but you must have called compute() at least once first.");
const size_t numVecs = X.getNumVectors ();
TEUCHOS_TEST_FOR_EXCEPTION(
numVecs != Y.getNumVectors (), std::runtime_error,
prefix << "X and Y have different numbers of vectors (columns). X has "
<< X.getNumVectors () << ", but Y has " << X.getNumVectors () << ".");
if (numVecs == 0) {
return; // done! nothing to do
}
// The local operator works on a permuted subset of the local parts
// of X and Y. The subset and permutation are defined by the index
// array returned by getLocalRows(). If the permutation is trivial
// and the subset is exactly equal to the local indices, then we
// could use the local parts of X and Y exactly, without needing to
// permute. Otherwise, we have to use temporary storage to permute
// X and Y. For now, we always use temporary storage.
//
// Create temporary permuted versions of the input and output.
// (Re)allocate X_ and/or Y_ only if necessary. We'll use them to
// store the permuted versions of X resp. Y. Note that X_local has
// the domain Map of the operator, which may be a permuted subset of
// the local Map corresponding to X.getMap(). Similarly, Y_local
// has the range Map of the operator, which may be a permuted subset
// of the local Map corresponding to Y.getMap(). numRows_ here
// gives the number of rows in the row Map of the local Inverse_
// operator.
//
// FIXME (mfh 20 Aug 2013) There might be an implicit assumption
// here that the row Map and the range Map of that operator are
// the same.
//
// FIXME (mfh 20 Aug 2013) This "local permutation" functionality
// really belongs in Tpetra.
if (X_.is_null ()) {
X_ = rcp (new local_mv_type (localMap_, numVecs));
}
RCP<local_mv_type> X_local = X_;
TEUCHOS_TEST_FOR_EXCEPTION(
X_local->getLocalLength () != numRows_, std::logic_error,
"Ifpack2::DenseContainer::apply: "
"X_local has length " << X_local->getLocalLength () << ", which does "
"not match numRows_ = " << numRows_ << ". Please report this bug to "
"the Ifpack2 developers.");
ArrayView<const local_ordinal_type> localRows = this->getLocalRows ();
Details::MultiVectorLocalGatherScatter<global_mv_type, local_mv_type> mvgs;
mvgs.gather (*X_local, X, localRows);
// We must gather the contents of the output multivector Y even on
// input to applyImpl(), since the inverse operator might use it as
// an initial guess for a linear solve. We have no way of knowing
// whether it does or does not.
if (Y_.is_null ()) {
Y_ = rcp (new local_mv_type (localMap_, numVecs));
}
RCP<local_mv_type> Y_local = Y_;
TEUCHOS_TEST_FOR_EXCEPTION(
Y_local->getLocalLength () != numRows_, std::logic_error,
"Ifpack2::DenseContainer::apply: "
"Y_local has length " << X_local->getLocalLength () << ", which does "
"not match numRows_ = " << numRows_ << ". Please report this bug to "
"the Ifpack2 developers.");
mvgs.gather (*Y_local, Y, localRows);
// Apply the local operator:
// Y_local := beta*Y_local + alpha*M^{-1}*X_local
this->applyImpl (*X_local, *Y_local, mode, as<local_scalar_type> (alpha),
as<local_scalar_type> (beta));
// Scatter the permuted subset output vector Y_local back into the
// original output multivector Y.
mvgs.scatter (Y, *Y_local, localRows);
}
void
SupportGraph<MatrixType>::
apply (const Tpetra::MultiVector<scalar_type,
local_ordinal_type,
global_ordinal_type,
node_type>& X,
Tpetra::MultiVector<scalar_type,
local_ordinal_type,
global_ordinal_type,
node_type>& Y,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
using Teuchos::FancyOStream;
using Teuchos::getFancyOStream;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcpFromRef;
using Teuchos::Time;
using Teuchos::TimeMonitor;
typedef scalar_type DomainScalar;
typedef scalar_type RangeScalar;
typedef Tpetra::MultiVector<DomainScalar, local_ordinal_type,
global_ordinal_type, node_type> MV;
RCP<FancyOStream> out = getFancyOStream(rcpFromRef(std::cout));
// Create a timer for this method, if it doesn't exist already.
// TimeMonitor::getNewCounter registers the timer, so that
// TimeMonitor's class methods like summarize() will report the
// total time spent in successful calls to this method.
const std::string timerName ("Ifpack2::SupportGraph::apply");
RCP<Time> timer = TimeMonitor::lookupCounter(timerName);
if (timer.is_null()) {
timer = TimeMonitor::getNewCounter(timerName);
}
{ // Start timing here.
