Performance Model Of OpenCL Programs For GPU

In recent years,general-purpose computing on GPU is widely used in many fields.As a parallel programming specification,OpenCL is suitable for writing GPU general computing programs.However,many programmers have difficulties in writing high-performance OpenCL programs.Therefore,it is necessary to build a performance model for the prediction of program performance,which provides a reference for the improvement of program performance.The thesis proposes a layered performance model by studying the factors of OpenCL performance towards GPU,which aims to be helpful for software developers to write high-performance programs.This model consists of transport layer,access layer and computing layer.Data transfer from the host to the global memory is the main factor that influences the performance in transport layer.The factors influencing the performance in the access layer are storage location,block conflict,coalesced access,etc.The factors in the computing layer are the form of thread organization,and the overlap of computation and access.In the memory and computing layers,the overall time can be estimated by calculating clock cycles of each module.The transport layer uses the linear regression equation derived from the test data to estimate time-needed.The layered model contributes to finding out the bottleneck of programs and provides a reference for the improving the performance of programs.The model is tested by vector addition and bitonic sorting.Results for test are very close to the predicted results.Based on the model,bitonic sorting is modified and then its performance has been improved.Testing results show that recommendations from this model are useful to improve the performance of programs.