ebook img

Handbook of Hardware/Software Codesign PDF

1314 Pages·42.528 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Handbook of Hardware/Software Codesign

HOPES: Programming Platform Approach for Embedded Systems Design SoonhoiHaandHanwoongJung Abstract HopeOfParallelEmbeddedSoftware(HOPES)isadesignenvironmentforem- bedded systemssupporting alldesign stepsfrombehavior specification tocode synthesis, including static performance estimation, design space exploration, and HW/SW cosimulation. Distinguished from other design environments, it introduces a novel concept of “programming platform” called Common Inter- mediate Code (CIC),which can beunderstood as ageneric execution model of heterogeneousmulti-processorarchitecture.IntheCICmodel,eachapplication is specified by a multi-mode Synchronous Data Flow (SDF) graph, called MTM-SDF. Each mode of operation is specified by an SDF graph and mode transition is expressed by an Finite-State Machine (FSM) model. It enables a designer to estimate the performance and resource demand by constructing staticschedules of the application with varying number of allocated processing elements at each mode. At the top level, a process network model is used to expressconcurrentexecutionofmultipleapplications.Aspecialprocess,called controltask,isintroducedtospecifythesystem-leveldynamismthroughanFSM modelinside.WithagivenCICmodelandasetofcandidatetargetarchitectures, HOPESperformsdesignspaceexplorationtochoosethebestHW/SWplatform, assuming that a hybrid mapping policy is used to map the applications to the processingelements.HOPESsynthesizesthetargetcodeautomaticallyfromthe CICmodelwiththemappinginformation.Theoveralldesignflowisverifiedby thedesignoftworeal-lifeexamples. Contents 1 Introduction............................................................... 2 2 CommonIntermediateCode(CIC)Model....................................... 7 2.1 ExtendedSDFModelforApplicationSpecification.......................... 9 S.Ha((cid:2))•H.Jung SeoulNationalUniversity,Gwanak-ro1,Gwanak-gu,Seoul,Korea e-mail:[email protected];[email protected] ©SpringerScience+BusinessMediaDordrecht2016 1 S.Ha,J.Teich(eds.),HandbookofHardware/SoftwareCodesign, DOI10.1007/978-94-017-7358-4_1-1 2 S.HaandH.Jung 2.2 DynamicBehaviorSpecificationattheTop-LevelSpecificationofthe CICModel........................................................... 14 3 DesignSpaceExplorationinHOPES........................................... 16 3.1 StaticSchedulingTechniqueofanMTM-SDFGraph......................... 18 3.2 DynamicMapping..................................................... 22 4 CICTranslator:AutomaticCodeSynthesisfromtheCICModel.................... 23 5 ExperimentalResults........................................................ 25 6 CurrentStatusandConclusion................................................ 28 References................................................................... 29 Acronyms API ApplicationProgrammingInterface BDF BooleanDataFlow CIC CommonIntermediateCode FSM Finite-StateMachine GA GeneticAlgorithm GUI GraphicalUserInterface KPN KahnProcessNetwork MTM ModeTransitionMachine NoC Network-on-Chip OS OperatingSystem PIM PlatformIndependentModel PSDF ParameterizedSynchronousDataFlow QoS QualityofService SADF Scenario-AwareDataFlow SDF SynchronousDataFlow SMP SymmetricMulti-Processing SysteMoC SystemCModelsofComputation WCRT Worst-CaseResponseTime 1 Introduction HOPES is under development as a generic design environment to support a wide rangeofembeddedsystemarchitecturesfromsystemonachip(SoC)tonetworked embeddedsystems.Startingfromatarget-independentbehaviorspecificationanda given set of candidate hardware architectures and available processing elements, we can explore the design space to find an optimal system configuration and mappingofapplications,andsynthesizethesoftwareandhardwarecomponentsin aunifiedframework.TheabstracttargetarchitectureassumedinHOPESconsistsof heterogeneousprocessingelementsthatareconnectedthroughanetwork.HOPES was originally introduced as a parallel programming environment for nontriv- ial heterogeneous multiprocessors with various design constraints on hardware cost, power, and real-time performance [15]. However, HW/SW codesign can HOPES:ProgrammingPlatformApproachforEmbeddedSystemsDesign 3 be naturally supported by HOPES, since a hardware IP can be regarded as a processingelement. As the system complexity incessantly grows with more processing