GPGPU status

[Axel]

On Jan 27, 7:52 am, Ondrej Marsalek <ondrej.... at gmail.com> wrote:
> Dear CP2K developers,

dear ondrej,

> I am interested in the status of GPGPU code in CP2K. So far, I have
> found only the very brief mention of single precision FFT using CUDA in
> the input manual and this e-mail from the CPMD archive:
>
> http://www.cpmd.org/pipermail/cpmd-list/2008-April/004330.html
>
> Could someone please give me a brief overview of the options one has in
> this area in CP2K? I would also like to know whether there is someone
> working on some sort of GPGPU code at this time.

yes. our group is working (to some degree) on it. however, the
outcome is not very promising. amdahl's law can be a real bitch.
the best we could get is a 2x speedup for a favorable problem.
the main issue is that you essentially will have to rewrite
cp2k _completely_ to make it useful for CUDA and get those
20x-100x speedups that everybody is going nuts over. if you
want to get an impression about how difficult that is going to
be, you can try compiling a single precision version of cp2k.
in anticipation of adding more CUDA support, we (= mostly teo)
put in some effort to revive this capability and so far we kept
it alive. however, only very few test calculations do finish,
and out of those many have large errors.

> I also have one more specific question. Could FFTCU be adapted to use
> double precision capable cards? I am interested in this because of

actually, the fft is one of the areas, where one can use
single precision with some success (and have a significant
speed gain over double precision). also, the use of double
precision data in the GPU incurs a significant performance hit.

> cluster calculations in open boundary conditions, where FFT seems to be
> the main bottleneck. Also, does anyone have any experience using FFTCU?

i certain that ben levine, the postdoc in out group working on
it, can fill you in on the details.

cheers,
   axel.

>
> Thanks a lot for any replies or comments,
> Ondrej