cuDNN R2 is available

[ducha-aiki]

CuDNN RC 2 is available for download. https://developer.nvidia.com/cuDNN
They say, it is 40% faster than original cuDNN. However, it breaks a compatibility with caffe-wrappers (more general ways to pass n-d tensors, additional coefficients, etc.) Does anyone know, NVIDIA guys going to support caffe in future, or rewriting cudnn-layers are up to us?

[immars]

Failed compilation with R2.

a piece from Theano:
https://github.com/Theano/Theano/blob/master/theano/sandbox/cuda/cudnn_helper.h
Theano/Theano@66f946d

[shelhamer]

While cuDNN R2 has a different interface, the change is quite slight and we will support it soon by a few changes to include/caffe/util/cudnn.hpp.

While cuDNN R2 is still a release candidate R1 will likely stay the default.

[Felix-neko]

Maybe you can make a temporary branch for RC2?

[shelhamer]

See #1731 for an in-progress cuDNN R2 branch.

The Caffe crew collaborates with the cuDNN team on roadmap planning and interface design but at the moment the Caffe coding happens at Berkeley and in the open source community.

[shelhamer]

Covered by #1731 / #1739 thanks to Simon Layton at NVIDIA.