Distributed high-level synthesis

Amarasinghe, V. Kosala I (2005) Distributed high-level synthesis. University of Southampton, Doctoral Thesis.

Record type: Thesis (Doctoral)

Abstract

High level synthesis describes the process by which a behavioural description of a system is translated automatically into a structural implementation that exhibits the desired behaviour. In general, many millions of alternative structural designs will exist for any nontrivial description, each exhibiting the desired behaviour, but possessing different electrical and physical attributes such as silicon area, data throughput time (latency) and power dissipation. The MOODS synthesis suite (Multiple Objective Optimisation in Data and control path Synthesis), as the name suggests, allows the designer to assert high level goals on the final design in terms of the attributes mentioned above.

Traditionally, the algorithms and data structures that form the core of a high-level synthesis system have been developed to run on single processor computers; they do not easily lend themselves to a scalable distributed (multi-processor) architecture, due to the complex interactions between the individual components. The principle purpose of this research has been to investigate and develop alternative approaches to synthesis (based around the simulated annealing approach) suitable for implementation on multi-processor machines.

The aim is to accelerate the synthesis process itself and to examine the ‘design space’ explored by a distributed synthesis system. Presented results illustrate an achieved increase in throughput, which is up to 10 times faster than when executed on a single processor. Results also demonstrate improvements in design quality of up to 17%.

This record has no associated files available for download.