Study And Realization Of The Trace-based Android 3D Workload Subsetting Method

In recent years,with the development of hardware and software in embedded systems,mobile phones and tablets have gradually replaced computers as the first choice for people to entertain and shop.Consumers pay more attention to the mobile GPUs.To find the optimal design parameters for mobile GPU,architects often need to simulate a large number of 3D workloads under a variety of architectural configurations.Simulaiton is time-consuming and greatly reduces the efficiency of the architecture design.In order to reduce the simulation time,3D workload subsetting is essential.Currently,the mainstream mobile system is Android.Therefore,learning from CPU benchmark subsetting and desktop 3D workload subsetting related research,this thesis propose the trace-based Android 3D workload subsetting method.Firstly,the CPU benchmark subsetting and desktop 3D workload subsetting reladted research are analyzed and find that they can not adapt to Android 3D workload subsetting.Then,in order to overcome the limitation of the existing subsetting method,the trace-based Android 3D workload subsetting method is proposed.This method includes four steps: trace generation,feature generation,representative frames selection and subset trace generation.Next,based on the open source project Apitrace and the GPU profiler ARM DS5 Streamline,the tool chain for the four steps is constructed.Finally,on Odroid XU4 board,the Antutu 3DBench,3DMark Ice Storm Extreme Demo,3DMark Ice Storm Extreme GT1,3DMark Ice Storm Extreme GT2 and SUV Traffic Racer are subsetted by this method.The average frame number of the subsets is 2.68% of its original 3D workload.The subsets run an average of 11.5 times faster than the original 3D workload.On Odroid XU4 board,Firefly RK3399 board,Cube Iwork 8 pad and Redmi Note 4X phone,the GPU frequency scaling experiment shows that the FPS of the subsets have high correlation(correlation coefficient = 99.5%+)to the FPS of the original 3D workload.