UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Processor architectures for synthetic aperture radar Meisl, Peter G. 1996

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


831-ubc_1996-0251.pdf [ 7.8MB ]
JSON: 831-1.0064828.json
JSON-LD: 831-1.0064828-ld.json
RDF/XML (Pretty): 831-1.0064828-rdf.xml
RDF/JSON: 831-1.0064828-rdf.json
Turtle: 831-1.0064828-turtle.txt
N-Triples: 831-1.0064828-rdf-ntriples.txt
Original Record: 831-1.0064828-source.json
Full Text

Full Text

Processor Architectures for Synthetic Aperture Radar by Peter G. Meisl B . A . S c . Electrical Engineering, University of British Columbia, 1990 A THESIS S U B M I T T E D I N P A R T I A L F U L F I L L M E N T O F T H E R E Q U I R E M E N T S F O R T H E D E G R E E O F M A S T E R O F A P P L I E D S C I E N C E in T H E F A C U L T Y O F G R A D U A T E STUDIES D E P A R T M E N T O F E L E C T R I C A L E N G I N E E R I N G We accept this thesis as conforming to the required standard T H E UNIVERSITY OF BRITISH C O L U M B I A March 1996 © Peter Meisl, 1996 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of f3 l e c i y * c » \ E>n^ t n ^ e / i The University of British Columbia Vancouver, Canada Date DE-6 (2/88) - l : Abstract This thesis examines processor architectures for Synthetic Aperture Radar (SAR). SAR is a remote sensing technique that requires large amounts of computation and memory to form images. Processor architectures are sought that exhibit high performance, are scalable, are flexible, and are cost effective to develop and build. Performance is taken to be the primary figure of merit. The three facets of systems design, namely algorithm, technology, and architecture, are each examined in the process of finding the best architecture implementations. The examination of the algorithms is begun by reviewing SAR processing theory with the intent of summarizing the background for typical SAR proc­essor performance requirements. A representative set of SAR algorithms is analyzed to determine and compare their computational requirements and to characterize them in terms of basic digital signal processing (DSP) operation types.,The algorithm partitioning options for parallel processing are classified and compared. The area of technology is addressed by surveying computing technologies that are relevant to SAR processing. The technologies considered cover computational devices, memory devices and interconnec­tions. The knowledge of SAR algorithms and computing technology are then combined to study design consider­ations for the memory and interconnection subsystems of SAR processors. The requirement for two dimensional access to large data arrays is found to be the main complicating factor in memory design. The judicious use of wide data path widths, caches, interleaving and fast memory is discussed as a solution to the memory latency problem. The applicability of the most commonly used interconnection networks is examined. Buses, meshes, and crossbars are all found to be effective in certain situations. A classification of architectural approaches adapted to describe current and future SAR processors is used as the framework for the architecture selection. Feasible implementations of the architectural approaches are identified and their suitability for SAR is analyzed. Three approaches are identified as the most promis­ing: networks of workstations, DSP chips, and field programmable gate arrays (FPGAs). A detailed exam­ination is made of these three approaches. Four variations of the DSP approach are considered: general purpose DSPs, vector DSPs, a proposed optimised DSP, and digital filter devices. Each approach offers a different trade-off between performance and flexibility. The most cost effective architectures approaches are found to be those based on general purpose or vector DSPs. An original heterogeneous design is pre­sented that combines the strengths of these two approaches. ii Table of Contents Abstract i i Table of Contents i i i Lis t of Tables v i List of Figures v i i Symbols and Acronyms ix Acknowledgment x i i i 1 Introduction 1 1.1 Background 1 1.2 Objectives 2 1.3 Outline 2 2 Overview of S A R Processing 4 2.1 SAR Processing Requirements 4 2.1.1 General SAR Background 4 