Abstract
This PhD dissertation proposes innovative methodologies to support the designing and the mapping process of embedded systems. Due to the increasing requirements, embedded systems have become quite complex, as they consist of several partially dependent heterogeneous components. Systematic Design Space Exploration (DSE) methodologies are required to support the near-optimal design of embedded systems within the available short time-to-market. In this target domain, the existing DSE approaches either require too much exploration time to find near-optimal designs due to the high number of parameters and the correlations between the parameters of the target domain, or they end up with a less efficient trade-off result in order to find a design within acceptable time. In this dissertation we present an alternative DSE methodology, which is based on systematic creation of scalable and near-optimal DSE frameworks. The frameworks describe all the available options of the exploration space in a ...
This PhD dissertation proposes innovative methodologies to support the designing and the mapping process of embedded systems. Due to the increasing requirements, embedded systems have become quite complex, as they consist of several partially dependent heterogeneous components. Systematic Design Space Exploration (DSE) methodologies are required to support the near-optimal design of embedded systems within the available short time-to-market. In this target domain, the existing DSE approaches either require too much exploration time to find near-optimal designs due to the high number of parameters and the correlations between the parameters of the target domain, or they end up with a less efficient trade-off result in order to find a design within acceptable time. In this dissertation we present an alternative DSE methodology, which is based on systematic creation of scalable and near-optimal DSE frameworks. The frameworks describe all the available options of the exploration space in a finite set of classes. A set of principles is presented which is used in the reusable DSE methodology to create a scalable and near-optimal framework and to efficiently use it to derive scalable and near-optimal design solutions within a Pareto trade-off space. The DSE reusable methodology is applied to several stages of the embedded system design flow to derive scalable and near-optimal methodologies. The first part of the dissertation is dedicated to the development of mapping methodologies for storing large embedded system data arrays in the lower layers of the on-chip background data memory hierarchy, and the second part to the DSE methodologies for the processing part of SW/HW architectures in embedded systems including the foreground memory systems. Existing mapping approaches for the background memory part are either enumerative, symbolic/polyhedral and worst case (heuristics) approximations. The enumerative approaches require too much exploration time, the worst case approximation lead to overestimation of the storage requirements, whereas the symbolic/polytope approaches are scalable and near-optimal for solid and regular iteration spaces. By applying the new reusable DSE methodology, we have developed an intra-signal in-place optimization methodology which is scalable and near-optimal for highly irregular access schemes. Scalable and near-optimal solutions for the different cases of the proposed methodology have been developed for the cases of non-overlapping and overlapping store and load access schemes. To support the proposed methodology, a new representation of the array access schemes, which is appropriate to express the irregular shapes in a scalable and near-optimal way, is presented. A general pattern formulation has been proposed which describes the access scheme in a compact and repetitive way. Pattern operations were developed to combine the patterns in a scalable and near-optimal way under all the potential pattern combination cases, which may exist in the application under study. In the processing oriented part of the dissertation, a DSE methodology is developed for mapping instance of a predefined target application domain onto a partially fixed architecture platform template, which consists of one processor core and several custom hardware accelerators. The DSE methodology consists of uni-directional steps, which are implemented through parametric templates and are applied without costly design iterations. The proposed DSE methodology explores the space by instantiating the steps and propagating design constraints which prune design options following the steps ordering. The result is a final Pareto trade-off curve with the most relevant near-optimal designs. As the scheduling and the assignment are the major tasks of both the foreground and the data path, near-optimal and scalable techniques are required to support the parametric templates of the proposed DSE methodology. A framework which describes the scheduling and assignment of the scalars into the registers and the scheduling and assignment of the operation into the function units of the data path is developed. Based on the framework, a systematic methodology to arrive at parametric templates for scheduling and assignment techniques which satisfy the target domain constraints is developed In this way, a scalable parametric template for scheduling and assignment tasks is created, which guarantees near-optimality for the domain under study. The developed template can be used in the Foreground Memory Management step and Data-path mapping step of the overall design flow. For the DSE of the domain under study, near-optimal results are hence achieved through a truly scalable technique.
show more
