Computation Partitioning and Reuse for Power Efficient High Performance Digital Signal Processing

The increasingly high demand of multimedia and portable electronics has necessitated the development of power efficient signal processing techniques. In this paper, we propose "computation partitioning and reuse" as an effective method for low power, high performance signal processing architectures. We show that two frequently used signal processing applications, Discrete Cosine Transform (DCT) and Finite Impulse Response (FIR) filters can obtain extremely high performance while satisfying the lower power criteria (compared to existing popular implementations) by adopting this technique. Additionally, it is often observed that in DSP systems a minor degradation in output quality (such as streaming video, image, etc.) can be tolerated since it is related to human perception. We use this concept to design architectures where power requirements can be further optimized by performing tradeoffs of energy requirements with output quality. We propose two techniques for DCT—(1) preferential computation of critical outputs, and (2) bitwidth adaptation to provide large power savings while maintaining system frequency requirements.