面向异构众核系统的统一编程框架研究

The heterogeneous many-core system has emerged as a promising development trend in the domain of high performance computing. How to reduce the programming difficulty, improve the program’s portability, and fully exploit the computing power is a major challenge in the heterogeneous many-core system. In order to tackle above challenge, the project studies on the new programming model theory and relative technology for heterogeneous many-core systems. ..First, abstract a general logical architecture of heterogeneous many-core system, which centers on the coprocessor and can adapt with the development trend of heterogeneous many-core system. Second, research a multi-task stream and multi-data stream programing model and corresponding parallel programming language which isolates with the details of the underlying hardware architecture and support software. Third, research a compiler which transforms the high level parallel code to the intermediate code. Finally, research a mapping scheme which ensures the efficient mapping and running of the intermediate code on the underlying heterogeneous computing components. The overall goal is to construct a unified heterogeneous programming framework which can provide an efficient parallel programing environment for programmers; adapt to the underlying hardware architecture changes and support the efficient mapping and running scheme in order to improve the programmability, portability, scalability, as well as performance of heterogeneous many-core system. ..The research of the project will have important influence on the basic theory of high performance computer system architecture, promote the widely use of the heterogeneous computing, and help creating the advanced computing platform that is listed in the Medium and Long Term Science and Technology Development Plan of China.

针对异构众核系统面临的编程难、程序移植难、系统资源得不到充分利用的挑战，本项目围绕异构众核系统编程模型展开研究，探索降低异构众核系统编程难度、充分发挥其计算能力、提高其通用性的编程模型相关理论及关键技术。项目以协处理器为中心抽象多样化发展态势的异构众核系统逻辑架构；研究面向计算单元的多任务流多数据流编程模型及并行描述语言；研究高层并行代码到跨平台中间代码的编译转换；研究中间代码到具体异构众核系统底层硬件映射机制。从而构建一个向上为用户提供屏蔽底层实现细节，简单高效跨平台的异构众核系统并行编程环境，向下能适应异构系统底层硬件架构变化，提供高效映射运行机制的统一编程框架，提高异构众核系统可编程性、软件可移植性、充分发挥异构众核系统高效计算潜能。本课题的研究将对高性能计算机系统结构的基础理论，对促进异构计算的广泛应用，以及《国家中长期科学和技术发展规划》提出的先进计算平台建设具有重要意义。

本项目针对异构众核系统面临的编程难、程序移植难、系统资源得不到充分利用的挑战，围绕异构众核系统编程模型展开研究，探索降低异构众核系统编程难度、充分发挥其计算能力、提高其通用性的编程模型相关理论及关键技术。本项目主要研究内容有：（1）以协处理器为中心的异构众核系统逻辑架构抽象；（2）面向计算核心的新型多任务流多数据流编程模型及相应的并行描述语言；（3）将高层并行代码转化成通用中间代码的编译转化技术；（4）高层多任务流多数据流向底层硬件计算单元的高效映射技术。. 针对以上四个研究内容，项目组通过对异构系统和异构众核处理器架构特征的研究，提出了以协处理器为中心面向大规模计算单元的统一架构抽象UAHM；在统一架构抽象的基础上，通过对上层应用特征的抽象研究设计了运行时无关的数据关联计算编程模型NoT，并设计实现了C语言兼容的数据关联计算描述语言Ruler，提供了用户友好、易学易用、运行时无关的高层统一并行编程接口；通过研究以OpenCL为中间语言的源到源编译技术，设计实现了高层语言Ruler到OpenCL的源到源编译器NoT-to-Ocl；通过整合线程映射机制、基于时间戳的数据管理机制等运行时映射执行优化技术，形成了面向异构众核系统的统一并行编程架构UPPA，实现高层编程向底层硬件架构的高效映射。通过在集成有GPU、MIC等异构处理器的不同异构平台上的对比测试，在多个测试用例下执行性能与人工编写且经过优化的测试用例代码相当，验证了统一编程框架原型系统的有效性和跨平台可移植性。