Research On Performance Optimization And Evaluation For FT Heterogeneous Processors

Performance evaluation method is the foundation of computer architecture research and development,and it helps architects delve deeper into computers.Researchers can use it to compare platforms,identify performance bottlenecks.Performance optimization is to improve program performance without affecting system performance.Performance optimization and performance evaluation are inseparable.As the microprocessor industry shifts to multi-core architectures,research into parallel computing is critical to ensuring the future progress of mainstream computer systems.Some current performance evaluation methods provide parallel programs,but only for general CPU architecture and GPU architecture,and there is no performance measurement method for energy efficient multi-core digital signal processor(DSP)architectures.The FT-M7032 is a CPU-DSPs heterogeneous processor that integrates 16ARMv8 CPU cores with four general-purpose multi-core DSP(GPDSP)clusters.This paper presents and describes a performance evaluation method for FT-M7032 heterogeneous processor.First,five test programs were selected for the FT-M7032 platform,the selection of which was inspired by the Berkeley dwarf classification.A microarchitecture-independent workload characterization(MICA)method and principal component analysis(PCA)method were used to demonstrate the diversity and coverage of the five test programs.Secondly,the selected test program is transplanted to FT-M7032.The data initialization part is in the CPU side,and the core code is implemented in the DSP side.After that,the transplanted program was further optimized,and the optimization method mainly used vectorization,the objective of optimization is to improve the acceleration ratio.Finally,through the series of experiments above,the performance and framework understanding is summarized,and the five applications cover a wide range of parallel communication modes and synchronization techniques,leading to some important architectural insights.This paper evaluates the performance of five test programs on the DSP kernel from two aspects of scalability and speed ratio.The experimental results show that these cores have good scalability in large-scale environment,and the optimized kernel also has a better acceleration ratio.