An in-depth description of an apparently forgotten matrix operation, the reversal operator, is developed. The properties of such an operation are also given, resulting in a new vector-matrix operation resembling the w...An in-depth description of an apparently forgotten matrix operation, the reversal operator, is developed. The properties of such an operation are also given, resulting in a new vector-matrix operation resembling the well-known ones of conjugation, transposition, and inversion. The reversal operator operates by ordering the object components where applied. Reversal is easy to perform as it is distributive regarding the vector sum and matrix product. Supplementary descriptions of matrix regions not often used in linear algebra, like the anti-diagonal concept, are also discussed. Some practical problems are given.展开更多
稀疏线性方程组求解等高性能计算应用常常涉及稀疏矩阵向量乘(SpMV)序列Ax,A2x,…,Asx的计算.上述SpMV序列操作又称为稀疏矩阵幂函数(matrix power kernel,MPK).由于MPK执行多次SpMV且稀疏矩阵保持不变,在缓存(cache)中重用稀疏矩阵,可...稀疏线性方程组求解等高性能计算应用常常涉及稀疏矩阵向量乘(SpMV)序列Ax,A2x,…,Asx的计算.上述SpMV序列操作又称为稀疏矩阵幂函数(matrix power kernel,MPK).由于MPK执行多次SpMV且稀疏矩阵保持不变,在缓存(cache)中重用稀疏矩阵,可避免每次执行SpMV均从主存加载A,从而缓解SpMV访存受限问题,提升MPK性能.但缓存数据重用会导致相邻SpMV操作之间的数据依赖,现有MPK优化多针对单次SpMV调用,或在实现数据重用时引入过多额外开销.提出了缓存感知的MPK(cache-awareMPK,Ca-MPK),基于稀疏矩阵的依赖图,设计了体系结构感知的递归划分方法,将依赖图划分为适合缓存大小的子图/子矩阵,通过构建分割子图解耦数据依赖,根据特定顺序在子矩阵上调度执行SpMV,实现缓存数据重用.测试结果表明,Ca-MPK相对于Intel OneMKL库和最新MPK实现,平均性能提升分别多达约1.57倍和1.40倍.展开更多
文摘An in-depth description of an apparently forgotten matrix operation, the reversal operator, is developed. The properties of such an operation are also given, resulting in a new vector-matrix operation resembling the well-known ones of conjugation, transposition, and inversion. The reversal operator operates by ordering the object components where applied. Reversal is easy to perform as it is distributive regarding the vector sum and matrix product. Supplementary descriptions of matrix regions not often used in linear algebra, like the anti-diagonal concept, are also discussed. Some practical problems are given.
文摘稀疏线性方程组求解等高性能计算应用常常涉及稀疏矩阵向量乘(SpMV)序列Ax,A2x,…,Asx的计算.上述SpMV序列操作又称为稀疏矩阵幂函数(matrix power kernel,MPK).由于MPK执行多次SpMV且稀疏矩阵保持不变,在缓存(cache)中重用稀疏矩阵,可避免每次执行SpMV均从主存加载A,从而缓解SpMV访存受限问题,提升MPK性能.但缓存数据重用会导致相邻SpMV操作之间的数据依赖,现有MPK优化多针对单次SpMV调用,或在实现数据重用时引入过多额外开销.提出了缓存感知的MPK(cache-awareMPK,Ca-MPK),基于稀疏矩阵的依赖图,设计了体系结构感知的递归划分方法,将依赖图划分为适合缓存大小的子图/子矩阵,通过构建分割子图解耦数据依赖,根据特定顺序在子矩阵上调度执行SpMV,实现缓存数据重用.测试结果表明,Ca-MPK相对于Intel OneMKL库和最新MPK实现,平均性能提升分别多达约1.57倍和1.40倍.
