GMP on the CELL processor

[Paul Zimmermann]

> Date: Wed, 12 Nov 2008 13:37:28 +0100
> From: David Monniaux <David.Monniaux at imag.fr>
> 
> Hi all,
> 
> I have an algorithm that I currently parallelize using OpenMP. 
> Basically, if I have n cores, I do the same operation on m lines of an 
> array, so I can split that into blocks of m/n lines each run on a core 
> (or finer adjustments taking into account the fact that each line does 
> not take the same time to process).
> 
> I was wondering whether it could be done on the CELL processor. I've 
> seen previous discussions on the same issue, but I think they looked at 
> the problem from a different angle. I'm not looking into parallelizing 
> GMP operations across the cores - just into
> 1) being able to run GMP operations in each core independently (this 
> requires being able to compile GMP for the SPU target); basically I 
> would assign m/n lines of the array to each of the n cores.
> 2) being able to transfer GMP data between the PPE and the SPE (this 
> looks like being able to do message passing with GMP data...).
> 
> Has somebody looked into that?

Arjen Lenstra had a student (Thomas Kunz) who did a Master Thesis on
"Factoring on a PlayStation 3". Part of it did consist in porting GMP
to the Cell (i.e., to the SPU). The main problem was the code size,
thus only part of GMP could be ported.

Paul Zimmermann