Knowledge-based transformation ordering

Transformations have been widely used in VLSI design, high level synthesis and DSP. We propose a two-step approach for transformation ordering which combines the use of optimization-intensive CAD techniques with knowledge-based user-driven search strategy. The first step is the development of basic building blocks which target small sets of transformations which are well suited for optimization intensive CAD treatment. Next, transformation orderings are developed using knowledge about mathematical laws, an application domain, and the relationship among transformations. Transformation ordering scripts combine several building blocks to form effective approaches for optimization of several design metrics in many common computational structures. As the highlight of the approach, we developed a method which efficiently simultaneously optimizes the throughput, latency, power, and area of linear computations