面向AR感知交互的高能效移动设备体系结构研究

One of the key components of AR (Augmented Reality) technology is perception and interaction, and they are usually executed by mobile devices. In contrast with the growing software functions of perception and interaction applications, the low-level hardware performance of mobile devices is stagnating, meanwhile, the power consumption of mobile devices is strictly limited. Thus, how to further improve the energy efficiency of mobile devices for perception and interaction applications becomes the bottleneck for AR. .One of the solutions is to optimize the architecture design based on software characteristics. AR perception and interaction applications have distinct characteristics, it brings up new requirements for user experience evaluation, and raises new challenges for architecture design. Currently, there is no such optimized mobile architecture design for AR perception and interaction applications. This study targets on this problem. We follow the principles of “analyzing design goal – designing benchmark suite – analyzing microarchitectural program execution characteristics – proposing and testing optimization methods”, and improve the energy efficiency in two levels, i.e., architecture design and system management levels. We propose a quality of experience model, perform design space exploration to reveal the most energy efficient design, explore synchronization management and hardware resource allocation policies to improve quality of experience, investigate task scheduling and system management schemes to optimize the usage of batteries, propose a customized energy efficiency management policy to further improve the user experience for individual users. This study will lay important hardware energy efficiency foundation for future AR development.

AR（增强现实）的关键技术之一是感知交互，它通常由移动设备来执行。感知交互的软件功能日益丰富，底层硬件却遭遇性能增长瓶颈，同时移动设备的功耗又受到严格限制。因此，如何针对感知交互进一步提升移动设备的能效成为制约AR技术发展的瓶颈。.解决办法之一是根据软件特点对体系结构进行优化。AR感知交互程序具有突出特征，它不仅对用户体验评价提出新的要求，也对体系结构设计提出新的挑战。但是，国内外目前还没有专门针对AR感知交互的移动设备体系结构设计。项目旨在填补这一空白，遵循“设计目标分析-基准程序开发-程序执行特点分析-优化方案提出与验证”的技术路线，从架构设计和系统管理两个层次进行能效优化。项目建立体验质量模型，通过设计空间探索来探究高能效架构，研究同步性管理与资源分配策略以提升体验质量，研究任务调度及系统管理策略以优化电池使用，研究定制化能效管理策略。项目研究为AR进一步发展奠定重要的硬件能效基础。

AR感知交互是人机交互中一类重要应用，为了达到较好的用户体验，它们需要很高的实时性，这无疑对系统性能提出了很高的要求，但是它们通常由移动设备执行，功耗受到严格限制，因此提高移动设备能效对于提升AR感知交互的用户体验具有重要意义。.项目遵循“基准程序开发-程序执行特点分析-优化方案提出并验证”的技术路线，开发了AR感知交互基准程序集，进行了设计空间探索，分析了程序对包括CPU、GPU以及深度学习加速器的异构架构的硬件资源需求。项目发现此类程序对并行度的要求较高，而对Cache的需求不高，因此设计了电源门控、硬件查找表等管理策略。项目发现内存访问是重要的性能瓶颈，设计了基于部件特点的内存调度策略，并提出了部件之间的同步性管理策略。此外，项目还提出了面向异构深度学习加速器群的任务调度方法，设计了面向稠密SLAM的近似计算及其控制方法，研究了基于太阳能的边缘计算架构，在提升性能的同时大幅降低了边缘计算对传统电池能力的消耗。