Implementations of Context Based Value Predictors
Implementations of Context Based Value Predictors
复制标题
基于上下文的值预测器的实现
DOI：
--
复制
发表时间：
1997
期刊：
影响因子：
--
通讯作者：
James E. Smith
中科院分区：
文献类型：
--
作者：
Yiannakis Sazeides;James E. Smith
基于价值预测的数据依赖性的执行范例被证明具有巨大的性能潜力。基于上下文的预测指标遵循特定上下文t（s quench of Value）的值，并在上下文重复时预测一个值。较早的研究是使用未结合的表确定价值可预测性的限制。指令。还检查了基于上下文的预测，例如上下文订单，也检查了该提议的实现。通过将其预测性能削减到“理想的”表中，可以评估其预测性能，从而对“理想”的预测进行了评估通过使用指令捕获准确的价值历史记录的表。准确性。探索了对IN PUT数据大小的基于基于的预测，并发现相对较小。
Execution paradigms that eliminate data dependences based on value prediction h ave been shown to have significant performance potential. High accuracy value prediction is es sential for the success of such paradigms. Recently it was shown that context-basedprediction can predict values with high accuracy. A context-based predictor learns the values that follow a particular contex t (s quence of values) and predicts one of the values when the context repeats. However, the goal of t he earlier study was to determine the limits of value predictability using unbounded tables. In this paper we discuss implementations of context-based value predicto rs using two level table organizations. The important new elements introduced in this paper are fix ed size tables and sharing of prediction information among instructions. Some other dimensions of context-based prediction such as context order, aliasing and hash functions are also examined. A variety o f implementations of the proposed value predictor are evaluated for SPECINT95 benchmarks by com paring their prediction performance to the “ideal” performance with unbounded tables. Such predicto rs obtain accuracies that approach that of “ideal” predictors. Aliasing is found to be a significant cau se for misprediction. Aliasing is mitigated by using tables that capture accurate value history per instru ction. Prediction rates for SPECFP95 benchmarks are evaluated and found to often exceed 90% accuracy. The pro pos d predictor is also compared against stride and last value predictors. Stride and last v alue do better than contextbased prediction for small tables but context-based prediction is superio r with increasing table size. The sensitivity of the context-based prediction to the size of the in put data is explored and found to be relatively insignificant.