Re: [Starpu-devel] Issues with asynchronous partitioning on temporary data

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

Jérôme Richard, on Wed 31 Aug 2016 13:13:44 +0200, wrote:
>     starpu_matrix_data_register(&tmprhs2_f_handles, -1, 0, Nr, Nr, Ntheta, 
> sizeof(double));
>     starpu_data_partition_plan(tmprhs2_f_handles, &f1, 
> &tmprhs2_f_handles_children[0]);
>     starpu_data_partition_submit(tmprhs2_f_handles, f1.nchildren, 
> &tmprhs2_f_handles_children[0]);

Ah, we didn't think about that case. starpu_data_partition_submit
currently assumes that there is always coherency to be done between the
value of the main handle and the sub-handles.  We'd just need to disable
that. What you could do for now to circumvent the issue is to just
submit a W empty task on tmprhs2_f_handles, to make starpu think it is
initialized:

void nop(void* buffers[], void* clArgs)
{
    printf("nop\n");
}

    starpu_codelet null_cl;
    starpu_codelet_init(&null_cl);
    null_cl.where = STARPU_CPU;
    null_cl.cpu_funcs[0] = nop;
    null_cl.nbuffers = 1;
    null_cl.modes[0] = STARPU_W;

    starpu_matrix_data_register(&tmprhs2_f_handles, -1, 0, Nr, Nr, Ntheta, 
sizeof(double));

    starpu_task_insert(&null_cl, STARPU_W, tmprhs2_f_handles, 0);

    starpu_data_partition_plan(tmprhs2_f_handles, &f1, 
&tmprhs2_f_handles_children[0]);    
    starpu_data_partition_submit(tmprhs2_f_handles, f1.nchildren, 
&tmprhs2_f_handles_children[0]);

There was also a bug in our code which wasn't properly updating
coherency for asynchronous partitioning for temporary datas. I have
fixed it in the trunk and 1.2 branch.

Samuel