On Tue, Jan 28, 2014 at 07:49:06PM +0000, Garth N. Wells wrote: > On 2014-01-28 19:23, Patrick Farrell wrote: > >Hi all, > > > >At the highest-level Python interface, the user calls > > > >solve(F == 0, u, bcs, ..). > > > >Lower down, the nonlinear solvers in Dolfin (NewtonSolver and > >PETScSNESSolver) take in objects of type NonlinearProblem. > >The code that glues these together is NonlinearDiscreteProblem, > >which is currently nested within NonlinearVariationalSolver, > >and private. > > > >Lately I've been using some advanced nonlinear solver options and > >block preconditioning techniques that I can only express by > >accessing the underlying SNES object. To set up the solver, > >I've had to duplicate the functionality of NonlinearDiscreteProblem > >in several different codes. > > > >My question: is there any reason why it's private and hidden? > > It looks like it was intended to be just a convenience class to > allow NonlinearVariationalSolver to interface with the nonlinear > solvers. > > >May I make it public, or shall I keep copying & pasting my Python > >reimplementation around? > > > > I don't have an objection to it being spun out, but it is of limited > practical use. I'd recommend implementing the NonlinearProblem > interface. A plus of this approach is that if your system is > symmetric you can use CG or MINRES, which you can't with > NonlinearDiscreteProblem. > > Write a class derived from NonlinearProblem to avoid code duplication.
I agree. There are three interfaces: 1. Low-level for nonlinear algebraic problems: NonlinearProblem 2. High-level for nonlinear variational problems: NonlinearVariationalProblem 3. Very high-level: solve() (3) uses (2) and (2) uses (1). The private class NonlinearDiscreteProblem is the way (2) uses (1). It must implement a subclass of NonlinearProblem. So in your case, either pick (1) or (2). I don't see that there should be an interface in between. Could you instead add the functionality you need to the interface (2)? Some access functions or parameters through which you can accomplish what you need? -- Anders _______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
