A method for floating-point to fixed-point conversion in fixed-point SoC based on statistical analysis
Yin Ai-chang · Journal of Circuits and Systems · 2007
The algorithms used by communication,voice and image processing systems are typically specified as floating-point operations.On the other hand,the minimization of cost,power consumption of the applications requires the implementation of floating-point algorithms in fixed-point architectures.Considering cost of hardware,performance,and power consumption on balance in the fixed-point SoC,the implementations of these algorithms usually use fixed-point approximations.In the bid to achieve the implementations,it requires the development of methods for the conversion of floating-point to fixed-point.In this paper,a method for floating-point to fixed-point conversion in fixed-point SoC is proposed.The parameters of conversion and the hardware are introduced to achieve the goal of floating-point to fixed-point conversion firstly.Secondly,the best parameters are calculated by the method for evaluating the quality of fixed-point approximate algorithms through the Signal to Noise Ratio(SNR) which is proposed in this paper.Then the best fixed-point approximate algorithm and the hardware accelerator are obtained through these calculated parameters.Simultaneously,the prototype design strategy of the hardware accelerated module in the single core SoC is also proposed.