权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

FoMR: Microarchitecture mechanisms for handling conditional branches that are (a) very hard to predict accurately or (b) impossible to predict accurately

FoMR：用于处理 (a) 很难准确预测或 (b) 不可能准确预测的条件分支的微架构机制

基本信息

批准号：
2011145
负责人：
Yale Patt
金额：
$ 10.5万
依托单位：
University of Texas at Austin
依托单位国家：
美国
项目类别：
Standard Grant
财政年份：
2020
资助国家：
美国
起止时间：
2020-07-15 至 2023-06-30
项目状态：
已结题

来源：
https://www.nsf.gov/awardsearch/showAward?AWD_ID=2011145&HistoricalAwards=false
关键词：
FoMR Microarchitecture mechanisms handling conditional

项目摘要

Computers are playing a continually increasing role in supporting a better quality of life, including targeted health care, autonomous vehicles, and weather prediction. Their effectiveness in doing so, however, depends on how fast these computers can execute the programs that do more accurate and quicker decisions and predictions. A computer program that predicts a tsunami will hit tomorrow is of no value if the computer produces its result three days from now. This speed and accuracy are tightly coupled with a basic logical step: How fast can a computer process conditional branch instruction, such as an if-then-else. Conditional branch instructions are commands in a computer program that direct the computer to choose between executing alternate tasks. This basic function can end up being a bottleneck in modern systems that have to make millions of decisions as part of complex models. This research addresses that bottleneck, and can greatly improve the capabilities of modern computer processors. To improve performance of the microarchitecture, assembly lines (aka pipelines) were introduced long ago to process each instruction. Like all assembly lines, most instructions benefit greatly from this assembly line. However, not so with conditional branch instructions, since they require the computer to decide at the front of the assembly line what to do next (aka branch prediction). The problem is that a wrong guess means trashing everything on the assembly line, which degrades performance enormously. This research minimizes that from happening by recognizing that conditional branches are of three types: those predicted accurately, those not predicted well today, but can benefit significantly from some the first type, by using the well-known Tagged Geometric history length branch predictor (TAGE). For the second type, TAGE is augmented with the results of information learned from machine learning. For the third type, not predict at all, but perform other tasks while waiting for the necessary information to reach the end of the assembly line instead of guessing incorrectly and then trashing all the useless work already performed as a result of the wrong guess.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

计算机在提高生活质量方面发挥着越来越重要的作用，包括有针对性的医疗保健、自动驾驶汽车和天气预报。然而，它们这样做的有效性取决于这些计算机执行程序的速度有多快，这些程序可以做出更准确、更快的决策和预测。一个预测明天海啸的计算机程序，如果是在三天后得出的结果，那么它就没有任何价值。这种速度和准确性与基本的逻辑步骤密切相关：计算机处理条件分支指令（如if-then-else）的速度有多快。条件分支指令是计算机程序中指示计算机在执行备选任务之间进行选择的命令。这个基本功能最终可能成为现代系统的瓶颈，因为现代系统必须作为复杂模型的一部分做出数百万个决策。这项研究解决了这个瓶颈，并且可以大大提高现代计算机处理器的能力。为了提高微体系结构的性能，很久以前就引入了装配线（又名管道）来处理每条指令。像所有的装配线一样，大多数指令都从这条装配线中受益匪浅。然而，条件分支指令不是这样，因为它们要求计算机在装配线的前端决定下一步要做什么（又名分支预测）。问题是，错误的猜测意味着生产线上的所有东西都要报废，这将极大地降低性能。本研究通过认识到条件分支有三种类型来最大限度地减少这种情况的发生：那些预测准确的，那些今天预测不佳的，但可以从第一种类型中获益，通过使用众所周知的标记几何历史长度分支预测器（TAGE）。对于第二种类型，TAGE使用从机器学习中学习到的信息结果进行增强。对于第三种类型，根本不进行预测，而是执行其他任务，同时等待必要的信息到达装配线的末端，而不是猜测错误，然后由于猜测错误而丢弃已经执行的所有无用的工作。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。