Notes on "Raising the Level of Abstraction: A Signal Processing System Design Course"

Prof. Brian L. Evans
  1. Title slide: This graduate class has been taught three times at UT Austin. It has roots in the Specification and Modeling of Reactive Real-Time Systems graduate class which was taught once at UC Berkeley in Fall 1996 by Prof. Edward Lee. (Seals should have been in color.)
  2. Introduction: The first three points address the issues in teaching signal processing systems. So much heterogeneity and complexity.
  3. Embedded Signal Processing Systems: Familiar graphical description of the software and hardware technologies in embedded sytems. This diagram could be a third-generation cell phone if you replace the microcontroller with a low-power general-purpose processor such as the ARM. TI TMS320CC6x in third-generation wireless systems
  4. Heterogenity in a System-Level Design Flow: This demonstrates the decoupling of system specification and implementation. The choice of specification model greatly impacts the candidate implementations. An imperative model restricts you to software (e.g. C) that has to be compiled. A discrete-event model of combinational logic leads to a hardware implementation. However, finite state machines and dataflow models are amenable to both software and hardware.
  5. System Modeling: In this slide, I make the connection between subsystems familiar to the audience with various model of computation. Composing models of computation yields complex systems.
  6. Specification Using Hierarchical Block Diagrams: An example of composing models of computation. This kind of composition could model a second-generation cell phone. The discrete-event model models events (such as call request) in continuous time. The FSM models the protocol when a call has been accepted (setup and speech transmission/reception modes). During reception, the compressed speech stream of data is decoded, which is modeled well as a dataflow model.
  7. Simulation and Sythesis: The key here is that models with finite state are preferred over models with infinite state in terms of simulation time and optimal scheduling time. The optimal scheduling for SDF is for scheduling onto a single processor. The problem is in general NP hard (not polynomial).
  8. Educational Objectives: Self-explanatory. Slide is mostly text.
  9. Conclusion: Self-explanatory. Slide is mostly text.

Last updated 06/18/99.