Re: [Starpu-devel] Strange behaviour using GPUs

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

Xavier Lacoste, le Mon 15 Jul 2013 08:29:09 +0200, a écrit :
> I regenerated the DGEMM models with StarPU r10552 and I still have the same 
> strange behaviour with CUDA Kernel : 
> http://img5.imageshack.us/img5/5117/j9o.png,
>  this explain why when I let StarPU be dynamic it can't find a good 
> scheduling.
> But when I force the scheduling I'm still far from what ParSEC can do with 
> nearly the same GPU scheduling.
> 
> I don't think there should be so much variations with the GEMM kernel on 
> GPU....

Sure.  I wonder where that comes from, as we don't have so big deviation
e.g. with the magma+plasma gemm on mirage:

performance model for cpu_impl_0
# hash          size            flops           mean (us)       stddev (us)   
          n
24c84a50        22118400        0.000000e+00    2.147992e+06    4.833901e+05  
  681
performance model for cuda_0_impl_0
# hash          size            flops           mean (us)       stddev (us)   
          n
24c84a50        22118400        0.000000e+00    5.777425e+03    1.662538e+02  
  6535

IIRC you said you had issues with
specifying the base size for the perfmodel regression?

Samuel