Use static linkage for CUDA runtime

[bdice]

Summary

• Remove cuda-cudart from run requirements in conda recipes

With static linking of the CUDA runtime (the default for C++ and Python packages), the runtime is embedded in the binaries and the cuda-cudart package is not needed at runtime.

Part of rapidsai/build-planning#235

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[jakirkham]

Thanks Bradley! 🙏

The commit history is pulling in a bunch of old commits from past releases. Could we please rebase this?

[robertmaynard]

Locations we directly link to shared cudart:

cpp/CMakeLists.txt:            CUDA::cudart
cpp/plugins/cucim.kit.cumed/CMakeLists.txt:            CUDA::cudart
cpp/plugins/cucim.kit.cuslide/CMakeLists.txt:                CUDA::cudart
cpp/plugins/cucim.kit.cuslide/CMakeLists.txt:                CUDA::cudart
cpp/plugins/cucim.kit.cuslide/tests/CMakeLists.txt:            CUDA::cudart
cpp/plugins/cucim.kit.cuslide2/CMakeLists.txt:                CUDA::cudart
cpp/plugins/cucim.kit.cuslide2/CMakeLists.txt:                CUDA::cudart
cpp/plugins/cucim.kit.cuslide2/tests/CMakeLists.txt:            CUDA::cudart
cpp/tests/CMakeLists.txt:            CUDA::cudart

[jakirkham]

Thanks Kyle! 🙏

This is looking better, but still needs a few things

[jakirkham]

This needs to be dropped too

Suggested change

[KyleFromNVIDIA]

No, we're still compiling against cudart, so we still need it as a host dependency. We've already removed it as a run dependency since we've statically linked it.

[jakirkham]

For static linkage this is already part of the NVCC compiler package

[KyleFromNVIDIA]

Depending on the NVCC Conda package for cudart is wrong. NVCC lives in $BUILD_PREFIX (the build environment), since that's where the executable for the building architecture lives, while we need to get cudart from $PREFIX (the host environment), since that's where the library for the built architecture lives.

This isn't just a theoretical cross-compilation problem. We have previously had real issues with this. See https://github.com/rapidsai/cudf/pull/22290/changes#r3184782324 - we had to add cuda-cudart-dev back as a host dependency because the location of compute-sanitizer was being computed relative to cudart.

The bottom line is that if we're going to build against cudart, we need to have a dependency on cuda-cudart-dev at build time.

[jakirkham]

Same here

Suggested change

[jakirkham]

Also please drop any other instances of cuda-cudart-dev like this one

cucim/dependencies.yaml

Line 218 in 2129c3b

- cuda-cudart-dev

[KyleFromNVIDIA]

We're still including cuda-cudart-dev in the cuda dependency set because it's still needed to build. We just don't include it in the runtime dependencies anymore.