|
Name |
|
Date |
Size |
#Lines |
LOC |
| .. | | - | - |
| bench/ | | 03-May-2024 | - | 40,393 | 34,727 |
| cmake/ | | 03-May-2024 | - | 169 | 141 |
| eval/ | | 03-May-2024 | - | 19,597 | 17,904 |
| include/ | | 03-May-2024 | - | 2,978 | 2,000 |
| models/ | | 03-May-2024 | - | 42,522 | 39,825 |
| scripts/ | | 03-May-2024 | - | 8,834 | 6,498 |
| src/ | | 03-May-2024 | - | 1,720,689 | 1,393,191 |
| test/ | | 03-May-2024 | - | 1,532,610 | 1,418,410 |
| third_party/ | | 03-May-2024 | - | 1,008 | 925 |
| tools/ | | 03-May-2024 | - | 11,260 | 9,801 |
| .bazelrc | D | 03-May-2024 | 1.4 KiB | 56 | 42 |
| .clang-format | D | 03-May-2024 | 521 | 18 | 17 |
| .gitignore | D | 03-May-2024 | 491 | 36 | 33 |
| Android.bp | D | 03-May-2024 | 432.9 KiB | 9,611 | 9,441 |
| BUILD.bazel | D | 03-May-2024 | 522.2 KiB | 13,207 | 12,652 |
| CMakeLists.txt | D | 03-May-2024 | 430.8 KiB | 9,079 | 8,692 |
| CONTRIBUTING.md | D | 03-May-2024 | 1.1 KiB | 29 | 20 |
| LICENSE | D | 03-May-2024 | 1.5 KiB | 32 | 24 |
| METADATA | D | 03-May-2024 | 737 | 20 | 19 |
| MODULE_LICENSE_BSD | D | 03-May-2024 | 0 | | |
| OWNERS | D | 03-May-2024 | 219 | 12 | 11 |
| README.md | D | 03-May-2024 | 6.9 KiB | 121 | 95 |
| TEST_MAPPING | D | 03-May-2024 | 3.8 KiB | 211 | 210 |
| WORKSPACE | D | 03-May-2024 | 3.2 KiB | 92 | 78 |
| build_defs.bzl | D | 03-May-2024 | 18.1 KiB | 471 | 440 |
| emscripten.bzl | D | 03-May-2024 | 1.2 KiB | 41 | 36 |
| preamble.js.lds | D | 03-May-2024 | 393 | 10 | 8 |
README.md
1# XNNPACK
2
3XNNPACK is a highly optimized library of floating-point neural network inference operators for ARM, WebAssembly, and x86 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 [TensorFlow Lite](https://www.tensorflow.org/lite), [TensorFlow.js](https://www.tensorflow.org/js), [PyTorch](https://pytorch.org/), and [MediaPipe](https://mediapipe.dev).
4
5## Supported Architectures
6
7- ARM64 on Android, Linux, macOS, and iOS (including WatchOS and tvOS)
8- ARMv7 (with NEON) on Android
9- ARMv6 (with VFPv2) on Linux
10- x86 and x86-64 (up to AVX512) on Windows, Linux, macOS, Android, and iOS simulator
11- WebAssembly MVP
12- WebAssembly SIMD
13- RISC-V (RV32GV and RV64GC)
14
15## Operator Coverage
16
17XNNPACK implements the following neural network operators:
18
19- 2D Convolution (including grouped and depthwise)
20- 2D Deconvolution (AKA Transposed Convolution)
21- 2D Average Pooling
22- 2D Max Pooling
23- 2D ArgMax Pooling (Max Pooling + indices)
24- 2D Unpooling
25- 2D Bilinear Resize
26- 2D Depth-to-Space (AKA Pixel Shuffle)
27- Add (including broadcasting, two inputs only)
28- Subtract (including broadcasting)
29- Divide (including broadcasting)
30- Maximum (including broadcasting)
31- Minimum (including broadcasting)
32- Multiply (including broadcasting)
33- Squared Difference (including broadcasting)
34- Global Average Pooling
35- Channel Shuffle
36- Fully Connected
37- Abs (absolute value)
38- Bankers' Rounding (rounding to nearest, ties to even)
39- Ceiling (rounding to integer above)
40- Clamp (includes ReLU and ReLU6)
41- Convert (includes fixed-point and half-precision quantization and
42 dequantization)
43- Copy
44- ELU
45- Floor (rounding to integer below)
46- HardSwish
47- Leaky ReLU
48- Negate
49- Sigmoid
50- Softmax
51- Square
52- Truncation (rounding to integer towards zero)
53- PReLU
54
55All operators in XNNPACK support NHWC layout, but additionally allow custom stride along the **C**hannel 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.
56
57## Performance
58
59### Mobile phones
60
61The table below presents **single-threaded** performance of XNNPACK library on three generations of MobileNet models and three generations of Pixel phones.
