基于线性变换的计算与数据动态分解方法

在并行优化编译器的并行识别过程中，许多串行代码无法找到全局一致的分解结果，数据重分布无可避免，有必要寻找一种有效的方法求解计算和数据的动态分解。该文研究了单个嵌套循环计算与数据分解算法以及分解结果表示方法，提出一种在多个嵌套循环间求解数据线性一致分布的动态分解算法，结合程序的结构分析和程序的控制流信息，用于通用串行代码的并行分解过程，可以同时给出串行代码的计算划分和数据分布结果。

Approach for Computation and Data Dynamic Decompositions Based on Linear Transformation

In the process of parallel recognition, many serial codes have no way to find a uniform decomposition result. Since data reorganization is unavoidable, how to design an effective dynamic decomposition algorithm is an issue to be researched. Based on the study of decomposition methods used in single loop nest and the representations of decomposition results, this paper proposes a dynamic decomposition algorithm which exploits coincidence linear data distribution results among loop nests. This algorithm also takes the structure and the control flows of serial programs into consideration for more accurate decomposition. It can be used in the process of parallelism decomposition for common serial codes. The computation partition and the data distribution result of serial codes are given by the algorithm.