Approximating Behavioral HW Accelerators through Selective Partial Extractions onto Synthesizable Predictive Models

Abstract

This work presents a method to selectively extract portions of a behavioral description to be synthesized as a hardware accelerator using High-Level Synthesis (HLS) onto different predictive models in order to trade-off the accuracy of the accelerators' outputs with area and power. Because the main aim of this work is to synthesize the newly approximated behavioral description, we investigate the use of different predictive models, mainly linear regression (LR) and multi-layer perceptron (MLP), highlighting the trade-offs of using one over the other. In addition, we further extend the search space by reducing the precision of the predictive models' coefficients, thus, leading to a wider range of solutions. Experimental results using a variety of benchmarks from different domains show that our proposed method works well compared to another state of the art approximate solution.