fast FwdMaxUnpooling GPU kernel

[ghost]

in file https://github.com/HyeonwooNoh/caffe/blob/master/src/caffe/layers/unpooling_layer.cpp
you asked: TODO(Yangqing): Is there a faster way to do unpooling in the channel-first

I tried also to implement a fast ForwardMaxUnPooling.
I only made the GPU version. See at the bottom of this file:
https://github.com/3DdeepAI/3DdeepAI/blob/3DdeepAI/NNkernels.h
kernel "FwdMaxUnpooling"

In my implementation there are less divisions: 2 instead of 6

https://github.com/HyeonwooNoh/caffe/blob/master/src/caffe/layers/unpooling_layer.cu

[ghost]

I updated the previous post and changed the code:
https://github.com/3DdeepAI/3DdeepAI/blob/3DdeepAI/NNkernels.h
Now it works. And it only uses 2 divisions (instead of 6 in your code)