Re: [Starpu-devel] [LU factorisation: gdb debug output]

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

Maxim Abalenkov, on jeu. 28 déc. 2017 19:22:16 +0200, wrote:
> Yes, you are right. I’m not using GPUs at the moment. It is purely aCPUs 
> code.
> My understanding was that once I register the data two times: (i) for panels
> and (ii) for tiles I will request double the memory:
> 
>     // Initialize panel and tile handles
>     for (int j = 0; j < A->nt; j++) {
> 
>         int naj = plasma_tile_nview(*A, j);
> 
>         // Initialise panel handle
>         starpu_matrix_data_register(&((A->panel_handles)[j]),
>                                     STARPU_MAIN_RAM,
>                                     (uintptr_t) plasma_tile_addr(*A, 0, j),
>                                     A->m, A->m, naj, eltsize);
> 
>         // Invalidate panel handle
>         starpu_data_invalidate(A->panel_handles[j]);
> 
>         // Initialise tile handle
>         for (int i = 0; i < A->mt; i++) {
> 
>             int ldai = plasma_tile_mmain(*A, i);
>             int mai  = plasma_tile_mview(*A, i);
> 
>             starpu_matrix_data_register(&((A->tile_handles)[j*A->mt + i]),
>                                         STARPU_MAIN_RAM,
>                                         (uintptr_t) plasma_tile_addr(*A, i, 
> j),
>                                         ldai, mai, naj, eltsize);
>         }
>     }
> 
> Please see the above code.

Note that starpu_matrix_data_register does not allocate memory, it only
tells starpu where the data is. tile_handles will only be sorts of
pointers into plasma_tile_addr.

> Therefore, now instead of “double” data registration, I would
> like to stick with tiles only, and provide panels as an abstract
> “container” entity.

That's already what double-registration does.

Samuel