[gmx-developers] Current status FFTW-3D

Erik Lindahl lindahl at cbr.su.se
Fri Aug 31 17:10:10 CEST 2007


On Aug 31, 2007, at 4:04 PM, Roland Schulz wrote:
> Am I mistaken, or does it at maximum reduce it by 1/6, because only  
> the
> FFT in one dimension can be done as real-to-complex? Still I will  
> try to
> add it and I think it should be easy because the code seems very  
> clean.
> But I think a larger improvement would be to allow also FFTW 3.x  
> because
> its performance seems to be a lot higher since it uses SSE/Altivec

It is indeed only the first dimension that is real-to-complex, but  
since the resulting data then only contains half as many points we  
will cut the execution time in the remaining dimensions by 50% too.

However, in practice I think the reduced bandwidth is more important.  
FFTW really shines when you do a full 3D transform on a single CPU,  
but as soon as we start the decomposition we'll have to use a series  
of 1D FFT calls, and then there are several other good  
implementations too.

> How do you handle these exceptions? Does every author has to agree?  
> Why
> do you not cross-license the software as GPL and
> GPL+desktop-science-exception (or any appropriate existing  
> license), so
> that the exception is for everyone equally and really only regarding

For the core parts of Gromacs we either have the copyright among the  
main authors who have written most of it (who have had to agree :-),  
or we have shared copyright/free distribution rights to non-trivial  
contributions and modifications. There are some of the analysis tools  
and other minor stuff we had to rip out, and this was also the reason  
we implemented support for Intel MKL and FFTPACK as alternatives to  
FFTW (which is GPL-only, although they seem to be generous with  
exceptions). If you have a really compelling case you're more than  
welcome to contact us too :-)

In practice we simply want to keep the options open to avoid the  
situation e.g. Linux is in now with GPL 2 vs. 3. We're in this for  
the science, but if company X very hypothetically stepped forward and  
promised to fund $1,000,000 for general Gromacs development provided  
they got permission to link the library into a commercial GUI  
application we would at least think about it, in particular if they  
were to offer the binaries for free to academic researchers.

Anyway - none of this should be an issue for the FFTs. We won't need  
to ask Steve for any exceptions since people unhappy with GPL can  
just use one of the slower/worse scaling FFT implementations.



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20070831/a6634864/attachment.html>

More information about the gromacs.org_gmx-developers mailing list