Signal processing applications usually encounter multi-dimensional real-time performance requirements and restrictions on resources, which makes software implementation complex. Although major advances have been made in embedded processor technology for this application domain -- in particular, in technology for programmable digital signal processors -- traditional compiler techniques applied to such platforms do not generate machine code of desired quality. As a result, low-level, human-driven fine tuning of software implementations is needed, and we are therefore in need of more effective strategies for software implementation for signal processing applications.

In this thesis, a number of important memory and performance optimization problems are addressed for translating high-level representations of signal processing applications into embedded software implementations. This investigation centers around signal processing-oriented dataflow models of computation. This form of dataflow provides a coarse grained modeling approach that is well-suited to the signal processing domain and is increasingly supported by commercial and research-oriented tools for design and implementation of signal processing systems.

Well-developed dataflow models of signal processing systems expose high-level application structure that can be used by designers and design tools to guide optimization of hardware and software implementations. This thesis advances the suite of techniques available for optimization of software implementations that are derived from the application structure exposed from dataflow representations. In addition, the specialized architecture of programmable digital signal processors is considered jointly with dataflow-based analysis to streamline the optimization process for this important family of embedded processors. The specialized features of programmable digital signal processors that are addressed in this thesis include parallel memory banks to facilitate data parallelism, and signal-processing-oriented addressing modes and address register management capabilities.

The problems addressed in this thesis involve several inter-related features, and therefore an integrated approach is required to solve them effectively. This thesis proposes such an integrated approach, and develops the approach through formal problem formulations, in-depth theoretical analysis, and extensive experimentation.