A specification invariant technique for operation cost minimisation in flow-graphs

In high-level synthesis, optimising area, time, and power in real-time applications are the prime objectives. A new model and technique are proposed, which minimise a weighted operation cost function for data-paths at an early stage in the synthesis process. The main target domain consists of lowly-multiplexed and hard-wired implementations of real-time DSP applications. The behavioral specification of an application is translated into a signal flow-graph, the operation cost (area/power) of which is minimised using algebraic transformations. A minimal set of elementary transformations are combined in a flexible way into composite transformations, which are used in an ascent/steepest descent search algorithm. Experiments show that this approach achieves optimal or close to optimal results with very few transformations. The technique is invariant to changes in the specification, as long as these retain the bit-true input/output behaviour