ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Tue, Apr 26, 2011 at 06:12:26PM +0200, Martin Sandve Alnæs wrote:
> On 26 April 2011 18:10, Johan Hake <johan.hake@xxxxxxxxx> wrote:
>
>     On Tuesday April 26 2011 09:01:35 Anders Logg wrote:
>     > On Tue, Apr 26, 2011 at 08:44:24AM -0700, Johan Hake wrote:
>     > > On Tuesday April 26 2011 08:42:32 Anders Logg wrote:
>     > > > On Tue, Apr 26, 2011 at 08:39:30AM -0700, Johan Hake wrote:
>     > > > > On Tuesday April 26 2011 08:33:11 Garth N. Wells wrote:
>     > > > > > On 26/04/11 16:31, Johan Hake wrote:
>     > > > > > > On Tuesday April 26 2011 08:16:29 Garth N. Wells wrote:
>     > > > > > >> On 26/04/11 16:07, Anders Logg wrote:
>     > > > > > >>> On Tue, Apr 26, 2011 at 03:59:52PM +0100, Garth N. Wells
>     wrote:
>     > > > > > >>>> On 26/04/11 15:55, Anders Logg wrote:
>     > > > > > >>>>> On Tue, Apr 26, 2011 at 03:45:22PM +0100, Garth N. Wells
>     wrote:
>     > > > > > >>>>>> On 26/04/11 13:51, Anders Logg wrote:
>     > > > > > >>>>>>> On Tue, Apr 26, 2011 at 02:00:50PM +0200, Anders Logg
>     wrote:
>     > > > > > >>>>>>>> It feels good that you trust me enough to handle it. ;-)
>     > > > > > >>>>>>>>
>     > > > > > >>>>>>>> Will add it sometime this afternoon and then we can
>     > > > > > >>>>>>>> revisit the JIT compiler caching.
>     > > > > > >>>>>>>
>     > > > > > >>>>>>> I'm getting confused here... Looking at preprocess.py in
>     > > > > > >>>>>>> UFL, I see
>     > > > > > >
>     > > > > > > this:
>     > > > > > >>>>>> It is confusing. Does the function 'preprocess' do
>     anything
>     > > > > > >>>>>> that the old FormData class didn't? It would be easier to
>     > > > > > >>>>>> follow if Form just had a member function form_data() that
>     > > > > > >>>>>> computes and stores data (like it used to), or if Form had
>     > > > > > >>>>>> a 'preprocess' function. Having the function preprocess
>     > > > > > >>>>>> return a new form is really confusing.
>     > > > > > >>>>>
>     > > > > > >>>>> I don't find that particularly confusing. It's the same as
>     > > > > > >>>>>
>     > > > > > >>>>>   refined_mesh = refine(mesh)
>     > > > > > >>>>
>     > > > > > >>>> Which is the whole problem. By creating a new object,
>     FormData
>     > > > > > >>>> is thrown away. The preprocessing should just compute some
>     > > > > > >>>> more data, just like we *don't* do
>     > > > > > >>>>
>     > > > > > >>>>   initialised_mesh = mesh.init(0)
>     > > > > > >>>>
>     > > > > > >>>> What was wrong with Martin's original design that
>     necessitated
>     > > > > > >>>> the change?
>     > > > > > >>>
>     > > > > > >>> As I explained, I thought it was better to have an explicit
>     > > > > > >>> call to preprocess since that makes it clear that one makes a
>     > > > > > >>> call to a function which may take some time to execute
>     > > > > > >>> (instead of just calling a member function which seems to
>     just
>     > > > > > >>> return some data).
>     > > > > > >>>
>     > > > > > >>> But as I say above: I added the caching back at some point
>     > > > > > >>> (maybe even the day after I removed it 2 years ago) so we
>     > > > > > >>> don't need to discuss why I removed it (as I realized myself
>     I
>     > > > > > >>> shouldn't have removed it and added it back a long time ago).
>     > > > > > >>>
>     > > > > > >>> What has me confused now is that the caching seems to be in
>     > > > > > >>> place but we still need the extra caching in FFC/DOLFIN and I
>     > > > > > >>> don't see why.
>     > > > > > >>
>     > > > > > >> Because preprocess returns a new form, e.g. define a form
>     > > > > > >>
>     > > > > > >>   a = u*v*dx
>     > > > > > >>   jit(a)
>     > > > > > >>
>     > > > > > >> Inside jit,
>     > > > > > >>
>     > > > > > >>    a.form_data() is None:
>     > > > > > >>        b = preprocess(a) # b now has data attached, but a
>     > > > > > >>        doesn't
>     > > > > > >>
>     > > > > > >>    else:
>     > > > > > >>        b = a
>     > > > > > >>
>     > > > > > >> Now 'b' has been preprocessed, and has form data attached, but
>     > > > > > >> 'a' doesn't. Calling 'jit(a)' again, the code will never enter
>     > > > > > >> the 'else' part of the clause because 'a' never gets any form
>     > > > > > >> data. Johan has added some code FFC that attaches the form
>     data
>     > > > > > >> of 'b' to 'a', but it is a bit clumsy.
>     > > > > > >
>     > > > > > > No, it was already attached. I just made ffc use it.
>     > > > > >
>     > > > > > Didn't you add the line
>     > > > > >
>     > > > > >     form._form_data = preprocessed_form.form_data()
>     > > > >
>     > > > > No, I added:
>     > > > >   preprocessed_form = form.form_data()._form
>     > > > >
>     > > > > I think the thing here is that form_data has always had a
>     > > > > preprocessed form. Someone (lets not point fingers!) thought that
>     > > > > was too much magic and added an
>     > > > >
>     > > > > explicit need to call:
>     > > > >   form = preprocess(form)
>     > > > >
>     > > > > in jit_compiler(). This made the design more complicated and also
>     > > > > introduced a cirucular dependency, as the return preprocessed form
>     > > > > need to know of its form_data, but the form_data already had a
>     > > > > reference to the preprocessed form. The latter is what I used in
>     the
>     > > > > one line I altered.
>     > > >
>     > > > No, it made the design cleaner since it makes clear something needs
>     to
>     > > > happen to get the metadata: a call to preprocess.
>     > > >
>     > > > Where did you add this line?
>     > >
>     > > I change
>     > >
>     > >   preprocessed_form = form
>     > >
>     > > to:
>     > >   preprocessed_form = form.form_data()._form
>     >
>     > Yes, but where?
>     >
>     > I've fixed the bug now in preprocess.py (attaching to both forms). Does
>     > that help?
>
>     In ffc.jit_form.
>
>     Your fix wont fix the circular dependency.
>
>     We also need to remove form_data from the preprocessed form.
>
>
> Or just use a weakref.
>
>
>     This means that
>     we need to return form_data from preprocess and maybe change its name to
>     compute_form_data.
>
>     Johan

Why is the circular dependency a problem? Anyway, I'm thinking now the
cleanest design would be

  form.compute_form_data()
  form_data = form.form_data()
  preprocessed_form = form_data.preprocessed_form

Here, the preprocessed_form would not store form_data.

--
Anders