NNPACK

NNPACK is an acceleration package for neural network computations. NNPACK aims to provide high-performance implementations of convnet layers for multi-core CPUs.

NNPACK is not intended to be directly used by machine learning researchers; instead it provides low-level performance primitives to be leveraged by higher-level frameworks, such as Caffe, Torch, MXNet, Theano, Tensorflow, and Mocha.jl.

Requirements

Host system

• Linux or OS X host system
• x86-64 processor with AVX2 instruction set
  • NNPACK is optimized for Intel Skylake, but can run on Haswell & Broadwell processors too
  • SSE2 instruction set can be targeted using --backend=psimd or --backend=scalar configuration options, but for performance reasons it is not recommended for production use
• ARMv7 processor with NEON instruction set
  • VFP instruction set (including ARMv6 systems with VFPv2) can be targeted using --backend=scalar configuration option, but for performance reasons it is not recommended for production use.
• ARMv8 (AArch64) processor

Cross-compilation options:

• Android with x86/x86-64 (SSE2), ARMv7 with NEON, or ARM64 architecture
• WebAssembly for next-generation Web browsers
• Emscripten/Asm.js to run inside any modern Web browser
• Portable Native Client to run inside Google Chrome (no packaging required)
• Native Client (x86-64) to run as a packaged Google Chrome App

Features

• Fast convolution algorithms based on Fourier transform and Winograd transform.

  • Forward propagation performance on Intel Core i7 6700K vs BVLC Caffe master branch as of March 24, 2016 (protobufs from convnet-benchmarks, integration via caffe-nnpack):

  Library	Caffe	NNPACK	NNPACK	NNPACK
  Algorithm	im2col + sgemm	FFT-8x8	FFT-16x16	Winograd F(6x6, 3x3)
  AlexNet:conv2	315 ms	129 ms	86 ms	N/A
  AlexNet:conv3	182 ms	87 ms	44 ms	70 ms
  AlexNet:conv4	264 ms	109 ms	56 ms	89 ms
  AlexNet:conv5	177 ms	77 ms	40 ms	64 ms
  VGG-A:conv1	255 ms	303 ms	260 ms	404 ms
  VGG-A:conv2	902 ms	369 ms	267 ms	372 ms
  VGG-A:conv3.1	566 ms	308 ms	185 ms	279 ms
  VGG-A:conv3.2	1091 ms	517 ms	309 ms	463 ms
  VGG-A:conv4.1	432 ms	228 ms	149 ms	188 ms
  VGG-A:conv4.2	842 ms	402 ms	264 ms	329 ms
  VGG-A:conv5	292 ms	141 ms	83 ms	114 ms
  OverFeat:conv2	424 ms	158 ms	73 ms	N/A
  OverFeat:conv3	250 ms	69 ms	74 ms	54 ms
  OverFeat:conv4	927 ms	256 ms	272 ms	173 ms
  OverFeat:conv5	1832 ms	466 ms	524 ms	315 ms

• Built-in expert-tuned kernels with very high performance:

  • Fast Fourier transform
  • Winograd transform
  • Matrix-matrix multiplication (GEMM)
  • Matrix-vector multiplication (GEMV)
  • Max-pooling.

• Multi-threaded SIMD-aware implementations of neural network layers.

• Implemented in C99 and Python without external dependencies.

• Extensive unit tests using C++ and Google Test.

• Supports Native Client target and outperforms native Caffe/CPU when running inside Chrome.

Layers

• Convolutional layer
  • Training-optimized forward propagation (nnp_convolution_output)
  • Training-optimized backward input gradient update (nnp_convolution_input_gradient)
  • Training-optimized backward kernel gradient update (nnp_convolution_kernel_gradient)
  • Inference-optimized forward propagation (nnp_convolution_inference)
• Fully-connected layer
  • Training-optimized forward propagation (nnp_fully_connected_output)
  • Inference-optimized forward propagation (nnp_fully_connected_inference)
• Max pooling layer
  • Forward propagation, both for training and inference, (nnp_max_pooling_output)
• ReLU layer (with parametrized negative slope)
  • Forward propagation, both for training and inference, optionally in-place, (nnp_relu_output)
  • Backward input gradient update (nnp_relu_input_gradient)
• Softmax layer
  • Forward propagation, both for training and inference, optionally in-place (nnp_softmax_output)

Building

NNPACK can be build on OS X and Linux.

