A Technique for FPGA Synthesis Driven by Automatic Source Code Analysis and Transformations

This paper presents a technique for automatic synthesis of high-performance FPGA-based computing machines from C language source code. It exploits data-parallelism present in source code, and its approach is based on hardware application of techniques for automatic loop transformations, mainly designed in the area of optimizing compilers for parallel and vector computers. Performance aspects are considered in early stage of design, before low-level synthesis process, through a transformation-intensive branch-and-bound approach, that searches design space exploring area-performance tradeoffs. Furthermore optimizations are applied at architectural level, thus achieving higher benefits with respect to gate-level optimizations, also by means of a library of hardware blocks implementing arithmetic and functional primitives. Application of the technique to partial and complete unrolling of a Successive Over-Relaxation code is presented, with results in terms of effectiveness of area-delay estimation, and speed-up for the generated circuit, ranging from 5 and 30 on a Virtex-E 2000-6 with respect to a Intel Pentium 3 1GHz.