elements integrated, design reuse of hardware platforms and IPs becomes the de facto practice to mitigate the difficulty of hardware validation. Then the HW/SW code- sign methodology is transformed to an embedded SW development methodology for a given hardware platform. Since the proportion of software components keeps increasing, HOPES puts more emphasis on the implementation of software components unlike our previous HW/SW codesign environment, PeaCE (Ptolemy extension as a Codesign Environment) [7]. While PeaCE focuses on the codesign of hardware and software modules that includes HW/SW partitioning, HW/SW cosynthesis, and HW/SW cosimulation, it takes little account of multi-processor architecturethatheavilyaffectstheparallelexecutionofsoftware. Asystematicdesignmethodologycanbeunderstoodasasequenceofstepsthat refine a higher level of abstraction to a lower level from initial specification to final implementation, which is summarized as the following phrase: “design is to represent”. Since refinement keeps the properties of the higher abstraction, how to specify the behavior is a key factor to distinguish various HW/SW codesign methods. Actor models that specify an application as a set of concurrent actors are widely adopted in the HW/SW codesign methodology since they express the potential parallelism of an application explicitly and parallelizing an application can be simply realized by mapping actors to the processing elements. Actors are connected to each other through channels that represent the flow of data samples between actors. Among many actor models, SDF is chosen as the baseline actor model of HOPES since it is a formal model that enables us to evaluate each design decision through static analysis. General introduction to data-flow models can be found in chapter (cid:2)“SysteMoC: A Data-Flow Programming Language for Codesign”. IntheSDFmodel[17],anapplicationisspecifiedwithadata-flowgraphwhere a node represents a function module, or a task, and an arc is a FIFO queue that deliversdatasamplesfromanoutputportofthesourcenodetoaninputportofthe destinationnode.Aninput(oranoutput)portisassociatedwithanintegernumber that indicates how many samples to consume (or to produce) per task execution; the number is called the sample rate of the port. Figure 1a shows a simple SDF graph representation of an application. A node becomes runnable only when all input arcs have no fewer samples queued than the specified sample rate. And the sample rates are fixed at run time in the SDF model. Then we can determine the mappingandschedulingoftheSDFgraph,whichistodeterminewhereandinwhat ordertoexecutethetasksonagiventargetarchitecture,atcompiletime.Foreach arc,wecandetermine therelativeexecution ratesbetween thesourcetaskandthe destinationtask,comparingtheoutputsamplerateofthesourceportandtheinput samplerateofthedestinationport.Forinstance,theexecutionrateoftaskC should betwicehigherthanthatoftaskAinFig.1a,inordertomakethenumberofsamples producedfromthesourcetaskthesameasthenumberofsamplesconsumedbythe 4 S.HaandH.Jung a b time 1 1 A B 2 PE1 A C C 2 1 1 PE2 B D D C D 1 1 c d 1 1 A B 1 AB AC BD CD 2 1 1 Buffer size 1 2 2 2 C D 1 1 Latency: 7 Fig.1 (a)AnexampleSDFgraphwithannotatedsampleratesonthearcs,(b)amappingand schedulingresultoftheSDFgraphontotwoprocessingelements,(c)anexampleSDFgraphthat hasabufferoverflowerror,and(d)thebufferrequirementofeacharcandtheestimatedlatency forthestaticschedulingresultof(b) destinationtask.AniterationofanSDFgraphisdefinedbythesetoftaskexecutions thatsatisfytherelativeexecutionratesoftaskswithminimumnumberofexecutions. AnSDFgraphissaidtobeconsistentiftherelativeexecutionratesoftasksare satisfiedforallarcs.IncasethereisanypossibilityofbufferoverflowonanyFIFO arcordeadlock,wecannotfindavalidscheduleofanSDFgraph.Figure1bshows anexampleofastaticschedulingresultonatargetarchitecturewithtwoprocessing elements.ForaconsistentSDFgraph,wecanrepeatthescheduleiterativelywithout bufferoverflow.Figure1cshowsanSDFgraphexamplethathasabufferoverflow error on arc AC by giving a wrong sample rate at the output port of node B. Such static analyzability is a very desirable feature for embedded system design since it enables us to detect a class of design errors by static analysis [17]. Moreover, once the mapping and scheduling decision is made, we can determine the buffer requirementsforallarcsandestimatethereal-timeperformanceoftheapplication. Wecaneasilycheckwhetherthedesignconstraintsonthehardwarerequirementand