Teuchos::TimeMonitor timeMon (*timer);
TEUCHOS_TEST_FOR_EXCEPTION(
! isComputed(), std::runtime_error,
"Ifpack2::SupportGraph::apply: You must call compute() to compute the "
"incomplete factorization, before calling apply().");
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::SupportGraph::apply: X and Y must have the same number of "
"columns. X has " << X.getNumVectors() << " columns, but Y has "
<< Y.getNumVectors() << " columns.");
TEUCHOS_TEST_FOR_EXCEPTION(
beta != STS::zero(), std::logic_error,
"Ifpack2::SupportGraph::apply: This method does not currently work when "
"beta != 0.");
// If X and Y are pointing to the same memory location,
// we need to create an auxiliary vector, Xcopy
RCP<const MV> Xcopy;
if (X.getLocalMV().getValues() == Y.getLocalMV().getValues()) {
Xcopy = rcp (new MV(X));
}
else {
Xcopy = rcpFromRef(X);
}
if (alpha != STS::one()) {
Y.scale(alpha);
}
RCP<MV> Ycopy = rcpFromRef(Y);
solver_->setB(Xcopy);
solver_->setX(Ycopy);
solver_->solve ();
} // Stop timing here.
++NumApply_;
// timer->totalElapsedTime() returns the total time over all timer
// calls. Thus, we use = instead of +=.
ApplyTime_ = timer->totalElapsedTime();
}
void
OverlappingRowMatrix<MatrixType>::
apply (const Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> &X,
Tpetra::MultiVector<scalar_type,local_ordinal_type,global_ordinal_type,node_type> &Y,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
using Teuchos::ArrayRCP;
using Teuchos::as;
typedef scalar_type RangeScalar;
typedef scalar_type DomainScalar;
typedef Teuchos::ScalarTraits<RangeScalar> STRS;
TEUCHOS_TEST_FOR_EXCEPTION(
X.getNumVectors() != Y.getNumVectors(), std::runtime_error,
"Ifpack2::OverlappingRowMatrix::apply: The input X and the output Y must "
"have the same number of columns. X.getNumVectors() = "
<< X.getNumVectors() << " != Y.getNumVectors() = " << Y.getNumVectors()
<< ".");
// FIXME (mfh 13 July 2013) This would be a good candidate for a
// Kokkos local parallel operator implementation. That would
// obviate the need for getting views of the data and make the code
// below a lot simpler.
const RangeScalar zero = STRS::zero ();
ArrayRCP<ArrayRCP<const DomainScalar> > x_ptr = X.get2dView();
ArrayRCP<ArrayRCP<RangeScalar> > y_ptr = Y.get2dViewNonConst();
Y.putScalar(zero);
size_t NumVectors = Y.getNumVectors();
const size_t numMyRowsA = A_->getNodeNumRows ();
for (size_t i = 0; i < numMyRowsA; ++i) {
size_t Nnz;
// Use this class's getrow to make the below code simpler
A_->getLocalRowCopy (i, Indices_ (),Values_ (), Nnz);
if (mode == Teuchos::NO_TRANS) {
for (size_t j = 0; j < Nnz; ++j)
for (size_t k = 0; k < NumVectors; ++k)
y_ptr[k][i] += as<RangeScalar> (Values_[j]) *
as<RangeScalar> (x_ptr[k][Indices_[j]]);
}
else if (mode == Teuchos::TRANS){
for (size_t j = 0; j < Nnz; ++j)
for (size_t k = 0; k < NumVectors; ++k)
y_ptr[k][Indices_[j]] += as<RangeScalar> (Values_[j]) *
as<RangeScalar> (x_ptr[k][i]);
}
else { // mode == Teuchos::CONJ_TRANS
for (size_t j = 0; j < Nnz; ++j)
for (size_t k = 0; k < NumVectors; ++k)
y_ptr[k][Indices_[j]] +=
STRS::conjugate (as<RangeScalar> (Values_[j])) *
as<RangeScalar> (x_ptr[k][i]);
}
}
const size_t numMyRowsB = ExtMatrix_->getNodeNumRows ();
for (size_t i = 0 ; i < numMyRowsB ; ++i) {
size_t Nnz;
// Use this class's getrow to make the below code simpler
ExtMatrix_->getLocalRowCopy (i, Indices_ (), Values_ (), Nnz);
if (mode == Teuchos::NO_TRANS) {
for (size_t j = 0; j < Nnz; ++j)
for (size_t k = 0; k < NumVectors; ++k)
y_ptr[k][numMyRowsA+i] += as<RangeScalar> (Values_[j]) *
as<RangeScalar> (x_ptr[k][Indices_[j]]);
}
else if (mode == Teuchos::TRANS) {
for (size_t j = 0; j < Nnz; ++j)
for (size_t k = 0; k < NumVectors; ++k)
y_ptr[k][numMyRowsA+Indices_[j]] += as<RangeScalar> (Values_[j]) *
