Dataflow Integration and Simulation Techniques for DSP System Design Tools

Thumbnail Image


umi-umd-4396.pdf (2.26 MB)
No. of downloads: 1407

Publication or External Link






System-level modeling, simulation, and synthesis using dataflow models of computation are widespread in electronic design automation (EDA) tools for digital signal processing (DSP) systems. Over the past few decades, various dataflow models and techniques have been developed for different DSP application domains; and many system design tools incorporate dataflow semantics for different objectives in the design process. In addition, a variety of digital signal processors and other types of embedded processors have been evolving continuously; and many off-the-shelf DSP libraries are optimized for specific processor architectures.

To explore their heterogeneous capabilities, we develop a novel framework that centers around the dataflow interchange format (DIF) for helping DSP system designers to integrate the diversity of dataflow models, techniques, design tools, DSP libraries, and embedded processing platforms. The dataflow interchange format is designed as a standard language for specifying DSP-oriented dataflow graphs, and the DIF framework is developed to achieve the following unique combination of objectives: 1) developing dataflow models and techniques to explore the complex design space for embedded DSP systems; 2) porting DSP designs across various tools, libraries, and embedded processing platforms; and 3) synthesizing software implementations from high-level dataflow-based program specifications.

System simulation using synchronous dataflow (SDF) is widely adopted in design tools for many years. However, for modern communication and signal processing systems, their SDF representations often consist of large-scale, complex topology, and heavily multirate behavior that challenge simulation -- simulating such systems using conventional SDF scheduling techniques generally leads to unacceptable simulation time and memory requirements. In this thesis, we develop a simulation-oriented scheduler (SOS) for efficient, joint minimization of scheduling time and memory requirements in conventional single-processor environments.

Nowadays, multi-core processors that provide on-chip, thread-level parallelism are increasingly popular for the potential in high performance. However, current simulation tools gain only minimal performance improvements due to their sequential SDF execution semantics. Motivated by the trend towards multi-core processors, we develop a novel multithreaded simulation scheduler (MSS) to pursue simulation runtime speed-up through multithreaded execution of SDF graphs on multi-core processors. Our results from SOS and MSS demonstrate large improvements in simulating real-world wireless communication systems.