Performance Analysis And Optimization Of Stencil On NVIDIA Kepler

Stencil-based kernels constitute the core of many compute-intensive applications in many fields.Recently,GPUs as the accelerators are widely deployed through thousands of workstations and supercomputers.The advantages of GPU are the high performance,low power consumption,and good programmability.Nvidia's CUDA platform is a widely adopted programming model for GPU.However,how to analyze and optimize the performance of applications with heterogeneous system architecture is the key problem needed to be solved.This thesis analyzed and tuned the performance of a 3D seven-point application of Stencil based the NVIDIA Kepler architecture GPU including Tesla K20,K40,and K80.Nvprof was used in this thesis to select the performance metrics and events during the execution of the application.In this work,the optimization technology improves the performance of 3D7 P based on Kepler GPU.The bandwidth of device memory is fully used after optimization.This thesis also studied the NVIDIA Jetson TX1 developer kit.TK1 is the first mobile processor to have the same architecture and features as a Kepler GPU and Maxwell GPU for TX1.It makes sense to compare Kepler GPU and TX1 for better understanding the Kepler and Maxwell architecture.The experience of Maxwell also has great significant meaning for Maxwell GPU in the computing platform in the future.