Investigating Performance Losses in High-Level Synthesis for Stencil Computations

With the aid of few directives and canonical forms, high-level synthesis (HLS) tools allow FPGA developers to describe their hardware designs in higher-level languages such as C or C++, thus enabling software engineers to exploit the powerful capabilities of FPGA-based accelerators. Hardware engineers may also benefit from automated RTL generation through HLS, as it can boost their productivity with shorter development cycles and simpler validation processes. However, the introduction of automation with its associated performance losses brings the viability of the current HLS-based approach into question: are we sacrificing the main advantage of FPGA designs, namely their performance, in return for higher productivity? This paper examines the performance-per-power penalties incurred in HLS designs in the context of stencil computations for fluid flow simulations, a prevalent class of applications that are difficult to accelerate because of their low arithmetic intensity. By using Xilinx's Vivado HLS tool to replicate a hand-crafted RTL implementation of a solution of Laplace's equation, this paper evaluates the impact of implicitly expressing parallelism, particularly if specific optimizations are not directly supported by the tool. In addition, by describing scenarios in which the HLS approach does and does not excel for stencil-based computations, this paper offers insights to assist hardware engineers in setting their expectations of the HLS approach and suggests alternative techniques to accomplish common tasks that fail or underperform using typical approaches.