Study and evaluation of automatic offloading for function blocks of applications

Abstract

Systems using graphical processing units (GPUs) and field-programmable gate arrays (FPGAs) have increased due to their advantages over central processing units (CPUs). However, such systems require the understanding of hardware-specific technical specifications such as Hardware Description Language (HDL) and compute unified device architecture (CUDA), which is a high hurdle. Based on this background, we previously proposed environment-adaptive software that enables automatic conversion, configuration and high-performance operation of existing code according to the hardware to be placed. As an element of this concept, we also proposed a method of automatically offloading loop statements of application source code for CPUs to GPUs and FPGAs. In this paper, we propose a method for offloading a function block, which is a larger unit, instead of individual loop statements in an application to achieve higher speed by automatically offloading to GPUs and FPGAs. We implemented the proposed method and evaluated it using current applications offloading to GPUs and FPGAs.