A novel method for the float-point to fixed-point conversion from statistical perspective
Jun Yang, Fan Zhou, Xinning Liu, Longxin Shi · International Journal of Electronics · 2008
Float-point to fixed-point conversion is required by many communication, voice and image applications, in order to achieve acceptable levels of performance and cost. This is particularly crucial in real-time applications that have strict timeline requirement on processing throughput and latency. Therefore, fixed-point conversion from float-point computation brings great challenges with the improvement at the same time: How to evaluate the accuracy of the fixed-point system quickly? This paper proposes a novel method for the float-point to fixed-point conversion and evaluates the precision of complex signal systems from statistical perspective. The proposed approach greatly reduces the complexity of the evaluation, and achieves optimum fixed-point implementation in a short time. The experiments in MP3 and the AAC decoding algorithm demonstrate that the accuracy of the fixed-point system is greatly improved by this method.