Re: [Starpu-devel] starpu 1.1 and execute_on_worker

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

Re-hello,

In fact I think that this part is simply missing in the loop :

else if (arg_type==STARPU_EXECUTE_ON_WORKER)

{

// the flag is decoded and set later when

// calling function _starpu_task_insert_create()

va_arg(varg_list, int);

}

XL.

----------------------------------------

Xavier Lacoste

xavier.lacoste@inria.fr

INRIA Bordeaux Sud-Ouest

200, avenue de la Vieille Tour

33405 Talence Cedex

Tél : +33 (0)5 24 57 40 69

Le 26 janv. 2015 à 17:50, Xavier Lacoste <xavier.lacoste@inria.fr> a écrit :

Hello,

I moved to branche 1.1 (from trunk 1.2) and get an error when submiting a task.

        ret =

            starpu_mpi_insert_task(sopalin_data->sopar->pastix_comm, cl,

                                   STARPU_VALUE, &sopalin_data, sizeof(z_Sopalin_Data_t*),

                                   STARPU_VALUE, &cblk,        sizeof(z_SolverCblk*),

                                   STARPU_VALUE, &blok,        sizeof(z_SolverBlok*),

                                   STARPU_VALUE, &fcblk,       sizeof(z_SolverCblk*),

                                   STARPU_R,     L_handle[itercblk],

                                   STARPU_COMMUTE|STARPU_RW,    *L_target_handle,

                                   STARPU_SCRATCH, starpu_loop_data->WORK_handle,

                                   STARPU_R,  starpu_loop_data->blocktab_handles[dst_proc],

                                   STARPU_EXECUTE_ON_WORKER, workerid,

                                   STARPU_CALLBACK,     z_prof_callback,

                                   STARPU_CALLBACK_ARG, sopalin_data->gemm_stats,

                                   STARPU_PRIORITY, prio,

                                   0);

The starpu_mpi_insert_task goes through the va_arg list several time and at line 469, when examining STARPU_EXECUTE_ON_WORKER, it goes line 473 (as if a buffer was the next argument) and then crash when trying to get the rank of what StarPU expect to be a data which is in fact the workerid...

This happen without MPI on curie (CCRT). I may be doing something wrong but the fact that the other va_list inspection succeed let me think there can be a mistake in the tests in this loop.

If I remove the STARPU_EXECUTE_ON_WORKER part it works again.

regards,

XL.

----------------------------------------

Xavier Lacoste

xavier.lacoste@inria.fr

INRIA Bordeaux Sud-Ouest

200, avenue de la Vieille Tour

33405 Talence Cedex

Tél : +33 (0)5 24 57 40 69