Representation and Scheduling of Scalable Dataflow Graph Topologies

Abstract

In dataflow-based application models, the underlying graph representations often consist of smaller sub-structures that repeat multiple times. In order to enable concise and scalable specification of digital signal processing (DSP) systems, a graphical modeling construct called "topological pattern" has been introduced in recent work. In this thesis, we present new design capabilities for specifying and working with topological patterns in the dataflow interchange format (DIF) framework, which is a software tool for model-based design and implementation of signal processing systems. We also present a plug-in to the DIF framework for deriving parameterized schedules, and a code generation module for generating code that implements these schedules. A novel schedule model called the scalable schedule tree (SST) is formulated. The SST model represents an important class of parameterized schedule structures in a form that is intuitive for representation, efficient for code generation, and flexible to support powerful forms of adaptation. We demonstrate our methods for topological pattern representation, SST derivation, and associated dataflow graph code generation using a case study centered around an image registration application.