MXNet white paper "Auto-Differentiating Linear Algebra"


This arxived white paper Auto-Differentiating Linear Algebra
is relevant for us. Focus is in the MXNet implementation but there is some useful discussion of general implementation issues. Important parts of the code are in C++ and using BLAS + LAPACK for CPU and CUBLAS+CUSolver for GPU.


mxnet.autograd is so convenience for me, it is very simple