Hybrid Vectorization and Parallelization for Matrix-Matrix Multiplication on Multi-core Platform

Matrix-matrix multiplication is commonly used as the core of operation in linear algebra computations and various applications such as finite element analysis and deep neural networks. Computations in large matrices are time-consuming and suffer time-delay due to memory bottlenecks. As a result, real-time computations of large matrices are not viable. This paper mitigates both issues by combining vectorization and parallelization of matrix-matrix multiplication with AVX functions and optimizing loops with OpenMP on a multi-core platform. Compared with native parallel algorithms, the proposed method gives a speed gain of 3.5-7.5 times with parallelism at 4-threads.