高性能、高可扩展和高可移植的Stencil代码生成和优化框架研究

stencil计算是科学应用中最重要的并行计算模式之一，其性能优化也是高性能计算领域的研究热点和难点之一。现有的stencil代码生成和优化技术对stencil计算场景支持有限，算法优化与特定体系结构紧密耦合，且在代码生成中缺少结合计算架构特性进行反馈优化。因此，生成的stencil 代码在高性能、高可扩展和高可移植等方面综合表现较差。本课题拟从问题域、计算抽象域和代码性能域三个层面，重点研究支持复杂stencil计算的面向领域编程语言、stencil计算和并行计算架构的抽象表示、stencil抽象计算架构到目标计算架构的高效映射方法、运行时评测指导的编译优化方法等关键技术，有望实现一套高性能、高可扩展和高可移植的stencil代码生成和优化框架。本课题可以为理解复杂stencil在不同并行计算架构下性能优化提供工程实践和理论探索，同时也有望为国家E级超级计算机研发进行软件技术储备。

Stencil computation is one of the most important parallel computing patterns in scientific applications, and its performance optimization is also one of the research hotspots and difficulties in the field of high performance computing. Existing stencil code generation and optimization technologies have the following drawbacks, such as limited support for complex stencil computation, optimizations tightly coupled with specific architectures, and lack of feedback optimization during code generation regarding the architecture features. Therefore, the generated stencil code performs poorly in terms of high performance, high scalability, and high portability. This project addresses the above drawbacks from three domains, such as problem domain, computation abstraction domain and code performance domain. Specifically, this project proposes the following research goals: 1) domain specific language for complex stencil computation, 2) abstract representation of stencil computation and parallel computing architectures, 3) the efficient mapping of the abstract representation of stencil computation to the target computing architecture, and 4) runtime profile-guided compilation optimization. Eventually, the project is expected to deliver a high-performance, highly scalable, and highly portable stencil code generation and optimization framework. This project can provide engineering practice and theoretical exploration for better understanding the performance optimization of complex stencil computation under different parallel computing architectures. It is also expected to reserve software technologies for national Exascale supercomputer research and development.