ufl team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Fri, May 13, 2011 at 10:16:22AM +0200, Martin Sandve Alnæs wrote:
> On 13 May 2011 09:53, Anders Logg <logg@xxxxxxxxx> wrote:
> > Thanks! I'll see if it works.
> >
> > Have you thought about what should happen if say degree is given for a
> > mixed element but only one of the subelements has an undefined degree?
>
> No. If you specify the degree, reconstruct will set it for all subelements.
> There is no other unique way. If you want to treat certain subelements
> in a hierarchy specially you'll just have to traverse the mixed elements
> yourself. I just added a reconstruct_from_elements(*elements) to
> MixedElement which you can use to get some checking when piecing
> together the mixed elements.
>
> Note that it is important that reconstructed elements have the same
> value shape. If used to reconstruct functions to reconstruct an expression,
> the result of a changing value shape would be completely undefined
> behaviour and there is no way to do that correctly. I've added some
> asserts for this, just so you know they are really there for a very good reason.

ok. I'm not saying I foresee that we will run into these cases very
often, but I wanted to point it out. I'm usually not the one bringing
up the nasty corner cases. ;-)

> > It might also be the case that we will completely miss such cases. I
> > think degree() for a mixed element returns the maximum degree of the
> > subelements and if one of them is None but not all, we will miss it
> > since max(None, q) = q.
>
> Seems so. But that's a problem for FFC/PyDOLFIN :)
> I don't know why you would want to put yourself in that situation...

I will try to avoid it. :-)

--
Anders


> Martin
>
> \> --
> > Anders
> >
> >
> > On Fri, May 13, 2011 at 09:03:27AM +0200, Martin Sandve Alnæs wrote:
> >> Done, did a manual test and checking in now:
> >>
> >> In [5]: print FiniteElement("CG", triangle,
> >> 2).reconstruct(family="DG", cell=quadrilateral, degree=3)
> >> <DG3 on a <quadrilateral cell in R2>>
> >>
> >> In [8]: print VectorElement("CG", triangle,
> >> 2).reconstruct(family="DG", cell=quadrilateral, degree=3)
> >> <Discontinuous Lagrange vector element of degree 3 on a <quadrilateral
> >> cell in R2>: 2 x <DG3 on a <quadrilateral cell in R2>>>
> >>
> >> In [9]: print TensorElement("CG", triangle,
> >> 2).reconstruct(family="DG", cell=quadrilateral, degree=3)
> >> <Discontinuous Lagrange tensor element of degree 3 and shape (2, 2) on
> >> a <quadrilateral cell in R2>>
> >>
> >> In [10]: print MixedElement(VectorElement("CG", triangle, 2),
> >> FiniteElement("CG", triangle, 2)).reconstruct(family="DG",
> >> cell=quadrilateral, degree=3)
> >> <Mixed element: (<Discontinuous Lagrange vector element of degree 3 on
> >> a <quadrilateral cell in R2>: 2 x <DG3 on a <quadrilateral cell in
> >> R2>>>, <DG3 on a <quadrilateral cell in R2>>)>
> >>
> >> I'll move this into the unit tests, but don't have more time for this today.
> >>
> >> Martin
> >>
> >> On 13 May 2011 00:33, Anders Logg <logg@xxxxxxxxx> wrote:
> >> > It seems to work fine, but some unit tests are failing since
> >> > reconstruct is only implemented for FiniteElement (not for mixed
> >> > elements etc).
> >> >
> >> > Could you add a suitable reconstruct for mixed elements? I'm not sure
> >> > how it should be done, in particular how the input should be
> >> > propagated to the subelements.
> >> >
> >> >
> >> >
> >> >
> >> > On Thu, May 12, 2011 at 10:37:08PM +0200, Martin Sandve Alnæs wrote:
> >> >> Be aware that form_data.elements and friends are _after_ mapping, so
> >> >> if you get to the dolfin.FunctionSpaces through them something must be
> >> >> changed. If you get them from the original functions everything should
> >> >> be fine though.
> >> >>
> >> >> Martin
> >> >>
> >> >> On 12 May 2011 22:30, Anders Logg <logg@xxxxxxxxx> wrote:
> >> >> > On Thu, May 12, 2011 at 10:22:51PM +0200, Martin Sandve Alnæs wrote:
> >> >> >> On 12 May 2011 22:21, Anders Logg <logg@xxxxxxxxx> wrote:
> >> >> >> > On Thu, May 12, 2011 at 08:39:55PM +0200, Anders Logg wrote:
> >> >> >> >> On Thu, May 12, 2011 at 07:39:39PM +0200, Martin Sandve Alnæs wrote:
> >> >> >> >> > I've pushed some changes to lp:~martinal/ufl/reconstruction mostly
> >> >> >> >> > implementing the ufl side of taking an (optional) element_mapping
> >> >> >> >> > as input to preprocess, but I want to test it a bit more before pushing
> >> >> >> >> > to trunk.
> >> >> >> >>
> >> >> >> >> ok. I'm currently working on sorting out some remaining cache problems
> >> >> >> >> in the FFC JIT compiler. That should hopefully take care of the
> >> >> >> >> remaining tests that fail in DOLFIN. When that works, I can try out
> >> >> >> >> the reconstruction.
> >> >> >> >
> >> >> >> > Is it safe to try out the new element replacement now?
> >> >> >>
> >> >> >> Yes, I even checked it in. element_mapping is None by default, just
> >> >> >> try adding it.
> >> >> >
> >> >> > ok, will try.
> >> >> >
> >> >> >
> >> >
> >