Re: [Starpu-devel] Question about sequential transfers in Starpu/SimGrid

[Samuel Thibault <samuel.thibault@inria.fr>]

Hello,

Lionel Eyraud-Dubois, on jeu. 09 nov. 2017 14:55:50 +0100, wrote:
>    I was reading core/simgrid.c to understand the way StarPU models GPU
> transfers, which ones are sequential. And I found the following lines of
> code (simgrid.c:643) :
> 
>     /* GPU-GPU transfers are sequential with any RAM->GPU transfer */
>     if (new_is_gpu_gpu
>             && old_transfer->dst_node == new_transfer->src_node
>             && old_transfer->dst_node == new_transfer->dst_node)
>         return 1;
>     if (old_is_gpu_gpu
>             && new_transfer->dst_node == old_transfer->src_node
>             && new_transfer->dst_node == old_transfer->dst_node)
>         return 1;
> 
> And this confuses me because this can only be true if new_transfer->src_node
> == new_transfer->dst_node (resp. for old_transfer in the second case), which
> I believe is not supposed to happen. Am I missing anything ?

It seems it should have been an || there indeed, now fixed.

Thanks!
Samuel