Is there a reason you're not trying to port this into an even more stack agnostic world without CUDA?