Re: [Starpu-devel] Problem with prio and tasks with several implementations

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

Cyril Bordage, le Thu 10 Oct 2013 20:35:36 +0200, a écrit :
> In _starpu_priority_pop_task, there is a loop over the implementations,
> but no break when a worker is choosen. I think a break is missing.

Ah, right, fixed it.

> In my application, I need to do a calibration step. You've guided me to
> the prio policy. But, prio cannot consider the model at all. So when I
> have several implementations, the first is always choosen and my task
> still remains uncalibrated. Is it not a good point, that prio policy
> chooses the best implementation (or an uncalibrated implementation) ?

It is. The idea would be more or less doing the same as in dmda, indeed.
Just nobody took the time to make prio do it (and I didn't know you had
several implementations for the same codelet).

Samuel