PairHMM Improvements for Modern Instruction Set Architectures

This paper focuses on optimization of the Genome Analysis Toolkit (GATK) open-source software using advanced instruction set architectures (ISAs), such as AVX-512 and ARM SVE2, to improve its performance and efficiency. The optimizations include the pairHMM algorithm for aligning reads to the reference genome, sliding antidiagonal compute, memoization, and mixed-precision calculations. Experiments conducted on chromosome 21 of the NA12878 dataset reveal that memory band-width has the most significant effect on run time, with mixed-precision approaches showing promising results. The proposed optimizations can extend the life of legacy hardware or boost the transition to new architectures while providing up to 30% speedup over the AVX2 baseline.