Publikation
DIVIAC: Library of Input Data Aware Approximate Dividers with Partial Exact Minimization
Chandan Jha; Sallar Ahmadi-Pour; Sajjad Parvin; Rolf Drechsler
In: 39th International Conference On VLSI Design. International Conference on VLSI Design (VLSID-2026), January 3-7, Pune, India, 2026.
Zusammenfassung
Approximate divider designs have been extensively
explored as they are widely used in image processing applications.
The design of approximate dividers is predominantly accomplished through the use of functional approximation, where the
Boolean functions of the subtractor sub-blocks in the dividers
are approximated by replacing them with a different Boolean
function. However, the prior works have explored only a few
Boolean approximations and evaluated the error metrics using
uniform distributions. This does not effectively explore the design
space and leads to suboptimal approximate divider designs. In
this work, we alleviate the limitations of prior works as follows:
Firstly, we perform an extensive and systematic design space
exploration to identify the Pareto-optimal approximate divider
designs, where each sub-block has been reduced through exact
minimization. Secondly, we do this for three input distributions,
namely uniform, normal, and exponential distributions. Lastly,
we also evaluate the design on the widely used image processing
applications for approximate dividers, namely background removal and change detection. We aim to make the Pareto-optimal
approximate divider designs available as open-source to stimulate
further research.
