处理器值预测技术研究  

Research on Processor Value Prediction

作　　者：黄立波[1,2] 杨凌 杨乾明 马胜[1] 王永文 隋兵才[1,2] 沈立[1] 徐炜遐 HUANG Li-bo;YANG Ling;YANG Qian-ming;MA Sheng;WANG Yong-wen;SUI Bing-cai;SHEN Li;XU Wei-xia(College of Computer Science and Technology,National University of Defense Technology,Changsha,Hunan 410073,China;Key Laboratory of Advanced Microprocessor Chips and Systems,Changsha,Hunan 410073,China)

机构地区：[1]国防科技大学计算机学院,湖南长沙410073 [2]先进微处理器芯片与系统重点实验室,湖南长沙410073

出　　处：《电子学报》2023年第12期3591-3618,共28页Acta Electronica Sinica

基　　金：国家重点研发计划(No.2021YFB0300300);国家自然科学基金(No.62272475,No.62172430);湖南省自然科学基金(No.2022JJ10064,No.2021JJ10052);湖南省科技创新计划项目(No.2022RC3065)。

摘　　要：当今的处理器性能与存储器带宽和延迟严重失衡的问题限制了计算系统的整体性能,而存储器的性能对制程工艺不敏感,在后摩尔时代下很难再通过集成电路制造工艺的迭代获得处理器性能收益,因此人们更多地想通过体系结构的创新获得更高性能的计算系统.处理器值预测技术是一种能在无需改变存储系统情况下有效缓解存储墙问题的解决方案,其通过预测性地打破数据真相关进而让更多的指令可以在乱序处理器中并行执行,而无需等待由于访存等操作造成的长周期指令执行.近年来,值预测在各个方面都有了实质性的进步,但现如今还没有商用处理器使用这一技术,这主要是由于值预测技术的使用还面临许多挑战:现有的处理器的流水线架构不能直接使用值预测技术;值预测所需的预测值传递机制需要额外的硬件资源开销;值预测器巨大的存储开销让其很难在片上实现;由于值预测错误时的性能惩罚大,因此预测准确率较低的值预测器会降低处理器性能.针对这些问题,本文以值预测技术为中心,围绕值预测技术相关的流水线架构、值预测器结构和错误恢复机制三个方面分别详细论述了国内外研究成果以及其对于各个问题挑战的解决策略.最后,本文对当今的处理器值预测技术进行了总结并对未来的研究方向进行了展望.The extreme imbalance between processor performance and memory bandwidth/latency limits the overall performance of computing systems.In the post-Moore era,it is challenging to obtain processor performance benefits through the iteration of the integrated circuit manufacturing process,and memory performance is not sensitive to the process.Therefore,people tend to obtain higher-performance computing systems through architectural innovation.Processor value prediction technology is a solution that can effectively alleviate the memory wall problem without changing the storage system.By speculatively breaking the true dependency of data,more instructions can be executed in parallel in an outof-order processor.There is no need to wait for the execution of long-cycle instructions caused by memory access,etc.In recent years,value prediction has made significant progress in various aspects.However,no commercial processors are using this technology,mainly because the development of value prediction technology still faces many challenges:the pipeline architecture of existing processors cannot directly use value prediction techniques;the register file read and write ports required for value prediction are physically challenging to implement;the huge storage overhead of the value predictor makes it difficult to implement on-chip;due to the significant performance loss when the value prediction is wrong,the value predictors with low prediction accuracy will reduce processor performance.In response to these problems,this paper focuses on value prediction technology.It discusses in detail the research at home and abroad and its solutions to problems and challenges around the value prediction pipeline architecture,value predictor structure,and misprediction recovery mechanism related to value prediction technology.Finally,this paper summarizes processor value prediction techniques and provides an outlook on future research directions.

关 键 词：值预测 数据依赖 处理器 流水线 存储墙 超标量 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]