real-timeperformance canbe satisfiedornot.For instance, thebuffer requirement and the estimated latency associated with the static scheduling result of Fig.1b is summarizedinFig.1d. WhiletheSDFmodelhastheaforementionedbenefitsfromitsstaticanalyzabil- ity,ithasseverelimitationstobeusedasageneralmodelforbehaviorspecification. First,itisnotpossibletospecifythedynamicbehaviorofanapplicationsincethe sample rate of a port may not change dynamically. Second, it does not allow the use of shared memory for inter-node communication since the access order to the shared memory may change depending on the execution order of nodes. So the synthesized code may require much larger memory than a manually written code thatusuallyusessharedvariablesforcommunicationbetweenfunctionmodules.To overcomethoselimitations,wehaveproposedseveralextensionstotheSDFmodel inHOPES.WeusetheFSMmodelincombinationwiththeSDFmodeltoexpress the dynamic behavior of an application [10]. Furthermore, a special actor, called libraryactor[20],isintroducedtohandlesharedresourcesefficientlywithoutside effects. HOPES:ProgrammingPlatformApproachforEmbeddedSystemsDesign 5 FollowingtheheritageofheterogeneousmodelingofPtolemy[3]andPeaCE[7], HOPES uses a process network model at the top level to express concurrent execution of multiple applications. An application is modeled as a single process at the top level while the internal behavior of an application is specified by the extendedSDFmodel.Thesystem-leveldynamicbehaviorisspecifiedbyacontrol task whose behavior is specified by the FSM model in the top-level task graph. TheoverallspecificationmodelofHOPESiscalledtheCICmodel,whichwillbe explainedinthenextsectionindetail. There is a clear distinction between HOPES and the other model-based design environments. As the name implies, the CIC model is not defined as a front-end specificationmodel,butanintermediatespecificationmodel,meaningthatHOPES design environment can accommodate various front-end specification models as long as the front-end specification model can be translated into the CIC model. In fact, the CIC model can be understood as an execution model of tasks at the OperatingSystem(OS)level.AttheOSlevel,thesystembehaviorisrepresentedas asetoftasksnomatterwhatthefront-endspecificationmodelis.Communication and synchronization between tasks and scheduling of tasks are heavily dependent upon the underlying software platform and hardware platform. In HOPES, we proposetodefinetheexecutionmodeloftasksattheOSlevelandenforcethesystem tokeepthesemanticsoftheexecutionmodel.Then,theCICmodelwillbeableto run on any hardware and software platform since the execution model is defined asplatform-independent.So,weintroduceanewnotionofprogrammingplatform that hides the underlying software and hardware platform from the application programmer.AsaprogrambasedonthevonNeumannexecutionmodelcanberun onanyvonNeumannprocessor,anyprogrambasedontheCICexecutionmodelcan berunonanytargetarchitecturethatkeepstheCICexecutionmodel,weenvision. SincetheCICmodelisbasedonformalmodelsofcomputation,wecanenjoythe benefitsofstaticanalyzabilityofthosemodelstoreducethedesigntimeandcost. EventhoughthesameSDFmodelisusedforbehaviorspecificationinHOPES andPeaCE,thegranularityofanodeisquitedifferent.InHOPES,anSDFnodeisa unitofmappingandschedulingattheOSlevel.Itimpliesthatthenodegranularity shouldbeaslargeasathreadtomakethethreadswitchingoverheadinsignificant. On the other hand, PeaCE assumes mixed granularity of SDF nodes in the initial specification of an application and clusters them to define a thread or a task at the code synthesis step. In our previous work [15], an SDF graph specification of PeaCE has been translated to the CIC model by clustering the nodes to increase the granularity while keeping the potential parallelism as much as possible. It is possibletospecifythesystembehaviorwiththeCICmodelmanually,regardingthe CICmodelasthefront-endspecification.Inthiscase,itistheresponsibilityofthe programmertodefinethegranularityofthenodetotrade-offtheparallelismandthe schedulingoverhead. Intheconventionalmodel-drivensoftwaredevelopmentapproaches,thesystem behavior is specified by a Platform Independent Model (PIM) that is translated into a platform-specific model (PSM) manually for a given hardware platform. Then the target code is generated from the translated PSM. Even though the CIC 6 S.HaandH.Jung model is a platform-independent model, there is no