[Starpu-devel] Applications allowed to do computations in their own cores ?

[Stephanie Even <Stephanie.Even@hpc-project.com>]

Hello,

I'm working on a classic example of matrice multiplication to test
StarPU on a 12 cores CPU and 1 GPU.  To extract performance as simply
as possible, I was trying to define 2 tasks, 1 on the CPU and another
1 on the GPU. The aim was to let MKL do the job of parallelization on
the 12 cores of CPU. But I have discovered that it is not possible
with StarPU according to page 17 of the manual, § 3.3 "Kernel threads
started by StarPU" which is flagged as TODO. The GFLOPS on the CPU are
only on sequential mode (despite StarPU displays 1200 % CPU. I guess
it is only busy waiting). Of course, I can define the appropriate
tasks manually, but it will be much more complicated but less
efficient as Intel engineer are probably quite pressured on it :-).


Is it possible to make StarPU allowing the user to use its own
parallelism ? It will be great. To be done in mediaGPU, for example ?


In advance, thank you,
-- 
Stéphanie EVEN                         
HPC Project                  
Techmart
5201 Great America Parkway #3241
Santa Clara, CA 95054