high performance computing on graphics processing units: hgpu.org

Query-by-content by means of similarity search is a fundamental operation for applications that deal with multimedia data. For this kind of query it is meaningless to look for elements exactly equal to the one given as query. Instead, we need to measure dissimilarity between the query object and each database object. The metric space model is a paradigm that allows modelling all similarity search problems. Metric databases permit to store objects from a metric space and efficiently perform similarity queries over them, in general, by reducing the number of distance evaluations needed. Therefore, the goal is to preprocess a particular dataset in such a way that queries can be answered with as few distance computations as possible. Moreover, for a very large metric database it is not enough to preprocess the dataset by building an index, it is also necessary to speed up the queries via high performance computing using GPU. In this work we show an implementation of a pure GPU architecture to build aPermutation Index used for approximate similarity search on databases of different data nature and to solve many queries at the same time. Besides, we evaluate the tradeoff between the answer quality and time performance of our implementation.