Redistribution Networks for Resampling

Resampling is an important building block in core signal processing methods such as particle filters or genetic algorithms. This work describes how to accelerate the redistribution part of resampling, in a parallelized form utilizing a processing network composed of low-complexity nodes with
$O(log_{2}^{2}(n))$
layers. We furthermore show how to use such networks for block-based processing of input vectors that are larger than the input of the network, allowing to tradeoff the best of both worlds: block-based processing with its low area requirements and network-based processing with its high speed. We present simulation results not only showing the performance gains compared to the trivial linear method but also showing that by using the proposed architecture one can achieve processing times on edge devices that recently required high-performance server clusters.