Lina: Timing-Constrained High-Level Synthesis Performance Estimator for Fast DSE

摘要

The adoption of Field-Programmable Gate Array (FPGA) for general use in the High-Performance Computing scenario has been limited by its complex development flow required to get optimised designs coupled with a time-consuming compilation. High-Level Synthesis (HLS) tools are adopted to improve programmability, however the developer must perform several iterations of optimisation schemes in order to achieve reasonable performance results, which is tedious and not trivial. Several works employ Design Space Exploration (DSE) through different optimisation possibilities, coupled with fast performance estimators to avoid the unacceptable compilation times. This paper presents Lina, an expansion of the Lin-Analyzer fast peformance estimator for C/C++ HLS including timing-constrained scheduling and an extended analysis for nested loops. Results over the PolyBench benchmark show that the average relative error dropped from 8.85% to 3.02% when loop unrolling and pipelining directives were considered. As a result Lina becomes a better estimator for non-perfect loop nests and for different timing constraints, which can be adopted as an additional design space exploration knob.