Research On Program Performance Evaluation Techniques For Hardware/Software Partitioning

Currently,CPU+FPGA-based reconfigurable computing systems are being widely used.Reconfigurable compiler for reconfigurable computing systems is becoming a hot topic in heterogeneous accelerated systems.Hardware/software automatic partitioning based on program performance evaluation is important and difficult research for reconfigurable compiler.Accurate performance evaluation techniques can provide supplementary information for the hardware/software automatic partitioning,making the partitioning results more efficient.Hardware/software partitioning is mostly based on the IR middle layer of code in reconfigurable compiler.So a program performance assessment based on IR middle layer of code is needed to provide supplementary information for the hardware/software partitioning.Due to the impact of the pipeline,branch prediction,cache and other factors,the execution time of the program is often nonlinear changes.While BP neural network has nonlinear mapping ability,therefore,this article researches BP neural network-based IR layer's program performance assessment methods.During the study,we find the program classification can improve the accuracy of evaluation.So this paper also proposes a program classification method.By judging the ratio of the number of if basic blocks executed to the total number of executions of the basic block in IR program,this method can determine whether this program is a control-dominated or a data-dominated.Besides,this method can also detect the situation of the program's cache hit,thus a more accurate assessment of program execution time is produced.Experimental results show that,for different types of programs,training different BP neural network model can obtain a more accurate evaluation.Compared with other application performance evaluation methods,in IR layer,program performance evaluation methods based on BP neural network has higher accuracy than the program performance evaluation method based on linear models.Compared with the program performance evaluation method based on the simulation,our program performance evaluation methods is equal in accuracy,but has advantage in speed.For hardware/software partitioning,our program performance evaluation method is so accurate and efficient to meet the requirements.