[gmx-users] fftw tweaking ?

[Erik Lindahl]

Marc Baaden wrote:

>Hi,
>
>I was checking on tweaking the fftw libraries for particular processor types.
>I couldn't verify if that topic came up already, as the search functiun on
>ww.gromacs.org it seems to be currently dysfunctional (probably a transition
>problem, Erik ?).
>
>
>I was pointed at this by the existence of a package p4fftwgel2 in Debian, the
>description of which says:
>
>
>Library for computing Fast Fourier Transforms on Intel P4
>
> This library computes Fast Fourier Transforms (FFT) in one
> or more dimensions. It is extremely fast. This package
> contains the documentation and the shared version of the
> libraries. To get the static library and the header files you
> need to install p4fftwgel-dev.
> .
> This library uses the same interface as fftw so should be a drop-in
> replacement, but is currently limited to single precision only.
> The code is tuned for Intel P4 processors and can be as much as three
> times as fast as vanilla fftw.
>
>
>Has anybody tried that ? Is it relevant for Gromacs performance ?
>(I guess it would only be needed for mdrun+pme, or maybe some analysis
>programs).
>
Hi,

I think it could be nice; Apple Computer also provide very optimized FFT 
routines with OS X 10.2.

The main reason why we aren't using them yet is that we still rely on 
the 3D-MPI-FFT routine in FFTW,
but we're working on replacing that with our own communication + 1D FFTs.

Don't expect any dramatic speed differences, though: Last time I 
benchmarked a PME run we spent about
10% doing the FFTs; the main culprit is the charge 
spreading/interpolation which really trashes the cache.

Cheers,

Erik