dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Sun, Jul 10, 2011 at 10:56:02AM +0100, Garth N. Wells wrote:
>
>
> On 10/07/11 10:08, Anders Logg wrote:
> > On Thu, Jul 07, 2011 at 02:44:20PM -0700, Johan Hake wrote:
> >> On Thursday July 7 2011 12:21:26 Anders Logg wrote:
> >>> On Thu, Jul 07, 2011 at 02:20:44PM +0200, Marie E. Rognes wrote:
> >>>> Is the plan for 1.0-beta to fix
> >>>>
> >>>> https://bugs.launchpad.net/ffc/+bug/787010
> >>>>
> >>>> and then release?
> >>>
> >>> Yes + decide on the interface for NonlinearVariationalProblem.
> >>>
> >>> I think that should be all.
> >>>
> >>> It would be good to hear more comments on the two suggestions:
> >>>
> >>> 1. (current)
> >>>
> >>>   NonlinearVariationalProblen(lhs, rhs, u, bcs, [J])
> >>>
> >>>   This is consistent with LinearVariationalProblem and the solve()
> >>>   functions; same order of arguments.
> >>>
> >>> 2. (Garth)
> >>>
> >>>   NonlinearVariationalProblen(lhs, u, bcs, [J])
> >>>
> >>>   This removes the unnecessary rhs argument which always has to be
> >>>   zero.
> >>>
> >>> I think there are good arguments for both but not very strong so it's
> >>> a matter of taste.
> >>
> >> If:
> >>
> >>   The point is that it makes the interface for all variational problems
> >>   (linear or nonlinear) the same:
> >>
> >> is the only reason, I go with Garth.
> >
> > Another reason is complications in the implementation, both in C++ and
> > the Python layer. Nothing I can't handle but it leads to complications.
> > I have started to implement it but it's looking messy.
> >
> > I also would like to keep the set_jacobian function. The constructors
> > will otherwise be a mess: there will be very many constructors with
> > varying order of arguments which is both error prone and tedious to
> > implement/maintain.
> >
>
> How is it error prone when each argument is a different type and can be
> check at compile time?

There would still be an error, even if it is caught by a check (at
compile-time or run-time depending on whether C++ or Python is used).

I claim it's less error prone (=less likely to cause errors, some of
which may be caught) since users will learn the one common signature
used by both Linear/NonlinearVariationalProblem and the solve() functions:

  lhs, rhs, solution, [bcs]

> > We would need the following variations of constructors for linear and
> > nonlinear variational problems:
> >
> >   a, L, u
> >   a, L, u, bc
> >   a, L, u, bcs
> >   F, u
> >   F, u, J
> >   F, u, bc
> >   F, u, bc, J
> >   F, u, bcs
> >   F, u, bcs, J
>
> > Instead of one common signature:
> >
> >   lhs, rhs, u, [bcs]
> >
> > Of course there will still need to be some variations to implement,
> > but those are only handling various ways to specify the boundary
> > conditions, either none, bc or bcs.
> >
> > The advantage is clarity: only needing to remember lhs, rhs, u, [bcs].
> >
>
> I would argue the opposite; requiring a pointless function argument is
> confusing.
>
> If it's too greater burden to provide all the above constructors, the
> number of constructors can be rationalised.

Exactly how do you propose to rationalize it?

--
Anders