It generates unused variable warnings, because now all cudnn routines call PyObject* accumTensorClass = getAccumTensorClass(args); even though only batch norm uses differently typed accum tensors. Could not figure out how to get around it, would appreciate if someone gave me pointers how to do that in cwrap.

Removed BN layer in torch/nn left backend in place.

I'm not sure what's going on in this change. It's changing the BatchNorm code path in C++, but stats will still be half on the Python side? Also, it's going to conflict with @ezyang's changes (to no longer use THVoidTensor)

it's a rebased branch to get Zach going with Volta fp16.
After ed's changes all go in, this can be rebased on top.

Yeah, but this is lacking Python changes. I can't see how will it not error out when you try to use BatchNorm

@apaszke, It does not touch current batch norm in any way, just allows to call cudnn batch norm with fp16 inputs from a custom python module. It will have to be changed after Ed's changes go in.

Replaced with #4021 due to cuDNN rewrite in aten.