as<RangeScalar> (x_ptr[k][i]);
}
else { // mode == Teuchos::CONJ_TRANS
for (size_t j = 0; j < Nnz; ++j)
for (size_t k = 0; k < NumVectors; ++k)
y_ptr[k][numMyRowsA+Indices_[j]] +=
STRS::conjugate (as<RangeScalar> (Values_[j])) *
as<RangeScalar> (x_ptr[k][i]);
}
}
}
void TomBlockRelaxation<MatrixType,ContainerType>::DoSGS(const Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>& X,
Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>& Xcopy,
Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>& Y) const
{
Scalar one=Teuchos::ScalarTraits<Scalar>::one();
int Length = A_->getNodeMaxNumRowEntries();
int NumVectors = X.getNumVectors();
Teuchos::Array<Scalar> Values;
Teuchos::Array<LocalOrdinal> Indices;
Values.resize(Length);
Indices.resize(Length);
// an additonal vector is needed by parallel computations
// (note that applications through Ifpack2_AdditiveSchwarz
// are always seen are serial)
Teuchos::RCP< Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node > > Y2;
if (IsParallel_)
Y2 = Teuchos::rcp( new Tpetra::MultiVector<Scalar,LocalOrdinal,GlobalOrdinal,Node>(Importer_->getTargetMap(), NumVectors) );
else
Y2 = Teuchos::rcp( &Y, false );
Teuchos::ArrayRCP<Teuchos::ArrayRCP<const Scalar> > x_ptr = X.get2dView();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<Scalar> > xcopy_ptr = Xcopy.get2dViewNonConst();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<Scalar> > y_ptr = Y.get2dViewNonConst();
Teuchos::ArrayRCP<Teuchos::ArrayRCP<Scalar> > y2_ptr = Y2->get2dViewNonConst();
// data exchange is here, once per sweep
if (IsParallel_) Y2->doImport(Y,*Importer_,Tpetra::INSERT);
// Forward Sweep
for(LocalOrdinal i = 0 ; i < NumLocalBlocks_ ; i++) {
// may happen that a partition is empty
if (Containers_[i]->getNumRows() == 0) continue;
LocalOrdinal LID;
// update from previous block
for(size_t j = 0 ; j < Containers_[i]->getNumRows(); j++) {
LID = Containers_[i]->ID(j);
size_t NumEntries;
A_->getLocalRowCopy(LID,Indices(),Values(),NumEntries);
for (size_t k = 0 ; k < NumEntries ; k++) {
LocalOrdinal col = Indices[k];
for (int kk = 0 ; kk < NumVectors ; kk++)
xcopy_ptr[kk][LID] -= Values[k] * y2_ptr[kk][col];
}
}
// solve with this block
// Note: I'm abusing the ordering information, knowing that X/Y and Y2 have the same ordering for on-proc unknowns.
// Note: Add flop counts for inverse apply
Containers_[i]->apply(Xcopy,*Y2,Teuchos::NO_TRANS,DampingFactor_,one);
// operations for all getrow's
ApplyFlops_ += NumVectors * (2 * NumGlobalNonzeros_ + 2 * NumGlobalRows_);
}// end forward sweep
// Reverse Sweep
Xcopy = X;
for(LocalOrdinal i = NumLocalBlocks_-1; i >=0 ; i--) {
// may happen that a partition is empty
if (Containers_[i]->getNumRows() == 0) continue;
LocalOrdinal LID;
// update from previous block
for(size_t j = 0 ; j < Containers_[i]->getNumRows(); j++) {
LID = Containers_[i]->ID(j);
size_t NumEntries;
A_->getLocalRowCopy(LID,Indices(),Values(),NumEntries);
for (size_t k = 0 ; k < NumEntries ; k++) {
LocalOrdinal col = Indices[k];
for (int kk = 0 ; kk < NumVectors ; kk++)
xcopy_ptr[kk][LID] -= Values[k] * y2_ptr[kk][col];
}
}
// solve with this block
// Note: I'm abusing the ordering information, knowing that X/Y and Y2 have the same ordering for on-proc unknowns.
// Note: Add flop counts for inverse apply
Containers_[i]->apply(Xcopy,*Y2,Teuchos::NO_TRANS,DampingFactor_,one);
// operations for all getrow's
ApplyFlops_ += NumVectors * (2 * NumGlobalNonzeros_ + 2 * NumGlobalRows_);
} //end reverse sweep
// Attention: this is delicate... Not all combinations
// of Y2 and Y will always work (tough for ML it should be ok)
if (IsParallel_)
for (int m = 0 ; m < NumVectors ; ++m)
for (size_t i = 0 ; i < NumMyRows_ ; ++i)
y_ptr[m][i] = y2_ptr[m][i];
}