Maximizing the throughput of high performance DSP applications using behavioral transformations

Meeting the stringent throughput requirements of high performance DSP applications is a challenging task. Extensive optimization of the computational structure is essential to satisfy these constraints. This paper proposes a new transformational approach for performance optimization. This approach consists of an ordered set of transformations, including algebraic transformations, loop unrolling, and retiming/pipelining, aimed at speeding up both recursive and non-recursive, us well as linear and non-linear applications. Impressive and close to optimal speed-up's have been obtained for a large range of benchmark examples.<>