Segmenting a signal based on a local property using multicore processors

Signal segmentation is a partitioning of a signal into contiguous regions (the segments), where each segment has similar values of local properties of interest. In this paper, we consider the problem of signal segmentation by classifying on one property, where a histeresis function is applied to the output to prevent noise or other temporally short disturbances from producing an over-abundance of segments. On serial processors, such segmentation is straightforwardly performed by going through the samples in order. However, on multicore processors the histeresis function presents a problem. Data cannot simply be divided between the cores due to the fact that the output of the histeresis function at any sample can depend on the entire history of the signal up to that time. We present a parallel algorithm based on the parallel scan (also called parallel prefix or prefix scan) pattern. Both the serial and parallel method was implemented and tested on a multicore processor. The two methods produced exactly the same segments, so the two methods result in exactly the same measurement results. Speedups of several fold over the serial performance were observed when using the parallel code.