嵌入式智能计算机计算能力评测方法  

Computing Capability Evaluation Method of Embedded Intelligent Computer

作　　者：马春燕[1] 陈晶 姚鼎 张涛[1] MA Chun-Yan;CHEN Jing;YAO Ding;ZHANG Tao(School of Software,Northwestern Polytechnical University,Xi’an 710072)

机构地区：[1]西北工业大学软件学院,西安710072

出　　处：《计算机学报》2023年第11期2279-2301,共23页Chinese Journal of Computers

摘　　要：计算能力评测方法是嵌入式智能计算机领域的研究热点之一.嵌入式智能计算机存在多种神经形态计算的优化方案和机器学习加速器,导致其评测难度大于通用计算机.基准测试是目前普遍采用的评测方法,但是在资源受限的嵌入式设备上,基准测试集和评价指标的复用能力有限,难以适应配置多样化的嵌入式智能系统;测试集中神经网络模型的计算强度存在一定的随机性,无法充分挖掘待测设备的计算潜力;评价指标不统一,难以对不同嵌入式智能计算机的计算能力进行对比分析.本文提出了一种基于神经进化算法的嵌入式智能计算机计算能力评测方法.首先,基于Roofline理论模型,融合计算潜力挖掘、资源适配和评价指标统一等优势,提出了一种适配各种嵌入式智能计算机的计算能力评测框架,并对其合理性进行分析;其次,提出了一种评测计算能力的神经网络模型生成算法,利用神经进化算法,使生成模型的计算强度逼近嵌入式智能计算机的计算强度上限,充分挖掘待测设备的计算潜力,使评测结果更客观;然后,采用环境固定的上位机作为对照,分别在待测设备和上位机交叉运行生成的神经网络模型,并以两次执行推断任务时的每秒浮点运算次数作为计算因子,给出计算能力评测的通用公式,可以实现不同嵌入式智能计算机计算能力的对比分析;最后,在Mindspore-cpu、Tensorflow-cpu和Mindspore-ascend310框架下评测华为Atlas200,相比基准测试中常用的5种神经网络模型,采用本文生成的神经网络模型的测评结果更合理,证实两个DaVinci核心的智能计算能力是八个Cortex-A55核心的42.37倍.Computing capability evaluation is one of the research hotspots in the field of embedded intel-ligent computing.It is an important means to test the execution speed and performance of intelligent com-puters when performing inference tasks.The rationality of its evaluation results directly affects the optimi-zation and improvement direction of embedded intelligent computers.Due to the various optimization schemes for neural form computing and machine learning accelerators,the evaluation of embedded intelli-gent computers is more difficult than that of general-purpose computers.Benchmark testing is currently a commonly used evaluation method,but on resource-limited embedded devices,the reuse ability of bench-mark test sets and evaluation indicators is limited,making it difficult to adapt to the diversified configura-tion of embedded intelligent systems.The computational intensity of the neural network models in the test set has a certain randomness,which cannot fully explore the computing potential of the device under test,and the evaluation indicators are not unified,making it difficult to compare and analyze the computing ca-pabilities of different embedded intelligent computers.To solve the problem that embedded intelligent de-vices have different configurations and cannot be performance tested through fixed models,this paper pro-poses a neural network model generation algorithm based on neural evolution algorithm to generate neural network models that can characterize the intelligent computing capabilities of embedded intelligent devices,and inversely infer the intelligent computing capabilities of embedded intelligent computers based on the complexity of the model.Firstly,based on the Roofline theory model,the advantages of integrating com-puting potential mining,resource adaptation,and unified evaluation indicators are used to propose a com-puting capability evaluation framework that can adapt to various embedded intelligent computers,and its rationality is analyzed.Secondly,a neural network model generation

关 键 词：神经进化算法 嵌入式智能计算机 计算能力评测 神经网络模型 Atlas200 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]