Re: [Starpu-devel] [Starpu-commits] r5692 - in trunk/src: core core/dependencies datawizard drivers/cpu drivers/cuda drivers/gordon drivers/opencl

[Xavier Lacoste <xavier.lacoste@inria.fr>]

Hello,

I'll try to right a simpler example.

XL.
Le 23 févr. 2012 à 15:42, Cyril Roelandt a écrit :

> On 02/21/2012 03:57 PM, Xavier Lacoste wrote:
>> 
> 
>> I have made few modification on the source code since I noticed the bug 
>> but I hope it is still there.
>> 
> 
> Indeed, the bug is still there.
> 
> I can confirm that your code works with r5691, but not with r5692.
> 
> I applied the following patch :
> 
> --- starpu_submit_tasks.c.bak   2012-02-22 17:39:19.000000000 +0100
> +++ sopalin/src/starpu_submit_tasks.c   2012-02-23 15:02:50.000000000 +0100
> @@ -647,7 +647,7 @@
>     matrix_child->offset = matrix_father->offset + current_pos*elemsize;
>   }
> }
> -#define USE_TASK_DEP
> +//#define USE_TASK_DEP
> 
> static void starpu_partition_data(Sopalin_Data_t * sopalin_data,
>                                   starpu_data_handle_t * L_handle,
> @@ -1237,7 +1237,9 @@
>            iterbloc < SYMB_BLOKNUM(itercblk+1);
>            iterbloc ++)
>         {
> +#ifdef USE_TASK_DEP
>           starpu_task_destroy(starpu_tasktab[iterbloc]);
> +#endif
> #ifndef STARPU_BLOCKTAB_GEMM
> #if (defined SPARSE_GEMM && defined CHOL_SOPALIN)
>           if (iterbloc > SYMB_BLOKNUM(itercblk))
> @@ -1254,7 +1256,9 @@
> #endif
> #endif
>     }
> +#ifdef USE_TASK_DEP
>   memFree_null(starpu_tasktab);
> +#endif
> #ifdef STARPU_BLOCKTAB_GEMM
> #ifdef STARPU_BLOCKTAB_SELFCOPY
>   {
> 
> in order to run your code without task dependencies. It works this way. 
> Since r 5692 changes a lot of code related to task dependencies, I think 
> we're getting close to understanding your bug :)
> 
> I would love to reproduce this bug with a simpler test case, but I have a 
> hard time figuring out how the dependencies work in starpu_submit_tasks(). 
> Would you mind writing a simple test case that declares the same kind of 
> dependencies and crashes ? It would really help us in the debugging process.
> 
> Regards,
> Cyril Roelandt.
>