high performance computing on graphics processing units: hgpu.org

Similarity search is a fundamental operation for applications that deal with multimedia data. For a query in a multimedia database it is meaningless to look for elements exactly equal to a given one as query. Instead, we need to measure the similarity (or dissimilarity) between the query object and each object of the database. The similarity search problem can be formally defined through the concept of metric space, which provides a formal framework that is independent of the application domain. In a metric database, the objects from a metric space can be stored and similarity queries about them can be efficiently answered. In general, the search efficiency is understood as minimizing the number of distance calculations required to answer them. Therefore, the goal is to preprocess the dataset by building an index, such that queries can be answered with as few distance computations as possible. However, with very large metric databases is not enough to preprocess the dataset by building an index, it is also necessary to speed up the queries by using high performance computing, as GPU. In this work we show an implementation of a pure GPU architecture to build the Pemutation Index, used for approximate similarity search on databases of different data nature. Our proposal is able to solve many queries at the same time.