ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Sat, Sep 04, 2010 at 10:54:27PM +0100, Garth N. Wells wrote:
>
>
> On 04/09/10 17:20, Johan Hake wrote:
> >On Saturday September 4 2010 08:55:42 Anders Logg wrote:
> >>On Sat, Sep 04, 2010 at 08:49:43AM -0700, Johan Hake wrote:
> >>>On Saturday September 4 2010 05:03:04 Anders Logg wrote:
> >>>>All of the problems seem to come from the Python tests and demos with
> >>>>std::bad_cast. Perhaps something with the SWIG layer?
> >>>
> >>>Yes I have noitced this before. Not sure why but I have not been able to
> >>>compile the wrapper code with -O3
> >>
> >>Isn't it just that the wrapper code (like the form files) can be very
> >>large (very many lines of code in a single function or file) so the
> >>compiler runs out of memory or takes a very long time to compute the
> >>optimizations?
> >
> >It might be that the dolfin wrapper file it too large. But I have tried
> >compiling the a much smaller wrapper code and it also bailed out for -O3.
> >
>
> I've never had a problem with -O3 when using the quadrature
> representation. It can be a problem for the tensor contraction, but
> by the time you run into this limitation the quadrature code is much
> faster anyway. Why I am seeing now for a particular problem is that
> Kristian's FFC optimisations so good that -O3 and -O0 make no
> difference because insertion is dominating.
>
> The tensor contraction would be more compiler friendly (and most
> likely efficient) if it didn't unroll loops, but used fixed size
> loops and let the compiler do tricks on those.

Perpaps. It would be relatively easy to make unrolling optional.

--
Anders