FSP: a Framework for Data Stream Processing Applications targeting FPGAs

Abstract

FPGA architectures are becoming popular because of their high performance-to-energy ratio. Nonetheless, their effective exploitation is often counterbalanced by a high programming effort, since most of the modern hardware description languages provide only low-level programming abstractions. This paper proposes FSP, a framework to productively support the development of Data Stream Processing applications on CPU+FPGA System-on-Chip devices (SoCs). By exploiting a code generation approach starting from a high-level DSL in Python, FSP generates an efficient OpenCL skeleton implementation of the parallel pipeline on FPGA and the library to be used by host programs to transfer inputs and collect results to/from the FPGA program. The experimental results showcase the effectiveness of FSP on an SoC equipped with an Intel Arria 10 FPGA by running two streaming benchmark applications.