[phenixbb] GPU acceleration of phenix.real_space_refine?

[Pavel Afonine]

Hi Oliver,

unlike reciprocal-space refinement (phenix.refine), real-space 
refinement (phenix.real_space_refine) is much easier to parallelize and 
it will benefit a lot from parallelization.

Making use of multiple processors is in my list of things to do. No GPU 
though.

Pavel


On 5/24/16 09:28, Oliver Clarke wrote:
> Hi all,
>
> phenix.real_space_refine takes a very long time to run on large 
> systems when using any of the more computationally expensive options 
> (e.g. morphing).
>
> I was wondering if there has been any consideration towards optimizing 
> this code (and maybe that of phenix.map_to_model) for GPU 
> acceleration? There is somewhat of a shift occuring in the EM 
> community towards using GPU workstations for data processing - if we 
> could use already purchased capabilities to accelerate model 
> refinement as well that would be a huge bonus.
>
> I saw a discussion on the bb from a while back concerning the same 
> question applied to phenix.refine, and the conclusion was that in that 
> case the benefits would be minimal, but I'm not sure whether the same 
> applies to real space operations.
>
> Cheers,
> Oli.
>
> PS has there been any consideration of including an option to allow 
> users to supply two half maps at run time, one "work" and one "test" 
> map for cross validation purposes? It is possible to do this with two 
> separate runs right now but having the option integrated would be 
> wonderful.