基于深度学习的稀疏矩阵向量乘运算性能预测模型

Computing Performance Prediction Model for Sparse Matrix Vector Multiplication Based on Deep Learning

稀疏矩阵向量乘(SpMV)是求解稀疏线性方程组的计算核心,被广泛应用在经济学模型、信号处理等科学计算和工程应用中,对于SpMV及其调优技术的研究有助于提升解决相关领域问题的运算效率。传统SpMV自动调优方法基于硬件平台的体系结构参数设置来提升SpMV性能,但巨大的参数设置量导致搜索空间变大且自动调优耗时大幅增加。采用深度学习技术,基于卷积神经网络,构建由双通道稀疏矩阵特征融合以及稀疏矩阵特征与体系结构特征融合组成的SpMV运算性能预测模型,实现快速自动调优。为提高SpMV运算时间的预测精度,选取特征数据并利用箱形图统计SpMV时间信息,同时在佛罗里达稀疏矩阵数据集上进行实验设计与验证,结果表明,该模型的SpMV运算时间预测准确率达到80%以上,并且具有较强的泛化能力。

Sparse Matrix Vector Multiplication(SpMV)is key to solving sparse linear equations.It is widely used in economic modeling,signal processing and other scientific and engineering tasks.The research on SpMV and its tuning technology can improve the computational efficiency of solving problems in related fields.Traditional SpMV automatic tuning methods improve the performance of SpMV based on the architecture parameter settings of the hardware platform,but the huge amount of parameter settings leads to a larger search space and a significant increase in the time consumption of automatic tuning.To implement fast and accurate automatic tuning,we use deep learning technology to construct a Convolutional Neural Network(CNN)model for SpMV computing performance prediction,which is built based on dual-channel sparse matrix feature fusion,sparse matrix feature fusion and architecture feature fusion.In order to improve the prediction accuracy of SpMV computing performance,feature data is selected and constructed.The box plot is used to count SpMV time information.Then the Florida sparse matrix dataset is selected for experimental design and verification.Experimental results show that the model displaying a prediction accuracy of SpMV computing time over 80%and strong generalization ability.