[Starpu-devel] OpenCL backend on CPU

[Sylvain HENRY <sylvain.henry@inria.fr>]

Hi,

I'm using StarPU's OpenCL backend on CPU (using AMD's or Intel's 
implementation). The problem is that StarPU considers the CPU device as 
remote and thus performs asynchronous data transfers while it could work 
on data in-place.

With AMD's OpenCL implementation, asynchronous data transfers are not 
performed, at least until a blocking command (clWaitForEvents or 
clFinish) is called. With Intel it works fine but an unnecessary data 
copy occurs.

I would like to avoid this data copy. What would be the best way to 
implement this in StarPU?

Cheers
Sylvain