Design and Verification of an open-source SFU model for GPGPUs

Abstract

General Purpose Graphic Processing Units (GPGPUs) are widely used in data-intensive applications, such as multimedia and high-performance computing. These technologies are currently used also to support safety-critical applications (e.g., in the automotive and industrial domains) to implement computer vision, sensor fusion, or machine learning algorithms, which often require the processing of complex transcendent or trigonometric functions. In these cases, an integrated special function unit in the GPGPU is utilized, which is intended to increase the performance in parallel operations. However, this complex module is not present in most of the available architectural and micro-architectural open-source models of GPGPUs, so limiting the characterization and analysis of applications using these units. In this work, we report about the design and functional verification of a Special Function Unit to execute transcendent and trigonometric operations in GPGPUs. We integrated the proposed module within an open-source GPGPU (FlexGripPlus) implementing the G80 micro-architecture. The experimental results show the significant improvements in performance and accuracy achievable by using these modules in parallel applications running in a GPGPU.