欢迎来到天天文库
浏览记录
ID:40717215
大小:1.21 MB
页数:7页
时间:2019-08-06
《GPU-based Acceleration of Deep Convolutional Neural Networks on Mobile Platforms》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、GPU-basedAccelerationofDeepConvolutionalNeuralNetworksonMobilePlatformsSeyyedSalarLatifiOskoueiHosseinGolestaniSharifUniversityofTechnologySharifUniversityofTechnologysalarlatifi@ee.sharif.eduhosseingolestani@ee.sharif.eduMohamadKachueeMatinHashemiSharifUniversityofTechnologySharifUniversit
2、yofTechnologykachueemohamad@ee.sharif.edumatin@sharif.eduHodaMohammadzadeSoheilGhiasiSharifUniversityofTechnologyUniversityofCalifornia,Davishoda@sharif.edughiasi@ucdavis.eduAbstractparallelprocessingorhardwareacceleration.Onserveranddesktopplatforms,therearemanylibrariessuchasCaffeMobilea
3、pplicationsrunningonwearabledevicesand[3],Torch[4],Theano[5],cuDNN[6]andcuda-convnetsmartphonescangreatlybenefitfromaccurateandscalable[7]whichemployGPUcomputingandSIMDprocessorex-deepCNN-basedmachinelearningalgorithms.Whilemo-tensionsforaccelerationofdeepCNNcomputations.bileCPUperformanced
4、oesnotmatchtheintensivecom-Onmobileplatforms(e.g.,wearabledevicesandsmart-putationalrequirementofdeepCNNs,theembeddedGPUphones),tothebestofourknowledge,theexistingCNNli-whichalreadyexistsinmanymobileplatformscanbelever-brariesarelimitedtotheprocessingpowerofmobileCPUsagedforaccelerationofC
5、NNcomputationsonthelocalde-[8].Asaresult,executionoflargedeepCNNsonmobileviceandwithouttheuseofacloudservice.Wepresentaplatformsseemscurrentlyunfeasibleintermsofbothcom-GPU-basedaccelerateddeepCNNengineformobileplat-putationalperformanceandpowerconsumption.Toaddressformswithupto60Xspeedup1
6、.thisissue,severalgroupshaveproposedvarioushardware-basedCNNaccelerators[9,10].Suchhardware-basedac-celerationenginesarenotpracticallydeployedinexisting1.Introductionmobilesystemsduetohighproductioncostsofhardwareandlargechiparearequirements.Moreimportantly,afixedManyapplicationssuchasimage
7、classification[1],hardwareacceleratorisnormallylimitedtoexecutionofspeechrecognition,andmovementactivityrecognition[2]onlyasetofpredefinedCNNs.requireutilizationofmachinelearningalgorithmsonmo-arXiv:1511.07376v1[cs.DC]23Nov2015Wepresentanovelsoftware-onlyacceler
此文档下载收益归作者所有