CAJ | 학술논문

针对小波分解计算速度慢、实际工程应用少的问题,采用图形处理器(GPU)作为计算平台,提出一种基于计算统一设备架构(CUDA)的细粒度高速并行小波分解算法.通过分析小波Mallat算法的并行性,并考虑GPU单个处理单元计算能力相对较弱的特点及CUDA的多层式存储器结构、多层式线程组织结构和单指令流多线程流(SIMT)体系结构,采用数据分组及轻量级线程任务分解的方式,提出了适合CUDA程序设计模型的高速并行小波分解算法,并将其用于电力系统谐波分析.实验证明,该算法相对于CPU串行小波分解和Matlab engine小波分解的计算耗时,最高可分别达到26倍和65倍的速度提升,且算法具有线性加速能力.
침대소파분해계산속도만、실제공정응용소적문제,채용도형처리기(GPU)작위계산평태,제출일충기우계산통일설비가구(CUDA)적세립도고속병행소파분해산법.통과분석소파Mallat산법적병행성,병고필GPU단개처리단원계산능력상대교약적특점급CUDA적다층식존저기결구、다층식선정조직결구화단지령류다선정류(SIMT)체계결구,채용수거분조급경량급선정임무분해적방식,제출료괄합CUDA정서설계모형적고속병행소파분해산법,병장기용우전력계통해파분석.실험증명,해산법상대우CPU천행소파분해화Matlab engine소파분해적계산모시,최고가분별체도26배화65배적속도제승,차산법구유선성가속능력.
As there is few applications of wavelet decomposition in actual engineering because of its low calculation speed,a fine-grained parallel wavelet decomposition algorithm based on CUDA(Compute Unified Device Architecture) is proposed,which takes the GPU(Graphic Processing Unit)as platform.The parallelity of the Mallat algorithm is analyzed.With the consideration of the poor performance of GPU processor and the CUDA framework of multilevel memory,multilevel thread organization and SIMT(Single-Instruction,Multiple-Thread),high-speed parallel wavelet decomposition algorithm is proposed for power system harmonic analysis,which applies the methodology of data grouping and lightweight thread task decomposing,suitable for the CUDA programming model.Experiments show that,the calculation speed is 26 times and 65 times faster compared with that of CPU serial wavelet decomposition and Matlab engine wavelet decomposition respectively,and the algorithm has the linear speedup capability.