dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Wed, Jan 19, 2005 at 05:16:32PM +0100, Johan Jansson wrote:

> > - We need to add a class NewPreconditioner that can be passed as an
> >   argument to NewGMRES, something like
> ...
> 
> Is this really the job of DOLFIN though? PETSc already has an
> infrastructure for preconditioners, shouldn't DOLFIN simply be a user
> of the linear solver system of PETSc? If something is missing in the
> way PETSc provides preconditioning, I think it makes more sense to add
> it to PETSc, and then use it in DOLFIN. Or what do you think?

Might be true. If we added a wrapper for preconditioners, the
preconditioner would probably need to work on PETSc data structures,
since it will be called from PETSc solvers.

VirtualMatrix has the same problem. The mult()-function has to take as
arguments two PETSc Vecs as arguments, since it will be called from
PETSc. However, I found it useful since it hides a large number of
PETSc calls.

Let's wait with the preconditioner until we know what's right.

/Anders