Lossy Polynomial Datapath Synthesis

Abstract

The design of the compute elements of hardware, its datapath, plays a crucial role in determining the speed, area and power consumption of a device. The building blocks of datapath are polynomial in nature. Research into the implementation of adders and multipliers has a long history and developments in this area will continue. Despite such efficient building block implementations, correctly determining the necessary precision of each building block within a design is a challenge. It is typical that standard or uniform precisions are chosen, such as the IEEE floating point precisions. The hardware quality of the datapath is inextricably linked to the precisions of which it is composed. There is, however, another essential element that determines hardware quality, namely that of the accuracy of the components. If one were to implement each of the official IEEE rounding modes, significant differences in hardware quality would be found. But in the same fashion that standard precisions may be unnecessarily chosen, it is typical that components may be constructed to return one of these correctly rounded results, where in fact such accuracy is far from necessary. Unfortunately if a lesser accuracy is permissible then the techniques that exist to reduce hardware implementation cost by exploiting such freedom invariably produce an error with extremely difficult to determine properties.

This thesis addresses the problem of how to construct hardware to efficiently implement fixed and floating-point polynomials while exploiting a global error freedom. This is a form of lossy synthesis. The fixed-point contributions include resource minimisation when implementing mutually exclusive polynomials, the construction of minimal lossy components with guaranteed worst case error and a technique for efficient composition of such components. Contributions are also made to how a floating-point polynomial can be implemented with guaranteed relative error.