I'm not seeing the connection. This work is about low-level optimization of matrix multiplication. The repo you linked seems to be about replacing back-propagated gradients with a cheaper estimate. What's the similarity you see between these two?
Wellll that means I can give dni another try :D