CPU-OpenMP和GPU-CUDA并行计算技术对矩阵乘法运算的加速效果分析

Accelerating Effect Analysis of Matrix Multiplication with CPU-OpenMP and GPU-CUDA Parallel Computing

本文对比了CPU-OpenMP和GPU-CUDA并行计算技术对不同阶矩阵乘法运算相对于CPU单线程计算的加速效果。结果表明,CPU-OpenMP并行的计算加速比与矩阵阶数无关,且低于所采用的线程数目。GPU-CUDA并行的计算加速比随矩阵阶数的增加显著增加,最大计算加速比可达570倍以上。相对于CPU单线程计算结果,CPU-OpenMP并行计算未产生误差,而GPU-CUDA并行计算会产生误差。结果表明,GPUCUDA并行适合高阶数矩阵乘法的加速计算,而CPU-OpenMP并行适合低阶数矩阵乘法的加速计算。

This paper compares the accelerating effects of CPU-OpenMP and GPU-CUDA parallel computing on the computation of different-order matrix multiplication over CPU single-thread. The results show that the computational speedup of CPU-OpenMP parallelism is independent of matrix order and lower than the number of threads used. GPU-CUDA parallel computing speedup ratio increases significantly with the increase of the matrix order, the maximum computational speedup up to 570 times. Relative to the CPU single-thread calculations, CPU-OpenMP parallel computing did not produce errors, and GPU-CUDA parallel computing will produce errors. The results show that GPU-CUDA parallel is suitable for accelerated computing of high-order matrix multiplication, while CPU-OpenMP parallel is suitable for accelerated computing of low-order matrix multiplication.