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

[Maxim Abalenkov <maxim.abalenkov@gmail.com>]

Hello Samuel,

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. Therefore, now instead of “double” data registration, I would like to stick with tiles only, and provide panels as an abstract “container” entity. They will also help me to get the task dependencies right. I just don’t know how to unpack the buffers for tiles and pivots. (Please see my previous email for an example). Can you provide a simple example please? I couldn’t find anything similar in the examples directory. Thank you.

—

Best wishes,

Maxim

Maxim Abalenkov \\ maxim.abalenkov@gmail.com
+44 7 486 486 505 \\ http://mabalenk.gitlab.io

On 28 Dec 2017, at 19:13, Samuel Thibault <samuel.thibault@inria.fr> wrote:

Hello,

Maxim Abalenkov, on jeu. 28 déc. 2017 18:05:22 +0200, wrote:

I hope all is well with you. I need your help please. I have successfully implemented a panel-wise version of the LU factorisation in StarPU using the "multiple views" approach. However, multiple views create a drawback of using twice as much memory compared to the tiles approach (when the matrix is divided into square or rectangular blocks). The memory for the tiles and the memory for the panel views.

Err, but IIRC you are not using GPUs? Multiple views do not bring more
memory consumption, they only bring different ways for looking at the
same data.

Samuel