Unifying wire and time scheduling for highlevel synthesis

Abstract

Wire scaling is becoming problematic for semi-global and global wires in sub 40nm ASIC/VLSI chips. We propose an algorithm that simultaneously minimizes the time T and wire-lengths W for highlevel synthesis from C to Verilog. The program is compiled to a graphs G of arithmetic/logical and memory operations. We developed a formal model wherein the optimal product W × T of a given G can be resolved. The proposed algorithm for general Gs works by recursively decomposing G to either grid-like or tree-like induced sub-graphs and combines their optimal solutions. We have implemented this algorithm in the LLVM compiler and obtained an HLS compiler that successfully minimizes both W × T of the resulting circuits.