Improving Parallel Processing of Matrix-Based Similarity Measures on Modern GPUs

Dynamic programming techniques are well-established and employed by various practical algorithms which are used as similarity measures, for instance the edit-distance algorithm or the dynamic time warping algorithm. These algorithms usually operate in iteration-based fashion where new values are computed from values of the previous iteration, thus they cannot be processed by simple data-parallel approaches. In this paper, we propose a way how to utilize computational power of massively parallel GPUs to compute dynamic programming algorithms effectively and efficiently. We address both the problem of computing one distance on large inputs concurrently and the problem of computing large number of distances simultaneously (e.g., when a similarity query is being resolved).