Re: [Starpu-devel] StarPU OpenMP+CUDA

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

alankrutha reddy, le Wed 03 Jul 2013 13:56:52 +0530, a écrit :
> On Wed, Jul 3, 2013 at 1:47 PM, Samuel Thibault 
> <samuel.thibault@ens-lyon.org>
> wrote:
> 
>     alankrutha reddy, le Wed 03 Jul 2013 13:44:52 +0530, a écrit :
>     > My scenario is as below:
>     >
>     > #pragma omp parallel for
>     > for(...)
>     > {
>     >         kernel<<<,>>>(...);
>     > }
> 
>     Err, I don't really understand: the code above would mean that all
>     OpenMP threads will make the GPU start a CUDA kernel, and thus encounter
>     a severe contention. Is this really what you want?
> 
> ya..

Well, this is completely outside of StarPU's way of submitting work to
the GPU.  For StarPU, you would rather make a GPU task which depends on
the result of the OpenMP task.

Samuel