syfi team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Fri, Nov 09, 2007 at 12:24:01AM +0100, Martin Sandve Alnæs wrote:
> 2007/11/8, Anders Logg <logg@xxxxxxxxx>:
> > Which form is "source element vector"? Is it the linear form
> >
> >   L = v*f*dx?
> 
> Yes.
> 
> > How is this computed by SyFi (tabulate_tensor)? It would be
> > interesting to compare with the FFC code.
> 
> I can post the code tomorrow. SyFi is faster than FFC in this case,
> while FFC is faster than SyFi for the Laplace matrix. Mass matrices
> don't differ. Also, the SyFi-generated quadrature can be faster than
> both FFC and "regular" SyFi code in some cases, in particular for the
> convection vector L = dot(dot(u,grad(u)),v)*dx.

ok, interesting. Look forward to seeing what the differences are.

> > I'm not sure what your timeframe is but maybe you should also include
> > timings for FErari-optimized FFC code (in which case you need to bug
> > me about activating the optimization again in FFC).
> 
> Bug bug :)
> 
> > We did a series of
> > benchmarks for comparing plain FFC code with FErari code a while back
> > and the local speedups were typically between a factor 1-4:
> >
> >     http://www.femcenter.org/pub/preprints/phiprint-2007-04.pdf
> 
> I'll read this tomorrow. Also, Ola is working on implementing a
> factorization algorithm I found in an article earlier, and if he
> manages to get it working we could test that as well or use it to
> replace the current optimization (a naive implementation of "common
> subexpression elimination"). The current naive algorithm gives speedup
> factors of around 0.9 to 3.5, not easily predictable.
> 
> Most of the article will be unchanged by these additional results, so
> if we can get them in time (don't know when "in time" is), we can
> include them.

ok.

/Anders