2.1.2 Azimuth Frequency 7 2.1.3 Resolution • 8 2.1.4 Range Cell Migration 9 2.1.5 Correlation Processing 10 2.1.6 Data Set Sizes 12 2.1.7 Summary 14 2.2 SAR Sensors 14 2.3 SAR Processors 16 2.4 Common SAR Algorithms 17 2.4.1 Time Domain Algorithm 19 2.4.2 SPECAN Algorithm 19 2.4.3 Range-Doppler Algorithm 21 2.4.4 Two Dimensional Frequency Domain Algorithms 22 2.4.5 Chirp Scaling Algorithm 23 2.4.6 Algorithm Variations 24 2.5 Computational Analysis of SAR Algorithms 25 2.5.1 Operation Types 25 2.5.2 Computational Requirements of SAR Algorithms 26 2.5.3 Model of Post-Processing Operations 36 2.5.4 Model of Range-Doppler Algorithm 38 2.6 Partitioning of SAR Processing 40 2.6.1 Considerations for Parallelism 41 2.6.2 Granularity of Parallelism 42 2.6.3 Classification of Approaches to Partitioning 43 2.6.4 An Example of Horizontal Data Partitioning: Range-Doppler Algorithm .........47 3 S A R Processor System Design 49 3.1 Design Methodology 49 3.2 System Requirements 50 3.3 System Design Considerations 51 3.4 Performance Prediction ; 52 iii 3.5 Architecture Selection Criteria 53 4 Architectural Approaches 55 4.1 Single Processor 55 4.2 Common Node Processor 55 4.3 Multiprocessor 57 4.4 Pipeline Processor 58 4.5 Multicomputer 59 4.6 SIMD Processor 60 4.7 Hardwired 61 5 Computing Technology 62 5.1 General Purpose Microprocessors 62 5.2 General Purpose DSPs 64 5.3 Special Purpose DSPs 65 5.3.1 Accelerated FFT Chips 66 5.3.2 Digital Filter Chips 67 5.4 Custom and Semi-Custom VLSI 67 5.5 Field Programmable Logic 68 5.6 Memory 68 5.7 Interconnect 70 6 Architecture Implementation Alternatives 73 6.1 Workstation 73 6.2 Accelerated General Purpose Computer 74 6.3 Supercomputer 75 6.4 Network of Workstations 75 6.5 DSP Uniprocessor and Multicomputer 77 6.6 Reconfigurable Computing Machine 78 6.7 Custom Algorithm Specific Processor 79 7 Memory System Design 81 7.1 High Performance Data Path Design 81 7.2 Memory Access in SAR Processing 82 7.3 Corner turns 82 7.3.1 Singe Processor Corner Turns 82 7.3.2 Multiple Processor Corner Turns 83 7.4 Memory Latency 84 7.4.1 Data Path Width 85 7.4.2 Caches 85 7.4.3 Interleaving 86 7.4.4 FastDRAMs 87 8 Interconnection Networks 89 8.1 System I/O 89 8.2 Inter-processor Communication 89 8.2.1 Bus 90 8.2.2 Mesh 91 8.2.3 Crossbar 92 8.2.4 Conclusion 93 iv 9 Examination of Architectures 94 9.1 Network of Workstations 94 9.1.1 NOW Model 94 9.1.2 Results 96 9.1.3 NOW Conclusions 97 9.2 General Purpose DSP Architecture 98 9.2.1 Processor 98 9.2.2 Memory 99 9.2.3 Interconnect 99 9.2.4 Sample Architecture 99 9.2.5 General Purpose DSP Conclusions 101 9.3 Vector DSP Architecture 101 9.3.1 Introduction to the LH9124 102 9.3.2 Single LH9124 Architectures 104 9.3.3 Range-Doppler Algorithm on a Single LH9124 Architecture 106 9.3.4 Multi-LH9124 Processor Architectures 107 9.3.5 Range Doppler Algorithm on Multi-LH9124 Architectures 108 9.3.6 Other Algorithms I l l 9.3.7 Vector DSP Conclusions 112 9.4 Optimized DSP Architecture 113 9.5 Digital Filter Chip Architecture 115 9.5.1 Architecture 115 9.5.2 Range and Azimuth Processing 117 9.5.3 Performance 118 9.5.4 Analysis 119 9.5.5 Digital Filter Conclusions 120 9.6 Fine Grained Parallel Architecture 121 9.7 A Heterogeneous Architecture 122 9.7.1 Vector Processor 124 9.7.2 Scalar Processor 125 9.7.3 Image Double Buffer 125 9.7.4 Data Movement 126 9.7.5 Implementation 126 9.7.6 Overall System 127 9.7.7 Range-Doppler Algorithm 128 9.7.8 Vector/Scalar Conclusions 130 9.8 FPGA Computing Machine 130 9.8.1 Standard Arithmetic 130 9.8.2 Distributed Arithmetic 131 9.8.3 FPGA Conclusions 133 9.9 Architecture Examination Conclusions 134 10 Conclusions and Future Work 138 10.1 Conclusions 138 10.2 Future Work 141 References 144 A Model of Range-Doppler Algorithm 150 List of Tables Table 2.1: Parameters of selected airborne radars 14 Table 2.2: Parameters of selected SAR satellites 15 Table 2.3: Attributes of some typical SAR processors 16 Table 2.4: Operation types 26 Table 2.5: Computational analysis of time domain algorithm 29 Table 2.6: Computational analysis of S P E C A N algorithm 30 Table 2.7: Computational analysis of range-Doppler algorithm 31 Table 2.8: Computational analysis of wave equation algorithm 32 Table 2.9: Computational analysis