Fixed-point Computation Infrastructure for Ptolemy

Researchers: Dr. Seehyun Kim---Visiting Scholar, LG Electronics Research Center
Advisor:Edward A. Lee

In terms of power consumption, hardware cost, and execution speed, fixed-point arithmetic is often preferable over floating point for embedded signal processing systems. Algorithms, however, are usually developed using an idealized arithmetic model, usually approximated by double-precision floating-point arithmetic on standard workstations. In practice, one can think of the idealized model as more abstract, and the fixed-point model as more concrete, closer to the physical world. Our approach is to make concrete the an abstract specification in the final phases of the ``algorithm-to-implementation'' design process. In more adaptive applications, fixed-point properties may be varied in the field, at run time, for example to conserve power.

In this project, an infrastructure of the fixed-point computation will be constructed for Ptolemy. Functional fixed-point blocks will be synthesized automatically from the corresponding floating-point blocks. Simulation and analysis will then be combined for performance assessment. Self-tuning systems, which dynamically change their own parameters to meet varying operational criteria, will also be explored. For example, the wordlength of a signal or the number of taps of a filter will be adjusted on-the-fly in order to adapt to a changing criterion, such as power dissipation or the signal-to-noise ratio.