need to translate it to a PSM since the CIC model can be executed in any hardware platform that supports the proposed execution model. The HOPES framework is distinguished from other model-based design frameworks that use a specific model of computation for behavior specification, which include Daedalus [19], DAL [21], CompSOC [6], andKoski [12].Ontheotherhand,asmentionedabove,HOPESdoesnotassume any specific model for behavior specification as long as it can be translated into the CIC model. Even though the CIC model is based on three different models of computation, its model composition rule is different from that of Ptolemy [3] whichallowshierarchicalcompositionofmodelswithoutlimitationonthedepthof hierarchyandonthekindsofmodels. Figure2showstheoveralldesignflowinHOPES.Theinputinformationconsists ofthefront-endspecificationofsystembehaviorandthesetofcandidateplatforms and hardware components. As explained above, the CIC model is generated manually or by an automatic translator from a different front-end specification of system behavior. We perform static analysis at the CIC level to detect the buffer overflow and deadlock errors for SDF specifications and to analyze the timing requirements that will be expressed in the proposed FSM model. The next step is to explore the architectural design space by selecting the hardware platform and processing elements and mapping the applications to the target architecture. Note that the profiling information of tasks for all kinds of candidate processing elementsisassumedtobegiven.TheDSEstepproducesahandfulofselectedtarget architectures and associated mapping results of applications. Note that if a target architecture is given as an input, it just determines the mapping of applications to thetargetarchitecture. Foreachcandidatesolution,theCICtranslatorgeneratesatargetCcodeforeach processor.WeneedtodevelopaseparateCICtranslatorforeachtargetarchitecture A set of CIC model applica(cid:2)ons (Behavior specifica(cid:2)on) Task profiling Sta(cid:2)c analysis informa(cid:2)on Mul(cid:2)-core host: Func(cid:2)onal Constraints Design space explora(cid:2)on Simula(cid:2)on (performance, resource, etc) (HW pla(cid:3)orm selec(cid:2)on & HSIM: HW/SW mapping/scheduling of A set of cosimula(cid:2)on applica(cid:2)on tasks candidate architectures IBM Cell CPU/GPU CIC translator Fig.2 DesignflowofHOPES HOPES:ProgrammingPlatformApproachforEmbeddedSystemsDesign 7 as we need a different compiler for a different von Neumann processor. A multi- core host processor is a base target platform for functional simulation. The CIC translator generates a multi-threaded C code for functional simulation. Another target platform that HOPES supports is the parallel simulator, called HSIM [26], that has been developed to simulate the target architecture without real hardware platform.Ahandfulofselectedarchitectureswillbeevaluatedmoreaccuratelyby HSIM simulation. Other target platforms that HOPES supports will be explained later. Therestofthischapterisorganizedasfollows.TheCICmodelwillbeexplained in the next section, which will be followed by Sect.3 that explains the mapping andschedulingtechniquesoftheCICmodel.Section4explainstheCICtranslator. PreliminaryexperimentalresultswillbediscussedinSect.5.Thecurrentstatusof HOPESdevelopmentispresentedwithconcludingremarksinSect.6. 2 CommonIntermediateCode(CIC)Model Asexplainedabove,theproposedCICmodeladoptsahierarchicalcompositionof different models of computation to express the system behavior at two different levels. At the top level, the CIC model expresses the system behavior with a process network. If an application can be specified by the extended SDF graph, theapplicationisencapsulatedasasupernodethatcontainsanextendedSDFgraph atthebottomlevel.Notethatthetop-levelprocessnetworkandtheextendedSDF modelthemselvescanbespecifiedinahierarchicalfashion. Thetop-levelprocessnetworkconsistsofCICtasksandchannelsasdepictedin Fig.3. There are two types of CIC tasks depending on the triggering condition of tasks:time-drivenanddata-driven.Atime-driventaskistriggeredbyapre-defined period that is given as a parameter. So it consumes the most recent sample from the input buffer channel. The input channels of a time-driven task are single-entry buffers that store the most recent data samples. An I/O task that interfaces with the outside is usually designated as a time-driven task. On the other hand, a data- driventaskistriggeredbythearrivalofdatasamplesontheinputports.Theinput