Modeling and Mapping of Optimized Schedules for Embedded Signal Processing Systems

Thumbnail Image


Publication or External Link






The demand for Digital Signal Processing (DSP) in embedded systems has been increasing rapidly due to the proliferation of multimedia- and communication-intensive devices such as pervasive tablets and smart phones. Efficient implementation of embedded DSP systems requires integration of diverse hardware and software components, as well as dynamic workload distribution across heterogeneous computational resources. The former implies increased complexity of application modeling and analysis, but also brings enhanced potential for achieving improved energy consumption, cost or performance. The latter results from the increased use of dynamic behavior in embedded DSP applications. Furthermore, parallel programming is highly relevant in many embedded DSP areas due to the development and use of Multiprocessor System-On-Chip (MPSoC) technology. The need for efficient cooperation among different devices supporting diverse parallel embedded computations motivates high-level modeling that expresses dynamic signal processing behaviors and supports efficient task scheduling and hardware mapping.

Starting with dynamic modeling, this thesis develops a systematic design methodology that supports functional simulation and hardware mapping of dynamic reconfiguration based on Parameterized Synchronous Dataflow (PSDF) graphs. By building on the DIF (Dataflow Interchange Format), which is a design language and associated software package for developing and experimenting with dataflow-based design techniques for signal processing systems, we have developed a novel tool for functional simulation of PSDF specifications. This simulation tool allows designers to model applications in PSDF and simulate their functionality, including use of the dynamic parameter reconfiguration capabilities offered by PSDF. With the help of this simulation tool, our design methodology

helps to map PSDF specifications into efficient implementations on field programmable gate arrays (FPGAs). Furthermore, valid schedules can be derived from the PSDF models at runtime to adapt hardware configurations based on changing data characteristics or

operational requirements. Under certain conditions, efficient quasi-static schedules can be applied to reduce overhead and enhance predictability in the scheduling process.

Motivated by the fact that scheduling is critical to performance and to efficient use of dynamic reconfiguration, we have focused on a methodology for schedule design, which complements the emphasis on automated schedule construction in the existing literature on dataflow-based design and implementation. In particular, we have proposed a dataflow-based schedule design framework called the dataflow schedule graph (DSG), which provides a graphical framework for schedule construction based on dataflow semantics, and can also be used as an intermediate representation target for automated schedule generation. Our approach to applying the DSG in this thesis emphasizes schedule

construction as a design process rather than an outcome of the synthesis process. Our approach employs dataflow graphs for representing both application models and schedules that are derived from them. By providing a dataflow-integrated framework for unambiguously representing, analyzing, manipulating, and interchanging schedules, the DSG facilitates effective codesign of dataflow-based application models and schedules for execution of these models.

As multicore processors are deployed in an increasing variety of embedded image processing systems, effective utilization of resources such as multiprocessor systemon-chip (MPSoC) devices, and effective handling of implementation concerns such as

memory management and I/O become critical to developing efficient embedded implementations. However, the diversity and complexity of applications and architectures in embedded image processing systems make the mapping of applications onto MPSoCs difficult. We help to address this challenge through a structured design methodology that is built upon the DSG modeling framework. We refer to this methodology as the DEIPS methodology (DSG-based design and implementation of Embedded Image Processing

Systems). The DEIPS methodology provides a unified framework for joint consideration of DSG structures and the application graphs from which they are derived, which allows designers to integrate considerations of parallelization and resource constraints together

with the application modeling process. We demonstrate the DEIPS methodology through cases studies on practical embedded image processing systems.