GPGPU Performance Estimation And Optimization

In recent years, GPU has been widely used in large-scale data-parallelcomputing. As a result, performance estimation of GPGPU(General PurposeGPU) programs become a hot research subject. Based on NVIDIA GPU andCUDA, this paper presents two kinds of performance estimation model,respectively for CUDA kernel and heterogeneous programming, with whichGPGPU developers can predict the performance of algorithms andunderstand the impact of design alternatives.The performance estimation model of kernel is mainly based on theinstruction execution mechanism of CUDA. First the GPU instruction issuemechanism is analyzed and a theoretical instruction throughput upper boundis proposed. On this basis, the factors that hold back GPGPU programs fromreaching the performance upper bound are taken into consideration.The heterogeneous performance estimation model is mainly about theasynchronous execution of CPU and GPU. In this model, the kernelexecution is regarded as a whole, which means this model is independent ofthe kernel performance model. Here the execution time of different parts areclassified into several categories and discussed respectively.Experiments are carried out for both models to prove their effectiveness.And approaches of GPGPU performance optimization based on the modelsare introduced.