OWL: Word-Length Optimization via Extreme Value Theory and Affine Arithmetic/Interval Arithmetic Model

S. Zhou and J. Bian (PRC)


Affine arithmetic, word-length, fixed-point


In the high level synthesis from languages such as C, word-length of variables is one of the key issues on VLSI design optimization. This paper proposes a new automated approach – OWL, for optimizing word-lengths of fixed-point designs. OWL is based on static analysis via extreme value theory and affine/interval arithmetic. It describes methods to minimize both the integer and fraction parts of fixed-point signals. For range analysis, it employs a semi-analytical approach with extreme value distribution model to identify the number of integer bits, while for precision analysis, it uses AAIA error model to find the optimum number of fraction bits. The reduction percentage of word-length is over 60% in integer part and over 45% in fraction part.

Important Links:

Go Back