Accéder au contenu.
Menu Sympa

starpu-devel - [Starpu-devel] Task pipelining commited to trunk

Objet : Developers list for StarPU

Archives de la liste

[Starpu-devel] Task pipelining commited to trunk


Chronologique Discussions 
  • From: Samuel Thibault <samuel.thibault@ens-lyon.org>
  • To: starpu-devel@lists.gforge.inria.fr
  • Subject: [Starpu-devel] Task pipelining commited to trunk
  • Date: Fri, 22 Aug 2014 19:48:27 +0200
  • 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>

Hello,

I've just committed to the trunk CUDA task pipelining. This is enabled
by default, but to get it, one has to use the STARPU_CUDA_ASYNC mode.

In combination with STARPU_NWORKER_PER_CUDA, this also permits to run
several kernels concurrently on Fermi cards (i.e. not only Kepler).

All in all, on a simple cholesky factorization with 8cpus+3gpus, for
instance, I got a performance bounce from ~1.3TF to ~1.55TF :)

Samuel



  • [Starpu-devel] Task pipelining commited to trunk, Samuel Thibault, 22/08/2014

Archives gérées par MHonArc 2.6.19+.

Haut de le page