ITR/NGS-Intelligent HW/SW Compilers for DSP Applications

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

We propose to develop a system to automate the design, implementation, and optimization of digital signal processing (DSP) applications for software (SW), hardware (HW), and SW/HW platforms with respect to multiple, possibly conflicting, performance metrics including speed, accuracy, area, and power. Ours is a new system of intelligent design and optimization tools that incorporate domainspecific knowledge of the mathematical structure of DSP computations and use this knowledge for automatic optimization and exploration of the design space, controlled by search and feedback optimization techniques. By including the 'high' mathematical level into the design automation, we achieve the following two main goals. The project will (1) overcome the known limitations of current optimizing compilers and synthesis tools that do not have access to high level structural information and are thus restricted to mechanical source code level manipulations, and will provide the framework to conquer for automation that step in the design flow that has traditionally been reserved to human ingenuity: intelligent algorithm manipulation and implementation to optimize desired performance metrics under given constraints.