[gmx-developers] GROMACS optimization for IBM Cell processor

[Maik Nijhuis]

On Wed, Jun 17, 2009 at 10:05:30PM +0200, Giovanni Lovato wrote:
> Hi all,
> at the University of Verona (Italy) we are starting to work on the  
> porting of MD algorithms to the IBM Cell processor. Before descending  
> into the code, is there already any sort of contribution/fork/patches/ 
> hints or anything else related to GROMACS optimization for Cell?

You can take a look at the following papers:

http://www.cs.unc.edu/~olivier/pdsec07.pdf

http://www.ncbi.nlm.nih.gov/pubmed/18615421

http://www.hicomb.org/papers/HICOMB2007-02.pdf

I have ported non-bonded kernels 010 and 112 to the SPEs in the Cell 
processor. My implementation supports any number of particles and uses 
the existing neighborlist. Since I'm storing the particle data 
(positions/forces) in main memory, I'm facing a memory bottleneck, and 
the speedup of these kernels is only 2 and 5, respectively, relative to 
the speed of the corresponding PPE altivec kernel (single-threaded).

Maik
-- 
maik at cs.vu.nl || http://www.cs.vu.nl/~maik