[ENHANCEMENT]: Enable PDL for all cuco kernels

[sleeepyjack]

The Programmatic Dependent Launch (PDL) mechanism allows for a dependent secondary kernel to launch before the primary kernel it depends on in the same CUDA stream has finished executing. Available starting with devices of compute capability 9.0, this technique can provide performance benefits when the secondary kernel can complete significant work that does not depend on the results of the primary kernel.

Cuco kernels should all support PDL.

[KaranPradhan266]

Hi @sleeepyjack, This looks very interesting. I don’t have prior hands on experience with CUDA or cuCollections yet, but I’m comfortable with modern C++ and would love to take this on as a learning opportunity.

Please let me know if there’s a preferred starting point or a specific kernel you’d recommend looking at first.

Also, I’m developing on macOS without local CUDA access, so I’d be relying on CI and reviewer feedback for runtime validation. Please let me know if that works.

[sleeepyjack]

Hi @KaranPradhan266,
Thanks for your interest!

Sure, we can figure this out together. Relying on CI should be fine I think. It would be great though if can get your hands on a Linux box just for compiling the code and catching compilation errors before going to CI. Note that each CI run might take an hour or so. You don't want to wait on CI for every little syntax error.

As for building/developing cuco, we have a correponding section in the project readme. We're happy to help with any questions.

I reached out to our core libraries team as they added PDL support a couple of months ago and have likely delevoped some utility macros we can make use of. I'll let you know once they got back to me.

In the meanwhile, could you read into the docs I shared in the issue description and familiarize with the PDL feature? That would be great.