RMM doesn't require the CUDA language to be enabled by consumers

[robertmaynard]

This allows consumers of libraries such as cudf that offer a C++ interface to not require the CUDA language be enabled.

[kkraus14]

@gpucibot merge

[gnaggnoyil]

I'm sorry, but if the goal is "to allow consuming targets that offers C++ interface only not requiring CUDA enabled", wouldn't it be a better solution to let the consuming target link rmm as private interface instead? The change here seems to indicate that rmm::rmm requires a C++ standard but doesn't require any CUDA ones, which seems not to fit the goal very well I think.

[jrhemstad]

I'm sorry, but if the goal is "to allow consuming targets that offers C++ interface only not requiring CUDA enabled", wouldn't it be a better solution to let the consuming target link rmm as private interface instead? The change here seems to indicate that rmm::rmm requires a C++ standard but doesn't require any CUDA ones, which seems not to fit the goal very well I think.

I'm not sure what precipitated this change, but strictly speaking, by design, RMM doesn't have any CUDA device code in its headers. It obviously has CUDA runtime calls like cudaMalloc, but no kernel launches nor device-side code that would require compiling with nvcc.

As such, I don't think it needs to require the CUDA language enabled, but just have the rmm::rmm target link against cudart. That said, I'm pretty willfully ignorant on build issues like this, so I may be wildly off base.

[gnaggnoyil]

I see. That makes sense. Thanks for your explanation!

[robertmaynard]

As such, I don't think it needs to require the CUDA language enabled, but just have the rmm::rmm target link against cudart. That said, I'm pretty willfully ignorant on build issues like this, so I may be wildly off base.

You are correct. The reason for this change is that before CMake 3.22 stating target_compile_features(X cuda_std_14) would error out if the consuming project didn't have the CUDA language enabled. This was an issue for RMM C++ consumers which didn't shouldn't need the CUDA language enabled.