A Performance Analysis of High-Level Synthesis for FPGAs

This repository stores the code and latex of my bachelor's thesis. The full thesis PDF can be found in hls_performance_thesis.pdf.

Abstract

It is becoming increasingly popular to use Field-Programmable Gate Arrays (FPGAs) as hardware accelerators in order to speed up certain parts of an algorithm. FPGAs promise more energy efficiency and increased performance compared to CPUs. They are, however, traditionally programmed with Hardware Description Languages (HDLs), which are notoriously hard to use. To accelerate their adoption in different contexts and domains, FPGAs can nowadays also be programmed with high-level languages, such as C/C++ or OpenCL, in a process called High-Level Synthesis (HLS). However, it can be a challenge to efficiently accelerate algorithms using HLS. In this thesis, we investigate the performance of HLS for a non-trivial case-study. To this end, we devise a performance comparison between a sequential CPU algorithm and its FPGA version, programmed with OpenCL for the Xilinx Vitis platform. Our case study is a string-searching algorithm using the FM-index method, which is able to efficiently locate substrings in arbitrarily long texts. Our naive reference implementation is a simple port from CPU to OpenCL. In search for more performance, on top of this naive OpenCL implementation, we also propose and evaluate two FPGA-specific optimizations, suggested by literature. Our empirical analysis shows that optimizations can greatly improve the performance of algorithms programmed with HLS for FPGAs. However, the performance of our HLS-based FPGA version for our case study could not match the performance of the CPU, despite being more work efficient.