Literature Study

Abstract

Over the last decade, a vast amount of research has taken place in the field of high level synthesis. Many of the research results that have been obtained are in principle applicable to the synthesis of high throughput signal processing applications. Still, many others are oriented mostly to other target application domains like microprocessor design or control-dominated synthesis. In this chapter, an overview of the relevant work will be given. The starting point for all synthesis approaches in our target domain is a flow graph which can contain loops and conditions and may exhibit looping precedences. According to the traditional high level synthesis school, the primitive operations of the flow graph are mapped onto primitive building blocks. This approach and its drawbacks will be discussed in Section 4.1. In more recent work, approaches in which an additional level of functional hierarchy is created, either in the flow graph, in the data path (the so called accelerator units) or in both have emerged. Systems which adopt this approach are CATHEDRAL-3 [Note 91b], PHIDEO [Lippens 91], HYPER [Chu 92] and the system from U.C. Irvine [Rao 92, Rundensteiner 93]. They will be discussed in detail in Section 4.2. Also the early approach in FACE to share mutually exclusive graphs in conditions [Hwang 88] has introduced already some form of operator merging but this is only done in a very limited scope. Note that hierarchy of functions and operations has been exploited also earlier already in a scheduling context (e.g. [Potkonjak 89]) but that is only part of the story.