dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Wed, Sep 21, 2011 at 07:06:02PM +0200, Anders Logg wrote:
> On Wed, Sep 21, 2011 at 04:54:29PM -0000, Garth Wells wrote:
> > On 21 September 2011 17:40, Johan Hake <787405@xxxxxxxxxxxxxxxxxx> wrote:
> > > On Wednesday September 21 2011 09:00:16 Garth Wells wrote:
> > >> We should look at this  - it will allow the removal of some ugly copies
> > >> in DOLFIN.
> > >
> > > Agree, but at the moment are boost::multi_arrays not so pleasent for the SWIG
> > > interface. Copying is enforced. Not sure this is a big deal for such a low
> > > lever functionality, but it is sad that it should be needed, if the underlying
> > > structure is just a c-array, which numpy wraps just fine.
> > >
> >
> > I think that if this type of data is used intensively at the Python
> > level, then there will be bigger performance issues to worry about
> > than data copying.
> >
> > The multi_array interface is somewhat secondary - the important point
> > is flattening the double pointer in the UFC interface.
>
> We should avoid using any other data structures than plain C arrays in
> the UFC interface.
>
> Or we should use STL data structures (or boost) throughout UFC, but
> that sounds more like UFC >= 3.0.

We should make a new UFC release to fix at least the bug in the
quadrature version of tabulate_tensor for interior_facet_integral.

Should we also change the double pointers into plain arrays?

This would mean quite subtstantial changes as the following variables
and arguments will need to change:

  cell::entity_indices
  cell::coordinates
  tabulate_tensor(..., w, ...)
  tabulate_tensor(..., quadrature_points, ...)

This will also require changes DOLFIN and FFC/SFC and (perhaps the
most time-consuming part...) in the UFC manual.

--
Anders