• Journal of Digital Contents Society
• /
• v.18 no.6
• /
• pp.1151-1156
• /
• 2017

It is common to achieve lower performance in traversing tree data structures in GPU than one expects. In this paper, we analyze the reason of lower-than-expected performance in GPU tree traversal and present that the warp divergences is caused by the branch instructions ("if${\ldots}$ else") which appear commonly in tree traversal CUDA codes. Also, we compare the parallel shifted sort algorithm which can reduce the number of warp divergences with a kd-tree CUDA implementation to show that the shifted sort algorithm can work faster than the kd-tree CUDA implementation thanks to less warp divergences. As the analysis result, the shifted sort algorithm worked about 16-fold faster than the kd-tree CUDA implementation for $2^{23}$ query points and $2^{23}$ data points in $R^3$ space. The performance gaps tend to increase in proportion to the number of query points and data points.

• Journal of Digital Contents Society
• /
• v.18 no.4
• /
• pp.739-745
• /
• 2017

This paper presents and discusses an implementation of the GPU shifted sorting method to find approximate k nearest neighbors which executes within "warp", the minimum execution unit in GPU parallel architecture. Also, this paper presents the comparison results with other two common nearest neighbor searching methods, GPU-based kd-tree and ANN (Approximate Nearest Neighbor) library. The proposed implementation focuses on the cases when k is small, i.e. 2, 4, 8, and 16, which are handled efficiently within warp to consider it is very common for applications to handle small k's. Also, this paper discusses optimization ways to implementation by improving memory management in a loop for the CUB open library and adopting CUDA commands which are supported by GPU hardware. The proposed implementation shows more than 16-fold speed-up against GPU-based other methods in the tests, implying that the improvement would become higher for more larger input data.