大数据驱动人工智能的兴起对数据密集型应用的加速提出了迫切的需求。软件定义硬件（SDH，Software Defined Hardware）因为兼具高能效和高灵活性的特点，成为加速数据密集型应用的最具潜力的技术之一。由于数据密集型应用的数据密集性、结构多样性和SDH硬件资源的复杂性、限制性，SDH编译映射算法呈现出编译时间长、访存冲突多、计算并行度低和扩展性差等问题。本项目针对以上共性和关键问题，在自己所取得的可重构计算芯片多参量协同映射技术等成果的坚实研究基础上，综合运用可重构计算、存储管理、深度学习、强化学习等多学科创新设计理论和方法，开展SDH架构多参量高精度性能模型建立、无访存冲突的内存划分优化、基于深度强化学习的布局布线优化等研究，最终提出SDH上数据密集型应用的高性能映射方法，在SDH架构编译基础研究方面有重大突破和重要创新成果，有助于打通大数据驱动人工智能软件和硬件之间的壁垒。

The rise of big data-driven artificial intelligence (AI) has put forward demand for the acceleration of data-intensive applications. Software defined hardware (SDH), targeting to build runtime reconfigurable hardware and software that enables near ASIC performance without sacrificing programmability, becomes one of the most potential technologies for accelerating these applications. Due to the data density and structural diversity of applications and the complexity and restriction of SDH hardware, the SDH mapping presents the problems of long compilation time, memory accessing conflict, low computational parallelism and poor scalability. To address these common and key problems, this project proposes a high performance mapping approach for data-intensive applications over SDH on the basis of the solid research of multi-parameter collaborative mapping technology over coarse-grained reconfigurable architecture. The mapping approach comprehensively utilizes the theory and method of multidisciplinary innovation from reconfigurable computing, storage management, deep learning and reinforcement learning. In detail, our project consists of three parts. First, the project establishes an accuracy performance model for SDH architecture. Then, a conflict-free data placement algorithm is well designed. Finally, a deep reinforcement learning based placement and routing algorithm is delicately studied. The project has significant breakthroughs for SDH compiling, which can help to break the barriers between the software and hardware of big data-driven AI.