Research On Performance Analysis Method Of Graphics Rendering On Mobile GPU Of ARM

Graphics rendering is one of the main functions of GPU.Graphics rendering technology is widely used in image display,video playback and game presentation.Performance analysis is an important part of the development and optimization of graphics application.Targeting at GPU,most of the existing performance analysis methods apply only to specific GPU architecture and do not apply to graphics rendering applicatlions.And the operations of these methods are complex.Based on the idea of performance decomposition,this thesis constructs a performance decomposition model for graphics rendering of ARM mobile GPU,and designs a bottleneck positioning tool to achieve simple and effective graphical rendering performance analysis.Firstly,this thesis modifies the source code of the performance data acquisition tool(DS-5 Streamline)to realize the collection of performance data.Secondly,by analyzing the types of software and hardware running pipeline and performance bottlenecks for graphics rendering under ARM Midgard architecture,this thesis defines the performance module and generates execution time formulas according to principle of the performance module.Thirdly,based on the hardware implementation principle and perfoance decomposition of ARM Midgard architecture,this thesis builds the key execution time formula of the module,and chooses the module with the longest execution time as the performance bottleneck module to establish the performance decomposition model and form the bottleneck positioning method.This thesis verifies the effectiveness of the graphic rendering performance analysis method from three aspects which are the accuracy of performance decomposition model,the accuracy and usefulness of performance bottleneck positioning.The verification of the accuracy of performance decomposition model shows that the average relative error between the graphic rendering time and the measured time is 1.67%in test cases which are Antutu Benchmark,GFXBench and MinMaxblending.The verification of the accuracy of performance bottleneck positioning shows that the performance bottleneck of the performance decomposition model is basically the same as the bottleneck of the ARM artificial experience 'bottleneck judgment method.The accuracy of performance bottleneck positioning is high.Based on the graphics application which source code is available,the verification of the validity of performance bottleneck positioning shows that the performance of the optimized application which use bottleneck positioning tool to locate performance bottlenecks is improved obviously,and the performance bottlenecks disappear.