of chirp scaling algorithm 33 Table 2.10: Approximate computational requirements for real-time processing 35 Table 2.11: Computational analysis of typical post-processing operations 36 Table 2.12: Range-Doppler model parameters 38 Table 5.1: High performance microprocessors and their FFT performance 63 Table 5.2: High performance DSPs and their FFT performance ..64 Table 5.3: FFT processor ICs 66 Table 5.4: High speed filter ICs 67 Table 5.5: Effect of memory chip size on chip and module count 69 Table 5.6: Bandwidth of some D R A M types ; 70 Table 5.7: Personal computer and workstation buses 71 Table 5.8: High speed buses and related interfaces 72 Table 5.9: Other high speed interfaces 72 Table 9.1: LH9124 execution times 103 Table 9.2: Range-Doppler algorithm implementation with a single LH9124 106 Table 9.3: Time domain processor performance 118 Table 9.4: Times for vector processor operations 129 Table 9.5: Scaling the distributed arithmetic filter 132 Table 9.6: Architecture trade-off 135 vi List of Figures Figure 2.1: General SAR geometry 5 Figure 2.2: Trajectory of point response 6 Figure 2.3: Block processing of SAR data 12 Figure 2.4: Main processing steps in some common SAR algorithms 18 Figure 2.5: Sample post-processing sequence 24 Figure 2.6: Computational analysis of time-domain algorithm 29 Figure 2.7: Computational analysis of S P E C A N algorithm 30 Figure 2.8: Computational analysis of range-Doppler algorithm 31 Figure 2.9: Computational analysis of wave equation algorithm 32 Figure 2.10: Computational analysis of chirp scaling algorithm 33 Figure 2.11: Linear scale plot of computation rate as a function of L ^ 34 Figure 2.12: Log scale plot of computation rate as a function of L a z 34 Figure 2.13: Plot of fraction of operations that are due to FFTs as a function of L a z 35 Figure 2.14: Plot of fraction of operations that are due to interpolations as a function of L a z .36 Figure 2.15: Computational analysis of post-processing operations 37 Figure 2.16: Post-processing computation requirements for varying problem sizes 37 Figure 2.17: Flow diagram of range-Doppler algorithm 39 Figure 2.18: Computational analysis of detailed range-Doppler algorithm 40 Figure 2.19: Symbols used to depict approaches to parallelism 44 Figure 2.20: Vertical partitioning 44 Figure 2.21: Horizontal partitioning 45 Figure 2.22: Vertical-horizontal partitioning 46 Figure 2.23: Range block size 47 Figure 2.24: Parallel operation rates and efficiency 47 Figure 2.25: Input buffer memory size 48 Figure 4.1: Typical common node architecture 56 Figure 4.2: Typical multiprocessor architecture 57 Figure 4.3: Typical pipeline architecture 58 Figure 4.4: Typical multicomputer architecture 59 Figure 4.5: Typical SIMD architecture 60 Figure 5.1: Performance for FFTs of various lengths 65 Figure 5.2: Memory chip sizes and their year of peak production 69 Figure 7.1: Sample double buffered caching scheme 86 Figure 7.2: Some interleaved memory schemes 87 Figure 7.3: Average data array access time in slow direction when using fast R A M s 88 Figure 8.1: Bus interconnection 90 Figure 8.2: 8k x 4k image corner turning time on various interconnection networks 91 Figure 8.3: Mesh interconnection 91 Figure 8.4: Crossbar interconnection 93 Figure 9.1: NOW SAR processor 95 Figure 9.2: Speed-up for ten workstation network vs network throughput and computation/communication ratio vs network throughput 96 Figure 9.3: Network throughput required vs. number of workstations 97 Figure 9.4: Speed-up for ten workstation network vs. network bandwidth with partially overlapped communications 97 Figure 9.5: General purpose DSP based mesh processor 100 Figure 9.6: LH9124 digital signal processor 103 Figure 9.7: LH9124 single chip architecture 104 Figure 9.8: Some sample LH9124 to memory interconnections 105 Figure 9.9: Some options for LH9124 parallelism 108 Figure 9.10: Performance achieved for various numbers of LH9124 processors on the range-Doppler algorithm 109 Figure 9.11: Simple architectures with corner turn memories 109 Figure 9.12: Performance of horizontal partitioning with a corner turn on multiple LH9124 processors with the range-Doppler algorithm 110 Figure 9.13: Performance of vertical-horizontal partitioning with a corner turn on multiple LH9124 processors with the range-Doppler algorithm 110 Figure 9.14: Hypothetical single chip SAR processor 114 Figure 9.15: Time domain SAR processor block diagram 116 Figure 9.16: Time domain complex filter block 117 Figure 9.17: Modification to complex filter for sample-rate conversion 118 Figure 9.18: Block diagram of the vector/scalar architecture 123 Figure 9.19: V S A based SAR processing system 127 Figure 9.20: Performance/generality trade-off of processor architectures 135 Figure 10.1: Classification of SAR processor architectures 140 Figure A . 