Re: [Starpu-devel] OpenMP StarPU

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

Roberto Ribeiro, le Tue 25 Sep 2012 16:50:31 +0100, a écrit :
> The reason is basically because i cant make it work. Using
> single_combined_worker and starpu_combined_worker_get_size() will always 
> return
> a single CPU. I've used your vector_scal.c example, compiled with the given
> command and the output is always: "running task with 1 CPUs".

Remember that the pheft scheduler benchmarks the parallel implementation
with varying numbers of cores, to check which ones are most efficient.
You can also try to use pgreedy instead.

> How do we do it then? we may still specify a single codelet, use:
>  .type = STARPU_FORKJOIN,
> .cpu_funcs = {CPU_func, NULL}
> cuda_funcs= {CUDA_func, NULL};
> and create subtasks with it?

Yes. You also need a performance model in order to use the pheft
scheduler, otherwise it might indeed get dumb and not dare parallel
execution, I don't remember.

Samuel