Objet : Developers list for StarPU
Archives de la liste
- From: Samuel Thibault <samuel.thibault@ens-lyon.org>
- To: Cyril Bordage <cbordage@illinois.edu>
- Cc: starpu-devel <starpu-devel@lists.gforge.inria.fr>
- Subject: Re: [Starpu-devel] Question about device performance
- Date: Thu, 26 Feb 2015 10:06:43 +0100
- List-archive: <http://lists.gforge.inria.fr/pipermail/starpu-devel/>
- List-id: "Developers list. For discussion of new features, code changes, etc." <starpu-devel.lists.gforge.inria.fr>
Cyril Bordage, le Wed 25 Feb 2015 23:49:37 +0100, a écrit :
> in StarPU, you use one dedicated thread for each device. I know it is
> because of performance issues.
Not really, it's mostly because the StarPU CUDA programming model was
that the application doesn't give back execution control to StarPU until
the kernel is finished.
With the STARPU_CUDA_ASYNC flag however, this could be lifted if the
application promises it never submits a kernel without the flag. This
shouldn't be hard to implement since we already have several workers run
by one thread for the multi-stream support.
Samuel
- [Starpu-devel] Question about device performance, Cyril Bordage, 26/02/2015
- Re: [Starpu-devel] Question about device performance, Samuel Thibault, 26/02/2015
Archives gérées par MHonArc 2.6.19+.