Fast trade-off evaluation for digital signal processing systems during wordlength optimization

This paper presents a novel precision analysis approach, which optimizes the fractional wordlengths of signals in digital signal processing systems. Quantization-operation-error model is proposed to formulate the quantization error bound propagations through modules. Based on that, explicit relationship between output accuracy and hardware cost is built up, and a greedy search algorithm is proposed to find the minimum implementation cost while meeting output error constraint. Moreover, the greedy search steps can form a near pareto-optimal front, which can be used for a fast trade-off evaluation between output accuracy and implementation cost, such as area, power and latency. Experimental results and comparisons with existing state-of-the-art methods demonstrate the efficiency of proposed approach.