Developing high-performance signal processing applications requires not only effective signal processing algorithms but also efficient software design methods that can take full advantage of the available processing resources. An increasingly important type of hardware platform for high-performance signal processing is a multicore central processing unit (CPU) combined with a graphics processing unit (GPU) accelerator. Efficiently coordinating computations on both the host (CPU) and device (GPU), and managing host-device data transfers are critical to utilizing CPU-GPU platforms effectively. However, such coordination is challenging for system designers, given the complexity of modern signal processing applications and the stringent constraints under which they must operate.

Dataflow models of computation provide a useful framework for addressing this challenge. In such a modeling approach, signal processing applications are represented as directed graphs that can be viewed intuitively as high-level signal flow diagrams. The formal, high-level abstraction provided by dataflow principles provides a useful foundation to investigate model-based analysis and optimization for new challenges in design and implementation of signal processing systems.

This thesis presents a new model-based design methodology and an evolution of three novel design tools. These contributions provide an automated design flow for high performance signal processing. The design flow takes high-level dataflow representations as input and systematically derives optimized implementations on CPU-GPU platforms.

The proposed design flow and associated design methodology are inspired by a previously-developed application programming interface (API) called the Hybrid Task Graph Scheduler (HTGS). HTGS was developed for implementing scalable workflows for high-performance computing applications on compute nodes that have large numbers of processing cores, and that may be equipped with multiple GPUs. However, HTGS has a limitation due to its relatively loose use of dataflow techniques (or other forms of model-based design), which results in a significant designer effort being required to apply the provided APIs effectively.

The main contributions of the thesis are summarized as follows:

(1) Development of a companion tool to HTGS that is called the HTGS Model-based Engine (HMBE). HMBE introduces novel capabilities to automatically analyze application dataflow graphs and generate efficient schedules for these graphs through hybrid compile-time and runtime analysis. The systematic, model-based approaches provided by HMBE enable the automation of complex tasks that must be performed manually when using HTGS alone. We have demonstrated the effectiveness of HMBE and the associated model-based design methodology through extensive experiments involving two case studies: an image stitching application for large scale microscopy images, and a background subtraction application for multispectral video streams.

(2) Integration of HMBE with HTGS to develop a new design tool for the design and implementation of high-performance signal processing systems. This tool, called HMBE-Integrated-HTGS (HI-HTGS), provides novel capabilities for model-based system design, memory management, and scheduling targeted to multicore platforms. HMBE takes as input a single- or multi-dimensional dataflow model of the given signal processing application. The tool then expands the dataflow model into an expanded representation that exposes more parallelism and provides significantly more detail on the interactions between different application tasks (dataflow actors). This expanded representation is derived by HI-HTGS at compile-time and provided as input to the HI-HTGS runtime system. The runtime system in turn applies the expanded representation to guide dynamic scheduling decisions throughout system execution.

(3) Extension of HMBE to the class of CPU-GPU platforms motivated above. We call this new model-based design tool the CPU-GPU Model-Based Engine (CGMBE). CGMBE uses an unfolded dataflow graph representation of the application along with thread-pool-based executors, which are optimized for efficient operation on the targeted CPU-GPU platform. This approach automates complex aspects of the design and implementation process for signal processing system designers while maximizing the utilization of computational power, reducing the memory footprint for both the CPU and GPU, and facilitating experimentation for tuning performance-oriented designs.