Research And Implementation Of Template Generation Method For CPU Benchmark Synthesis

In the process of processor design and performance evaluation,benchmark are often used instead of real applications to explore the design of microarchitecture and test the overall performance of processors.The method of distilling load characteristics from real applications and generating benchmark based on characteristics is called benchmark synthesis.The template program needed in the synthesis is designed and implemented,which is divided into single characteristic micro template and multi characteristics macro template.The simplified template library is removed by SOM + K-means algorithm.The accuracy and speed of benchmark synthesis based on template are improved.The micro template with single characteristic in this thesis has 7 kinds of micro architecture independent characteristics,such as instruction mixing,register dependence distance,critical path length,spatial and temporal locality of data,space and time locality of instruction,branch transition rate,serial instruction distribution and so on,and a total of 220 dimensions.The generation method of micro template and how to reduce the coupling between characteristics is introduced in detail,and the simulation platform plug-in is added to extract the micro architecture independent characteristics of the template program,and the feasibility of the verification method is verified.However,the error between the benchmark and the original program synthesized using only single characteristic micro template is very large.In order to improve the accuracy and speed,based on the objective coupling between characteristics,the generation method of multi characteristics macro template is also implented,including the register dependent distance and critical path length macro template,the data locality macro template,the instruction locality and branch transition rate macro template.In addition,the template library is formed with the micro template and the macro template together,and the size of the template library is reduced by the SOM + K-means algorithm,and the accuracy and speed of benchmark synthesis are further improved.The feasibility of the template generation method is verified on the simulation platform,and the Android system startup and embedded benchmark Mibench are used as the original load program,and the corresponding benchmark is synthesized with the template generated in this thesis.The error of the Android system startup and the Mibench and the synthetic benchmark on the micro architecture independent characteristic is 9.98%,and the micro architecture independent characteristic is 9.72%,which all meet the expected goal of this thesis.