基于集成学习的编译优化参数选择方法研究  

Compilation Optimization Parameters Selection Method Based on Ensemble Learning

作　　者：刘慧 姚金阳[1] 赵荣彩 LIU Hui;YAO Jinyang;ZHAO Rongcai(State Key Laboratory of Mathematical Engineering and Advanced Computing,Zhengzhou 450001,China;College of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,China)

机构地区：[1]数学工程与先进计算国家重点实验室,河南郑州450001 [2]河南师范大学计算机与信息工程学院,河南新乡453007

出　　处：《信息工程大学学报》2019年第3期319-327,共9页Journal of Information Engineering University

基　　金：国家重点研发计划资助项目(2016YFB0200503)

摘　　要：基于机器学习的迭代编译方法是当前迭代编译领域的研究热点,其挑战在于学习算法选择、优化参数搜索及程序特征表示等问题。基于集成学习技术提出一种优化参数组合自动预测方法ELOPS。使用约束多目标粒子群优化算法对优化空间进行搜索,找到样本函数最佳优化参数组合;通过特征—类相关性度量方法对函数特征进行抽取;最后通过由函数特征和最佳参数组合形成的样本构建集成学习模型,对新程序优化参数进行预测。基于两种平台,在NPB测试集和大型科学计算程序上与现有方法进行对比。实验结果表明,ELOPS方法比现有方法具有更好的预测性能。Machine learning based iterative compilation is a research hotspot in the field of iterative compilation,with its research challenge lying in learning algorithm selection,parameter search optimization and program feature representation.Based on the ensemble learning technique,we propose an automatic optimization parameter predictive method ELOPS.First,we search the optimal space by the constraining multi-objective PSO algorithm,and find the best optimization parameters for the samples.Second,we extract the features of the function through feature-class relevance measure.Finally,the learning model is constructed by the samples composed of program features and optimization parameters,and the new program optimization parameter combination can be predicted.We compare ELOPS with the existing methods when predicting for NPB benchmark and large scientific computing program.The results show that in most cases,the ELOPS method can obtain the best performance.

关 键 词：编译优化参数 集成学习 优化空间搜索 特征抽取 

分 类 号：TP314[自动化与计算机技术—计算机软件与理论]