稀疏矩阵在C66x上的应用及优化  

Application and optimization of sparse matrix vector multiplication on C66x

作　　者：黄旭东[1] 洪泽 陈振娇[1] Huang Xudong;Hong Ze;Chen Zhenjiao(China Electronics Technology Group Corporation No.58 Research Institute,Wuxi 214035,China)

机构地区：[1]中国电子科技集团公司第五十八研究所,江苏无锡214035

出　　处：《电子技术应用》2024年第11期23-27,共5页Application of Electronic Technique

摘　　要：随着大数据的爆炸式发展,稀疏矩阵已经成为机器学习和边缘计算的重要一环。在机器学习领域,数据集的稀疏矩阵化既可以保存信息又可以节省内存,已成为不可避免的趋势。SpMV(稀疏矩阵向量乘)作为稀疏矩阵计算中的核心,其迭代求解过程的空间复杂度和时间复杂度具有重要研究意义。分析稀疏矩阵C00、CSR、ELLPACK和DIA压缩格式,改变稀疏矩阵的稀疏度和非零元素的分布,得出COO读取数据、CSR进行计算的SpMV通用性更强。利用C66x的VLIW指令构架,采用软件流水的方式对SpMV_CSR算法进行指令并行优化,利用SIMD单指令多数据指令集对SpMV_CSR算法完成数据并行优化。实验结果表明,优化后的SpMV_CSR算法相较于优化前的加速比平均达到5倍以上。With the explosive development of big data,sparse matrix has become an important part of machine learning and edge computing.In the field of machine learning,sparse matrix of data sets can not only save information but also save memory,which has become an inevitable trend.Sparse matrix vector multiplication(SpMV)is the core of sparse matrix computation.The space complexity and time complexity of its iterative solution process have important research significance.Analyze the compres‐sion format of sparse matrix C00,CSR,ELLPACK and DIA,change the sparsity of sparse matrix and the distribution of non-zero elements,and conclude that the SpMV read by COO and calculated by CSR is more universal.Utilizing the VLIW instruction ar‐chitecture of C66x,using software pipelining to manage SpMV_CSR algorithm for instruction parallel optimization,utilizing SIMD single instruction multiple data instruction set for SpMV_CSR algorithm completes data parallel optimization.The experi‐mental results indicate that the optimized SpMV_CSR algorithm has an average acceleration ratio of over 5 times compared to be‐fore optimization.

关 键 词：稀疏矩阵 SpMV CSR C66x 软件流水 SIMD 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]