[CMake] Explicitly link with CCCL (standalone or CTK)

[hcho3]

Related: conda-forge/xgboost-feedstock#167

[trivialfis]

Do we want to replace thrust as well?

[jakirkham]

Thanks Hyunsu! 🙏

Had a couple questions below

[hcho3]

@jakirkham It seems that for CUDA 11.8, CTK doesn't have the CMake config for CCCL:

$ find /usr/local/cuda/lib64/cmake/
/usr/local/cuda/lib64/cmake/
/usr/local/cuda/lib64/cmake/libcudacxx
/usr/local/cuda/lib64/cmake/libcudacxx/libcudacxx-config.cmake
/usr/local/cuda/lib64/cmake/libcudacxx/libcudacxx-header-search.cmake
/usr/local/cuda/lib64/cmake/libcudacxx/libcudacxx-config-version.cmake
/usr/local/cuda/lib64/cmake/thrust
/usr/local/cuda/lib64/cmake/thrust/README.md
/usr/local/cuda/lib64/cmake/thrust/FindTBB.cmake
/usr/local/cuda/lib64/cmake/thrust/thrust-config-version.cmake
/usr/local/cuda/lib64/cmake/thrust/thrust-config.cmake
/usr/local/cuda/lib64/cmake/cub
/usr/local/cuda/lib64/cmake/cub/cub-config-version.cmake
/usr/local/cuda/lib64/cmake/cub/cub-config.cmake

Do we need CTK 12.0+ to access the built-in CCCL? cc @bdice

[jakirkham]

That sounds plausible

Can we just fallback to Thrust & CUB in that case?

[jakirkham]

Thanks Hyunsu! 🙏

Added a suggestion below

[bdice]

Do we need CTK 12.0+ to access the built-in CCCL? cc @bdice

The better way to get CCCL is via GitHub or conda, rather than the CTK: https://github.com/NVIDIA/cccl?tab=readme-ov-file#github

That way you can use a CCCL version that works with all the supported CTK versions, rather than having to make exceptions for every CUDA version in the build scripts. I think you need at least CCCL 2.2.0 to make things work properly with CMake, which shipped with CUDA 12.3.0. If you fetch CCCL from GitHub via CPM (or use a conda package of CCCL) you can get a more recent version that will support both CUDA 11 and 12. Currently RAPIDS is using CCCL 2.5.0.