Re: [Starpu-devel] GPU issue with r12137

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

Xavier Lacoste, le Thu 03 Apr 2014 13:44:00 +0200, a écrit :
> I runned the test several time because i didn't believe this can be the
> cause....
> 
> Indeed doing :
> 
> diff --git a/src/datawizard/coherency.c b/src/datawizard/coherency.c
> index 6e67156..8e82163 100644
> --- a/src/datawizard/coherency.c
> +++ b/src/datawizard/coherency.c
> @@ -172,13 +172,13 @@ static int worker_supports_direct_access(unsigned 
> node,
> unsigned handling_node)
>                         enum starpu_node_kind kind = starpu_node_get_kind
> (handling_node);
>                         /* GPUs not always allow direct remote access: if 
> CUDA4
>                          * is enabled, we allow two CUDA devices to
> communicate. */
> -                       return
> -#if 0
> +                       return kind ==
> +#if 1
>                                 /* CUDA does not seem very safe with 
> concurrent
>                                  * transfer queueing, avoid queueing from 
> CPUs
> */
> -                               kind == STARPU_CPU_RAM ||
> +                               STARPU_CPU_RAM ||
>  #endif
> -                               kind == STARPU_CUDA_RAM;
> +                               STARPU_CUDA_RAM;
>                 }
>  #else
>                         /* Direct GPU-GPU transfers are not allowed in 
> general
> */
> 
> solves the issue.

Uh?!

Since kind can not be both STARPU_CPU_RAM || STARPU_CUDA_RAM, it
basically means always returning 0.  What is really disturbing is that
all it does is disabling GPU-GPU direct transfers.  You are testing with
just one GPU, aren't you?

Samuel