Accelerating floating-point to fixed-point data type conversion with evolutionary algorithms

摘要

The choice of the data type representation has significant impacts onudthe resource utilisation, maximum clock frequency and power consumptionudof any hardware design. Although arithmetic hardwareudunits for the fixed-point format can improve performance and reduceudenergy consumption, the process of tuning the right bit length isudknown as a time-consuming task, since it is a combinatorial optimisationudproblem guided by the accumulative arithmetic computationuderror. A novel evolutionary approach to accelerate the process of convertingudalgorithms from the floating-point to fixed-point format is presented.udResults are demonstrated by converting three computingintensiveudalgorithms from the mobile robotic scenario, where datauderror accumulated during execution is influenced by external factors,udsuch as sensor noise and navigation environment characteristics. Theudproposed evolutionary algorithm accelerated the conversion processudby up to 2.5 × against the state-of-the-art methods, allowing evenudfurther bit-length optimisations.