Re: [Libmesh-devel] [Libmesh-users] Matrix free solving

[Jed B. <je...@59...>]

On Mon 2008-10-20 12:49, John Peterson wrote:
> On Mon, Oct 20, 2008 at 12:34 PM, Jed Brown <je...@59...> wrote:
> > class Matrix {
> > public:
> >  // not_implemented for all matrix operations (mult, multtranspose, etc)
> > protected:
> >  Mat A_petsc;
> >  Epetra::RowMatrix A_epetra;
> > };
> 
> I see what you are getting at, but this violates all the encapsulation
> we have worked to achieve in the linear solver interfaces up to this
> point.  I'm not saying that's intrinsically bad, it's just not the
> path we've taken.  Unless it were absolutely required, I don't think
> the amount of work required for redesigning the whole class hierarchy
> in this way is warranted.

It was just thought.  But mandating that a PetscLinearSolver needs a
PetscMatrix means that your MatShell must derive from PetscMatrix and
::mat() must return a PETSc Mat.  This means you can't use the same
implementation with a different solver.

If you care about supporting different solvers (personally I don't since
I'm happy with PETSc's interface to external packages) then you wouldn't
want your proposed hierarchy (unless I'm missing some trick).  An
alternative is to make ShellMatrix a factory, producing a PetscMatrix or
EpetraMatrix.  This is basically the same, but you'll probably end up
duplicating matrix-sized storage in the (rare) case that the matrix is
used by a PETSc solver in one place and by a Trilinos solver elsewhere.

Jed