A Method for Integrating Allocation and Binding in Modulo Scheduling using Complex Sub-circuits

• Applicant: Professor Dr.-Ing. Peter Zipf
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems
• Term: from 2019 to 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 420658696

Many current and future problems with a technical context, from control engineering to the use of neural networks, already require computing power that can only be supplied by specialized hardware systems. Their design is thus increasingly becoming a challenge. At present, for example in the automotive industry, a model-based design approach with corresponding tools is used, while on the other hand a long-standing development pursues the use of conventional programming languages; in both cases the design is to be made more abstract, simpler and faster from a behavioral description, and in addition flexibility is to be achieved through a certain independence from the target hardware. At present, however, there are still major deficits with regard to the optimization methods and algorithms used in both approaches, which very often prevent direct usability of the generated hardware models, and thus of the design process as a whole.The project presented here addresses three basic problems of the current design processes (toolflows): (1) The previously unfavourable restrictions of the design space due to resource allocation specifications which are not adequately adapted to the conditions of the target system are to be greatly reduced by means of an integrated cost estimation for suitable sub-circuits. A heuristic to be developed is intended to limit the growing runtime due to the associated modeling. (2) The great complexity of practice-relevant models has so far prevented efficient resource sharing and a design space exploration based on it. The model complexity is to be reduced by working with repeating sub-circuits. The problem of identifying suitable sub-circuits itself must be solved and integrated into the allocation process. (3) Due to the limited applicability of modulo scheduling to realistic problem sizes and its currently inadequate combination with allocation and binding, the potential of this method has not yet been fully exploited. A new ILP formulation, which integrates allocation and binding based on the selected sub-circuits, should bring an improvement here.Some of the applicant's preparatory work in these areas indicates that the approach of identifying and using suitable sub-circuits proposed here can lead to a significant improvement in the achievable results. Overall, an important step in design automation and its practical application can be established.

• DFG Programme: Research Grants