1: Computational analysis of range-Doppler algorithm (Part 1 of 3) 151 Figure A.2: Computational analysis of range-Doppler algorithm (Part 2 of 3) 152 Figure A.3: Computational analysis of range-Doppler algorithm (Part 3 of 3) 153 via Symbols and Acronyms Y antenna elevation angle X wavelength e squint angle instantaneous squint angle 9v antenna azimuth beamwidth azimuth time variable azimuth time at which sensor is closest to target \ beam center offset time AT) exposure time & scatterer reflectivity \ pulse duration \ coherent integration time BD Doppler bandwidth BP processed Doppler bandwidth c speed of light CVM complex vector multiply operation CVRVM complex vector / real vector multiply operation fo carrier frequency Ai azimuth frequency •Ale Doppler centroid / , range sampling rate A/x pulse bandwidth FilterCC filtering operation with complex data and complex coefficients FilterCR filtering operation with complex data and real coefficients FilterRR filtering operation with real data and real coefficients FFTld one dimensional FFT operation Sa azimuth processing efficiency Sr range processing efficiency GR ground range h platform altitude H height of antenna L a Z azimuth reference function length Lr range reference function length K a azimuth frequency rate Kr range frequency rate L antenna length nu azimuth matched filter update interval N az size of data array in azimuth Nfilter interpolation filter length Nr size of data array in range NRC maximum range curvature ix NRW maximum range walk Ns maximum azimuth shift between blocks OPref number of operations required to compute each point of azimuth matched filter p ( t ) baseband representation of transmitted pulse rQ range at time of closest approach rc range when target passes through center of beam 8r spatial range resolution R (r|) distance between platform and target Rm range at mid-swath Root root operation Round rounding operation RVS real vector scalar operation Sr correlation replica s[x (x) transmitted pulse SR slant range t range time variable Tp pulse repetition period TB azimuth time bandwidth product v platform velocity w (ri) antenna pattern in azimuth Wr extent of radar beam in slant range bx spatial azimuth resolution A C azimuth compression A L U arithmetic logic unit ARPA Advanced Research Projects Agency ASIC application specific integrated circuit A T M asynchronous transfer mode BPF block floating point CCRS Canadian Center for Remote Sensing CFFT complex FFT CISC complex instruction set computer C L B configurable logic block C N common node COTS commercial off-the-shelf C P U central processing unit C T M corner turn memory dB deciBel D E M digital elevation model D R A M dynamic R A M DSP digital signal processing or digital signal processor EDO extended data out ERS earth resources satellite FFT fast Fourier transform IFFT inverse fast Fourier transform FIR finite impulse response FLOP floating point operation F M frequency modulation F P G A field programmable gate array GP general purpose H D L hardware description language H W hardwired IC integrated circuit IIR infinite impulse response JPL Jet Propulsion Lab M A C multiply accumulate M B mega-byte Mb/s mega-bits per second MB/s mega-bytes per second MCW/s mega-complex words per second M C multicomputer M C M multi-chip module M D A MacDonald Dettwiler and Associates M I M D multiple instruction multiple data MISD multiple instruction single data M P mulitprocessor MPI message passing interface MW/s mega-words per second N A S A National Aeronautics and Space Administration NOW network of workstations PCI peripheral component interconnect OP operation (arithmetic) PE