Innovative Unit-Vector-Block Storage Format of Sparse Matrix and Vector

Sparse Matrix-Vector Multiplication (SpMV) operation is widely used in many iterative numerical algorithms, such as high performance computing, machine learning, and graph processing. However, peak performance is limited due to the fact that the commonly used algorithms alternate between compute-bound and memory-bound steps. This paper introduces a new storage format UVB (Unit Vector Block) for sparse matrix and vector, which needs less memory capacity compared to current popular formats for sparse vector and matrix such as COO, CSR, DIA, ELL and etc. Moreover, a new SpMV algorithm based on this new format can make full use of Intel
^®
AVX-512 instruction set, and decrease the execution time by up to 99% compared with the most popular LIBSVM kernel.