//=============================================================================
int Ifpack_IKLU::ApplyInverse(const Epetra_MultiVector& X,
Epetra_MultiVector& Y) const
{
if (!IsComputed())
IFPACK_CHK_ERR(-2); // compute preconditioner first
if (X.NumVectors() != Y.NumVectors())
IFPACK_CHK_ERR(-3); // Return error: X and Y not the same size
Time_.ResetStartTime();
// NOTE: L_ and U_ are based on SerialMap_, while Xcopy is based
// on A.Map()... which are in general different. However, Solve()
// does not seem to care... which is fine with me.
//
// AztecOO gives X and Y pointing to the same memory location,
// need to create an auxiliary vector, Xcopy and apply permutation.
vector<int> invq( NumMyRows_ );
for (int i=0; i<NumMyRows_; ++i ) {
csrnN_->perm[ csrnN_->pinv[i] ] = i;
invq[ cssS_->q[i] ] = i;
}
Teuchos::RefCountPtr<Epetra_MultiVector> Xcopy = Teuchos::rcp( new Epetra_MultiVector(X.Map(),X.NumVectors()), false );
Teuchos::RefCountPtr<Epetra_MultiVector> Ytemp = Teuchos::rcp( new Epetra_MultiVector(Y.Map(),Y.NumVectors()) );
for (int i=0; i<NumMyRows_; ++i) {
for (int j=0; j<X.NumVectors(); ++j) {
Xcopy->ReplaceMyValue( invq[i], j, (*X(j))[i] );
}
}
if (!UseTranspose_)
{
// solves LU Y = X
IFPACK_CHK_ERR(L_->Solve(false,false,false,*Xcopy,*Ytemp));
IFPACK_CHK_ERR(U_->Solve(true,false,false,*Ytemp,*Ytemp));
}
else
{
// solves U(trans) L(trans) Y = X
IFPACK_CHK_ERR(U_->Solve(true,true,false,*Xcopy,*Ytemp));
IFPACK_CHK_ERR(L_->Solve(false,true,false,*Ytemp,*Ytemp));
}
// Reverse the permutation.
for (int i=0; i<NumMyRows_; ++i) {
for (int j=0; j<Y.NumVectors(); ++j) {
Y.ReplaceMyValue( csrnN_->perm[i], j, (*(*Ytemp)(j))[i] );
}
}
++NumApplyInverse_;
#ifdef IFPACK_FLOPCOUNTERS
ApplyInverseFlops_ += X.NumVectors() * 2 * GlobalNonzeros_;
#endif
ApplyInverseTime_ += Time_.ElapsedTime();
return(0);
}
