Compiling SA-C programs to FPGAs: Performance results

Abstract

At the first ICVS, we presented SA-C (“sassy”), a singleassignment variant of the C programming language designed to exploit both coarse-grain and fine-grain parallelism in computer vision and image processing applications. This paper presents a new optimizing compiler that maps SA-C source code onto field programmable gate array (FPGA) configurations. The compiler allows programmers to exploit FPGAs as inexpensive and massively parallel processors by writing high-level source code rather than hardware-level circuit designs. We present several examples of simple image-based programs and the optimizations that are automatically applied to them during compilation, and compare their performance on FPGAs and Pentiums of similar ages. From this, we determine what types of applications benefit from current FPGA technology, and conclude with some speculations on the future development of FPGAs and their expanding role in computer vision systems.