Template-based hardware-software codesign for high-performance embedded numerical accelerators

Abstract

Sophisticated algorithms for control, state estimation and equalization have tremendous potential to improve performance and create new capabilities in embedded and mobile systems. Traditional implementation approaches are not well suited for porting these algorithmic solutions into practical implementations within embedded system constraints. Most of the technical challenges arise from design approach that manipulates only one level in the design stack, thus being forced to conform to constraints imposed by other levels without question. In tightly constrained environments, like embedded and mobile systems, such approaches have a hard time efficiently delivering and delivering efficiency. In this work we offer a solution that cuts through all the design stack layers. We build flexible structures at the hardware, software and algorithm level, and approach the solution through design space exploration. To do this efficiently we use a template-based hardware-software development flow. The main incentive for template use is, as in software development, to relax the generality vs. efficiency/performance type tradeoffs that appear in solutions striving to achieve run-time flexibility. As a form of static polymorphism, templates typically incur very little performance overhead once the design is instantiated, thus offering the possibility to defer many design decisions until later stages when more is known about the overall system design. However, simply including templates into design flow is not sufficient to result in benefits greater than some level of code reuse. In our work we propose using templates as flexible interfaces between various levels in the design stack. As such, template parameters become the common language that designers at different levels of design hierarchy can use to succinctly express their assumptions and ideas. Thus, it is of great benefit if template parameters map directly and intuitively into models at every level. To showcase the approach we implement a numerical accelerator for embedded Model Predictive Control (MPC) algorithm. While most of this work and design flow are quite general, their full power is realized in search for good solutions to a specific problem. This is best understood in direct comparison with recent works on embedded and high-speed MPC implementations. The controllers we generate outperform published works by a handsome margin in both speed and power consumption, while taking very little time to generate.