Re: [Starpu-devel] Using STARPU_COMMUTE with starpu_mpi_task_build()

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Xavier Lacoste, le Wed 26 Feb 2014 16:56:05 +0100, a écrit :
> Le 26 févr. 2014 à 16:30, Samuel Thibault <samuel.thibault@ens-lyon.org> a 
> écrit :
> > Xavier Lacoste, le Wed 26 Feb 2014 15:43:26 +0100, a écrit :
> >> I tried to use STARPU_COMMUTE with MPI but the option gives an error 
> >> when building MPI task.
> >> Indeed it does not appear in the test "if (arg_type==STARPU_R || 
> >> arg_type==STARPU_W || arg_type==STARPU_RW || arg_type==STARPU_SCRATCH || 
> >> arg_type==STARPU_REDUX" in starpu_mpi_task_insert.c
> >> Is this the only thing missing or is it more complicated ?
> > 
> > On the top of my head I think it's a matter of testing against
> > arg_type & ~STARPU_COMMUTE to get it working at least. Getting it
> > optimized is probably another question, could you show us your use case?
> 
> I have several task preforming a GEMM updating a data that can be executed 
> in parallel.
> In shared memory mode I just used 
> starpu_data_set_sequential_consistency_flag() and listed all my 
> dependencies, but as it was more complicated to get the list of tasks in 
> distributed, I tried STARPU_COMMUTE.
> 
> So, in my algorithm, I have several GEMM tasks unlocking a GETRF+TRSM 
> unlocking an other list of GEMMs etc.

Ok, and are all the GEMMs happening on the same MPI node?

Thinking about it, I tend to think that we don't want to optimize the
case where it would not be on the same MPI node, i.e. simply ignore the
STARPU_COMMUTE flag for all the MPI stuff (but of course keep it for the
actual task insertion, to get the commute effect inside MPI nodes.

> Is that what you wanted to know ?

Yes :)

Samuel