-
Notifications
You must be signed in to change notification settings - Fork 407
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
HIP: Fix concurrency #6479
HIP: Fix concurrency #6479
Conversation
Can one of the admins verify this patch? |
OK to test. |
The |
It is someone from AMD that wrote that code... @IanBogle do you know if now there is a way to get the sustainable number of wavefronts / CU instead of the maximum of WF / CU? |
I don't know off the top of my head, but I'm asking around. What do you mean by "sustainable number of wavefronts/CU"? Is it sustainable in terms of occupancy or some other characteristic? |
Yes sustainable in term of occupancy. @arghdos wrote that code a while back kokkos/core/src/HIP/Kokkos_HIP.cpp Lines 70 to 72 in 41cf2e5
I wonder if there is a new property that would give us that number now. |
I don't know of any property that exposes this number, but I'll check with @arghdos when he's back in the office. I have checked recent versions of the file that is referenced in the code snippet, and it is still looks accurate. |
Thanks for looking @IanBogle! |
Unrelated OpenMPTarget failure
|
This PR fixes
i.e., Kokkos hardodes in
Kokkos_HIP.cpp
the maximum number of waves per CU, and the goal of this PR is to make computations of the concurrency consistent with this hardcoded number.Longer term, should we try to work towards removing this hardcoded value, and ping somebody from AMD to bring this issue to their attention and ask if there is a way for the hip function to ideally return directly the appropriate value for
maxThreadsPerMultiProcessor
?