《通信工程专业英语（第二版）》课件Unit 16

Unit16TextADigitalSignalProcessingDigitalSignalProcessingDigitalsignalprocessing(DSP)istheuseofdigitalprocessing,suchasbycomputers,toperformawidevarietyofsignalprocessingoperations.Thesignalsprocessedinthismannerareasequenceofnumbersthatrepresentsamplesofacontinuousvariableinadomainsuchastime,space,orfrequency.TextTextWordsNotesDiscussionHomeDigitalsignalprocessingandanalogsignalprocessingaresubfieldsofsignalprocessing.DSPapplicationsincludeaudioandspeechsignalprocessing,sonar,radarandothersensorarrayprocessing,spectralestimation,statisticalsignalprocessing,digitalimageprocessing,signalprocessingfortelecommunications,controlofsystems,biomedicalengineering,seismicdataprocessing,amongothers.TextTextWordsNotesDiscussionHomeDigitalsignalprocessingcaninvolvelinearornonlinearoperations.Nonlinearsignalprocessingiscloselyrelatedtononlinearsystemidentificationandcanbeimplementedinthetime,frequency,andspatio-temporaldomains.TextTextWordsNotesDiscussionHomeTheapplicationofdigitalcomputationtosignalprocessingallowsformanyadvantagesoveranalogprocessinginmanyapplications,suchaserrordetectionandcorrectionintransmissionaswellasdatacompression.DSPisapplicabletobothstreamingdataandstatic(stored)data.TextTextWordsNotesDiscussionHomeTextTextWordsNotesDiscussionHome1.SignalsamplingTheincreasinguseofcomputershasresultedintheincreaseduseof,andneedfor,digitalsignalprocessing.Todigitallyanalyzeandmanipulateananalogsignal,itmustbedigitizedwithananalog-to-digitalconverter.Samplingisusuallycarriedoutintwostages,discretizationandquantization.Discretizationmeansthatthesignalisdividedintoequalintervalsoftime,andeachintervalisrepresentedbyasinglemeasurementofamplitude.Quantizationmeanseachamplitudemeasurementisapproximatedbyavaluefromafiniteset.Roundingrealnumberstointegersisanexample.TextTextWordsNotesDiscussionHomeTheNyquist–Shannonsamplingtheoremstatesthatasignalcanbeexactlyreconstructedfromitssamplesifthesamplingfrequencyisgreaterthantwicethehighestfrequencyofthesignal,butthisrequiresaninfinitenumberofquantizationlevels.Inpractice,thesamplingfrequencyisoftensignificantlyhigherthantwicethatrequiredbythesignal'slimitedbandwidth.TextTextWordsNotesDiscussionHomeTheoreticalDSPanalysesandderivationsaretypicallyperformedondiscrete-timesignalmodelswithnoamplitudeinaccuracies(quantizationerror),"created"bytheabstractprocessofsampling.Numericalmethodsrequireaquantizedsignal,suchasthoseproducedbyananalog-to-digitalconverter(ADC)1.Theprocessedresultmightbeafrequencyspectrumorasetofstatistics.Butoftenitisanotherquantizedsignalthatisconvertedbacktoanalogformbyadigital-to-analogconverter(DAC).TextTextWordsNotesDiscussionHome2.DomainsInDSP,engineersusuallystudydigitalsignalsinoneofthefollowingdomains:timedomain(one-dimensionalsignals),spatialdomain(multidimensionalsignals),frequencydomain,andwaveletdomains.Theychoosethedomaininwhichtoprocessasignalbymakinganinformedassumption(orbytryingdifferentpossibilities)astowhichdomainbestrepresentstheessentialcharacteristicsofthesignal.Asequenceofsamplesfromameasuringdeviceproducesatemporalorspatialdomainrepresentation,whereasadiscreteFouriertransformproducesthefrequencydomaininformation,thatis,thefrequencyspectrum.TextTextWordsNotesDiscussionHome3.TimeandspacedomainsThemostcommonprocessingapproachinthetimeorspacedomainisenhancementoftheinputsignalthroughamethodcalledfiltering.Digitalfilteringgenerallyconsistsofsomelineartransformationofanumberofsurroundingsamplesaroundthecurrentsampleoftheinputoroutputsignal.Therearevariouswaystocharacterizefilters;forexample:TextTextWordsNotesDiscussionHome1)A"linear"filterisalineartransformationofinputsamples;otherfiltersare"non-linear".Linearfilterssatisfythesuperpositioncondition,i.e.ifaninputisaweightedlinearcombinationofdifferentsignals,theoutputisasimilarlyweightedlinearcombinationofthecorrespondingoutputsignals.TextTextWordsNotesDiscussionHome2)A"causal"filterusesonlyprevioussamplesoftheinputoroutputsignals;whilea"non-causal"filterusesfutureinputsamples.Anon-causalfiltercanusuallybechangedintoacausalfilterbyaddingadelaytoit.TextTextWordsNotesDiscussionHome3)A"time-invariant"filterhasconstantpropertiesovertime;otherfilterssuchasadaptivefilterschangeintime.TextTextWordsNotesDiscussionHome4)A"stable"filterproducesanoutputthatconvergestoaconstantvaluewithtime,orremainsboundedwithinafiniteinterval.An"unstable"filtercanproduceanoutputthatgrowswithoutbounds,withboundedorevenzeroinput.TextTextWordsNotesDiscussionHome5)A"finiteimpulseresponse"(FIR)filterusesonlytheinputsignals,whilean"infiniteimpulseresponse"filter(IIR)usesboththeinputsignalandprevioussamplesoftheoutputsignal.FIRfiltersarealwaysstable,whileIIRfiltersmaybeunstable.TextTextWordsNotesDiscussionHomeAfiltercanberepresentedbyablockdiagram,whichcanthenbeusedtoderiveasampleprocessingalgorithmtoimplementthefilterwithhardwareinstructions.Afiltermayalsobedescribedasadifferenceequation,acollectionofzeroesandpolesor,ifitisanFIRfilter,animpulseresponseorstepresponse.TextTextWordsNotesDiscussionHomeTheoutputofalineardigitalfiltertoanygiveninputmaybecalculatedbyconvolvingtheinputsignalwiththeimpulseresponse.TextTextWordsNotesDiscussionHome4.FrequencydomainSignalsareconvertedfromtimeorspacedomaintothefrequencydomainusuallythroughtheFouriertransform.TheFouriertransformconvertsthesignalinformationtoamagnitudeandphasecomponentofeachfrequency.OftentheFouriertransformisconvertedtothepowerspectrum,whichisthemagnitudeofeachfrequencycomponentsquared.TextTextWordsNotesDiscussionHomeThemostcommonpurposeforanalysisofsignalsinthefrequencydomainisanalysisofsignalproperties.Theengineercanstudythespectrumtodeterminewhichfrequenciesarepresentintheinputsignalandwhicharemissing.TextTextWordsNotesDiscussionHomeInadditiontofrequencyinformation,phaseinformationisoftenneeded.ThiscanbeobtainedfromtheFouriertransform.Withsomeapplications,howthephasevarieswithfrequencycanbeasignificantconsideration.TextTextWordsNotesDiscussionHomeFiltering,particularlyinnon-realtimeworkcanalsobeachievedbyconvertingtothefrequencydomain,applyingthefilterandthenconvertingbacktothetimedomain.Thisisafast,O(nlogn)operation,andcangiveessentiallyanyfiltershapeincludingexcellentapproximationstobrickwallfilters.TextTextWordsNotesDiscussionHomeTherearesomecommonlyusedfrequencydomaintransformations.Forexample,thecepstrumconvertsasignaltothefrequencydomainthroughFouriertransform,takesthelogarithm,thenappliesanotherFouriertransform.Thisemphasizestheharmonicstructureoftheoriginalspectrum.Frequencydomainanalysisisalsocalledspectrum-orspectralanalysis.TextTextWordsNotesDiscussionHome5.Z-planeanalysisDigitalfilterscomeinbothIIRandFIRtypes.FIRfiltershavemanyadvantages,butarecomputationallymoredemanding.WhereasFIRfiltersarealwaysstable,IIRfiltershavefeedbackloopsthatmayresonatewhenstimulatedwithcertaininputsignals.TheZ-transformprovidesatoolforanalyzingpotentialstabilityissuesofdigitalIIRfilters.ItisanalogoustotheLaplacetransform,whichisusedtodesignanalogIIRfilters.TextTextWordsNotesDiscussionHome6.WaveletInnumericalanalysisandfunctionalanalysis,adiscretewavelettransform(DWT)isanywavelettransformforwhichthewaveletsarediscretelysampled.Aswithotherwavelettransforms,akeyadvantageithasoverFouriertransformsistemporalresolution:itcapturesbothfrequencyandlocationinformation.TextTextWordsNotesDiscussionHome7.ApplicationsThemainapplicationsofDSPareaudiosignalprocessing,audiocompression,digitalimageprocessing,videocompression,speechprocessing,speechrecognition,digitalcommunications,digitalsynthesizers,radar,sonar,financialsignalprocessing,seismologyandbiomedicine.Specificexamplesarespeechcompressionandtransmissionindigitalmobilephones,roomcorrectionofsoundinhi-fiandsoundreinforcementapplications,weatherforecasting,economicforecasting,seismicdataprocessing,analysisandcontrolofindustrialprocesses,medicalimagingsuchasCATscansandMRI,MP3compression,computergraphics,imagemanipulation,hi-filoudspeakercrossoversandequalization,andaudioeffectsforusewithelectricguitaramplifiers.TextTextWordsNotesDiscussionHome8.ImplementationDSPalgorithmshavelongbeenrunongeneral-purposecomputersanddigitalsignalprocessors.DSPalgorithmsarealsoimplementedonpurpose-builthardwaresuchasapplication-specificintegratedcircuit(ASICs)2.Additionaltechnologiesfordigitalsignalprocessingincludemorepowerfulgeneralpurposemicroprocessors,field-programmablegatearrays(FPGAs)3,digitalsignalcontrollers(mostlyforindustrialapplicationssuchasmotorcontrol),andstreamprocessors.TextTextWordsNotesDiscussionHomeDependingontherequirementsoftheapplication,digitalsignalprocessingtaskscanbeimplementedongeneralpurposecomputers.TextTextWordsNotesDiscussionHomeOftenwhentheprocessingrequirementisnotreal-time,processingiseconomicallydonewithanexistinggeneral-purposecomputerandthesignaldata(eitherinputoroutput)existsindatafiles.Thisisessentiallynodifferentfromanyotherdataprocessing,exceptDSPmathematicaltechniques(suchastheFFT)areused,andthesampleddataisusuallyassumedtobeuniformlysampledintimeorspace.Forexample:processingdigitalphotographswithsoftwaresuchasPhotoshop.TextTextWordsNotesDiscussionHomeHowever,whentheapplicationrequirementisreal-time,DSPisoftenimplementedusingspecializedmicroprocessorssuchastheDSP56000,theTMS320,ortheSHARC.Theseoftenprocessdatausingfixed-pointarithmetic,thoughsomemorepowerfulversionsusefloatingpoint.ForfasterapplicationsFPGAsmightbeused.Beginningin2007,multicoreimplementationsofDSPshavestartedtoemergefromcompaniesincludingFreescale4andStreamProcessors,Inc.Forfasterapplicationswithvastusage,ASICsmightbedesignedspecifically.Forslowapplications,atraditionalslowerprocessorsuchasamicrocontrollermaybeadequate.AlsoagrowingnumberofDSPapplicationsarenowbeingimplementedonembeddedsystemsusingpowerfulPCswithmulti-coreprocessors.TextTextWordsNotesDiscussionHomesequencen.数列，序列；顺序；连续domainn.[计]域名；管辖范围

