FPGA Implementation of a Point Cloud Processing KNN Algorithm Used in GCN Network

KNN (k-nearest neighbor) algorithm is an important method which exhibits great performance in many fields. It is a commonly used step in graph convolutional networks (GCN) when graph structure is not available. However, only a few works have been proposed to implement KNN part of GCN through hardware solutions. In this paper, we present an FPGA implementation of k-nearest neighbor construction part of a point cloud processing GCN. Unlike several related works, the distance calculation part and sorting part of our design executes in parallel. By employing a dedicated calculation procedure and optimizing previous architectures, we successfully reduce resource consumption of our design. The proposed KNN unit which implements KNN part of 3D-GCN only consumes 5907 LUTs, 1687 FFs and 18 DSPs. Our work is resource-efficient compared to previous methods and achieves 24x performance speed up over Core i5 CPU, operating at a much lower clock frequency.