The new version of Tencent Youtu ncnn is suitable for domestic CPUs, and the maximum speed is increased by 70 times

With the continuous promotion of independent information technology innovation and Application , domestic PC、 domestic OS And software and hardware equipment are becoming more and more mature . In order to better assist domestic enterprises CPU stay AI Software ecology starts from “ You can use ” To “ To use ”, As the first high-performance neural network forward computing open source framework launched by Tencent Youtu Laboratory ,ncnn Recently in China CPU Godson and D1 A more comprehensive adaptation and performance optimization are carried out ,ncnn Join hands with Godson and Quanzhi technology , We got through together AI Application and domestic production CPU Barriers between hardware .

Godson

It is a general-purpose computer independently developed by the Institute of computing, Chinese Academy of Sciences CPU, Adopt autonomy LoongISA Command system , compatible MIPS Instructions

D1

It is the first Quanzhi technology based on RISC-V Instruction set chips , Integrated with ALI Flathead 64 position C906 The core , Support RVV

This time ncnn The updated 20210720 edition , It's done risc-v And mips framework cpu The adaptation of , And make use of risc-v vector And mips msa Vector acceleration extensions , The performance of most common operators is optimized . stay ncnn Incidental benchmark In the test ,ncnn In dragon core CPU The upper speed increases the most 18.64 times , stay D1 The upper speed increases the most 70 times , To satisfy the AI Basic requirements for end-to-end reasoning deployment .

ncnn In dragon core CPU Test data on , The maximum speed is increased 18.64 times

ncnn In Quanzhi Technology D1 Test data on , The maximum speed is increased 70 times

Godson 2k Send it to the development board for use ncnn Deploy yolov5 Check the effect of the algorithm

Full ambition D1 Use on development board ncnn Deploy nanodet Check the effect of the algorithm

ncnn 20210720 Other updates to the version

• Support x86 avx-only cpu Optimization acceleration
• Mathematical functions log/exp/tanh arm Optimize
• promote ncnn Quantify the multithreading efficiency of the tool
• Repair some phones gpu Inferential memory leaks and other bugfix wait
• Support Godson autonomic instruction set architecture loongarch

  Test platform -1

Godson 2K1000,2 Threads ,mips framework , Turn on msa

ncnn In dragon core CPU Upper adaptation test data

  Test platform -2

Full ambition D1,1 Threads ,risc-v framework , Turn on v Expand

ncnn In Quanzhi Technology D1 Adaptation test data

Last , Welcome to ncnn Project home page , read Readme Join in ncnn Technical communication QQ Group , Communicate with front-line engineers and many technical leaders .

See below for details ：

ncnn 20210720 edition Download the address or click to read the original

(linux/windows/macos/android/ios/webassembly,cpu+gpu)

https://github.com/Tencent/ncnn/releases/tag/20210720

ncnn Open source project Access address

https://github.com/Tencent/ncnn