[Starpu-devel] Performance issues when specifying const and __attribute__ ((output))

[Mehdi AMINI <mehdi.amini@silkan.com>]

Hi,

I'm still playing with the GCC plugin.

I have a quite simple code :

   /* Start timer. */
   timer_start();

   p4a_launcher_main(s, ny);
   p4a_launcher_main_1((double *)A, p, q, nx, ny);
   for(i = 0; i <= nx-1; i += 1) {
      double r_0 = r[i];
      p4a_launcher_main_2((double *)A, s, i, ny, r_0);
   }
   /* Stop and print timer. */
   #pragma starpu unregister s
   #pragma starpu unregister q
   timer_stop_display();

A,r,s,p,q are arrays registered to StarPU. The p4a_launcher_xxxx 
functions are tasks that launch very simple CUDA kernel. A CPU 
implementation is also given but is avoided at runtime using 
STARPU_NCPUS environment variable.



I tried to make use of "const" qualifier and "__attribute__((output))" 
on task parameter to hint StarPU about the data transfer.

However it seems to impact badly the runtime since using only one GPU I 
get the following execution time:


$ STARPU_NCUDA=1 STARPU_NCPUS=0 ./bicg_starpu_naive
192.2
$ STARPU_NCUDA=1 STARPU_NCPUS=0 ./bicg_starpu
746.5

Nearly four times slower !
Using CUDA_PROFILE=1 environment variable I confirmed that the GPU time 
is very close for both execution, and that the only difference is the 
few more data transfers for the fastest naive version.

Is it a known issue or do you have hints to explain this behavior?

Thanks,

Mehdi