Content area
The power of modern multi-core and many-core platforms is an excellent fit for meeting the performance needs of embedded software applications. However, there are many ways to map these applications to a specific multi-core or many-core architecture, leading to a large design space that requires extensive analysis. This analysis is needed to weigh the trade-offs of a set of design goals that need to be optimized. Therefore, there is a need for efficient design methods that can identify the Pareto-optimal application mappings while considering well-defined design constraints.
This thesis proposes approaches to address these challenges and integrate them into SYSTEMCODESIGNER’s Electronic System Level (ESL) design flow. The contributions presented in this thesis aim to facilitate the definition of dataflow applications by combining the model-based design of dataflow applications with block diagrams in MATLAB/Simulink to define the input application for the subsequent steps of the SYSTEMCODESIGNER methodology. Additionally, Design Space Exploration (DSE) methods were proposed to accelerate the exploration of mappings of dataflow-based applications to multi-core and many-core architectures. Compared to other DSE approaches, the methods proposed in this dissertation improve the quality of the found Pareto fronts.
Conversion of periodic discrete block diagrams into the data flow language SysteMoC
Chapter 3 presents a framework for automatically converting periodic discrete blockbased MATLAB/Simulink models into data-driven SystemC Models of Computation (SysteMoC) models. The main goal is to integrate MATLAB/Simulink as a front-end for the ESL design flow of SYSTEMCODESIGNER.
The proposed conversion method offers a significant advantage over other state-ofthe-art works by providing complete coverage of periodic discrete MATLAB/Simulink models that automatically generate executable program code for each generated actor without requiring an explicit library definition. This conversion method was employed to define input applications in Chapters 4 and 5 of this thesis, enabling the multiobjective DSE approaches to optimize the cost, throughput, and memory requirements of periodic discrete block diagrams. Furthermore, the proposed fully automated framework enhances productivity and reduces the risk of manual errors when coding dataflow-based actor networks. Finally, the proposed conversion method proved to be correct by reproducing the same sequence of values at each output port with the same sequence of values at each input port of the original block diagram and the corresponding SysteMoC actor network.