dolfin team mailing list archive

["Garth N. Wells" <gnw20@xxxxxxxxx>]

On 01/06/10 10:29, Anders Logg wrote:
> On Mon, May 31, 2010 at 01:58:08PM +0100, Garth N. Wells wrote:
> > On 31/05/10 13:50, Anders Logg wrote:
> > > On Mon, May 31, 2010 at 01:17:15PM +0100, Garth N. Wells wrote:
> > > > At the moment, a linear solver is not associated with a particular
> > > > matrix. This makes it awkward and dangerous to re-use
> > > > preconditioners, factorisations, etc. I suggest that we change the
> > > > interface of the linear solver classes to accept a matrix at
> > > > construction, and that the solvers maintain a (smart) pointer to the
> > > > matrix. Opinions?
> > >
> > > It feels awkward to associate a linear solver with a specific system.
> >
> > Why is that? All our la backends do it. It seems natural to me since
> > solving a linear system involves a number of data structures which
> > are specific to the matrix.
>
> Would would the interface look like?
>
>    LinearSolver solver(A);

This is more or less how how PETSc goes about creating solvers 
(associating the matrix with the solver).

>    solver.parameters["preconditioner"] = "amg";
>
>    solver.solve(x, b); ?
>    x = solver.solve(b); ?

We would need to make sure that we avoid a copy constructor call for the 
latter.

> The other option would be to introduce a class LinearSystem:
>
>    LinearSystem system(A, b);

This is more or less how how AztecOO (Trilinos) goes about creating 
solvers, except LinearSystem is just a holder, which used to create a solver

    LinearSystem system(A, x, b);
    KrylovSolver solver(system);

>    x = system.solve();
>
> That would be analogous to VariationalProblem:
>
>    VariationalProblem problem(a, L);
>    u = problem.solve()
>
> Looks like there are three options:
>
>    1. LinearSolver / VariationalSolver
>    2. LinearSystem / VariationalProblem
>    3. Both 1 and 2

Perhaps both. LinearSystem/VariationalProblem hold data, and 
LinearSolver/VariationalSolver do something.

Garth

> --
> Anders
>
>
> > Garth
> >
> > > One option would be to have both Problem and Solver classes for linear
> > > systems.
> > >
> > > This is also related to the following blueprint that addresses the
> > > issue of the inconsistency between the interfaces for linear systems
> > > and variational problems:
> > >
> > >    https://blueprints.launchpad.net/dolfin/+spec/solver-interfaces