Analýza a řešení výzev implementace vyhledávání k nejbližších sousedů pro GPU akcelerátory

Hanák, Drahomír

Similarity search is a commonly used technique in databases for finding objects si- milar to a query. It finds applications in content-based retrieval of complex objects like images, information retrieval, and statistical learning. Our thesis focuses on the imple- mentation and optimization of the k nearest neighbours (kNN) algorithm on a GPU, a commonly used technique in similarity search. We analyze and evaluate several existing GPU kNN implementations in various configurations and propose the best algorithm for each configuration. We also suggest optimizations of k-selection. In particular, we suggest a small k-selection approach, which achieves up to 80% of peak theoretical throughput on a typical configuration used in many applications of kNN and is faster than the current state-of-the-art. We implemented a fused algorithm, which solves kNN without mate- rializing the distance matrix, and a large k-selection, which outperforms an optimized, parallel sorting of the whole database by a significant margin. 1

guest :: login Digital Repository
		Search		Submit		Help		About