The RASSP Digest - Vol. 3, September 1996

VHDL Hybrid Models

by Fred Rose

Abstract

Hybrid modeling is the capability of mixing high-level performance constructs and functional components in a common analysis environment. A coordinated research effort between the Honeywell Technology Center (HTC) and the Center for Semicustom Integrated Systems at the University of Virginia (UVa) is addressing issues related to hybrid modeling.

1. Introduction

The primary objective of hybrid modeling is to handle the complex task of translating data and control flow between models at different levels of abstraction and interpretation. The necessary information content at interface boundaries varies with the model abstraction level. As models become more detailed the amount of information in the interface increases. Differences in information content cannot be handled purely via translation (such as translating an integer representation of data into a bit vector). Rather, information might need to be synthesized or otherwise generated when interfacing to a more detailed model. When interfacing to a less detailed model, information might need to be discarded, encapsulated, or abstracted. This must all occur seamlessly within a single model context, where multilevel models interconnected with the hybrid interface components operate in a integrated fashion.

Within the RASSP methodology, hybrid modeling supports the spawning of mini-spirals to develop critical or high risk items. Based upon risk, pieces of the overall design may be at differing maturity levels. This is where the concept of hybrid modeling becomes important. As the overall system model is developed, certain high risk areas need further development. From a modeling perspective, this will be accomplished with more detailed models. Using modeling terminology, the system level uninterpreted (performance) model will have certain components replaced with the more detailed, interpreted (functional) models. To continue the RASSP philosophy of virtual prototypes, methods must enable the mixture of different model abstractions.

The hybrid modeling capability permits incorporation of detailed, functional or behavioral models into the overall system performance model. This allows the designer to explore and determine effects of the detailed design which would not otherwise be well characterized or understood. The RASSP concept of model/design reuse means that major parts of a new design will be constructed from well understood components, where these components may be processors, MCMs, or modules. Therefore a good characterization of these components should exist for use in uninterpreted modeling. However, characterizations for high risk or new components are usually initial approximations. As that new component design becomes more refined, these initial approximation may change. That change could have a ripple effect throughout the entire design. Hybrid modeling permits heterogeneous simulation and analysis to occur within the VHDL simulation environment to establish good estimates of behavior early in the design cycle.

The hybrid model allows validation of certain assumptions and dependencies. While a detailed model of the critical portion might run stand-alone, and the timing results back-annotated into the performance model, this may not work in all cases. For instance, many classes of contention, such as memory, network, and processor scheduling and overhead, might only be resolved with the majority of the system modeled, providing accurate workload characterizations. Hybrid models permit this early in the design cycle, without requiring the entire system be modeled at an equivalently detailed level.

2. Approach

The approach to the hybrid modeling problem has been to define a taxonomy for hybrid models, develop a common structure for hybrid architectures, and develop a set of generic VHDL architectures and library elements to support a wide range of hybrid modeling applications.

The classes of hybrid modeling are defined by those model attributes which fundamentally alter the development and implementation of the hybrid interface. The hybrid modeling space is partitioned according to the following characteristics:

  1. The hybrid model objective
  2. The timing and synchronization mechanism of the model
  3. The nature of the interpreted element
  4. The data transformation mode
  5. The data type of the interpreted signals

The details of the hybrid taxonomy are contained in [1].

Honeywell has developed a library of hybrid architectures to enhance our existing VHDL Performance Modeling Library (PML) [2]. The commercialization of the PML is discussed in an accompanying article.

The typical hybrid architecture is shown in Figure 1. As with the PML, an attempt was made to design the architectures as generic as possible. Unfortunately this is more difficult for hybrid architectures than for performance architectures. The functional, or interpreted, component interface can be significantly different from one model to the next. Performance models, on the other hand, typically use a similar signal structure.

The approach of developing hybrid architectures for existing PML components was chosen for several reasons. The insertion of functional model components into a performance model is a likely methodology step under the RASSP methodology. Also the existing PML gives a solid basis for developing an initial library of hybrid architectures for validation purposes. HTC is concentrating on the implementation aspects of hybrid models while UVa is focusing on some specific research issues [1] such as models with sequential interpreted elements with known and unknown inputs, and adding hybrid capability to the ADEPT environment. Additionally, UVa has developed a set of components to allow the integration of the UVa ADEPT models and HTC PML models within one simulation environment.

3. Results

Releases of the Hybrid Modeling Library (HML) have been made in December 1995, and May 1996. The HML is released in conjunction with the PML. The HML contains hybrid architectures for the following PML components: indevice, outdevice, iodevice, pipeline, processor, and a communication element. The HML also contains the PML/ADEPT interface components. Examples of interpreted components are also contained in the HML to provide examples of complete hybrid architectures. The May release of the library also contains a hybrid interface generation toolkit. The purpose is to make the development of hybrid architectures as easy as possible for the user. Additionally the HML will be integrated with the commercial PMW and other capabilities of the Omniview tools are being investigated with respect to hybrid architecture generation.

Verification of the HML against real world design examples is ongoing. Results of this work will be available at a later date.

The PML, and associated HML, are available to RASSP contractors through Omniview. They will both be available through the commercial Performance Modeling Workbench.

For more information about the PML and ADEPT, reference the HTC (http://www.htc.honeywell.com/projects/rassp/) and UVa (http://csis.eee.virginia.edu/rassp/adept.html) Web pages, which are available off the E&F Web page (http://rassp.scra.org/).

References

[1] Meyassed, M., et al, “A Framework for the Development of Hybrid Models,” RASSP Conference Proceedings, Washington, D.C., July, 1995.

[2] Steeves, T., F. Rose, T. Carpenter, J. Shackleton, O.von der Hoff, “Evaluating Distributed Multiprocessor Designs,’’ RASSP Conference Proceedings, Washington, D.C., July, 1995.

Fred Rose
Honeywell Technology Center
3660 Technology Drive
Minneapolis, MN 55418
rose_fred@htc.honeywell.com

Newsletter Index
The RASSP Digest - Vol. 3, September 1996
newsletter/html/96sep/news_15.html