基于可重构的多并行计算模式视觉识别系统研究

Computer vision has become as one of the most promising areas of embedded applications. However, balancing the huge computation and communication demands of the algorithms with the stringent size, power and memory resource constaints of embedded platforms have created incredible challenges. Parallelization has emerged as one of the solutions to overcome these challenges. ..We aim to achieve an efficient embedded image recognition system with robustness to complicated environments containing multiple objects. The proposed solution is an improvement and expansion of conventional SIMD/MIMD architectures. By combining reconfigurable technology with SIMD/MIMD parallel computing models, and taking advantage of the flexibility and configurability of a reconfigurable architecture, we improve the performance on all three different levels of image recognition algorithms at an expense of the minimal hardware resources. In addition, by employing the data organization adaptable to different algorithm levels, the solution can effectively release the pressure on the memory by object recognition algorithms. Implementing hardware and software co-design methodology, an evaluation platform for object recognition system is also very necessary.

目标识别算法是当前嵌入式领域最具应用前景的视觉算法之一，然而设计出既能满足识别算法性能，又能平衡性能与面积、功耗间矛盾的硬件架构仍是未解决的难题。单指令多数据/多指令多数据(SIMD/MIMD)架构因其高能效、低面积、灵活性强的特点而引起关注。然而，识别算法中的高层计算处理的数据结构不规则，包含大量循环中的判断，限制了指令执行的SIMD/MIMD架构充分发挥效率。本项目通过对SIMD/MIMD架构改进和拓展，拟在同一块硬件上分时支持SIMD模式、多SIMD簇模式及可重构模式。主要内容包括：利用可重构的可配置性和灵活性最大限度挖掘并行度，提高高层算法的执行效率；设计支持可重构的NoC网络，实现多计算模式间的快速切换，减小多计算模式的引入对架构性能的影响； 研究适应不同层次计算的数据组织方法，从而缓解识别算法对存储带宽的压力；此外，本项目拟搭建目标识别系统评估平台，探索和评估硬件设计方案。

本项目设计可重构架构，实现视觉目标识别算法中低/中/高层计算的SIMD/ MIMD加速。除设计该可重构架构以外，主要课题内容包括：分析目标识别算法计算数据结构以及仿存模式特点，并据此设计主计算单元RPU片上二维数据缓存以及管理机制，增加数据重用效率。同时改善主计算单元互联结构，提高目标识别算法的数据调用灵活性以及调用效率。设计可重构配置字压缩方法，配置字分层缓存结构，发挥配置的可重用性，提高可重构配置效率。实现目标识别算法低/中/高层计算的SIMD/MIMD加速方法：将规使用整数据且循环执行的数据密集型低/中层计算以SIMD方式交RPU执行；将富含判断分支的控制密集型高层计算交μPU执行；硬件支持μPU与RPU任务的流水化，形成MIMD计算模式。实验证明，与其他加速器架构比较，本项目在目标识别算法低/中/高层计算均取得2x~3x倍的性能加速收益。本项目在目标识别算法低/中/高层计算的可重构SIMD/MIMD加速实现上有创新性，发表期刊以及会议论文10篇，申请专利12篇其中3篇获得授权，培养博士2名，硕士4名。