FFI cudaLaunchCooperativeKernel?

[staticimport]

Is there any reason why doing cooperative kernels wouldn’t work? I’m seeing unexpected results in some cases vs standard kernels launch, but I’m also seeing much better performance with it too.

[dfm]

I don't know enough off the top of my head, but @ezhulenev might have useful feedback. If you can minimize the issues you're having to something that you think should work, but doesn't, we can always help debug!

[ezhulenev]

Based on CUDA documentation it should work, I don't see any constraints mentioned anywhere. Capturing cudaLaunchCooperativeKernel should work similar to regular launches. @nouiz to confirm.

[nouiz]

I can confirm that it should work. But that pre-hopper, it is less optimized.
Also the coop launch will trigger the highest stream priority by default.