面向众核计算平台的性能自适应优化方法与关键技术研究

As the diversity of the architecture of many-core processors, autotuning is emerging as a critical strategy for achieving portable performance across a broad range of architectures. However, optimization space explosion, costly tuning overhead and difficult to achieve high performance are main obstacle to its popularity. In order to address these challenges, we put forward the concept of adaptive performance tuning mechanism. And get down to do research on the method of adaptive tuning which combines hardware characteristics with application features for computer vision applications. The main research contents are follows: (1) Abstraction and definition of parallel optimization pattern related with the domain knowledge and hand optimization experience. (2) Construction of performance optimization guiding model based on parallel optimization pattern. (3) Contruction of auto-tuning framework guided by performance optimization guiding model. Finally, we will propose a new methods and technological system for taking full advantage of the computing capability of many-core processors. So as to that we can simplify parallel programming and achieve the performance portability for computer vision algorithms on many-core processors with high performance. The research results of this project will also play an important role on promoting the development of many-core processors.

随着众核计算平台架构的日益多样性，性能自适应优化技术成为实现性能可移植的重要方法。然而，面向众核计算平台的性能自适应优化机制面临着优化空间爆炸、自适应调优开销过高、自适应优化难以实现最佳性能等问题。为此，本课题将针对计算机视觉应用领域，围绕性能自适应优化的相关要素，在融合领域特征和手工优化经验的并行优化模式的抽象和定义、基于并行优化模式的性能优化指导模型、及其制导的性能自适应优化框架三个方面获得突破，最终形成能够充分发挥众核处理器计算能力的性能自适应优化关键技术方法体系，实现计算机视觉应用在不同众核计算平台上的高性能及性能移植，在推动我国众核处理器的应用和发展中发挥重要作用。

随着众核计算平台架构的日益多样性，性能自适应优化技术成为实现性能可移植的重要方法。然而，面向众核计算平台的性能自适应优化机制面临着优化空间爆炸、自适应调优开销过高、自适应优化难以实现最佳性能等问题，本项目与图像处理算法、FFT算法为研究对象，主要研究内容包括：1）基于模板的代码自动生成方法。首次提出了一种基于模板的FFT高性能汇编代码自动生成方法，能够根据体系架构特征自动生成针对X86和ARM计算平台的高性能FFT代码；2）基于搜索的性能自适应优化方法。针对FFT算法的分解特征和硬件体系结构特征，通过宽度优先搜索和深度优先搜索结合的方法，确定FFT的最佳分解方式。3）非规则算法在GPU计算平台上的自适应优化框架。针对非规则算法在GPU计算平台上由于线程间负载不均衡导致的性能瓶颈，打破传统GPU编程和优化模式，通过Persistent thread、多粒度并行、任务队列等优化方法的使用，大幅提高了非规则算法在GPU计算平台上的性能。通过以上内容的研究，本项目基本形成了面向众核计算平台的性能优化方法体系。