Further Optimizations and Analysis of Smith-Waterman with Vector Extensions

Sequence alignment based on dynamic programming, e.g., Smith-Waterman (SW), remains central to bioinformatics both in various standalone scenarios and as a function in other critical bioinformatics tasks. Much work has been done in optimizing SW for various accelerators; CPU-based platforms, however, are possibly still the most accessible computational resource for bioinformatics and our focus here. CPU acceleration necessarily involves applying vector extensions; a recent CPU/SW study resulted in the creation of an extensive autotuning framework. We add to this corpus a number of contributions: (i) a new optimization based on a novel mapping of the scoring matrix together with interleaving data coming from the substitution matrix; (ii) an implementation that combines disjoint optimizations from previous studies; (iii) a variable (8/16) bit width implementation; (iv) a finding of the benefits of learned compiler hyperparameters; (v) performance comparisons with respect to different usage scenarios; and (vi) a results showing that, despite appearing memory bound, multicore SW remains CPU bound. Among the most significant findings are that for protein alignments these approaches result in performance superior to that of the Parasail library's approaches, with the added benefit of determinism and robustness. Versus the deterministic Parasail+diag the improvement is 3.9×; versus the non-deterministic Parasail+scan and Parasail+striped, it is 1.9 × and 1.5 ×, respectively.