Support NVidia CUDA compilation with nvcc

[fhunleth]

Pelemay uses nvcc to generate CUDA code and would require it when targeting devices with NVidia GPUs (like the Jetson Nano)

The path to nvcc should be set in the NVCC environment variable.

[zacky1972]

Thank you.

Usually, nvcc is located at /usr/local/cuda/bin at least in case of Ubuntu x86_64.

[mobileoverlord]

Is this an option that could be brought into the tool chain via CTNG?

[fhunleth]

I'm quite sure that the answer is no. Given that it's NVidia, I bet that it's a set of precompiled binaries and header files that need to be plucked out of their SDK. It would be way better to have a separate "toolchain" archive for it, but I don't know how to do that.

[zacky1972]

I mistook it.

To cross-compile .cu code by nvcc, we need to set -ccbin and -Xcompiler options to a cross-compiler of C++, and its compiling options, respectively.

Does Nerves set CXX environment variable?

[fhunleth]

Yes, CXX is set and the compiler options are in CXXFLAGS. Here's where it's set: https://github.com/nerves-project/nerves_system_br/blob/master/nerves_env.exs#L124-L136.

[zacky1972]

That's great! Thus, I can implement Pelemay to work Jetson, even if on Nerves in future.