62
63| Model | Pixel, ms | Pixel 2, ms | Pixel 3a, ms |
64| ----------------------- | :-------: | :---------: | :----------: |
65| FP32 MobileNet v1 1.0X | 82 | 86 | 88 |
66| FP32 MobileNet v2 1.0X | 49 | 53 | 55 |
67| FP32 MobileNet v3 Large | 39 | 42 | 44 |
68| FP32 MobileNet v3 Small | 12 | 14 | 14 |
69
70The following table presents **multi-threaded** (using as many threads as there are big cores) performance of XNNPACK library on three generations of MobileNet models and three generations of Pixel phones.
71
72| Model | Pixel, ms | Pixel 2, ms | Pixel 3a, ms |
73| ----------------------- | :-------: | :---------: | :----------: |
74| FP32 MobileNet v1 1.0X | 43 | 27 | 46 |
75| FP32 MobileNet v2 1.0X | 26 | 18 | 28 |
76| FP32 MobileNet v3 Large | 22 | 16 | 24 |
77| FP32 MobileNet v3 Small | 7 | 6 | 8 |
78
79Benchmarked on March 27, 2020 with `end2end_bench --benchmark_min_time=5` on an Android/ARM64 build with Android NDK r21 (`bazel build -c opt --config android_arm64 :end2end_bench`) and neural network models with randomized weights and inputs.
80
81### Raspberry Pi
82
83The table below presents **multi-threaded** performance of XNNPACK library on three generations of MobileNet models and three generations of Raspberry Pi boards.
84
85| Model | RPi Zero W (BCM2835), ms | RPi 2 (BCM2836), ms | RPi 3+ (BCM2837B0), ms | RPi 4 (BCM2711), ms | RPi 4 (BCM2711, ARM64), ms |
86| ----------------------- | :----------------------: | :-----------------: | :--------------------: | :-----------------: | :------------------------: |
87| FP32 MobileNet v1 1.0X | 3937 | 299 | 114 | 72 | 76 |
88| FP32 MobileNet v2 1.0X | 1987 | 187 | 79 | 41 | 44 |
89| FP32 MobileNet v3 Large | 1658 | 158 | 67 | 38 | 41 |
90| FP32 MobileNet v3 Small | 487 | 50 | 23 | 13 | 14 |
91| INT8 MobileNet v1 1.0X | 2598 | 169 | 61 | 29 | 24 |
92| INT8 MobileNet v2 1.0X | 1487 | 109 | 40 | 20 | 17 |
93
94Benchmarked on Oct 15, 2021 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. INT8 inference was evaluated on per-channel quantization schema.
95
96## Publications
97
98- Marat Dukhan "The Indirect Convolution Algorithm". Presented on [Efficient Deep Learning for Compute Vision (ECV) 2019](https://sites.google.com/corp/view/ecv2019/) workshop ([slides](https://drive.google.com/file/d/1ZayB3By5ZxxQIRtN7UDq_JvPg1IYd3Ac/view), [paper on ArXiv](https://arxiv.org/abs/1907.02129)).
99- Erich Elsen, Marat Dukhan, Trevor Gale, Karen Simonyan "Fast Sparse ConvNets".
100 [Paper on ArXiv](https://arxiv.org/abs/1911.09723), [pre-trained sparse
101 models](https://github.com/google-research/google-research/tree/master/fastconvnets).
102- Marat Dukhan, Artsiom Ablavatski "The Two-Pass Softmax Algorithm".
103 [Paper on ArXiv](https://arxiv.org/abs/2001.04438).
104- Yury Pisarchyk, Juhyun Lee "Efficient Memory Management for Deep Neural Net Inference".
105 [Paper on ArXiv](https://arxiv.org/abs/2001.03288).
106
107## Ecosystem
108
109### Machine Learning Frameworks
110
111- [TensorFlow Lite](https://blog.tensorflow.org/2020/07/accelerating-tensorflow-lite-xnnpack-integration.html).
112- [TensorFlow.js WebAssembly backend](https://blog.tensorflow.org/2020/03/introducing-webassembly-backend-for-tensorflow-js.html).
113- [PyTorch Mobile](https://pytorch.org/mobile).
114- [MediaPipe for the Web](https://developers.googleblog.com/2020/01/mediapipe-on-web.html).
115- [Alibaba HALO (Heterogeneity-Aware Lowering and Optimization)](https://github.com/alibaba/heterogeneity-aware-lowering-and-optimization)
116- [Samsung ONE (On-device Neural Engine)](https://github.com/Samsung/ONE)
117
118## Acknowledgements
119
120XNNPACK is a based on [QNNPACK](https://github.com/pytorch/QNNPACK) library. Over time its codebase diverged a lot, and XNNPACK API is no longer compatible with QNNPACK.
121