Add the sm_89 arch to the default CUDA binaries #107585
Without this, support for newer NVIDIA cards will not be compiled with pre-compiling the CUDA binaries.
We'll have to upgrade to a newer CUDA version to support this architecture. I'll put it on my todo list.
Hmm it works on my computer. I guess it might be because I have a newer cuda-toolkit (12.1) perhaps?
EDIT: Gah, the page didn't refresh automatically so I didn't see that you had replied already stating the same thing.
The configuration key naming is a bit misleading: it is called cuda11 but from the find_package perspective and some other semantic meaning it actually means the latest cuda. We should consider renaming the option, but it is to be done outside of the scope of this patch.
Good news: CUDA 12 is installed on all currently available Rocky and Windows workers.
Bad news: it does not seem we can avoid having 3 version of CUDA. With the current configuration the sm_36 and sm_37 are silently (!) missing.
I would need to work further on the build pipeline on the bot side to allow specifying cuda10, cuda11, and cuda12. And somehow do it in a way that does not interfere with LTS builders.
Marking as requested changes to indicate that the current code should not be merged.
EDIT: Not dully silently, there is a print about it, but it is so easy to miss..
The Windows build succeeded after the "Rebuild" on the buildbot (it was some intermittent problem on the worker).
Both Linux and Windows now have full set of CUDA kernels.
From my side the code is good to go.
Deleting a branch is permanent. Although the deleted branch may exist for a short time before cleaning up, in most cases it CANNOT be undone. Continue?