Fast Parallel Suffix Array on the GPU

Leyuan Wang, Sean Baxter, John D. Owens

Abstract: We implement two classes of suffix array construction algorithms on the GPU. The first, skew, makes algorithmic improvements to the previous work of Deo and Keely to achieve a speedup of 1.45 \(\times \) over their work. The second, a hybrid skew and prefix-doubling implementation, is the first of its kind on the GPU and achieves a speedup of 2.3–4.4 \(\times \) over Osipov’s prefix-doubling and 2.4–7.9 \(\times \) over our skew implementation on large datasets. Our implementations rely on two efficient parallel primitives, a merge and a segmented sort. We also demonstrate the effectiveness of our implementations in a Burrows-Wheeler transform and a parallel FM index for pattern searching.