channels of data-driven tasks are assumed to be FIFO queues. A data-driven task Task : (cid:2)me-driven (t)or Port: Fixed sample rate (f)or data-driven (d) unspecified (u)for varying sample rates Input task (q) Compute (b) Output task (t) task (d) (t) Channel: FIFO queue (q)or buffer (b) Fig.3 CICtaskgraph 8 S.HaandH.Jung TASK_INIT{ /* task initialization code */ }; TASK_GO { /* generic API for data read from an input port */ MQ_RECEIVE(port_name, data, size); ... /* generic API for system service request */ SYS_REQ(command, argument_list); ... /* generic API for data write to an output port */ MQ_SEND(port_name, data, size); } TASK_WRAPUP { /* task wrapup code */ }; Fig.4 CICtaskcodetemplate basically follows the semantics of the Kahn Process Network (KPN) model that performsblockingreadandnon-blockingwriteaccesstothechannels. As shown in Fig.4, the code template of a CIC task consists of three sections, enclosed by three keywords, TASK_INIT, TASK_GO, and TASK_WRAPUP. As the name implies, the TASK_INIT section is executed when the task is initialized and the TASK_WRAPUP section is executed just before it is terminated. The TASK_GO function is the main body that will repeat until the task is terminated. A CIC task accesses a channel with target-independent generic APIs, MQ_SEND andMQ_RECEIVE.TheMQ_RECEIVEAPIperformsblockingreadoperationto the associated input port while the MQ_SEND API performs non-blocking write operation to the associated output port. Since the CIC model is defined at the OS level, the CIC task assumes that there is a supervisor that schedules the CIC tasks andprovidessupervisoryservicestotheCICtasks.Thus,wedefineanothergeneric API, SYS_REQ, that requests a service to the supervisor. The first argument of theSYS_REQAPIdefinestheservicecommandwhoselistwillbeshownlater.In principle,aCICtaskdoesnotuseplatform-specificAPIsforportability.Thegeneric APIswillbetranslatedintotarget-specificAPIsatthecodegenerationstep.Wemay defineaCICtaskthatusesplatform-specificAPIsforefficientimplementationatthe expenseofportability. Thenumberofdatasamplesconsumedorproducedperexecutionofataskcan bespecifiedexplicitlyforeachinputoroutputport.Thesamplerateisspecified,if it is fixed and not changing at run time. Otherwise, the sample rate is assumed to bevaryingatruntime.Iftheinputsampleratesofallinputportsarespecified,the data-driven task becomes an SDF task that follows the execution semantics of the SDFmodel.IfalltasksinaCICsubgraphareSDFtasks,theCICsubgraphbecomes anSDFsubgraph.SincetheSDFmodelhasmanymeritsfromstaticanalyzability, itishighlyrecommendedtoidentifySDFsubgraphsasmuchaspossibleatthetop leveluntilnomoreSDFsubgraphcanbeidentified.Andeachsubgraphisreplaced byasupernodeatthetopleveltomakeitatwo-levelhierarchicalgraph.Toalleviate HOPES:ProgrammingPlatformApproachforEmbeddedSystemsDesign 9 thedifficultyofidentifyingtheSDFsubgraphautomatically,itisrecommendedto specifyanapplicationwiththeextendedSDFmodelmanuallyinsideasupernode. 2.1 ExtendedSDFModelforApplicationSpecification Inthissubsection,weexplainacoupleofextensionsthataremadetoovercomethe limitations of the SDF model while preserving the benefits of static analyzability. ThefirstextensionistousetheFSMmodeltoexpressthedynamicbehaviorofthe application.Thesecondistointroduceaspecialactor,calledlibraryactor,toallow taskstoshareHWorSWresources. 2.1.1 DynamicBehaviorSpecification There exist several approaches that have been proposed to increase the expression capabilityoftheSDFmodeltosupportintra-applicationdynamism.Oneapproach is to extend the SDF model itself. Dynamic Data Flow (DDF) and Boolean Data Flow(BDF)aretwoexamplesofthisapproachwheretheyintroducespecialkinds of nodes that may have varying sample rates [2]. Since BDF was proven to be Turing equivalent and DDF is a super set of BDF, their expression capability is maximal in theory. But they compromise some benefits of static analysis and efficientimplementation. Another approach is to express the dynamism of an application as a set of modes that the application takes and each mode is specified by an SDF graph. This approach assumes that the number of possible dynamic behaviors, or modes, is finite. Then the dynamic behavior can be expressed as dynamic mode change. Thereareseveralwaystospecifymodechange.InParameterizedSynchronousData Flow (PSDF), the dynamic behavior of a task is modeled by parameters, and the modechangeisrealizedbychangingtheparametersatruntimebeforestartingan