ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Fri, Jan 22, 2010 at 09:35:23AM +0100, Kristian Oelgaard wrote:
>
>
> 2010/1/22 Anders Logg <logg@xxxxxxxxx>:
> >On Thu, Jan 21, 2010 at 11:26:47PM +0100, Anders Logg wrote:
> >>The MetaData.ufl example is failing in the quadrature representation.
> >>
> >>Kristian, could you have a look and see what goes wrong. Then I can
> >>continue to make the incremental/combined integral stuff work.
> >
> >On second thought, maybe we should skip the whole thing with having
> >different representations inside the same tabulate_tensor function
> >for the following reasons:
> >
> >1. It leads to complications in the code generation.
>
> Very true.
>
> >2. It probably doesn't make much difference since if quadrature is
> >required for one part of an integral, then it is likely the cost is
> >low for including the other terms inside the existing quadrature loop.
>
> True.
>
> >3. The situation will not arise often since UFL will anyway group
> >terms with the same dx inside the same integral and then all will have
> >the same representation.
>
> I'm not sure this is correct, if the metadata is different the integrals will not be grouped.

That's correct. So it has an effect when someone explicitly wants to
treat a term with a specific representation, but not when FFC tries to
auto-select the representation for each term since then all the terms
will have metadata "auto" and be grouped together.

I still think we have a strong case for a simplification. Agree?

> >So the cases where this is useful are rare (because of the UFL
> >grouping), it won't give any speedup and it's complicated to
> >implement...
> >
> >We will still have the possibility of different representations for
> >integrals on different subdomains and for cell, interior and exterior
> >facet integrals.
>
> Yes, and this is probably the case where one will use it the most anyway.
>
> >If this sounds ok, I can add some checks in analysis.py that force all
> >UFL integrals that will be grouped as part of the same UFC integral to
> >use the same representation.
>
> Yes, such that if any quadrature is present they will all be quadrature.
>
> >Btw, does quadrature handle different quadrature degree for different
> >integrals that become part of the same tabulate_tensor? If not, we can
> >force metadata to be the same for all integrals that belong together.
>
> I don't know how the meta data is set up now, but it used to be able to do:
>
> a = v*u*dx + inner(grad(v), grad(u))*dx -->
>
> loop ip range(4):
>  ....
>
> loop ip range(9):
>  ....
>
> Inside one tabulate_tensor() funtion.

ok, if there are no objections, I can have a go at adding some checks
for having the same metadata in all terms UFL integrals that
correspond to the same UFC integral (same domain type and domain number).

--
Anders

Attachment: signature.asc
Description: Digital signature