Research and Implementation of Performance Analysis Tool for CUDA Programs with Directive

In recent years, the rapid expansion of graphics processing unit (GPU) as well as the computer unified device architecture (CUDA) technology proposed by NVIDIA pushes forward the application of GPU in the field of high performance computing (HPC). In this paper, GPU's architecture and CUDA programming model are introduced first. According to the method of parallel program performance analysis in CPU cluster mode, a performance analysis tool for CUDA programs based on directive is designed and implemented. Experiment results validate the validity of this performance analysis tool on different GPU hardware platforms.