Natural Language Based Power Domain Partitioning

摘要

The increased importance of power consumption as a design factor is now undeniable. Power aware design flows are increasingly targeting high abstraction levels (e.g. ESL), where optimization gains are bigger. The designers are thus required to define the power intent already at these levels. Here the major challenge is to perform power domain partitioning. However, this is a fully manual step based on reading and understanding the system specification, and it has to be performed before the Virtual Prototype (VP) is built. This paper presents an approach to aid architects in specifying power intent by suggesting coarse-grained power domain partitioning schemes, as the VP is built. The approach starts with structural and behavioral information being extracted from the system specification using Natural Language Processing (NLP) techniques. Then, a semantic network map is created which depicts the hierarchical structure and the abstract block level dependencies that can be used as a foundation for the VP. Finally, a partitioning scheme is derived from the application of an extendable set of analytic rules. Experimental results on an encoding system demonstrate the applicability and efficacy of the proposed approach.