Rajeev Jain
Associate Professor, EE Dept., UCLA
Chief Scientist, Angeles Design Systems
Etan Cohen
VP Engineering,
Angeles Design Systems
November 7, 1996
Wang Room, 535 Cory Hall
4:00-5:00 p.m.
CAD tools for DSP and communications have focussed on simulating systems using different scheduling paradigms - data-flow and cycle-driven. The data-flow paradigm allows easy system specification, especially for multi-rate systems, while cycle-driven description allow specification of non data-flow (control) parts of the system. As lower level network control protocols are integrated with communications functions in single chips the need for simulating control and data-flow together is becoming critical. However, in contrast to classical multi-domain simulation (or co-simulation) efforts, actual system design experience shows that the real need is not combining simulation, but exchanging the analysis results from different simulations. Capturing the analysis data from iterative simulation in one domain and providing that in encapsulated form to other domains is a more important requirement that is not currently addressed.
A second area that is not supported by CAD tools is system optimization. In contrast to logic design and simulation where substantial research effort in generic optimization techniques (espresso, mis) has led to commercial tools such as design compiler, there is little effort in optimization techniques for DSP and communication systems. The efforts have traditionally been limited to specific DSP functions such as filter optimization, which are important but cover a very narrow area. Much worse when a function such as a filter is embedded in a larger system, such as a receiver, the optimization constraints apply to th system, not just the filter and the filter optimization tools become useless unless the designer can define the filter design constraints as a function of the system design constraints. Multi-variate optimization techniques using generic techniques such as binary search and centroid algorithm can actually be applied to this system optimization problem quite successfully as will be demonstrated with a simple example. Thus multi-domain analysis and system optimization are two areas of current interest in system design.
Implementation tools for DSP and communication have largely focussed on HDL generation (VHDL or Verilog) and code compilation. If the system design is not optimized within implementation constraints and performance requirements, then much of this optimization has to be done after the HDL or firmware code generation. This negates the productivity gains of the code generation. Consequently, really aggressive design are still implemented manually. A possible solution to this problem is creating system simulation, analysis and optimization techniques which permit appropriate abstraction of implementation constraints and optimization criteria at the system design level. This should also be a focus of the system optimization tools mentioned above to evolve a successful implementation strategy - rather than just providing code generation and ocde compilation tools.
This talk will provide an overview of the above problems and is intended to stimulate discussion on solutions rather than present any "ideal" solutions.