MAIN MENUPREVIOUS MENU---------------------------------SearchNext DocumentNext ResultPrevious ResultPrevious DocumentPrintOptimization of Signal Processing Software for Control SystemImplementationShuvra S. Bhattacharyya and William S. LevineAbstract— Signal processing plays a fundamental role inthe design of control systems — the portion of a digitally-implemented control system between the sensor outputs andthe actuator inputs is precisely a digital signal processor (DSP).Consequently, effective techniques for design and optimizationof signal processing software are important in achieving efficientcontroller implementations.Motivated by these relationships, this paper reviews tech-niques for modeling signal processing functionality in a mannerthat exposes aspects of application structure that are usefulfor mapping the functionality into efficient implementations.The paper then introduces some representative techniques thatoperate on such models to systematically derive optimizedimplementations from them.I. INTRODUCTIONToday almost all controllers are implemented digitally.In many complex or geographically distributed systems thecontroller operates over an automated communication net-work. Such embedded networked control systems presentnew challenges to the control system designer. This paperaddresses the interaction between two of these challenges.As in any control system it is essential to have the controlleroperate in real time. Each input must be there at the instantit is needed. Delayed or missing data can cause disasters.Hence, the timing of the computations must be preciselyspecifiable.Several issues operate to make the timing questionhard. Delays due to communication problems are relativelywell understood and techniques for preventing them devel-oped [1], [2], [3], [4]. However, there is also the possibilityof delays due to the software. The likelihood of such delaystends to increase as the complexity of the control algorithmsincreases. Because computing hardware is so cheap andinexpensive sensors are becoming more and more available,there is increasing pressure on the control system designer toincorporate more and more sophisticated functions into thecontroller. In addition, the control computer is likely to betime shared among many different control loops. The portionof a digitally-implemented control system between the sensoroutputs and the actuator inputs is a digital signal processor(DSP). This is well known [5]. This means that techniquesdeveloped for optimizing DSP software apply equally well tocontroller software. However, the objectives and criteria aredifferent. Specifically, in many signal processing applicationsS. S. Bhattacharyya is with the Department of Electrical and ComputerEngineering, and Institute for Advanced Computer Studies, University ofMaryland, College Park, MD, 20742, USA,[email protected]. S. Levine is with the Department of Electrical and Computer Engineer-ing, and Institute for Systems Research, University of Maryland, CollegePark, MD, 20742, USA,[email protected] main issue is minimizing the power needed. In mostcontrol applications the power used for signal processing isnegligible compared to that used for control. The automobileis a good example.Various researchers have studied the problem of real-time, software implementation of controllers, where a singleprocessor must be time-shared across multiple controllersin such a way that all controllers reliably keep up withtheir respective sampling periods. Caspi and Maler present arecent overview of such techniques in [6].The developments in this paper are largely complemen-tary to the existing body of work on real-time controllerimplementation. When each controller is viewed as a signalprocessing system, and attacked by state-of-the-art tech-niques for optimizing signal processing software, the task ofreal-time coordination becomes easier because the individualcontrollers consume less resources, (e.g., their worst-case ex-ecution times are significantly improved). This enables morefunctionality to be mapped to a given processing platform, orallows a cheaper platform — with slower processors — to beemployed for a given set of functionality. The advantages ofapplying a signal processing design flow are especially usefulwhen complex controllers, such as multirate controllers, areinvolved.II. MODELING OF SIGNAL PROCESSINGSOFTWARESignal processing system design is increasingly carried outthrough block diagram based environments. Block diagramrepresentations are a natural match for the signal flow graphdescriptions that are used by algorithm designers. In thecontext of efficient implementation, block diagram represen-tations of signal processing algorithms are attractive becausethey can be associated with coarse-grain dataflow semanticsthat expose opportunities for hardware and software opti-mization.Dataflow is a model of computation in which applicationsare represented as directed graphs whose vertices correspondto computations and whose edges specify logical channelsthrough which the output values of computations becomethe input values for subsequent computations. Dataflowprogramming is related to the actor model of concurrentcomputation [7], and vertices in a dataflow graph are oftenreferred to as actors of the graph. A dataflow actor canexecute whenever it has sufficient data on its input edgesto perform meaningful computation. Upon execution, therequired input data values are consumed from the input edgesProceedings of the 2006 IEEEConference on Computer Aided Control Systems DesignMunich, Germany, October 4-6, 2006ThB02.20-7803-9797-5/06/$20.00 ©2006 IEEE 1562and the resulting output data values are produced onto oneor more output edges.Unlike sequential programming languages, where execu-tion of operations follows the ordering of statements inthe program, a dataflow representation specifies nothingabout the ordering of operations apart from the precedenceconstraints that are implied by the flow requirements of theedges (the source actor of an edge must execute before thesink actor is allowed to consume the data that is producedby the source).Dataflow actors can have arbitrary complexity. In thedesign of signal processing systems, examples of practicalactors range from simple computations, such as addition andmultiplication, to signal processing blocks, such as FIR orIIR filters, and even complete subsystems, such as audio andvideo coders. For optimization of signal processing