Re: [Starpu-devel] About dag.dot

[Xavier Lacoste <xl64100@gmail.com>]

Le 13 mars 2014 à 18:01, Samuel Thibault <samuel.thibault@ens-lyon.org> a 
écrit :

> Hello,
> 
> Xavier Lacoste, le Thu 13 Mar 2014 12:01:08 +0100, a écrit :
>> An other question : what are sync_task_seq_cons ?
> 
> It actually comes from starpu_data_unregister.  I have now changed the
> code, so that information gets propagated there.
> 
>> When generating a DAG using STARPU_GENERATE_TRACE=1 I get several separated
>> trees on my plot while I'm expecting one single tree.
>> Is that the normal behaviour ?
> 
> No: we're supposed to be able to track data dependencies, even when they
> happen to be useless (e.g. A->B but task B was submitted after task A
> finished), but perhaps there is a bug there. How are you tasks
> submitted?  Would it be possible to run your application?
Yes, maybe it comes from the fact I use nested task ?
You can run it on plafrim/mirage :
You can copy ~lacoste/ricar git repository
cd ricar/src
source loadEnv.mpi (You can edit this file to use your own starpu version)
STARPU_GENERATE_TRACE=1 ./example/bin/simple -iparm IPARM_STARPU API_YES 
-iparm IPARM_DEFAULT_ORDERING API_NO -iparm IPARM_ORDERING_CMIN 20 -t 1 -mm 
matrix/young4c.mtx

to rebuild PaStiX you can run make examples
If you don't want to rebuild PaStiX you can copy only the binary file.

> 
>> A last one, I expected to see comutable task as several GEMM tasks 
>> unlocking
>> one HETRF_TRSM task but this is not what I get... Is that expected ?
> 
> AIUI that should have been so, yes. I guess that might be related with
> the potential bug mentioned above.
> 
> Samuel