A parallel algorithm with embedded load balancing for autocorrelation matrix computation

The computation of autocorrelation matrix is used heavily in several areas including signal and image processing, where parallel and application-specific architectures are also being increasingly used. Therefore, an efficient scheme to compute autocorrelation matrix on parallel architectures has tremendous benefits. In this paper, a parallel algorithm for the computation of autocorrelation matrix on 2-D mesh, is presented. The computation requirements for the elements of the autocorrelation matrix is highly skewed and the proposed algorithm attempts to balance the computation load, without requiring an external load balancing algorithm or processor. In this sense, the load balancing is embedded within the algorithm. The exact number of computation steps are derived. The time complexity of the proposed algorithm is shown to be within twice the optimal (or lower bound). It is also shown to have twice the speedup of a straight-forward parallel algorithm.