XNNPACK

XNNPACK is a highly optimized library of floating-point neural network inference operators for ARM, WebAssembly, and x86 (SSE2 level) platforms. XNNPACK is not intended for direct use by deep learning practitioners and researchers; instead it provides low-level performance primitives for accelerating high-level machine learning frameworks, such as MediaPipe, TensorFlow Lite, and TensorFlow.js.

Supported Architectures

• ARM64 on Android and Linux
• ARMv7 (with NEON) on Android and Linux
• WebAssembly MVP
• WebAssembly SIMD (experimental)
• x86 and x86-64 (up to AVX512) on Android, Linux, and macOS

Operator Coverage

XNNPACK implements the following neural network operators:

• 2D Convolution (including grouped and depthwise)
• 2D Deconvolution (AKA Transposed Convolution)
• 2D Average Pooling
• 2D Max Pooling
• 2D ArgMax Pooling (Max Pooling + indices)
• 2D Unpooling
• 2D Bilinear Resize
• Add (including broadcasting, two inputs only)
• Subtract (including broadcasting)
• Divide (including broadcasting)
• Maximum (including broadcasting)
• Minimum (including broadcasting)
• Multiply (including broadcasting)
• Global Average Pooling
• Channel Shuffle
• Fully Connected
• Clamp (includes ReLU and ReLU6)
• HardSwish
• PReLU

All operators in XNNPACK support NHWC layout, but additionally allow custom stride along the Channel dimension. Thus, operators can consume a subset of channels in the input tensor, and produce a subset of channels in the output tensor, providing a zero-cost Channel Split and Channel Concatenation operations.

Performance

Mobile phones

The table below presents single-threaded performance of XNNPACK library on two generations of MobileNet models and three generations of Pixel phones.

Model	Pixel, ms	Pixel 2, ms	Pixel 3a, ms
MobileNet v1 1.0X	81	93	88
MobileNet v2 1.0X	48	58	54

Benchmarked on October 9, 2019 with end2end_bench --benchmark_min_time=5 on an Android/ARM64 build (bazel build -c opt --config android_arm64 :end2end_bench) and neural network models with randomized weights and inputs.

Raspberry Pi

The table below presents multi-threaded performance of XNNPACK library on three generations of MobileNet models and three generations of Raspberry Pi boards.

Model	RPi 2 (BCM2836), ms	RPi 3+ (BCM2837B0), ms	RPi 4 (BCM2711), ms
MobileNet v1 1.0X	342	122	79
MobileNet v2 1.0X	199	82	47
MobileNet v3 Large	166	71	42
MobileNet v3 Small	53	24	15

Benchmarked on December 12, 2019 with end2end_bench --benchmark_min_time=5 on a Raspbian Buster build with CMake (./scripts/build-local.sh) and neural network models with randomized weights and inputs.

Publications

• Marat Dukhan "The Indirect Convolution Algorithm". Presented on Efficient Deep Learning for Compute Vision (ECV) 2019 workshop (slides, paper on ArXiv).
• Erich Elsen, Marat Dukhan, Trevor Gale, Karen Simonyan "Fast Sparse ConvNets". Paper on ArXiv, pre-trained sparse models.

Acknowledgements

XNNPACK is a based on QNNPACK library. Unlike QNNPACK, XNNPACK focuses entirely on floating-point operators, and its API is no longer compatible with QNNPACK.