Hardware acceleration of the pair HMM algorithm for DNA variant calling

With the advent of several accurate and sophisticated statistical algorithms and pipelines for DNA sequence analysis, it is becoming increasingly possible to translate raw sequencing data into biologically meaningful information for further clinical analysis and processing. However, given the large volume of the data involved, even modestly complex algorithms would require a prohibitively long time to complete. Hence it is urgent to explore non-conventional implementation platforms to accelerate genomics research. In this thesis, we present a Field-Programmable Gate Array (FPGA) accelerated implementation of the Pair Hidden Markov Model (Pair HMM) forward algorithm, the performance bottleneck in the HaplotypeCaller, a critical function in the popular Genome Analysis Toolkit (GATK) variant calling tool. We introduce the PE ring structure which, thanks to the fine-grained parallelism allowed by the FPGA, can be built into various configurations striking a trade-off between Instruction-Level Parallelism (ILP) and data parallelism. We investigate the resource utilization and performance of different configurations. Our solution can achieve a speed-up of up to 487x compared to the C++ baseline implementation on CPU and 1.56x compared to the previous best hardware implementation