Forward-looking objective functions: concept and applications in high level synthesis

The effectiveness of traditional CAD optimization algorithms is proportional to the accuracy of the targeted objective functions. However, behavioral synthesis tools are not used in isolation; they form a strongly connected design flow where each tool optimizes its own objective function without considering the consequences on the optimization goals of the subsequently applied tools. Therefore, efforts to optimize one aspect of a design often have unforeseen negative impacts on other phases of the design process. Our objective is to establish a systematic way of developing and validating new types of objective functions that consider the effects on subsequently applied synthesis steps. We demonstrate the generic forward-looking objective function (FLOF) strategy on three main steps in behavioral synthesis: (i) transformation, (ii) scheduling, and (iii) register assignment. We show how the FLOF can be used in the first two phases to reduce the total number of registers required in the third phase.