cudnnSetConvolutionNdDescriptor now requires cudnn storage type
cudnnAddTensor no longer allows us to specify add method
cudnnConvolutionBackwardData and cudnnConvolutionBackwardFilter now take three additional arguments (algo, workSpace, and workSpaceSizeInBytes). The documentation states that if we set these to CUDNN_CONVOLUTIONBWD{DATA|FILTER}_ALGO_0, NULL, and 0, respectively, we should have identical behavior as before
Tested on Ubuntu 15.04 and OS X 10.11.1