High-level synthesis of a genomic database search engine

Genomic database search is an I/O-bound problem, so avoiding unnecessary I/O transactions is a key consideration for improving search throughput. Many approximate search tools such as NCBI BLAST perform a database scan for each query, lacking a mechanism to avoid access to portions of the database that offer no potential for a match. In this paper we present an approach for using an FPGA-based pattern filter to convert each search query into a set of potential database matches that reduces the average portion of the database accessed per query. The approach is based on a hardware design for a pattern filter that can achieve a sustained recognition rate of one pattern per cycle. We used Vivado HLS to design the filter. Despite the presence of loop-carried dependencies, the final design meets the maximum possible throughout as constrained by the code's arithmetic intensity and available memory bandwidth. In this paper we describe the filter implementation and our code tuning methodology.