An Optimisation Methodology for Mapping a Diffusion Algorithm for Vision into a Modular and Flexible Array Architecture

This chapter addresses the architecture mapping of a complex non-linear diffusion algorithm for use in vision applications. It shows the way through which manually applied transformations allow the final architecture to be modular and have a low communication bandwidth. The algorithm exhibits several parameters and has to be partly programmable. Real-time execution for a particular parameter set is, however, required. Starting point for this design is a set of affine recurrence equations. For this abstract specification, difficult task of finding a globally optimized architecture with fully matched throughput is tackled while avoiding an explosion of the search space. Real-time signal processing (RSP) applications as in video, speech, and image processing require a mapping methodology, which is optimized to throughput and not to latency. This demand contrasts with most other current array synthesis approaches that allow transforming regular algorithms from affine (ARE) or uniform recurrence equation (URE) form into efficient regular array architectures (RAA's). To allow pipelined processing elements (PEs) and multi-dimensional algorithms, changes are needed to the uniform mapping and projection approach. This paper discusses a manually applied formal design methodology resulting in a modular and programmable regular architecture implementing an image diffusion algorithm. It involves initial algorithm transformations, 1D or 2D placement mapping, multi-dimensional scheduling, local processing element (PE) design, and scheduling and low-level scheduling. All these steps are embedded in an iterative decision process with pruning due to ordering of the decisions and gradual refinement of the cost functions used to differentiate alternatives.