dolfin team mailing list archive

["Marie E. Rognes" <meg@xxxxxxxxx>]

On 06/23/11 15:38, Neilen Marais wrote:
> Hi Anders.
>
> On Thu, Jun 23, 2011 at 3:10 PM, Anders Logg<logg@xxxxxxxxx>  wrote:
> > > I would just like to know how bc.apply() goes about approximating
> > > BC_expr. Does it take the estimated degree of BC_expr into account, or
> > > does it only take the degree of V into account? And is there any way
> > > to control this?
> > It only depends on the dofs of V. For each degree of freedom of V, it
> > evaluates that degree of freedom on your expression.
> OK, but how are those DOFs defined? For interpolatory elements that
> should just be a point-wise evaluation of the function you're
> interpolating. But the theoretical DOFs can also be defined as
> integrals over edges, faces, volumes depending on how the basis is
> constructed. Do higher degree Nedelec elements still result only in
> point-wise evaluation (one point per DOF), or are integrals evalued?

The degrees of freedom are the degrees of freedom defined by FIAT on the 
reference elements,
mapped in a sensible manner onto the physical elements.  Hence, If FIAT 
defines a degree of  freedom using a (numerical) integration rule, so 
does DOLFIN.

Specifically, you can examine how the degrees of freedom are evaluated 
by looking at the function evaluate_dof(...) in the FFC-generated code 
(or browsing the FIAT source ;-) )

And in particular, for higher order Nedelecs (H(curl), of the first 
kind), the interior degrees of freedom involve integral moments. These 
should not be evaluated if you only do stuff on boundary edges though.

--
Marie


> I ask this since I changed the interpolant from a python class
> subclassed from dolfin.Expression to actual JIT-able Expressions, and
> two things happened.
>
> Firstly, the numerical results change by a maginitude larger than can
> be explained by machine precision effects, but not enough (I think) to
> indicate incorrect expressions; it looks more like what I'd expect
> from a different quadrature degree.
>
> Secondly, using the expressions with no form compiler optimisations
> lead to execution several times slower than using the python class.
> Even without optimisation enabled, I'd expect the JIT expressions to
> be much faster than the pure python expressions. When I enable form
> optimisation it is indeed faster than python expression equivalent.
> This seems to indicate to me that extra expression evaluations are
> happening? How could I check?
>
> Thanks
> Neilen
>
> > --
> > Anders
> _______________________________________________
> Mailing list: https://launchpad.net/~dolfin
> Post to     : dolfin@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~dolfin
> More help   : https://help.launchpad.net/ListHelp