processing element PL pipeline processor PRF pulse repetition frequency P V M parallel virtual machine R A M random access memory RASSP rapid prototyping of application specific signal processors RC range compression R C M range cell migration R C M C range cell migration correction RF radio frequency R F M reference function multiply RISC reduced instruction set computer RT real-time SAR synthetic aperture radar xi SIMD single instruction multiple data SIR shuttle imaging radar SISD single instruction single data SP single processor S P E C A N spectral analysis SRA sample rate adjustment S R A M static R A M SRC secondary range compression VASP versatile array signal processor VLSI very large scale integration V M E versa module european V S A vector scalar architecture W E wave equation Acknowledgment I would like to thank my supervisors, Dr. M.R. Ito and Dr. I. Cumming, for their support. Many thanks are also due to all those in the High Performance Computing Lab who helped provide an enjoyable and con­structive work environment. I also thank the Natural Sciences and Engineering Research Council of Can­ada, the University of British Columbia, and MacDonald Dettwiler and Associates for their financial support. xiii 1 Introduction 1.1 Background Synthetic Aperture Radar (SAR) is a remote sensing technique for obtaining high resolution images of the earth's surface. SAR sensors are typically flown in airplanes or satellites. Since SAR is a radar sensor, it has the advantage that it can collect images at night and through cloud cover. There has been great growth in the field of satellite SAR in the last few years. Until quite recently, only one SAR satellite had ever existed: Seasat. However, now several SAR-capable satellites are in operation with several more coming. The study of SAR processing is particularly relevant to Canada's interests. Canada's vast geography has made it necessary to become a world leader in remote sensing. Canada has been at the forefront of research in SAR and SAR processing since the first SAR satellite. This was continued in 1995 with the launch of Canada's Radarsat satellite, which is a state-of-the-art SAR satellite. One of the traditional problems with SAR as a remote sensing tool is the huge amount of signal processing that is required to form an image from the raw data. Originally, optical techniques were used to process SAR data. Optical methods were superseded by digital processing in the late 1970s. Digital SAR process­ing systems have usually been either very slow or very complex and expensive. However, technology has advanced a great deal since the first digital SAR processors were built. The processing power of computer components has been catching up with the requirements posed by SAR processing, opening up new possi­bilities for SAR processor designs. This thesis explores some of these new possibilities. SAR processors are complex high performance signal processing systems. The designer of such a system must address each of the three fundamental facets of signal processing systems: algorithms, component technology, and architectures. Algorithm The characteristics of the SAR processing algorithm must be understood in detail. This includes the structure, computational requirements, and issues affecting the partitioning for parallel processing. Technology The current state of technology must be understood. This includes making a survey of the available components that might be used in building a SAR processor. 1. Introduction 1 Architecture The high computational requirements of SAR processing call for sophisticated computer architectures and parallel processing techniques. The various types of signal processor architectures must be exam­ined and the most promising ones identified. Each of these three areas offers the designer many choices and trade-offs. The system design process seeks an optimal combination of choices in each of the areas. 