Dataflow-Based, Cross-Platform Design Flow for DSP Applications

Dataflow methods have been widely explored over the years in the digital signal processing (DSP) domain to model, design, analyze, implement, and optimize DSP applications, such as applications in the areas of audio and video data stream processing, digital communications, and image processing. DSP-oriented dataflow methods provide formal techniques that facilitate software design, simulation, analysis, verification, instrumentation and optimization for exploring effective implementations on diverse target platforms. As the landscape of embedded platforms becomes increasingly diverse, a wide variety of different kinds of devices, including graphics processing units (GPUs), multicore programmable digital signal processors (PDSPs), and field programmable gate arrays (FPGAs), must be considered to thoroughly address the design space for a given application. In this chapter, we discuss design methodologies, based on the core functional dataflow (CFDF) model of computation, that help engineers to efficiently explore such diverse design spaces. In particular, we discuss a CFDF-based design flow and associated design methodology for efficient simulation and implementation of DSP applications. The design flow supports system formulation, simulation, validation, cross-platform software implementation, instrumentation, and system integration capabilities to derive optimized signal processing implementations on a variety of platforms. We provide a comprehensive specification of the design flow using the lightweight dataflow (LWDF) and targeted dataflow interchange format (TDIF) tools, and demonstrate it with case studies on CPU/GPU and multicore PDSP designs that are geared towards fast simulation, quick transition from simulation to the implementation, high performance implementation, and power-efficient acceleration, respectively.