A Fast and Accurate Cost Model for FPGA Design Space Exploration in HPC Applications

Heterogeneous High-Performance Computing (HPC) platforms present a significant programming challenge, especially because the key users of HPC resources are scientists, not parallel programmers. We contend that compiler technology has to evolve to automatically create the best program variant by transforming a given original program. We have developed a novel methodology based on type transformations for generating correct-by-construction design variants, and an associated light-weight cost model for evaluating these variants for implementation on FPGAs. In this paper we present a key enabler of our approach, the cost model. We discuss how we are able to quickly derive accurate estimates of performance and resource-utilization from the design's representation in our intermediate language. We show results confirming the accuracy of our cost model by testing it on three different scientific kernels. We conclude with a case-study that compares a solution generated by our framework with one from a conventional high-level synthesis tool, showing better performance and power-efficiency using our cost model based approach.