SIFOpt - Fixed-Point Implementations of Calculations from Floating-Point Descriptions

Abstract

Signal and image processing and control algorithms are typically developed using double-precision floating point arithmetic, in a high-level software design environment such as Matlab/Simulink, IDL, or C/C++. The avionics system designer must implement these computations in a platform that satisfies the constraints of on-board data processing such as size, weight, power consumption and radiation hardness. The ability to implement math-intensive signal processing and control algorithms on simple, integer-only hardware would greatly increase the avionics system designer's options. This paper presents a methodology and software tool for generating fixed point computations, enabling integer-only processors to achieve numerical accuracy comparable to floating point, faster and with reduced size, weight, cost, and power consumption. The methodology has been successfully demonstrated on image focal plane array calibration, Fourier transform hyper-spectral imager data conversion, wavelet-based feature selection, and various pattern recognition tasks.