Accéder au contenu.
Menu Sympa

starpu-devel - Re: [Starpu-devel] Constructing codelet-chain loop

Objet : Developers list for StarPU

Archives de la liste

Re: [Starpu-devel] Constructing codelet-chain loop


Chronologique Discussions 
  • From: Samuel Thibault <samuel.thibault@ens-lyon.org>
  • To: Albert Sidelnik <sidelnik@gmail.com>
  • Cc: starpu-devel@lists.gforge.inria.fr
  • Subject: Re: [Starpu-devel] Constructing codelet-chain loop
  • Date: Thu, 19 Apr 2012 17:44:43 +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>

Albert Sidelnik, le Thu 19 Apr 2012 10:05:13 -0500, a écrit :
> 1) By setting regenerate = 1 for all tasks, my assumption is that
> when tasks are regenerated, they preserve their tag dependences.

They do

> 2) If each task has only one predecessor that it is dependent on, I
> can resolve that dependence directly through a starpu_tag_notify()
> call in order for that task to execute

In case of a tag associated to a task, I'm not sure of the semantic.
Better use a tag not associated to a task, as in the
tests/main/subgraph_repeat_regenerate_tag.c test I have just added.

> 3) The task graph is constructed a priori (I never need to resubmit the
> task)

Yes.

> For the previous test case, if I were not to utilize tags (and
> starpu_tag_notify), and only rely on resubmitting tasks similar to the
> test case in main/subgraph_repeat_regenerate.c, things work It's only
> when dealing with the tags, that things fail.

Yes, it's the restart part that is missing.

Samuel





Archives gérées par MHonArc 2.6.19+.

Haut de le page