Proposal: Stop cccl include paths conflicting with cuda-cccl when used with nvcc

[bdice]

Solution to issue cannot be found in the documentation.

• I checked the documentation.

Issue

Proposal

Install CCCL headers into a nested ${PREFIX}/include/cccl/ directory to resolve include-order conflicts

Discussed with @vyasr @robertmaynard.

Problem Summary

The current cccl conda package installs its headers directly into ${PREFIX}/include. Meanwhile, the CUDA Toolkit (CTK) ships its own copy of CCCL into, e.g., ${PREFIX}/targets/x86_64-linux/include via cuda-cccl. Since both paths are added as system include directories (-isystem) like...

-isystem /usr/local/cuda/targets/x86_64-linux/include -isystem $PREFIX/include

...the compiler ends up picking the CTK copy first, causing include‐order issues. This manifests in issues like those seen here with RMM (rapidsai/rmm#1930 (comment)), where the goal is to use the conda-provided cccl but the older cuda-cccl copy is found instead.

The current issue is unavoidable without changing compiler flags because nvcc will always have an -isystem for its headers in the targets/... directory. However, we can't change the priority of ${PREFIX} without breaking normal operations in the conda sysroot, so we need to use a different include path.

This issue results in unreliable header selection: Projects depending on the conda‐installed cccl can inadvertently pick up the wrong headers from the CTK.

Also worth noting here is that CCCL is unique among CTK libraries: It’s the only package we install twice (once in ${PREFIX}/include, once in ${PREFIX}/targets/…/include) and is an implicit include of the nvcc compiler. This is a special case and doesn't affect other packages.

Proposed Solution

Change install path in the cccl-feedstock recipe so that headers go under ${PREFIX}/include/cccl/ instead of directly into ${PREFIX}/include.

Rationale: Conda will no longer treat the nested include/cccl as an implicit system include. CMake can add it with -I (user include) rather than -isystem, giving the cccl package higher search priority than the CTK copy in cuda-cccl.

Recipe updates: In meta.yaml, patch the CMake install rules and/or build.sh install steps to create the cccl/ subdirectory and install headers there.

We should confirm with CCCL maintainers that this directory layout is acceptable, and if changes should be upstreamed in some way.

Backward‐Compatibility / Migration

CMake consumers of CCCL should not require any changes except reconfiguring / doing a clean build.

Consumers should continue to #include <foo.h> by updating include paths to have ${PREFIX}/include/cccl and should not change to #include <cccl/foo.h>.

Installed packages

n/a

Environment info

n/a

[bdice]

The same changes have been applied to CCCL in CUDA 13.0, for similar reasons to those described above. We now have an additional reason to make this change, for alignment with the CUDA Toolkit layout.