ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Feb 01, 2010 at 11:04:44PM +0100, Kristian Oelgaard wrote:
>
>
> On 1 February 2010 22:57, Marie Rognes <meg@xxxxxxxxx> wrote:
> >Kristian Oelgaard wrote:
> >>
> >>
> >>On 1 February 2010 22:32, Marie Rognes <meg@xxxxxxxxx> wrote:
> >>>
> >>>Kristian Oelgaard wrote:
> >>>>
> >>>>
> >>>>On 1 February 2010 22:02, Marie Rognes <meg@xxxxxxxxx> wrote:
> >>>>>
> >>>>>Kristian Oelgaard wrote:
> >>>>>>
> >>>>>>
> >>>>>>On 1 February 2010 21:37, Anders Logg <logg@xxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>>Impressive! :-)
> >>>>>>
> >>>>>>It sure is, many things could be handled in a much simpler way if it
> >>>>>>wasn't for memory and runtime considerations of the optimisations. If
> >>>>>>I
> >>>>>>get
> >>>>>>around to implementing the other optimisation strategy, we could make
> >>>>>>life
> >>>>>>simpler in these classes and only let the symbolics stuff work on
> >>>>>>moderately
> >>>>>>complex forms.
> >>>>>>
> >>>>>
> >>>>>Sorry if I messed something up. I just thought it was better to have
> >>>>>something working (but suboptimal) than something broken (and optimal
> >>>>>;)
> >>>>>)
> >>>>
> >>>>Agree, the code looks OK.
> >>>>
> >>>>>>>Is it release time or are we waiting for Kristian to do some magic
> >>>>>>>with evaluate_derivatives for Hdiv/curl?
> >>>>>>
> >>>>>>Actually, I'm waiting for Marie to fix this one too :)
> >>>>>
> >>>>>I'm taking a look, but it is not likely to happen within the next
> >>>>>couple
> >>>>>of
> >>>>>hours.
> >>>>>
> >>>>>I'm lost in the vectors: What is 'derivatives' and 'values' supposed to
> >>>>>be
> >>>>>holding? (Math, please ;)
> >>>>
> >>>>'values' is the argument to evaluate_basis_derivatives that will hold
> >>>>the
> >>>>return values (derivatives on the physical cell) from the function. It
> >>>>is
> >>>>the users responsibility to make sure it has the correct size.
> >>>>'derivatives' is the derivatives on the reference element, which is
> >>>>computed like it is done inside FIAT (multiplication of coefficients,
> >>>>dmats
> >>>>and basis values of the expansion).
> >>>>
> >>>
> >>>Ok. And the values are ordered so that first all derivatives for value
> >>>dimension 0, then all derivatives for value dimension 1, etc.?
> >>
> >>Yes. There is a comment at the top of evaluatebasisderivatives.py, but it
> >>might be a good idea to put in the generated code too?
> >>
> >>>Currently only the first two (out of four?) 'derivatives' are mapped for
> >>>BDM
> >>>_1on triangle?
> >>>
> >>>     derivatives[0] = (1.0/detJ)*((J_00*tmp_ref0 + J_01*tmp_ref1));
> >>>     derivatives[1] = (1.0/detJ)*((J_10*tmp_ref0 + J_11*tmp_ref1));
> >>
> >>I pushed a fix for this a while back, I was hoping that was the error, but
> >>I still get wrong results.
> >
> >
> >Sorry, my bad -- something strange happened.
> >
> >And I guess you have checked the indices a couple of hundred times?
> >
> >How about the famous UFC/FIAT reference element difference?
>
> I think I found it, I was mapping values from FIAT incorrectly. Stupid, stupid me.
> I'll just clean things up then Anders can make a release.
>
> Kristian

Yes! Yes! Yes! :-)

I'll start with FIAT, UFC, UFL and FFC. Then recompile all forms in
DOLFIN and wait until tomorrow with the release of DOLFIN.

We need to hurry up before Garth finds a new bug. He seems determined
to delay the release. ;-)

--
Anders

Attachment: signature.asc
Description: Digital signature