Torch FFI-bindings for NNPACK

NNPACK is a fast CPU implementation of convolution operations for training ConvNets:

https://github.com/Maratyszcza/NNPACK

The bindings are fully working and tested against nn version. Only single precision supported. Make sure you have AVX2 compatible Skylake/Broadwell/Haswell CPU.

Limitations of NNPACK:

• there is no scale parameter on accGradParameters call

Installation

Follow installation steps at https://github.com/Maratyszcza/NNPACK to generate libnnpack.so and place where LD_LIBRARY_PATH can find it.

Then do

luarocks install https://raw.githubusercontent.com/szagoruyko/nnpack.torch/master/nnpack-scm-1.rockspec

Conversion between nnpack and nn

Similar to cudnn.convert in cudnn.torch easy backend switching is supported. To switch to nnpack just do:

nnpack.convert(net, nnpack)

There will be no memory copy, just metatables will be swapped.

Credits

Thanks to @Maratyszcza for adding the option to generate shared NNPACK library.