Model-Based Design Space Exploration for Approximate Image Processing on FPGA

Increasing demands on computational performance are outpacing the technological improvements in computer hardware. Approximate computing provides a recent approach to bridge this gap by exploiting the error resilience of applications and trading in quality for less resource usage. Research in this field has introduced numerous approximation methods. Combining multiple methods can further increase the benefits for complex systems. Because of error propagation and potential interactions between components, the approximation parameters cannot be optimized individually and therefore the design space grows exponentially across all approximations. Focusing on FPGA-based systems, we propose a methodology to explore this design space using a multi-objective genetic algorithm guided by appropriate models which estimate resource demands and anticipated quality degradation without time-consuming synthesis. The effectiveness of our approach is experimentally evaluated on a typical image color processing pipeline considering multiple approximation methods. Our results show that the methodology is able to find a wide range of Pareto-optimal solutions, among which the desired quality-resource trade-off can be chosen.