void
LocalSparseTriangularSolver<MatrixType>::
localApply (const MV& X,
MV& Y,
const Teuchos::ETransp mode,
const scalar_type& alpha,
const scalar_type& beta) const
{
using Teuchos::RCP;
typedef scalar_type ST;
typedef Teuchos::ScalarTraits<ST> STS;
if (beta == STS::zero ()) {
if (alpha == STS::zero ()) {
Y.putScalar (STS::zero ()); // Y := 0 * Y (ignore contents of Y)
}
else { // alpha != 0
A_crs_->template localSolve<ST, ST> (X, Y, mode);
if (alpha != STS::one ()) {
Y.scale (alpha);
}
}
}
else { // beta != 0
if (alpha == STS::zero ()) {
Y.scale (beta); // Y := beta * Y
}
else { // alpha != 0
MV Y_tmp (Y, Teuchos::Copy);
A_crs_->template localSolve<ST, ST> (X, Y_tmp, mode); // Y_tmp := M * X
Y.update (alpha, Y_tmp, beta); // Y := beta * Y + alpha * Y_tmp
}
}
}
void
Chebyshev<MatrixType>::
applyImpl (const MV& X,
MV& Y,
Teuchos::ETransp mode,
scalar_type alpha,
scalar_type beta) const
{
using Teuchos::ArrayRCP;
using Teuchos::as;
using Teuchos::RCP;
using Teuchos::rcp;
using Teuchos::rcp_const_cast;
using Teuchos::rcpFromRef;
const scalar_type zero = STS::zero();
const scalar_type one = STS::one();
// Y = beta*Y + alpha*M*X.
// If alpha == 0, then we don't need to do Chebyshev at all.
if (alpha == zero) {
if (beta == zero) { // Obey Sparse BLAS rules; avoid 0*NaN.
Y.putScalar (zero);
}
else {
Y.scale (beta);
}
return;
}
// If beta != 0, then we need to keep a copy of the initial value of
// Y, so that we can add beta*it to the Chebyshev result at the end.
// Usually this method is called with beta == 0, so we don't have to
// worry about caching Y_org.
RCP<MV> Y_orig;
if (beta != zero) {
Y_orig = rcp (new MV (Y));
}
// If X and Y point to the same memory location, we need to use a
// copy of X (X_copy) as the input MV. Otherwise, just let X_copy
// point to X.
//
// This is hopefully an uncommon use case, so we don't bother to
// optimize for it by caching X_copy.
RCP<const MV> X_copy;
bool copiedInput = false;
if (X.getLocalMV().getValues() == Y.getLocalMV().getValues()) {
X_copy = rcp (new MV (X));
copiedInput = true;
}
else {
X_copy = rcpFromRef (X);
}
// If alpha != 1, fold alpha into (a copy of) X.
//
// This is an uncommon use case, so we don't bother to optimize for
// it by caching X_copy. However, we do check whether we've already
// copied X above, to avoid a second copy.
if (alpha != one) {
RCP<MV> X_copy_nonConst = rcp_const_cast<MV> (X_copy);
if (! copiedInput) {
X_copy_nonConst = rcp (new MV (X));
copiedInput = true;
}
X_copy_nonConst->scale (alpha);
X_copy = rcp_const_cast<const MV> (X_copy_nonConst);
}
impl_.apply (*X_copy, Y);
if (beta != zero) {
Y.update (beta, *Y_orig, one); // Y = beta * Y_orig + 1 * Y
}
}