1.2 Objectives The main objective of this thesis is to examine the alternatives for SAR processor architectures and find the best ones based on current technology. The goal is to come up with SAR processor designs that • exhibit high performance, • are scalable, • are flexible, and • are cost-effective to develop and build. The approach used to meet these requirements is to look at each of the three areas of algorithm, technol­ogy, and architecture in turn, and then try to combine them into design solutions. Performance is used as the primary criteria in the examination of architectures. Since there is a wide range of relevant computing technologies, a similarly wide range of architectures must be examined. No wide ranging up-to-date treatment of the alternatives for SAR processor architec­tures is currently available. This thesis fills that gap and provides a basis for continued work in SAR proc­essor architectures. 1.3 Outline The structure of this thesis is summarized below. Section 1 provides the background, objectives and outline for the thesis. Section 2 addresses the algorithm aspect of SAR systems design. An overview of SAR processing is given with an emphasis on the issues that affect the performance requirements of a high-speed processor. Basic SAR theory, sensors, and algorithms are covered. The computational requirements of some common SAR 1. Introduction 2 algorithms are derived and compared. Finally, the partitioning of SAR processing for parallel implementa­tion is examined. Section 3 gives some background material that is necessary to set up the framework within which the study of architectures will occur. Among other things, the criteria by which the SAR architectures will be judged are established. Section 4 gives a high level classification of architectural approaches. This is based on the regular parallel processor classifications, but is adapted to suit SAR processors. Section 5 identifies the technology areas that are the key drivers behind SAR processor design, and sur­veys the state-of-the-art in these areas. The architectures from Section 4 and the technologies from Section 5 are combined to find some possible SAR processor architectures. These possibilities are discussed in Section 6 and the most promising ones are identified for further study. Section 7 examines some of the issues related to memory systems design for SAR processors. Memory access patterns during SAR processing are characterized. Various techniques for designing fast memory systems for SAR processors are presented. Interconnection networks for parallel SAR processors are considered in Section 8. The most promising ones are analyzed and compared. Section 9 performs a detailed examination of some possible SAR processor architecture designs. Their suitability for SAR processing is determined and, where possible, the results of a quantitative analysis are presented. The conclusions of the thesis are given in Section 10. 1. Introduction 3 2 Overview of SAR Processing This section gives a general overview of the SAR processing problem. The topics addressed include basic S A R theory (Section 2.1), types of SAR sensors (Section 2.2), types of S A R processors (Section 2.3), common SAR processing algorithms (Section 2.4), the computational requirements of SAR algorithms (Section 2.5), and the partitioning of SAR processing for parallel processing (Section 2.6). 2.1 SAR Processing Requirements This section is a summary of the main elements of SAR processing theory that are useful in deriving the requirements for SAR processor architectures. This information is widely reported in the literature but is not easily usable because it is either presented as part of a larger theoretical treatment of SAR or is given for only a specific algorithm implementation. This section attempts to provide a concise reference to the key equations of SAR processing theory. No attempt is made to give a thorough treatment of SAR imag­ing. SAR has been described and studied in numerous papers and several books [27] [36] [34] [87]. The dis­cussion in this section is based on [27] and [28]. 