A Dataflow Framework For Developing Flexible Embedded Accelerators A Computer Vision Case Study.

The focus of this dissertation is the design and the implementation of a computing platform which can accelerate data processing in the embedded computation domain. We focus on a heterogeneous computing platform, whose hardware implementation can approach the power and area efficiency of specialized designs, while remaining flexible across the application domain. The multi-core architectures require parallel programming, which is widely-regarded as more challenging than sequential programming. Although shared memory parallel programs may be fairly easy to write (using OpenMP, for example), they are quite hard to optimize; providing embedded application developers with optimizing tools and programming frameworks is a challenge. The heterogeneous specialized elements make the problem even more difficult. Dataflow is a parallel computation model that relies exclusively on message passing, and that has some advantages over parallel programming tools in wide use today: simplicity, graphical representation, and determinism. Dataflow model is also a good match to streaming applications, such as audio, video and image processing, which operate on large sequences of data and are characterized by abundant parallelism and regular memory access patterns. Dataflow model of computation has gained acceptance in simulation and signal-processing communities. This thesis evaluates the applicability of the dataflow model for implementing domain-specific embedded accelerators for streaming applications.