Optimization of the n-dimensional sliding window inter-channel correlation algorithm for multi-core architecture

TL;DR

This paper presents an optimized and parallelized GPU algorithm for calculating inter-channel correlation in sliding windows, significantly reducing processing time for large data sets.

Contribution

The paper introduces a novel optimization reducing operations in overlapping windows and a GPU parallel implementation for faster correlation computation.

Findings

Achieved 60x speedup on GPU compared to CPU implementation.

Optimized algorithm reduces computational complexity in overlapping window areas.

Demonstrated effectiveness on large 12 MPixel images.

Abstract

Calculating the correlation in a sliding window is a common method of statistical evaluation of the interconnect between two sets of data. And although the calculation of a single correlation coefficient is not resource-intensive and algorithmically complex, sequential computation in a large number of windows on large data sets can take quite a long time. In this case, each value in the data, falling into different windows, will be processed many times, increasing the complexity of the algorithm and the processing time. We took this fact into account and optimized the correlation calculation in the sliding window, reducing the number of operations in the overlapping area of the windows. In addition, we developed a parallel version of the optimized algorithm for the GPU architecture. Experimental studies have shown that for a 7x7 correlation window sliding in one pixel increments, we were able to accelerate the processing of an 12 MPixel image pixels on the GPU by about 60 times compared to the serial version running on the CPU. The article presents an optimized version of the algorithm, a scheme for its parallelization, as well as the results of experimental studies.