Performance of an Occam/transputer implementation of interval arithmetic

Abstract

<p>Rounded interval arithmetic is very easy to implement by means of directed rounding arithmetic operators. Such operators are available in the IEEE floating point arithmetic of the transputer. When a few small pieces of assembly language code are used to access the directed rounding operators, the four basic rounded interval arithmetic operators can easily be expressed in the programming language Occam.</p><p>The performance of this implementation is assessed and it is shown that the time consuming part of the calculation are not the directed rounding floating point operations as one might have expected. Most of the time is spent with transport of operands to and from the on-chip floating point unit and the procedure call/parameter passing overhead. Based on this experience the implementation is improved. This implementation runs with 0.15 MIOPS (Million Interval Operations Per Second) or 0.30 MFLOPS on an example interval calculation proposed by Moore. Furthermore, it is demonstrated that an advanced interval language compiler may provide a performance of 0.30 MIOPS or 0.59 MFLOPS on this example calculation.</p>