[Next] [Previous] [Up] [Top]

1.0 Introduction

1.2 Code Generation Domains


A Domain in Ptolemy consists of a set of Blocks and Targets, and associated Schedulers that conform to a common computational model. By "computational model" we mean the operational semantics governing how Blocks interact with one another. Furthermore, all Blocks and Targets of a code generation Domain target the same language; for example, Blocks that generate code for the Motorola 56000 using the SDF model of computation form their own domain*1. A Scheduler will exploit knowledge of these semantics to order the execution of the Blocks. SDF and DDF are domains related to one another as illustrated in figure 3.

Stars and Targets are shown within each domain. The inner Domain (SDF) in figure 3 is an illustration of a sub-Domain, which implements a more specialized model of computation than the outer Domain (DDF). Hence all its Stars and Targets can also be used with the outer Domain. Schedulers can be associated with more than one Domain, but a Scheduler for a sub-Domain is not necessarily valid within the outer Domain.

For code generation, Domains are further subdivided according to the language synthesized. Hence, an SDF domain synthesizing C code is a domain that we call CGC (code generation in C). An SDF domain synthesizing assembly code for the Motorola DSP56000 family is called the CG56 domain. We have also developed SDF domains that synthesize assembly code for the Motorola DSP96000 family (CG96) and the Sproc multiprocessor DSP from Star Semiconductor. Finally, a Silage code generation domain is being used to couple to hardware synthesis tools developed at Berkeley [2].

As a simple example of how Blocks, Schedulers, and Targets can be mixed and matched, consider a set of Blocks that generate assembly language code for Motorola DSP56000 family processors. We might choose to use any of several Targets; examples of Targets that have been implemented include one that runs the assembled code on a simulator on the workstation, one that describes an S-bus card with a single 56000 processor on a workstation, and one that describes a set of four interconnected processors on a single card. It is also possible to define targets that have not been built. In these cases the generated code runs on functional simulations of the processors in the Thor domain in Ptolemy [19]. Most targets have parameters that select what scheduler is to be used; we have several single- and multiple-processor Schedulers that use different algorithms for determining partitioning and order of execution of stars. These schedulers have no processor-specific information; they "ask" the Target to determine commun


Software Synthesis for Single-Processor DSP Systems Using Ptolemy - 04 SEP 94

[Next]