Abstract:
To optimize word-length determination for fixed point implementation of digital
signal processing and communication algorithms, the idea is to use the same floating
point code with minimum changes for fixed point implementation and the use of a
tool/environment which provides a complete from abstraction to low level design
support. Transforming from floating point to fixed point involves throwing away
fraction bits hence introducing noise to the system referred as quantization noise.
The process of word-length determination involves trial and error. During the
process of determination of word-lengths the objective is to minimize the error
introduced into the system by the quantization and hardware cost. In this thesis
multi objective problem is studied. An IIR filter is first developed in floating point and
then fixed point implementation is simulated for different word lengths. A genetic
and evolutionary approach is used to determine the optimum fixed point word
length, which gives a trade off curve of multiple solutions from which a designer can
choose.
