众核体系架构并行计算模型与算法自适应调优框架研究

项目名称： 众核体系架构并行计算模型与算法自适应调优框架研究

项目编号： No.61272136

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 张云泉

作者单位： 中国科学院计算技术研究所

项目金额： 82万元

中文摘要： 众核架构已成为处理器体系结构发展新趋势，其多粒度并行性、复杂存储层次和有限存储带宽等都对并行算法的性能调优及性能可移植提出巨大挑战。FFT、稠密矩阵计算和OpenCV是三类不同的典型应用核心算法库，如何有效提高三类算法在众核上的运行和开发效率成为一个迫切需要解决的问题。本项目拟针对三类典型核心算法在众核上的性能可移植和高效实现开展研究,主要研究内容包括：1）新的众核并行计算模型研究。对众核架构的多粒度计算和多层次存储等关键特性进行建模，用以指导众核并行算法设计和分析；2）基于计算与访存模式分离指导语句的并行编程框架研究。利用并行计算模型和基于手工调优形成的优化方法链进行双向验证，提出一套基于计算和访存模式分离的制导语句的并行程序设计框架；3）算法自适应性能调优框架原型研究。将调优框架从过去的仅仅是平台自适应，扩展为平台自适应和算法自适应两个维度，进一步提升自适应优化的适用范围和灵活性。

中文关键词： 众核体系架构；并行计算模型；算法自适应；数学库基础算法；OpenCV

英文摘要： Many-core architecture has become the new development trend of processor architecture,Its multi-granularity parallelism,complicated memory hierarchy, and limited memory access bandwidth, etc., poses great challenges on the performance tuning and portability of parallel algorithm. FFT, dense linear algebra subroutine and OpenCV libraries are three very different and classic kernel function libraries. How to improve the running efficiency and development efficiency of these three kinds of libraries on many-core architecture becomes a more and more urgent problem. We plan to carry out researches on the performance portability and high productivity of these three kinds of classic algorithms on many-core architecture. Our major researches including 1) research on a new many-core parallel computational model characterizing with the complex computing hierarchy and memory hierarchy that can provide theoretical analysis results for parallel algorithm design. 2) research on a new parallel programming framework prototype based on splitted computation pattern and memory access pattern directives.Through exhausted validating on the optimization methods chain and our new proposed computational model, we will study on the most powerful methods and used these results to establish a suit of computation pattern directives and mem

英文关键词： manycore architectures；parallel computational model；algorithm self-adaptive；basic linear algebra algorithms；OpenCV