CUDNN support for Convolution Layers

[pavanky]

Wraps perceptrons, convolutional and softmax layers

Other

• Integrate with cuDNN when possible
• Add a fallback version using cuDNN is not available.
• Investigate alternatives and analyze performance

[futurely]

facebook/fbcunn exploits cuDNN through a limited number of DNN specific operations implemented by the wrapper cudnn.torch. This is probably the least invasive method to integrate with cuDNN.

The fallback CUDA backend of torch is cunn.

[UniqueFool]

FWIW, this may be worth looking at (Mozilla Public License 2.0): https://github.com/hughperkins/DeepCL
More specfically: https://github.com/hughperkins/DeepCL/tree/master/cl

[pavanky]

This is going to be solved upstream: arrayfire/arrayfire#1868