I am pushing to petsc-3.3 and petsc-dev for MatConjugate to be a no-op (like VecConjugate already is) in the case of complex numbers so you do not need the #ifdef below Will be there in minutes Barry On Jul 18, 2012, at 1:42 PM, Jose E. Roman wrote:

El 18/07/2012, a las 20:07, Jed Brown escribió:

...

On Wed, Jul 18, 2012 at 11:09 AM, Jose E. Roman <[email protected]> wrote: Er... well, now that I see better: it is not always safe, it will error out if mat->ops->conjugate is not available, for instance in BAIJ.

There is MatMultHermitianTranspose_SeqBAIJ so that would only an issue for parallel BAIJ, but you were just going to use VecConjugate, right? In Matlab notation, A'*x = (x'*A)', so you only need VecConjugate.

No, the problem with mat->ops->conjugate is in the "explicit" transpose case. For instance this code: ierr = MatTranspose(svd->OP,MAT_INITIAL_MATRIX,&svd->A);CHKERRQ(ierr); #if defined(PETSC_USE_COMPLEX) ierr = MatConjugate(svd->A);CHKERRQ(ierr); #endif

You were suggesting to remove the #ifdef, right? What I was saying is that in real scalars the MatConjugate operation would fail in some matrix types (parallel BAIJ), but this operation is only necessary in complex scalars. So I will not remove the #ifdef

Jose