Fine grain precision scaling for datapath approximations in digital signal processing systems

Abstract

Finding optimal word lengths in digital signal processing systems has been one of the primary mechanisms for reducing complexity. Recently, this topic has been explored in a broader approximate computing context, where architectures allowing for fine-grain control of hardware or software accuracy have been proposed. One of the obstacles for adoption of fine-grain scaling techniques is that they require determining the precision of all intermediate values at all possible operation points, making simulation-based optimization infeasible. In this chapter, we study efficient analytical heuristics to find optimal sets of word lengths for all variables and operations in a dataflow graph constrained by mean squared error type of metrics. We apply our method to several industrial strength examples. Our results show a more than 5,000x improvement in optimization time compared to an efficient simulation-based word length optimization method with less than 10% estimation error across a range of target quality metrics.