Multiple objective optimisation of data and control paths in a behavioural silicon compiler

The objective of this research was to implement an `intelligent' silicon compiler that provides the ability to automatically explore the design space and optimise a design, given as a behavioural description, with respect to multiple objectives. The objective has been met by the implementation of the MOODS Silicon Compiler. The user submits goals or objectives to the system which automatically finds near optimal solutions. As objectives may be conflicting, trade-offs between synthesis tasks are essential and consequently their simultaneous execution must occur. Tasks are decomposed into behaviour preserving transformations which, due to their completeness, can be applied in any sequence to a multi-level representation of the design. An accurate evaluation of the design is ensured by feeding up technology dependent information to a cost function. The cost function guides the simulated annealing algorithm in applying transformations to iteratively optimise the design. The simulated annealing algorithm provides an abstractness from the transformations and designer's objectives. This abstractness avoids the construction of tailored heuristics which pre-program trade-offs into a system. Pre-programmed trade-offs are used in most systems by assuming a particular shape to the trade-off curve and are inappropriate as trade-offs are technology dependent. The lack of pre-programmed trade-offs in the MOODS system allows it to adapt to changes in technology or library cells. The choice of cells and their subsequent sharing are based on the user's criteria expressed in the cost function, rather than being pre-programmed into the system. The results show that implementations created by MOODS are better than or equal to those achieved by other systems. Comparisons with other systems highlighted the importance of specifying all of a design's data as the lack of data misrepresents the design leading to misleading comparisons. The MOODS synthesis system includes an efficient method for automated design space exploration where a varied set of near optimal implementations can be produced from a single behavioural specification. Design space exploration is an important aspect of designing by high-level synthesis and in the development of synthesis systems. It allows the designer to obtain a perspicuous characterization of a design's design space allowing him to investigate alternative designs.