STRIDE-N Kernels and number of iterations

Question

STRIDE-N Kernels and number of iterations

Closed this issue 2 years ago · 4 comments

After some discussion with @BrodyWilliams, we agreed that the number of iterations for the STRIDE-N Kernels for all backends except MPI, OPENSHMEM, and xBGAS are incorrect. If N is the number set by the user to be performed by each thread, the backends are performing N / stride iterations. Only the 3 mentioned before are performing the correct number of iterations. (Differing from the image that was extracted from the 1st CT paper)

Also, the number of iterations used to calculate the GAMs for each kernel call is still N, not the actual number of iterations

pbbdasilva commented 2 years ago

Yes

Answer 1 · 2022-10-10T11:52:33.000Z

Is the fix for this in PR #19 ?

Answer 2 · 2022-10-10T14:44:14.000Z

Excellent, tracking the PR back to the original issue.

Answer 3 · 2022-10-10T17:05:17.000Z

Merged in PR #19