论文部分内容阅读
本文提出一个基于分布式局存的SIMD机器的循环分割理论体系以优化运算中所需要的数据传输．该体系使用矩阵表示迭代空间、数据空间和数组存取式．我们引入数据传输概念，并建立一个简单有效的数据传输模型来评估数据在全局内存和局部内存之间的传输开销．最后，对于给定的循环嵌套，我们给出一个循环分割算法以获得优化循环块，使得循环嵌套中所需要的数据传输开销最小，并且大大减少了数据传输和计算的同步开销．实验结果证明了该算法的有效性．
In this paper, we propose a cyclic segmentation theory system based on distributed inventoried SIMD machines to optimize the data transmission in computing. The system uses a matrix to represent iterative space, data space, and array access. We introduce the concept of data transmission and establish a simple and efficient data transfer model to evaluate the data transmission overhead between global memory and local memory. Finally, for a given loop nesting, we give a loop partitioning algorithm to get an optimized loop block, which minimizes the data transfer overhead required for loop nesting and greatly reduces the data transfer and computational synchronization overhead. Experimental results show the effectiveness of the algorithm.