CUDA and CMAKE_SYSROOT: not picking up libraries from the sysroot

Hi,

We have a cross-compilation toolchain (Linux, aarch64) that uses CUDA. Following the recommended pattern for Linux toolchains, we set the CMAKE_SYSROOT to point to the targetfs containing the /include and /lib from the target platform.

For C++ code this works fine and it uses the libraries from the targetfs. However it does not work for CUDA libraries, because CMake is appending -L /usr/local/cuda/targets/aarch64-linux/lib instead of -L <targetfs>/usr/local/cuda/targets/aarch64-linux/lib. I.e the CMAKE_CUDA_HOST_IMPLICIT_LINK_DIRECTORIES is taken from the host filesystem instead of from the targetfs set by CMAKE_SYSROOT.

Therefore we need ugly symlinks from targetfs to /usr/local to make this work. Is there a flag we can use to tell CMake to use the CUDA libraries that it can find inside the targetfs specified by CMAKE_SYSROOT?

Thanks!