subfieldn.子域；子字段；分支estimationn.估计；评价；判断seismicadj.地震的；震撼世界的linearadj.直线的，线形的WordsWordsTextWordsNotesDiscussionHomenonlinearadj.非线性的compressionn.压缩，压紧，浓缩，紧缩discretizationn.离散化quantizationn.量子化；数字化；量化approximateadj.&v.近似的；接近于roundingadj.圆的，环绕的，凑整的WordsTextWordsNotesDiscussionHomeintegern.整数theoremn.[数]定理；定律derivationn.引出，导出；衍生quantizevt.使量子化enhancementn.增强；增加；提高；改善invariantadj.&n.无变化的；不变式WordsTextWordsNotesDiscussionHomeadaptiveadj.适应的；有适应能力的convolvev.卷，盘旋，缠绕cepstrumn.对数倒频谱，对数逆谱，倒频谱logarithmn.对数capturevt.俘获；夺取；夺得；引起（注意、想像、兴趣）synthesizern.合成者，合成物；合成器，综合器arithmeticn.算术，计算；算法WordsTextWordsNotesDiscussionHome[1]ADC：将模拟信号转换成数字信号的电路，称为模数转换器（简称a/d转换器或adc,analogtodigitalconverter），A/D转换的作用是将时间连续、幅值也连续的模拟量转换为时间离散、幅值也离散的数字信号，因此，A/D转换一般要经过取样、保持、量化及编码4个过程。在实际电路中，这些过程有的是合并进行的。模数转换器的种类很多，按工作原理的不同，可分成间接ADC和直接ADC。间接ADC是先将输入模拟电压转换成时间或频率，然后再把这些中间量转换成数字量，常用的有中间量是时间的双积分型ADC。直接ADC则直接转换成数字量，常用的有并联比较型ADC和逐次逼近型ADC。TextWordsNotesDiscussionNotesHome[2]application-specificintegratedcircuit（ASIC）：在集成电路界ASIC被认为是一种为专门目的而设计的集成电路。是指应特定用户要求和特定电子系统的需要而设计、制造的集成电路。ASIC的特点是面向特定用户的需求，ASIC在批量生产时与通用集成电路相比具有体积更小、功耗更低、可靠性提高、性能提高、保密性增强、成本降低等优点。TextWordsNotesDiscussionHome[3]FPGA（Field－ProgrammableGateArray），即现场可编程门阵列，它是在PAL、GAL、CPLD等可编程器件的基础上进一步发展的产物。它是作为专用集成电路（ASIC）领域中的一种半定制电路而出现的，既解决了定制电路的不足，又克服了原有可编程器件门电路数有限的缺点。

TextWordsNotesDiscussionHome[4]Freescale：飞思卡尔半导体（FreescaleSemiconductor）是全球领先的半导体公司，全球总部位于美国德州的奥斯汀市。专注于嵌入式处理解决方案。飞思卡尔面向汽车、网络、工业和消费电子市场，提供的技术包括微处理器、微控制器、传感器、模拟集成电路和连接。飞思卡尔的一些主要应用和终端市场包括汽车安全、混合动力和全电动汽车、下一代无线基础设施、智能能源管理、便携式医疗器件、消费电器以及智能移动器件等。在全世界拥有多家设计、研发、制造和销售机构。TextWordsNotesDiscussionHomeQuestionsfordiscussion1.InwhataspectscanDSPbeused?2.Whatarediscretizationandquantization?3.Accordingtothispassage,whatdoesfilteringmean?Whatkindsoffiltersdoyouknow?4.Whatcommonlyusedfrequencydomaintransformations?TextWordsNotesDiscussionHomeAnswerstoquestionsfordiscussionTextWordsNotesDiscussionHome1.InwhataspectscanDSPbeused?DSPcanbeusedinsuchaspectsasaudioandspeechsignalprocessing,sonar,radarandothersensorarrayprocessing,spectralestimation,statisticalsignalprocessing,digitalimageprocessing,signalprocessingfortelecommunications,controlofsystems,biomedicalengineering,seismicdataprocessing,amongothers.TextWordsNotesDiscussionHome2.Whatarediscretizationandquantization?Discretizationmeansthatthesignalisdividedintoequalintervalsoftime,andeachintervalisrepresentedbyasinglemeasurementofamplitude.Quantizationmeanseachamplitudemeasurementisapproximatedbyavaluefromafiniteset.TextWordsNotesDiscussionHome3.Accordingtothispassage,whatdoesfilteringmean?Whatkindsoffiltersdoyouknow?Filteringinthispassagereferstoamethodthroughwhichtheinputsignalcanbeenhanced,namely,themostcommonprocessingapproachinthetimeorspacedomain.Youmayknowthefollowingfilters:1)A"linear"filterisalineartransformationofinputsamples;otherfiltersare"non-linear".Linearfilterssatisfythesuperpositioncondition,i.e.ifaninputisaweightedlinearcombinationofdifferentsignals,theoutputisasimilarlyweightedlinearcombinationofthecorrespondingoutputsignals.2)A"causal"filterusesonlyprevioussamplesoftheinputoroutputsignals;whilea"non-causal"filterusesfutureinputsamples.Anon-causalfiltercanusuallybechangedintoacausalfilterbyaddingadelaytoit.TextWordsNotesDiscussionHome3)A"time-invariant"filterhasconstantpropertiesovertime;otherfilterssuchasadaptivefilterschangeintime.4)A"stable"filterproducesanoutputthatconvergestoaconstantvaluewithtime,orremainsboundedwithinafiniteinterval.An"unstable"filtercanproduceanoutputthatgrowswithoutbounds,withboundedorevenzeroinput.5)A"finiteimpulseresponse"(FIR)filterusesonlytheinputsignals,whilean"infiniteimpulseresponse"filter(IIR)usesboththeinputsignalandprevioussamplesoftheoutputsignal.FIRfiltersarealwaysstable,whileIIRfiltersmaybeunstable.TextWordsNotesDiscussionHome4.Whatcommonlyusedfrequencydomaintransformations?Therearesomecommonlyusedfrequencydomaintransformations.Forexample,thecepstrumconvertsasignaltothefrequencydomainthroughFouriertransform,takesthelogarithm,thenappliesanotherFouriertransform.Thisemphasizestheharmonicstructureoftheoriginalspectrum.Frequencydomainanalysisisalsocalledspectrum-orspectralanalysis.TextWordsNotesDiscussionHomeThankyou!Unit16TextBDigitalSignalProcessorDigitalSignalProcessorAdigitalsignalprocessor(DSP)isaspecializedmicroprocessor(oraSIPblock),withitsarchitectureoptimizedfortheoperationalneedsofdigitalsignalprocessing.TextTextWordsNotesDiscussionHomeThegoalofDSPsisusuallytomeasure,filterand/orcompresscontinuousreal-worldanalogsignals.Mostgeneral-purposemicroprocessorscanalsoexecutedigitalsignalprocessingalgorithmssuccessfully,butdedicatedDSPsusuallyhavebetterpowerefficiencythustheyaremoresuitableinportabledevicessuchasmobilephonesbecauseofpowerconsumptionconstraints.DSPsoftenusespecialmemoryarchitecturesthatareabletofetchmultipledataand/orinstructionsatthesametime.TextTextWordsNotesDiscussionHome1.OverviewDigitalsignalprocessingalgorithmstypicallyrequirealargenumberofmathematicaloperationstobeperformedquicklyandrepeatedlyonaseriesofdatasamples.Signals(perhapsfromaudioorvideosensors)areconstantlyconvertedfromanalogtodigital,manipulateddigitally,andthenconvertedbacktoanalogform.ManyDSPapplicationshaveconstraintsonlatency;thatis,forthesystemtowork,theDSPoperationmustbecompletedwithinsomefixedtime,anddeferred(orbatch)processingisnotviable.TextTextWordsNotesDiscussionHomeMostgeneral-purposemicroprocessorsandoperatingsystemscanexecuteDSPalgorithmssuccessfully,butarenotsuitableforuseinportabledevicessuchasmobilephonesandPDAsbecauseofpowerefficiencyconstraints.Aspecializeddigitalsignalprocessor,however,willtendtoprovidealower-costsolution,withbetterperformance,lowerlatency,andnorequirementsforspecializedcoolingorlargebatteries.TextTextWordsNotesDiscussionHomeThearchitectureofadigitalsignalprocessorisoptimizedspecificallyfordigitalsignalprocessing.Mostalsosupportsomeofthefeaturesasanapplicationsprocessorormicrocontroller,sincesignalprocessingisrarelytheonlytaskofasystem.SomeusefulfeaturesforoptimizingDSPalgorithmsareoutlinedbelow.TextTextWordsNotesDiscussionHome2.Architecture(1)SoftwareBythestandardsofgeneral-purposeprocessors,DSPinstructionsetsareoftenhighlyirregular;whiletraditionalinstructionsetsaremadeupofmoregeneralinstructionsthatallowthemtoperformawidervarietyofoperations,instructionsetsoptimizedfordigitalsignalprocessingcontaininstructionsforcommonmathematicaloperationsthatoccurfrequentlyinDSPcalculations.BothtraditionalandDSP-optimizedinstructionsetsareabletocomputeanyarbitraryoperationbutanoperationthatmightrequiremultipleARMorx86instructionstocomputemightrequireonlyoneinstructioninaDSPoptimizedinstructionset.TextTextWordsNotesDiscussionHomeOneimplicationforsoftwarearchitectureisthathand-optimizedassembly-coderoutinesarecommonlypackagedintolibrariesforre-use,insteadofrelyingonadvancedcompilertechnologiestohandleessentialalgorithms.Evenwithmoderncompileroptimizationshand-optimizedassemblycodeismoreefficientandmanycommonalgorithmsinvolvedinDSPcalculationsarehand-writteninordertotakefulladvantageofthearchitecturaloptimizations.TextTextWordsNotesDiscussionHome1)Instructionsets·multiply–accumulates(MACs,includingfusedmultiply–add,FMA)operations·usedextensivelyinallkindsofmatrixoperations·convolutionforfiltering·dotproduct·polynomialevaluation·FundamentalDSPalgorithmsdependheavilyonmultiply–accumulateperformance·FIRfiltersTextTextWordsNotesDiscussionHome·FastFouriertransform(FFT)·Instructionstoincreaseparallelism:·SIMD[1]·VLIW[2]·superscalararchitecture·Specializedinstructionsformoduloaddressinginringbuffersandbit-reversedaddressingmodeforFFTcross-referencingTextTextWordsNotesDiscussionHome·Digitalsignalprocessorssometimesusetime-stationaryencodingtosimplifyhardwareandincreasecodingefficiency.·Multiplearithmeticunitsmayrequirememoryarchitecturestosupportseveralaccessesperinstructioncycle·Specialloopcontrols,suchasarchitecturalsupportforexecutingafewinstructionwordsinaverytightloopwithoutoverheadforinstructionfetchesorexittestingTextTextWordsNotesDiscussionHome2)Datainstructions·Saturationarithmetic,inwhichoperationsthatproduceoverflowswillaccumulateatthemaximum(orminimum)valuesthattheregistercanholdratherthanwrappingaround(maximum+1doesn'toverflowtominimumasinmanygeneral-purposeCPUs,insteaditstaysatmaximum).Sometimesvariousstickybitsoperationmodesareavailable.·Fixed-pointarithmeticisoftenusedtospeeduparithmeticprocessing·Single-cycleoperationstoincreasethebenefitsofpipeliningTextTextWordsNotesDiscussionHome3)Programflow·Floating-pointunitintegrateddirectlyintothedatapath·Pipelinedarchitecture·Highlyparallelmultiplier–accumulators(MACunits)·Hardware-controlledlooping,toreduceoreliminatetheoverheadrequiredforloopingoperationsTextTextWordsNotesDiscussionHome(2)Hardware1)MemoryarchitectureDSPsareusuallyoptimizedforstreamingdataandusespecialmemoryarchitecturesthatareabletofetchmultipledataand/orinstructionsatthesametime,suchastheHarvardarchitectureorModifiedvonNeumannarchitecture,whichuseseparateprogramanddatamemories(sometimesevenconcurrentaccessonmultipledatabuses).TextTextWordsNotesDiscussionHome2)AddressingandvirtualmemoryDSPsfrequentlyusemulti-taskingoperatingsystems,buthavenosupportforvirtualmemoryormemoryprotection.Operatingsystemsthatusevirtualmemoryrequiremoretimeforcontextswitchingamongprocesses,whichincreaseslatency.TextTextWordsNotesDiscussionHome·Hardwaremoduloaddressing·Allowscircularbufferstobeimplementedwithouthavingtotestforwrapping·Bit-reversedaddressing,aspecialaddressingmode·usefulforcalculatingFFTs·Exclusionofamemorymanagementunit·Memory-addresscalculationunitTextTextWordsNotesDiscussionHome3.HistoryPriortotheadventofstand-aloneDSPchipsdiscussedbelow,mostDSPapplicationswereimplementedusingbit-sliceprocessors.TheAMD2901bit-slicechipwithitsfamilyofcomponentswasaverypopularchoice.TherewerereferencedesignsfromAMD,butveryoftenthespecificsofaparticulardesignwereapplicationspecific.Thesebitslicearchitectureswouldsometimesincludeaperipheralmultiplierchip.ExamplesofthesemultiplierswereaseriesfromTRWincludingtheTDC1008andTDC1010,someofwhichincludedanaccumulator,providingtherequisitemultiply–accumulate(MAC)function.TextTextWordsNotesDiscussionHomeIn1976,RichardWigginsproposedtheSpeak&SpellconcepttoPaulBreedlove,LarryBrantingham,andGeneFrantzatTexasInstrument'sDallasresearchfacility.Twoyearslaterin1978theyproducedthefirstSpeak&Spell,withthetechnologicalcenterpiecebeingtheTMS5100,theindustry'sfirstdigitalsignalprocessor.Italsosetothermilestones,beingthefirstchiptouseLinearpredictivecodingtoperformspeechsynthesis.TextTextWordsNotesDiscussionHomeIn1978,Intelreleasedthe2920asan"analogsignalprocessor".Ithadanon-chipADC/DACwithaninternalsignalprocessor,butitdidn'thaveahardwaremultiplierandwasnotsuccessfulinthemarket.In1979,AMIreleasedtheS2811.Itwasdesignedasamicroprocessorperipheral,andithadtobeinitializedbythehost.TheS2811waslikewisenotsuccessfulinthemarket.TextTextWordsNotesDiscussionHomeIn1980thefirststand-alone,completeDSPs–theNECµPD7720andAT&TDSP1–werepresentedattheInternationalSolid-StateCircuitsConference'80.BothprocessorswereinspiredbytheresearchinPSTN[3]telecommunications.TextTextWordsNotesDiscussionHomeTheAltamiraDX-1wasanotherearlyDSP,utilizingquadintegerpipelineswithdelayedbranchesandbranchprediction.TextTextWordsNotesDiscussionHomeAnotherDSPproducedbyTexasInstruments(TI)[4],theTMS32010presentedin1983,provedtobeanevenbiggersuccess.ItwasbasedontheHarvardarchitecture,andsohadseparateinstructionanddatamemory.Italreadyhadaspecialinstructionset,withinstructionslikeload-and-accumulateormultiply-and-accumulate.Itcouldworkon16-bitnumbersandneeded390nsforamultiply–addoperation.TIisnowthemarketleaderingeneral-purposeDSPs.TextTextWordsNotesDiscussionHomeAboutfiveyearslater,thesecondgenerationofDSPsbegantospread.Theyhad3memoriesforstoringtwooperandssimultaneouslyandincludedhardwaretoacceleratetightloops,theyalsohadanaddressingunitcapableofloop-addressing.Someofthemoperatedon24-bitvariablesandatypicalmodelonlyrequiredabout21nsforaMAC.MembersofthisgenerationwereforexampletheAT&TDSP16AortheMotorola56000.TextTextWordsNotesDiscussionHomeThemainimprovementinthethirdgenerationwastheappearanceofapplication-specificunitsandinstructionsinthedatapath,orsometimesascoprocessors.Theseunitsalloweddirecthardwareaccelerationofveryspecificbutcomplexmathematicalproblems,liketheFourier-transformormatrixoperations.Somechips,liketheMotorolaMC68356,evenincludedmorethanoneprocessorcoretoworkinparallel.OtherDSPsfrom1995aretheTITMS320C541ortheTMS320C80.TextTextWordsNotesDiscussionHomeThefourthgenerationisbestcharacterizedbythechangesintheinstructionsetandtheinstructionencoding/decoding.SIMDextensionswereadded,VLIWandthesuperscalararchitectureappeared.Asalways,theclock-speedshaveincreased,a3nsMACnowbecamepossible.TextTextWordsNotesDiscussionHome4.ModernDSPsModernsignalprocessorsyieldgreaterperformance;thisisdueinparttobothtechnologicalandarchitecturaladvancementslikelowerdesignrules,fast-accesstwo-levelcache,(E)DMAcircuitryandawiderbussystem.NotallDSPsprovidethesamespeedandmanykindsofsignalprocessorsexist,eachoneofthembeingbettersuitedforaspecifictask,ranginginpricefromaboutUS$1.50toUS$300.TexasInstrumentsproducestheC6000seriesDSPs,whichhaveclockspeedsof1.2GHzandimplementseparateinstructionanddatacaches.(tobecontinued)TextTextWordsNotesDiscussionHome(Cont