Precision Modeling of Floating-Point Applications for Variable Bitwidth Computing

摘要

This paper presents an innovative, application-adaptive modeling methodology for variable precision floating-point (FP) computing. Implementing floatingpoint applications on custom devices can be costly. Since implementation cost can be traded off with precision (represented by mantissa bitwidth) for some FP applications, automatically obtaining the appropriate reduced bitwidth will lead to the low-cost implementation of FP computation. The methodology introduced in this paper targets model-based bitwidth optimization. Arithmetic models developed using the methodology can be used in optimization algorithms to compute the minimal bitwidth set. This differs from most existing methodologies which are simulation based. The methodology makes use of a behavioral representation of the application and a behavioral profiling process to obtain the necessary parameters. Two examples are introduced to demonstrate the usage and accuracy of the modeling methodology.