iterationofaschedule[1].Anapplicationisspecifiedbyatupleofgraphs,initgraph andbodygraph,wherethebodygraphspecifiestheapplicationbehaviorandtheinit graphsetstheparametervaluestochangethemodebeforeanewiterationstarts. The other way is to combine the SDF model with another computation model, usually FSM to express the mode change. In the *-chart approach [5], each state of a finite state machine contains an SDF graph inside to make a hierarchical composition of SDF and FSM models. The state change in the FSM can be understood as the mode change of the SDF model. In the SystemC Models of Computation (SysteMoC) approach [9], a task is associated with an FSM that determines the execution behavior of the task. FSM-based SADF, shortly FSM- SADF, is a restricted form of Scenario-Aware Data Flow (SADF) that specifies eachmodeofoperation,calledscenario,withanSDFgraph[22,23].AnSDFtask mayhavemultipleversionsofdefinitiondependingonthemodeofoperationwhile a special control actor, called detector, that has an FSM inside sends the control informationtonormalSDFtaskstochangethemodeofoperation. HOPESusesasimilarapproachasFSM-SADF;anSDFtaskmayhavemultiple behaviors and a tabular specification of an FSM, called Mode Transition Machine 10 S.HaandH.Jung (MTM),describesthemodetransitionrulesfortheSDFgraph.AnMTMisdefined as a tuple {Modes, Variables, Transitions} where Modes and Variables represent a set of modes and a set of mode variables respectively, and Transitions is a set of transitions that consists of the current mode, a Boolean function of conditions, and the next mode. A Boolean function of transition condition is defined by a simplecomparisonoperationbetweenamodevariableandavalue.AnMTM-SDF specification of an H.264 decoder is shown in Fig.5. The H.264 decoder has two modes of operation: I-frame and P-frame. In the I-frame mode, the sample rate of each port in red boxes becomes zero while the sample rate of each port in blue boxesbecomeszerointheP-framemode.TheMTMisquitesimplesinceitneeds to distinguish two modes of operation by a single mode variable. Remind that the granularity of a CIC task is large and the dynamic behavior inside a task is not visibleattheCIClevel.Thus,anMTMisnotcomplexingeneralforstream-based applications. Mode transition is enabled by setting the mode variable so as to satisfy the transitioncondition.Butactualmodetransitionoccursonlyattheiterationboundary of the SDF schedule. Since an SDF graph has a well-defined notion of iteration andeachtaskknowshowmanytimesitshouldbeexecutedineachiteration,mode transitioncanbeperformedautonomouslybyindividualtaskswithoutglobaltiming synchronization. A mode variable can be set by the hidden supervisor, which will bediscussedinthenextsubsection.Oradesignatedtaskmaysetthemodevariable. A stream-based application usually starts with parsing a header information that determinesthemodeofoperation,followedbyprocessingastreamofdata.Inthis case,theSDFtaskthatparsestheheaderinformationisdesignatedasaspecialtask thatmaychangethemodevariable.Tosatisfytherestrictionthatthemodetransition occurs at the iteration boundary, the designated task should be the first task in the SDFschedule.IntheH.264decoderofFig.5,RealFileHisdesignatedasthespecial taskthatdeterminesthemodeofoperation. TheinternalbehaviorofanSDFtaskshouldbedefinedmanuallydependingon themodeofoperation.ThecodeskeletonofanMTM-SDFtaskisshowninFig.6; a task first checks the current mode of its MTM before starting the next iteration. Ifitisdesignatedasaspecialtask,itmaychangethemodevariable.Basedonthe modeofoperation,thesampleratesofSDFgraphcanbechanged.Forinstance,the sampleratesfortheinputandoutputarcsofIntraPredY/U/V tasksbecomeallzero for the P-frame mode and the sample rates for the output arcs of InterPredY/U/V tasksbecomezerofortheI-framemodeofoperation.Notethatthefeedbackinput arcsofInterPredY/U/Vtasksdonotchangethesampleratessincetheyneedtostore thepreviousframefedbackfromtheDeblocktaskevenintheI-framemode. 2.1.2 LibraryTask Inadditiontodynamicbehaviorspecification,anotherextensionismadetotheSDF graph by introducing a special task, called library task, to allow the use of shared resources inthe SDF model. A librarytaskis asharable and mappable object that definesasetoffunctioninterfacesinside.Figure7showsanSDFgraphthatconsists of three normal SDF tasks (T1–T3) and three library tasks (L4–L6). Connection

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.