Acceleration of band algorithms using fast dense kernels for large Band matrices

点击翻译按钮获取中文摘要

Acceleration of band algorithms has been tested using a co-processor which enables high-performance dense matrix multiplication. The algorithms include Gaussian elimination for general band matrices, and Householder tridiagonalization for symmetric band matrices. Because the band algorithms reduce the amount of storage space in memory by utilizing the specific geometric structure of the matrices and the co-processor uses a separate memory space, data transfer and conversion between memory and co-processor are essential.