Designing Efficient SIMD Kernels for High Performance Sequence Alignment.

The amount of genetic information being generated and stored has seen a surge in recent years due to current technological improvements. Processing this vast amount of data requires efficient genomic pipelines. Hence, key tasks like sequence alignment between two or more sequences are implement on modern hardware using the Smith-Waterman algorithm for local alignment and the Needleman-Wunsch algorithm for global alignment. However, most current solutions fall short because they offer rigid implementations that are tuned for specific platforms and cannot easily be ported from one hardware to another. In this work, we design a flexible implementation for the key alignment kernels and show that a systematic approach to optimizing both the local and global sequence algorithms can be achieved. In addition, we show that using our approach the high performance implementation of the computation using SIMD instructions can achieve close to peak performance on most modern CPU platforms (from vendors like Intel, AMD and ARM) and across a wide range of SIMD vector lengths.