CompatHelper: bump compat for "CUDA" to "5" #155
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
|
Benchmark results for commit dedaad6 (comparing to bca8c30):
|
Codecov ReportPatch and project coverage have no change.
Additional details and impacted files@@ Coverage Diff @@
## master #155 +/- ##
=======================================
Coverage 32.17% 32.17%
=======================================
Files 11 11
Lines 889 889
=======================================
Hits 286 286
Misses 603 603 ☔ View full report in Codecov by Sentry. |
|
The change in registers is due to the removal of a quirk in JuliaGPU/CUDA.jl@760c2bd#diff-d1deb39259ffe01ecfe5db7c6c2331af73b81772539dd8ef49076d58fa2529d2, but AFAICT that actually improves performance. I still have to find the culprit of the other slowdowns, but I can't reproduce it locally. |
maleadt
force-pushed
the
compathelper/new_version/2023-09-20-00-30-53-609-4285891187
branch
from
16416fe
to
47ee618
Compare
|
Painfully bisected to JuliaGPU/CUDA.jl#2025. Which doesn't make sense, as we manually synchronize the stream, and thus do not rely on non-blocking synchronization. The only point nonblocking synchronization is used, is when allocating the arrays during the set-up of each benchmark. Maybe we're just introducing lots of noise there right before each benchmark? Getting rid of that seems to greatly improve timings, so hopefully that fixes the issue here. |
maleadt
force-pushed
the
compathelper/new_version/2023-09-20-00-30-53-609-4285891187
branch
from
47ee618
to
dedaad6
Compare
maleadt
deleted the
compathelper/new_version/2023-09-20-00-30-53-609-4285891187
branch
This pull request changes the compat entry for the
CUDApackage from3.5, 4to3.5, 4, 5.This keeps the compat entries for earlier versions.
Note: I have not tested your package with this new compat entry. It is your responsibility to make sure that your package tests pass before you merge this pull request.