A Mixed Hierarchical Algorithm for Nearest Neighbor Search

The k nearest neighbor (kNN) search is a computationally intensive application critical to fields such as image processing, statistics, and biology. Recent works have demonstrated the efficacy of k-d tree based implementations on multi-core CPUs. It is unclear, however, whether such tree based implementations are amenable for execution in high-density processors typified today by the graphics processing unit (GPU). This work seeks to map and optimize kNN to massively parallel architectures such as the GPU. Our approach synthesizes a clustering technique, k-means, with traditional brute force methods to prune the search space while taking advantage of data-parallel execution of kNN on the GPU. Overall, our general case GPU version outperforms a singlethreaded CPU by factors as high as 108.