| 本文已被：浏览 1498次 下载 3403次 |
码上扫一扫！ |
|
|
| 面向申威异构架构的并行代码自动生成 |
|
陶小涵1,2, 朱雨1,2, 庞建民1,2, 赵捷1,2, 徐金龙1,2
|
|
1.信息工程大学, 河南 郑州 450001;2.数学工程与先进计算国家重点实验室, 河南 郑州 450001
|
|
| 摘要: |
| 异构架构逐渐成为高性能计算领域的主流架构, 但相较于同构多核架构, 其硬件结构及存储层次更为复杂, 程序编写更为困难. 先进的优化编译器可以协助程序开发人员实现更为高效的代码, 降低程序开发复杂度. 多面体编译模型通过抽象分析将程序抽象成空间多面体表示形式, 能够将多种循环变换与硬件映射相结合, 并面向特定体系结构生成相应的代码. 设计实现了一个面向国产申威异构架构的并行代码自动生成系统, 采用“源-源”编译模式, 基于多面体编译模型实现. 系统针对申威异构架构特点将程序计算过程进行硬件部署, 同时实现数据传输与内存空间的自动管理. 实验基于Polybench测试集中线性代数相关用例进行测试. 结果表明, 利用代码自动生成系统生成的异构并行代码能够在申威异构平台上正确运行, 并能够有效发挥申威异构平台的性能, 基于申威异构平台利用64线程加速计算的平均加速比达到了539.16倍. |
| 关键词: 申威异构架构 多面体模型 并行计算 代码生成 |
| DOI：10.13328/j.cnki.jos.006688 |
| 分类号:TP311 |
| 基金项目:国家自然科学基金(61702546) |
|
| Parallel Code Generation for Sunway Heterogeneous Architecture |
|
TAO Xiao-Han1,2, ZHU Yu1,2, PANG Jian-Min1,2, ZHAO-Jie1,2, XU Jin-Long1,2
|
|
1.Information Engineering University, Zhengzhou 450001, China;2.State Key Laboratory of Mathematical Engineering and Advanced Computing, Zhengzhou 450001, China
|
| Abstract: |
| Heterogeneous architectures are dominating the realm of high-performance computing. However, these architectures also complicate the programming issue due to its increasingly complex hardware and memory hierarchy compared to homogeneous architectures. One of the most promising solutions to this issue is making use of optimizing compilers which can help programmers develop high-performance code executable on target machines, thereby simplifying the difficulty of programming. The polyhedral model is widely studied due to its ability to generate effective code and portability to various targets, which is realized by first converting a program into its intermediate representation and then combining the compositions of loop transformations and hardware binding strategies. This paper presents a source-to-source parallel code generator targeting the domestic, heterogeneous architecture of the Sunway machine using the polyhedral model. In particular, the computation is deployed automatedly onto the Sunway architecture and memory management, minimizing the amount of data movements between the management processing element and computing processing elements of the target. The experiments are conducted on 13 linear algebra applications extracted from the Polybench Benchmarks. The experimental results show that the proposed approach can generate effective code executable on the Sunway heterogeneous architecture, providing a mean speedup of 539.16× on 64 threads over the sequential implementation executed on a management processing element. |
| Key words: Sunway heterogeneous architecture polyhedral model parallel computing code generation |