Install ninja build system

sudo apt-get install ninja-build || brew install ninja

Install PeachPy assembler and confu configuration system

[sudo] pip install --upgrade git+https://github.com/Maratyszcza/PeachPy
[sudo] pip install --upgrade git+https://github.com/Maratyszcza/confu

Then clone NNPACK, install dependencies, configure, and build

git clone https://github.com/Maratyszcza/NNPACK.git
cd NNPACK
confu setup
python ./configure.py
ninja

Cross-compilation for Android

• Download and setup Android NDK
• Add ndk-build to PATH variable
• Navigate to NNPACK directory and setup dependencies (confu setup)
• Build NNPACK with ndk-build build system.

Cross-compilation for Emscripten/WebAssembly

• Download and setup upstream version of Emscripten SDK
• Using emsdk, download, build and activate incoming version of Emscripten and Binaryen, and setup environment variables. $EMSCRIPTEN should specify the path to activated Emscripten environment.
• Configure NNPACK with --target=wasm option.

Cross-compilation for Emscripten/Asm.js

• Download and setup Emscripten SDK
• Using emsdk, download, build and activate one of the environments, and setup environment variables. $EMSCRIPTEN should specify the path to activated Emscripten environment.
• Configure NNPACK with --target=asmjs option.

Cross-compilation for Portable Native Client

• Download and setup Native Client SDK
• Set NACL_SDK_ROOT variable to a versioned SDK directory (e.g. /opt/nacl_sdk/pepper_49).
• Configure NNPACK with --target=pnacl option.

Cross-compilation for Native Client

• Download and setup Native Client SDK
• Set NACL_SDK_ROOT variable to a versioned SDK directory (e.g. /opt/nacl_sdk/pepper_49).
• Configure NNPACK with --target=x86_64-nacl-newlib (recommended) or --target=x86_64-nacl-gnu option.

Testing

NNPACK contains extensive test suite for transformation and neural network layers.

After configuration type ninja smoketest to run a set of quick tests, or ninja test to additionally NNPACK layers with parameters from AlexNet, VGG-A, and Overfeat-Fast networks (this will take a while).

Packaging

Binary packages need to distribute two files: include/nnpack.h and lib/libnnpack.a (also lib/libnnpack.so or lib/libnnpack.dylib if NNPACK was configured with shared library support).

Bindings

Deep Learning Frameworks

• Caffe2 natively supports NNPACK
• MXNet - supports NNPACK for inference in convolutional layers, fully-connected, and max-pooling layers. See MXNet wiki for configuration instructions and performance benchmarks).
• szagoruyko/nnpack.torch - integration of NNPACK into Lua Torch via ffi
• tiny-dnn - header-only deep learning framework in C++11, which natively supports NNPACK.
• darknet-nnpack - fork of Darknet framework with NNPACK support.
• Maratyszcza/caffe - up-to-date integration of NNPACK (convolutional, fully-connected, max-pooling, and ReLU layers) into Caffe based on nnpack-pr branch in ajtulloch/caffe.
• Maratyszcza/caffe-nnpack - older and unmaintained integration of NNPACK (convolutional layers only) into Caffe.
• See also discussion in Issue #1

Languages and Environments

• node-nnpack - Node.js bindings
• peterhj/libnnpack - Rust bindings

Users

• Facebook uses NNPACK in production.
• Prisma uses NNPACK in the mobile app.

Acknowledgements

The library is developed by Marat Dukhan of Georgia Tech with extensive advice from Nicolas Vasilache and Soumith Chintala of Facebook Artificial Intelligence Research. Andrew Tulloch of Facebook Artificial Intelligence Research contributed Caffe integration. We thank Andrew Lavin for fruitful discussions on Winograd transform-based implementations. NNPACK is a research project at Richard Vuduc's HPC Garage lab in the Georgia Institute of Technology, College of Computing, School of Computational Science and Engineering.

This material is based upon work supported by the U.S. National Science Foundation (NSF) Award Number 1339745. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect those of NSF.