一种面向OpenCL架构的矩阵-向量乘并行算法与实现

Matrix-vector Multiplication Parallel Algorithm and Implementation for OpenCL Architecture

矩阵-向量乘法算法的时间复杂度大,传统计算方法的实时性和跨平台性难以保证.本文提出一种基于开放式计算语言(Open Computing Language,OpenCL)的矩阵-向量乘并行算法,矩阵-向量乘法过程被分解成若干具有不同粒度的子任务.根据相应的并行度,每个工作组进行矩阵中的行块与列向量的乘积,每个工作项进行行块中行向量与列向量的乘积,并把计算任务分别分配到计算单元和处理单元进行处理.实验结果表明,与基于CPU的串行算法、基于OpenMP并行算法和基于统一计算设备架构(Compute Unified Device Architecture,CUDA)并行算法性能相比,矩阵-向量乘并行算法在OpenCL架构下NVIDIA图形处理器(Graphic Processing Unit,GPU)计算平台上分别获得了20. 86倍、6. 39倍和1. 49倍的加速比.验证了提出的并行优化方法的有效性和性能可移植性.

The time complexity of matrix-vector multiplication algorithm is large,and the real-time and cross-platform performance of traditional computing methods is difficult to guarantee. This paper presents a matrix-vector multiplication parallel algorithm based on Open Computing Language( OpenCL),and the matrix-vector multiplication process is decomposed into several subtasks with different granularity. According to the corresponding degree of parallelism,each work-group carries on the product of the rowblock in the matrix and the column vector,each work-item carries on the product of the rowvector in the rowblock and the column vector,and assigns the computation task separately to the compute unit and the processing element for processing. The experimental results showthat compared with the performance of the serial algorithm based on CPU,parallel algorithm based on OpenMP and parallel algorithm based on Compute Unified Device Architecture( CUDA),the matrix-vector multiplication parallel algorithm obtains 20. 86 times,6. 39 times and 1. 49 times speedup in the NVIDIA GPU computing platform under the OpenCL architecture respectively. The validity and performance portability of the proposed parallel optimization method are verified.