dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Jan 29, 2007 at 04:30:36PM +0100, Garth N. Wells wrote:
> Anders Logg wrote:
> > Yes, look at the demo in src/demo/pde/dg/. (It contained some test
> > code before that I just removed.)
> > 
> > Yes, you can take the avg() and jump() of an arbitrary function (but
> > there may be some limitations that I can't recall right now).
> > 
> > It currently only works in 2D. In 3D, it only works in principle but
> > not in practice... :-)
> > 
> > What is currently holding us back from getting everything in place is
> > 
> >     (1) Decide on a numbering scheme for the mesh and write this down
> >         firmly in the manual
> > 
> >     (2) Implement this in FFC and DOLFIN
> > 
> >     (3) Porting FFC to generate UFC code
> >
> 
> There is another problem. The approach using the tensor contraction is 
> not practical, especially for vector equations. FFC generate 96 cases 
> (in 3D) which takes way too long and produces very large output.
> 
> Garth

I think we'll be able to get this down to 16 cases, but I agree it is
still a lot of code.

The switch itself is not a problem, it's throwing all the code into a
single function that is the problem. We could fix this by generating
16 functions and just leaving the switch + function calls in eval().

/Anders