可配置可扩展处理器编译器设计

发布时间：2018-06-08 00:31

本文选题：可配置处理器 + 传输触发体系结构(TTA)　；参考：《天津大学》2012年硕士论文

【摘要】：随着集成电路制造技术的进步,各种可定制处理器在嵌入式设计中得到了广泛的应用。针对不同目标处理器结构生成优化的代码成了编译器的一个艰巨而重要的工作。当前,一种新型的可定制处理器结构——传输触发架构(Transport Trigger Architecture, TTA)由于其灵活性、可扩展、高性能、低成本和低功耗等特点在科研机构和工业界中获得了广泛的关注。本文主要研究了TTA处理器的编译器后端优化部分,主要包括了指令筛选和指令调度。本文采用图变换的方式进行DAG的预处理,解决了通常模式匹配中出现的由于拓扑结构不同,无法完成匹配的问题。通过按照一些特定的规则,在进行指令匹配之前进行DAG变换,使得变换之后的DAG结构与功能一一对应,从而使得功能相同的部分能够完全完成匹配。并且通过特殊的建模方法,使得存在交集的待匹配节点集合也可以同时匹配。之后采用启发式的遗传算法搜索到近似的最优解。在完成指令筛选之后,本文提出了一种基本块范围内的融合了遗传算法和禁忌搜索的新型搜索算法模型。在确定了指令间的相互依赖关系以及资源约束后,通过在遗传算法的架构下进行禁忌搜索近似最优的指令调度方案来完成指令调度。实验结果表明与其他常规算法相比该算法不仅能够产生较高质量的并行代码,还能更快的完成收敛。
[Abstract]:With the development of IC manufacturing technology, various customizable processors have been widely used in embedded design. Generating optimized code for different target processor architectures has become a difficult and important task for compilers. At present, a new customizable processor architecture, Transport trigger Architecture (TTA), has gained wide attention in scientific research institutions and industry because of its flexibility, scalability, high performance, low cost and low power consumption. In this paper, the compiler back-end optimization of TTA processor is studied, including instruction filtering and instruction scheduling. In this paper, the DAG is preprocessed by graph transformation, which solves the problem of different topology in common pattern matching. Unable to complete the matching problem. According to some specific rules, the DAG transformation is carried out before the instruction matching, so that the transformed DAG structure corresponds to the function one by one, so that the part with the same function can complete the matching completely. And through the special modeling method, the set of nodes to be matched with intersection can be matched at the same time. Then the heuristic genetic algorithm is used to search the approximate optimal solution. After the instruction selection is completed, a new search algorithm model combining genetic algorithm and Tabu search is proposed in this paper. After the interdependence between instructions and resource constraints are determined, instruction scheduling is accomplished by using Tabu search approach to approximate optimal instruction scheduling scheme under the framework of genetic algorithm (GA). Experimental results show that this algorithm can not only produce high quality parallel code, but also complete convergence faster than other conventional algorithms.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP332

【参考文献】