基于GPU的大规模栅格数据分块并行处理方法

摘要

数学形态学运算是栅格数据处理的重要方法,具有较高的计算复杂度、并行度等特点,较容易发挥GPU众核高度并行执行的优势,以提高其计算效率。然而,有限的GPU全局存储器限制了其在大规模数据中的应用。文中在分析现有栅格数据并行方法的基础上,基于通用并行计算架构CUDA,设计一种适应大规模数据的分块处理方法。文中以经典的膨胀算法为例对分块处理方法进行测试。实验结果表明：与传统的CPU串行处理方法相比,该方法可以显著提高数据处理速度。%Mathematical morphology operations are important methods in the field of raster data processing,with high degree of computa-tional complexity and parallelism. GPU-based hypercore parallel computing method can significantly improve the calculation speed. How-ever,GPU’ s global memory limits its application in large scale data. Present a block-based method for large scale data,based on a gener-al purpose parallel computing architecture,after analyzing the present parallel method for raster data. The new method is specifically tested with the classical dilation algorithm. Experimental results show that the calculation speed of the new method is faster than that of tradition-al sequence algorithm based on CPU.