ffc team mailing list archive

["Garth N. Wells" <g.n.wells@xxxxxxxxxxxxxxx>]

Quoting Anders Logg <logg@xxxxxxxxx>:

> Garth, I tried the form you sent me for testing (CahnHilliard.form).
> 
> Here are the results, before and after optimizations:
> 
>         |  FFC default mode    FFC BLAS mode (-f blas)
> ------------------------------------------------------
> Before  |  4min 32s, 851MB     5min 31s, 1200MB
> After   |  2min 27s,  59MB     1min 13s,  60MB
> 
> Memory usage reduced by a factor 20 which is pretty good... :-)

Looks great.

> 
> I turned off the computation of integrals for the benchmarks (or it
> would have taken forever). Rob has a nice idea for optimization of the
> integral computation so that should be fixed in the near future.
> 
> The Cahn-Hilliard form is really demanding with lots of coefficients
> and terms. Is it possible to choose piecewise constant representation
> for some of the coefficients to reduce compile-time?

Some piecewise constants could be used to imply reduced integration. It's true
that very few FE codes would use exact integration for such a complex problem.
Most would just look at the order of the test and trial functions in determining
the integation scheme. I see that reduced integration is in the ffc TODO list,
which would offer a nice fix.

Garth

> 
> /Anders
> 
> _______________________________________________
> FFC-dev mailing list
> FFC-dev@xxxxxxxxxx
> http://www.fenics.org/cgi-bin/mailman/listinfo/ffc-dev
>