基于MPI的匹配方体并行计算研究

MPI based parallel computation of matching cube

双目立体视觉的匹配方体计算过程可以进行SIMD类型的并行计算,基于MPI通信环境将视差值的计算任务分配到不同的计算节点上,然后将各节点计算所获得的DSI图像汇集在根节点上,最终通过数据规整快速获得所需的匹配方体。同时建立了该并行算法基于处理器时钟周期的相对精确的计算时间复杂度模型,用于分析不同计算平台上的性能。由于计算过程中数据相关性较低,因此在基于MPI与Myrinet网络的Linux集群计算平台上获得了较好的加速比。

According to PCAM, a parallel algorithm was figured out to accelerate the computation of matching-cube for stereoscopic vision. The computation of matching-cube was divided by vertical coordinate into many sub-computations, which minimized the communication between computing nodes. By assigning these computation jobs of different disparity values to multiple computing nodes and gathering all these DSI to root node, a matching cube was obtained. A relatively accurate computational time complexity modal was built on CPU cycles to analyze the performance on different platforms, which was very important to real-time applications. As the data-dependence during computation was very low, a nearly linear speedup could be obtained on MPI cluster parallel platform.