Raw wrappers to NVIDIA's nvrtc library. See:
https://docs.nvidia.com/cuda/nvrtc/index.html
In short, you can give this a string of CUDA code, and it will give back the compiled object.
By default, this looks in /usr/local/cuda/ for headers and libraries. If that
is not appropriate for your system, you can set the CUDA_PATH environment
variable.