Fixed-point accuracy evaluation in the context of conditional structures

Abstract

The automation of fixed-point conversion requires generic methods to study accuracy degradation. Accuracy evaluation is often based on simulation approaches, at the cost of an important execution time. This paper proposes a new approach using fast analytical noise power propagation considering conditional structures. These structures are generated from programming language statements such as if-then-else or switch. The proposed model takes two key points into account in fixed-point design: first, an alternative processing of noise depending on the condition; second, decision errors generated by quantization noise affecting the condition. This method is integrated in the fixed-point conversion process and uses path probabilities of execution alternatives. This work extends existing analytical approaches for fixed-point conversion. Experiments of our analytical method show that it has a fairly accurate noise power estimation compared to the real accuracy degradation.