dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Sun, Sep 13, 2009 at 10:23:01PM +0200, Johan Hake wrote:
> On Sunday 13 September 2009 20:32:18 Anders Logg wrote:
> > On Fri, Sep 11, 2009 at 07:49:36AM -0000, Johan Hake wrote:
> > > Blueprint changed by Johan Hake:
> > >
> > > Whiteboard changed to:
> > >
> > > Might also want LinearSystem and VariationalProblem as placeholders for
> > > A, b and a, L, bcs. There would then be three ways to solve a problem:
> > >
> > >   x = solve(A, b)
> > >   x = solver.solve(A, b)
> > >   x = LinearSystem(A, b).solve()
> > >
> > > Johan Hake:
> > > Why cannot LinearSolver and VariationalSolver act as this place holder?
> >
> > Since (A, b) defines a linear system but not a linear solver. The
> > solver attributes should be things like preconditioner, iterative
> > method, tolerances etc.
> >
> > I'm not sure we really need LinearSystem and VariationalProblem. It's
> > something Marie requested. Let's discuss it further.
>
> Ok, but if not VariationalSolver could act as a placeholder we would not need
> to distinguish between this and a LinearSolver. Then we could just overload
> the solve method for LinearSolver to also take a variational problem.

I don't think a solver should be associated with any particular system.

> Would it make sense to have a LinearSystem and a VaritionalProblem, each
> having a LinearSolver? The LinearSolver has a solve method which overloads for
> both A,b and a,L?

That makes more sense, but I think we should keep it as simple as possible:

1. LinearSolver::solve(x, A, b)

2. solve(x, A, b, parameters={})

3. LinearSystem::solve(x, parameters={})

(3) would be implemented by calling (2).

So in the simplest case one can do

  x = solve(A, b)  # Python
  solve(x, A, b)  // C++

Or send parameters to control the solver. If one needs more control or
don't like recreating the solver object in each solve, one can create
a solver object.

--
Anders

Attachment: signature.asc
Description: Digital signature