Precision analysis for hardware acceleration of numerical algorithms

• Main Author: Boland, David Peter
• Other Authors: Constantinides, George
• Published: Imperial College London 2012
• Subjects: 004
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.544305

The precision used in an algorithm affects the error and performance of individual computations, the memory usage, and the potential parallelism for a fixed hardware budget. However, when migrating an algorithm onto hardware, the potential improvements that can be obtained by tuning the precision throughout an algorithm to meet a range or error specification are often overlooked; the major reason is that it is hard to choose a number system which can guarantee any such specification can be met. Instead, the problem is mitigated by opting to use IEEE standard double precision arithmetic so as to be ‘no worse’ than a software implementation. However, the flexibility in the number representation is one of the key factors that can be exploited on reconfigurable hardware such as FPGAs, and hence ignoring this potential significantly limits the performance achievable. In order to optimise the performance of hardware reliably, we require a method that can tractably calculate tight bounds for the error or range of any variable within an algorithm, but currently only a handful of methods to calculate such bounds exist, and these either sacrifice tightness or tractability, whilst simulation-based methods cannot guarantee the given error estimate. This thesis presents a new method to calculate these bounds, taking into account both input ranges and finite precision effects, which we show to be, in general, tighter in comparison to existing methods; this in turn can be used to tune the hardware to the algorithm specifications. We demonstrate the use of this software to optimise hardware for various algorithms to accelerate the solution of a system of linear equations, which forms the basis of many problems in engineering and science, and show that significant performance gains can be obtained by using this new approach in conjunction with more traditional hardware optimisations.