使用 SIMD 和多线程增强 TensorFlow.js WebAssembly 后端

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"本文最初发表在 TensorFlow 官博，经原作者授权，InfoQ 中文站翻译并分享。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"今年三月，我们为 TensorFlow.js "},{"type":"link","attrs":{"href":"https:\/\/blog.tensorflow.org\/2020\/03\/introducing-webassembly-backend-for-tensorflow-js.html","title":"","type":null},"content":[{"type":"text","text":"引入"}]},{"type":"text","text":"了一个新的 WebAssembly（Wasm）加速后端。今天，我们很高兴的宣布一个重大的性能更新：从 TensorFlow.js 2.3.0 版本开始，通过 "},{"type":"link","attrs":{"href":"https:\/\/github.com\/google\/XNNPACK","title":"","type":null},"content":[{"type":"text","text":"XNNPACK"}]},{"type":"text","text":"（一个高度优化的神经网络运算符库）利用"},{"type":"link","attrs":{"href":"https:\/\/github.com\/WebAssembly\/simd","title":"","type":null},"content":[{"type":"text","text":" SIMD（向量）指令"}]},{"type":"text","text":"和"},{"type":"link","attrs":{"href":"https:\/\/github.com\/WebAssembly\/threads","title":"","type":null},"content":[{"type":"text","text":"多线程"}]},{"type":"text","text":"，我们的 Wasm 后端速度提高了 10 倍。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"基准测试"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"SIMD 和多线程为我们的 Wasm 后端带来了重大的性能提升。下面是 Google Chrome 浏览器的基准测试，展示了 "},{"type":"link","attrs":{"href":"https:\/\/github.com\/tensorflow\/tfjs-models\/tree\/master\/blazeface","title":"","type":null},"content":[{"type":"text","text":"BlazeFace"}]},{"type":"text","text":" 的改进。BlazeFace 是一个具有 10 万个参数和大约 2000 万次乘加运算的轻量级模型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"（表中所列出的时间为每次推断的毫秒数）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":"

设备

普通 JS

WebGL

Wasm

Wasm+SIMD

Wasm+SIMD+线程

Pixel 4

368

28

28

15.9

N\/A*

第 6 代 ThnkPad X1，Linux

301.0

25.0

15.5

7.3

4.1

MacBook Pro 15 2019

209.1

22.7

13.3

7.9

4.0

在 MacBook Pro 15 2019 上运行 Node v.14

201.2

N\/A

25.5

15.2

N\/A**"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"更大的模型，如 MobileNet V2，这是一个拥有 350 万个参数和大约 3 亿个乘加运算的中性模型，可以获得更快的加速："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":"

设备	普通 JS	WebGL	Wasm	Wasm+SIMD	Wasm+SIMD+线程
Pixel 4	1628	76.7	182	82	N\/A*
第 6 代 ThinkPad X1，Linux	1489.4	44.8	122.7	34.6	12.4
MacBook Pro 15 2019	893.5	19.6	98.4	30.2	10.3
在 MacBook Pro 15 2019shang yunxing Node v.14	1404.3	N\/A	290.0	64.2	N\/A**"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"注："},{"type":"link","attrs":{"href":"https:\/\/www.chromestatus.com\/feature\/5724132452859904","title":"","type":null},"content":[{"type":"text","text":"由于移动浏览器中的多线程支持仍在进行中"}]},{"type":"text","text":"，故 Pixel 4 不支持 TF.js 多线程 Wasm 后端的基准测试。iOS 版本的 SIMD 支持也仍在开发中。TF.js 多线程 Wasm 后端的节点支持即将推出。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}