Tree-based Nearest Neighbor Search (NNS) is hard to parallelize on GPUs. However, newer Nvidia GPUs are equipped with Ray Tracing (RT) cores that can build a spatial tree called Bounding Volume Hierarchy (BVH) to accelerate graphics rendering. Recent work proposed using RT cores to implement NNS, but they all have a hardware-imposed constraint on the type of distance metric, which is the Euclidean distance. We propose and implement two approaches for generalized distance computations: filter-refine, and monotone transformation, each of which allows non-euclidean nearest neighbor queries to be performed in terms of Euclidean distances. We find that our reductions improve the time taken to perform distance computations during the search, thereby improving the overall performance of the NNS.

翻译：基于树的最近邻搜索在GPU上难以并行化。然而，新型英伟达GPU配备了光线追踪核心，可构建名为包围体层次结构（BVH）的空间树以加速图形渲染。近期工作提出利用光追核心实现最近邻搜索，但均存在硬件对距离度量类型的约束，即仅支持欧氏距离。我们提出并实现了两种广义距离计算方法：过滤-精化与单调变换，每种方法均允许将非欧氏距离最近邻查询转化为基于欧氏距离的查询。实验表明，我们提出的约简方法能有效缩短搜索过程中的距离计算耗时，从而提升最近邻搜索的整体性能。