基于路预测访问的低功耗高速缓存设计研究

发布时间：2018-03-21 09:07

本文选题：高速缓存　切入点：低功耗　出处：《浙江大学》2012年硕士论文　论文类型：学位论文

【摘要】：随着智能手机、平板电脑等移动设备的盛行,低功耗已成为处理器必不可少的属性,高速缓存作为处理器的主要功耗来源之一,其低功耗设计的需要日益突出。本文提出并研究了三种路预测访问方案,以较少的附加资源,减少高速缓存访问的功耗,同时降低带来的性能损失。本文的主要研究内容和创新点包括： 1、基于指令的高速缓存分类路预测。在充分研究各种指令和路预测信息源关系的基础上,将指令根据所需预测源的不同进行划分,并结合处理器不同流水线阶段的指令信息,从多个预测源中选取所需的预测路信息。之后研究程序中各种指令的比例,以较少的附加预测资源和性能损失节省较大的功耗,同时通过实验分析各种预测源中资源与预测效率、功耗、性能之间的关系,以得到最佳的预测资源配置。该方案能够很好的降低缓存功耗,但附加资源稍大,性能损失也较多。 2、基于循环的高速缓存单路径路预测。首先根据预测基于历史的性质,分析了分类路预测的缺点,提出了基于循环的单路径路预测。该方案以循环作为路预测开启的先决条件,只在遇到循环且该循环已被记录的情况下,方才开启路预测。之后通过实验分析了单路径路预测系统中资源分配与预测效率的关系,并证实了该预测方案亦能很好的降低缓存功耗,且附加资源和性能损失皆大幅下降。 3、基于循环的高速缓存多路径路预测。通过在循环中使用多个路径的方法,解决了循环中因分支指令存在而造成的多个路径之间相互冲突的现象,并分析程序中分支指令与循环的数量,提出最佳的路径个数,并通过实验验证在资源足够的情况下,相同资源配置的多路径路预测优于单路径路预测,尤其是性能损失得到了更大的降低。
[Abstract]:With the popularity of mobile devices such as smart phones and tablets, low power consumption has become an essential attribute of processors. Cache is one of the main power sources of processors. In order to reduce the power consumption of cache access, we propose and study three kinds of path predictive access schemes to reduce the power consumption of cache access. At the same time, reduce the performance loss. The main contents and innovations of this paper include:. 1. Cache classified path prediction based on instruction. On the basis of fully studying the relationship between various kinds of instructions and path prediction information sources, the instructions are divided according to the different prediction sources needed, and the instruction information of different stages of pipeline processor is combined. The required prediction path information is selected from multiple prediction sources. After that, the proportion of instructions in the program is studied to save large power consumption with less additional prediction resources and performance loss. At the same time, the relationship between resource and prediction efficiency, power consumption and performance is analyzed through experiments to obtain the best predictive resource allocation. This scheme can reduce cache power consumption well, but the additional resources are slightly larger, and the performance loss is more. (2) Cache single path prediction based on cycle. Firstly, according to the properties of prediction based on history, the shortcomings of classifying path prediction are analyzed, and a single path prediction based on cycle is proposed. Only when the cycle is encountered and the cycle has been recorded, the path prediction is opened. Then, the relationship between resource allocation and prediction efficiency in a single path prediction system is analyzed through experiments. It is proved that the proposed scheme can also reduce cache power consumption, and the loss of additional resources and performance are greatly reduced. 3. Cache multipath prediction based on loop. By using multiple paths in the loop, the conflict between multiple paths caused by branch instructions in the loop is solved. The number of branch instructions and loops in the program is analyzed, and the optimal number of paths is put forward. The experimental results show that the multi-path prediction of the same resource allocation is better than that of single-path prediction in the case of sufficient resources. In particular, the performance loss has been further reduced.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP333

【参考文献】