The current version of CUDA (11.1.1-1) doesn't provide the "version.txt" file as it used to. This is a problem for clang, as it's currently [1] the only way it determines the installed CUDA version. Clang fallbacks to CUDA 7.0 [2] and is unable to compile .cu files because it can not find the libdevice file [3].

$ clang --cuda-device-only --cuda-path=/opt/cuda --cuda-gpu-arch=sm_35 -c test.cu
clang-11: error: cannot find libdevice for sm_35. Provide path to different CUDA installation via --cuda-path, or pass -nocudalib to build without linking with libdevice.

The "fix" consist on adding a "version.txt" file on the cuda package (full path: /opt/cuda/version.txt) with the following content: "CUDA Version 11.1.1". Of course, the version information could be derived from the PKGBUILD metadata. With the file, clang is able to compile cuda files:

$ clang --cuda-device-only --cuda-path=/opt/cuda --cuda-gpu-arch=sm_35 -c test.cu
clang-11: warning: Unknown CUDA version 11.1. Assuming the latest supported version 10.1 [-Wunknown-cuda-version]

This should be easily validated with `clang -v` output:

Found CUDA installation: /opt/cuda, version X.XX

Creating the file locally is trivial, but it takes a lot of time to find the real cause. If the feature it not accepted, at least I hope this information is useful for other Arch users.

Thanks.

[1]: There is already a patch merged upstream recovering the cuda version from the "cuda.h" header, making this feature unnecessary once clang 12 is released: https://github.com/llvm/llvm-project/commit/e7fe125b776bf08d95e60ff3354a5c836218a0e6

[2]: https://github.com/llvm/llvm-project/blob/llvmorg-10.0.1/clang/lib/Driver/ToolChains/Cuda.cpp#L147-L155

[3]: https://github.com/llvm/llvm-project/blob/llvmorg-11.0.0/clang/lib/Driver/ToolChains/Cuda.cpp#L165-L166