Boda-RTC - Productive Generation of Portable Efficient Code for Convolutional Neural Networks on Mobile Computing Platforms

《Boda-RTC - Productive Generation of Portable Efficient Code for Convolutional Neural Networks on Mobile Computing 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、Boda-RTC:ProductiveGenerationofPortable,EfficientCodeforConvolutionalNeuralNetworksonMobileComputingPlatformsMatthewW.MoskewiczForrestN.IandolaKurtKeutzerUniversityofCalifornia,Berkeley{moskewcz,forresti,keutzer}@eecs.berkeley.eduAbstract—Thepopularityofneuralnetworks(NNs)spansanddeplo

2、ymentofsystemsthatincludeNNs,itisdesirabletoacademia[1],industry[2],andpopularculture[3].Inparticular,nurtureadiverseenablingecosystemoftoolsandapproaches.convolutionalneuralnetworks(CNNs)havebeenappliedtomanyInparticular,wefeelitisdesirabletosupportmanyhardwareimagebasedmachinelearni

3、ngtasksandhaveyieldedstrongandsoftwareplatformstoenablenewapplicationsacrossresults[4].Theavailabilityofhardware/softwaresystemsforefficienttraininganddeploymentoflargeand/ordeepCNNmanyareas,includingmobile,IoT,transportation,medical,modelsiscriticalforthecontinuedsuccessofthefield[5][1

4、].andothers.Imaginethat,foragiventask,high-performanceEarlysystemsforNNcomputationfocusedonleveragingexistingvendorlibrariesexistforatleastoneplatform.Currently,fordenselinearalgebratechniquesandlibraries[6][7].CurrentCNNs,thevendorisNvidia,theplatformisMaxwell,andapproachesuselow-lev

5、elmachinespecificprogramming[8]thelibraryiscuDNN[9].Whynotsimplyusethatvendor’sand/orclosed-source,purpose-builtvendorlibraries[9].Inthiswork,wepresentanopensourcesystemthat,comparedtoplatformandlibrariesforthetaskandbesatisfied?Oneissueexistingapproaches,achievescompetitivecomputationa

6、lspeedisquitesimple:inindustrialusecases,choiceofplatformmaywhileachievingsignificantlygreaterportability.Weachievethisbedictatedbybusinessconcerns.Further,thosesamebusinessbytargetingthevendor-neutralOpenCLplatform[10]usingaconcernsmayprecludedependenceonanysinglevendor.Forcode-genera

7、tionapproach.Wearguethatourapproachallowsexample,theflagshipSamsungGalaxyS7mobilephoneshipsforboth:(1)therapiddevelopmentofnewcomputationalkernelsforexistinghardwaretargets,and(2)therapidtuningofexistingintwoversions:oneusingaSamsung-proprietaryExynoscomputationalkernelsfornewhardwaret

8、arget