面向大尺寸滑动窗口应用的并行计算模型

摘要

There exists a large gap between the data input speed and processing speed in large-size sliding-window applications. To shorten this gap, a parallel processing scheme was proposed, which achieved high data reusability and parallelism with memory resources as few as possible and memory access control logics as simple as possible. The scheme combined the advantages of parallelism among different sliding-windows and parallelism among different data in a single window. For different windows, they were divided into groups and mapped into multiple processing elements. For the data in a single window, multi-module memory structure was introduced to buffer them, where module assignment and addressing scheme was designed for conflict-free parallel access. Experimental results on FPGA show that this approach can improve the processing speed significantly without incurring too muCh memory resources and too complicated memoy access control logics.%大尺寸滑动窗口的应用在数据输入速度与处理速度之间存在较大差距.为了缩短差距,提出了一种并行计算模型,使用尽可能少的存储资源与尽可能简单的存储器读写控制逻辑实现了尽可能高的数据重用性与并行性.该模型将不同滑动窗口之间的并行处理与单个窗口内不同数据之间的并行处理结合起来:对于不同窗口,按列进行分组并映射到多个处理单元上并行处理;对于单个窗口内的数据,使用多体存储结构进行缓存,并设计了存储体分配机制与寻址函数以实现多个数据的无冲突并行访问.在FPGA上的实验结果表明:提出的计算模型在没有明显增加存储资源使用代价与读写控制逻辑复杂性的情况下大大提高了处理速度.