bugs in the adv benchmark ?

[zjin-lcf]

Dumping the output of the adv benchmark shows that the results are different significantly for each run. Can you reproduce that ?

Thanks

[stgeke]

That’s expected. We use a random input.

…

On 14 Nov 2020, at 13:42, zjin-lcf ***@***.***> wrote:  Dumping the output of the adv benchmark shows that the results are different significantly for each run. Can you reproduce that ? Thanks — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[zjin-lcf]

When a seed is used for initializing input data, the outputs still differ significantly for each run.

In addition, do you think the range is right for s_cubInterpT as shown below ? I am not familiar with your domain. Could you please provide a CPU reference implementation for comparison/verification ?

Thanks

"""
@shared dfloat s_cubD[p_cubNq][p_cubNq];
@shared dfloat s_cubInterpT[p_Nq][p_cubNq];

for(int j = 0; j < p_cubNq; ++j; @inner(1))
  for(int i = 0; i < p_cubNq; ++i; @inner(0)) {
    const int id = i + j * p_cubNq;
    if (id < p_Nq * p_cubNq)
      s_cubInterpT[0][id] = cubInterpT[id];  // s_cubInterpT[0][id]  where id ranges from 0 to  p_cubNq, not from 0 to p_Nq * p_cubNq

"""