ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Wed, Dec 09, 2009 at 09:32:23AM -0800, Johan Hake wrote:
> On Wednesday 09 December 2009 09:17:17 Anders Logg wrote:
> > On Wed, Dec 09, 2009 at 06:09:56PM +0100, Anders Logg wrote:
> > > On Wed, Dec 09, 2009 at 08:56:51AM -0800, Johan Hake wrote:
> > > > On Wednesday 09 December 2009 08:39:17 Anders Logg wrote:
> > > > > On Wed, Dec 09, 2009 at 08:25:24AM -0800, Johan Hake wrote:
> > > > > > On Wednesday 09 December 2009 08:13:20 Anders Logg wrote:
> > > > > > > On Wed, Dec 09, 2009 at 08:04:48AM -0800, Johan Hake wrote:
> > > > > > > > On Wednesday 09 December 2009 01:59:58 Anders Logg wrote:
> > > > > > > > > On Tue, Dec 08, 2009 at 11:52:38PM -0800, Johan Hake wrote:
> > > > > > > > > > On Tuesday 08 December 2009 17:09:09 Anders Logg wrote:
> > > > > > > > > > > We're working on simplifying the FFC code (and parts of
> > > > > > > > > > > UFL). Both have grown out of hand and with limited
> > > > > > > > > > > resources (like Martin and Johan leaving for other
> > > > > > > > > > > ventures) we need to make the codebase easier to
> > > > > > > > > > > maintain.
> > > > > > > > > > >
> > > > > > > > > > > As part of this, we've made some changes to how UFL and
> > > > > > > > > > > FFC handle the preprocessing of forms. In particular,
> > > > > > > > > > > we've removed the caching of so called form data in a
> > > > > > > > > > > form. Ideally, this should not lead to any problems but
> > > > > > > > > > > there may be regressions such as slow JIT compilation as
> > > > > > > > > > > a result of not caching certain data so keep an eye out
> > > > > > > > > > > for this.
> > > > > > > > > >
> > > > > > > > > > Do you know why Martin did include these in the first
> > > > > > > > > > place?
> > > > > > > > >
> > > > > > > > > Yes, we discussed it quite a bit but I believe he was never
> > > > > > > > > happy with the solution. We had things like form_data being
> > > > > > > > > cached in a form and form_data also storing a modified
> > > > > > > > > version of the form so there were things like
> > > > > > > > >
> > > > > > > > >   self.form.form_data().form
> > > > > > > > >
> > > > > > > > > being used in the form. I believe the new design is cleaner.
> > > > > > > >
> > > > > > > > Ok.
> > > > > > > >
> > > > > > > >   v = TestFunction
> > > > > > > >   u = TrialFunction
> > > > > > > >
> > > > > > > >   M0 = assemble(u*v*dx)
> > > > > > > >
> > > > > > > >   v = TestFunction
> > > > > > > >   u = TrialFunction
> > > > > > > >
> > > > > > > >   M1 = assemble(u*v*dx)
> > > > > > > >
> > > > > > > > I think that he wanted to prevent recompilation of M1, as it is
> > > > > > > > basically the same code that is generated as for M0. But due to
> > > > > > > > some internal and necessary they are treated differently. I
> > > > > > > > think that the form_data thing was a workaround for that. Is
> > > > > > > > this correct?
> > > > > > >
> > > > > > > I don't think form_data has anything to do with this.
> > > > > > >
> > > > > > > The above code should not lead to any new code generation as both
> > > > > > > M0 and M1 should have the same signatures. If not, there's a bug
> > > > > > > in jit.py in FFC. The code for M1 should be picked up from disk
> > > > > > > cache.
> > > > > >
> > > > > > Ok, I see that this is not the case for Arguments (BasisFunctions),
> > > > > > but it is for Coefficients (Functions):
> > > > > >
> > > > > >    f = Function(V)
> > > > > >    u = TestFunction(V)
> > > > > >
> > > > > >    m0 = assemble(v*f*dx)
> > > > > >
> > > > > >    f = Function(V)
> > > > > >
> > > > > >    m1 = assemble(v*f*dx)
> > > > > >
> > > > > > will now trigger a recompilation.
> > > > > >
> > > > > > Just take a repr on the two 'f' and see.
> > > > >
> > > > > Should be fixed now. I've added a call to preprocess before computing
> > > > > the form signature.
> > > > >
> > > > > Does it work now?
> > > >
> > > > Did you push it?
> > >
> > > Yes.
> > >
> > > >   el = FiniteElement("CG",triangle,1)
> > > >   f = Coefficient(el)
> > > >
> > > >   repr(f*dx)
> > > >
> > > >   f = Coefficient(el)
> > > >
> > > >   repr(f*dx)
> > > >
> > > > repr does here return different results, and JitObject.signature() does
> > > > only use repr on the passed form.
> > >
> > > repr should return a different string because there are two different
> > > instances of f.
> > >
> > > But JITObject gets the preprocessed form and should therefore see an f
> > > with number (count) 0 since it should have been renumbered.
> > >
> > > I'll take a closer look.
> >
> > It looks to me like the signature is correct so the second assembly
> > should trigger a reload from (disk) cache.
> >
> > Does it seem to work on your end?
>
> Yes! I did only look at the signature method :P
>
> However when I tried to assemble:
>
>   assemble(f*dx)
>
> where f is a dolfin.Function assemble complained of no mesh provided. Is this
> correct, as f has a FunctionSpace?

I think this has been the case before. We should be able to make it so
that the mesh argument is not needed if it can be figured out.

I won't have time to fix that now but will do it at some point in the
future. (But feel free to fix it.)

--
Anders

Attachment: signature.asc
Description: Digital signature