P-SOBI: A Parallel Implementation for Second Order Blind Identification Algorithm

Abstract

Blind source separation (BSS) is one of the most significant frontier research fields of modern signal processing. However, the signal to be processed is often a mixture of many different signals and it is very hard to separate target signal from the mixed signals. The Second Order Blind identification (SOBI) algorithm is widely used in blind source separation (BSS) processing and has achieved many satisfactory results. And in some real time applications such as human machine interfaces (BMI) based on Brain Computer Interface (BCI) systems, realtime processing is often required. But the iteration period of SOBI algorithm takes too much time and makes it hard to meet the requirement of real time. In this paper, we introduce our P-SOBI algorithm which is a parallel implementation of the original SOBI algorithm. We use Intel MKL library to accelerate and simplify the matrix operation in the algorithm. For large matrices, we use MPI to divide and handle them in parallel. Apart from that, we utilize the compiler options and some fine-grained optimization strategies to boost the algorithm. Our results indicate that, our proposed P-SOBI algorithm has a better performance compared with the original SOBI algorithm and reduce the execution time significantly. And in more detail, it can achieve a speedup of 12.9x times compared with the original SOBI algorithm. Thus our proposed P-SOBI algorithm has great potential for real time processing.