A Methodology and Tool for Automated Transformational High-Level Design Space Exploration

design loop, which covers a large number of design steps (called synthesis loop in figure 1). Therefore, a single cycle of the design loop is quite expensive to perform. Our approach to face this problem is to cut the design loop on a high level of abstraction by integrating a high-level estimation step. This results in a design loop which is tight and settled on high level of abstraction (called estimation loop in figure 1). For this, the evaluation of an optimization step can be done efficiently. Nevertheless, an exhaustive exploration of the design space leads to unacceptable high effort of the exploration process. This paper concentrates on algorithmic approaches on controlling the selection and application of transformations, which allows an effective exploration of the design space in an automated fashion. The paper is organized as follows: After a look at the state of the art in chapter 2, chapter 3 gives an overview on our design space exploration methodology. Chapter 4 presents some classical algorithmic search strategies adapted for transformation control, as well as our own modular approach. Chapter 5 summarizes the results of their experimental evaluation and application. In most of the approaches known from literature, which support (some) high-level transformations, selection and application of transformations is done manually by the Figure 1. Transformational design space exploration layout level level level level algorithmic register-transfer logic hardware estimation design loop synthesis design loop specification high-level synthesis high-level transformations layout synthesis logic synthesis high-level estimation Abstract Objective of the methodology presented in this paper is to perform design space exploration on a high level of abstraction by applying high-level transformations. The paper concentrates on algorithmic approaches on controlling the iterative process of transformation selection. A novel modular algorithm for transformation control is presented and its effectiveness is experimentally validated. In combination with a large set of transformation algorithms and mechanisms for high-level estimation of transformation quality, there results a methodology for automated high-level design space exploration. All the concepts are summarized in a software tool called ExTra (Design Space Exploration Using Transformations). Finally, the results of the application of ExTra to the JPEG encoding algorithm are presented.