grain: mir, LLVM, GPU, CUDA, dynamic neural networks

[Per Nordlöw]

I just discovered

https://github.com/ShigekiKarita/grain

which seems like a very ambitious and active project for making 
dynamic neural networks run on the GPU using D in front of mir 
and CUDA.

Are there any long-term goals around this project except for the 
title?

It would great if someone (author) could write a little 
background-knowledge (tutorial) around the subject of dynamic 
neural networks that assists all the details in the examples at

https://github.com/ShigekiKarita/grain/tree/master/example

Further, could parts of grain be refactored out into some generic 
CUDA-library for use in domains other than dynamic neural 
networks?