2.1.1 General SAR Background The general geometry of a SAR imaging system is shown in Figure 2.1. As the SAR platform travels over the surface of the earth it transmits pulses of radio frequency (RF) energy towards the ground. The pulses are repeated at the pulse repetition frequency (PRF). The SAR antenna is aimed in a direction approxi­mately perpendicular to the direction of motion and downwards at an angle y from the nadir, y is called the elevation or nadir angle. The SAR platform travels along its flight path at a velocity v and at an altitude of h. The angle between the center line of the antenna and the direction broadside to the platform's motion is called the squint or look angle, 6. The energy transmitted by the antenna illuminates a footprint on the ground, the area of which is determined by the size of the antenna, the range, and the radar wavelength. As the platform moves, the radar images a strip along the ground called the swath. The radar pulse is scattered by objects on the ground and some of the energy is reflected back towards the radar antenna where it arrives after a time delay proportional to the distance between the radar and target. The range of a target from the radar is r = t • c/2, where t is the time variable in the range direction, c is the speed of light and the factor of two accounts for round trip propagation delay. The range between the radar and a target at the time of their closest approach is denoted by rQ. The range between the radar and target when the target passes through the center of the antenna beam is denoted by rc and is equal to r 0 when 9 = 0. The dimension along the radar's path is called the azimuth direction and the azimuth time variable is denoted 2. Overview of SAR Processing 4 by T|. This means that the position of the radar in azimuth relative to the origin is equal to v • r\. The time Figure 2.1: General SAR geometry The received SAR signal is digitized and stored in the memory of the signal processor as a two dimen­sional array of samples. One dimension of the array represents distance in the slant range direction between the sensor and the target. The other dimension represents the azimuth direction. Each line of data with a common azimuth index represents data collected during one pulse repetition interval. A number of assumptions are commonly made in analysing SAR processing. It is convenient to consider only the case of a single point scatterer. This is allowable since the SAR system is linear and a complete image is made up of a superposition of point scatterers. In this section rectilinear geometry is assumed, that is, it is assumed the platform is travelling in a straight line. From Figure 2.1 it can be seen that the instantaneous range to a target with range of closest approach r 0 is: Since the range of a target depends on its position in azimuth, the returned echoes from a target migrate in range as the radar footprint sweeps over the target. The drift of the returns from a target in range is called range cell migration (RCM). The trajectory of a point target response in the received signal memory is 2. Overview of SAR Processing 5 shown in Figure 2.2 [28]. The figure depicts the case of a non-zero squint angle. R C M must usually be accounted for in SAR processing. R C M Correction (RCMC) constitutes one of the main complications in the processing of SAR data. If R C M C were not necessary, then SAR processing would be substantially simplified. received signal 2xJo 4-range migration curve 2R(r|-n„,r0)/c 1 1— n ° Tlo+11c Figure 2.2: Trajectory of point response The pulse transmitted by the radar can be expressed as s(x(%) = Re[p{%) e ) , (2.2) where p (%) is the complex baseband representation of the pulse, and / 0 is the carrier frequency. The wavelength of the carrier is X = c/f0. Typically a linear frequency modulation (FM) pulse with a rectan­gular amplitude of duration xp is used. An F M pulse or chirp is superior to using a simple rectangular pulse since it allows much greater range resolution to be attained for the same peak transmitter power. An F M chirp can be represented by p (T) = recti - \ - e , (2.3) where Kr is the frequency modulation rate. The bandwidth of the chirp is A / x = Kr-xp. The signal received by the antenna from a point scatterer in response to a transmitted pulse is a delayed version of the same pulse. The delay time for the echo to return to the antenna is 2 U U . (2.4) 2. Overview of SAR Processing 6 Using equations 2.2 and 2.4, the received SAR signal from a point scatterer at (r)0, tQ) in complex base­band form can be represented as -•/•••4-it-/o-*(Tl-T)o,r0) ( 2 • (T| - Tl 0 , r 0 ) ^ d(i\,x) = o ' - w ( T i - n 0 - r i ) - p [ t : J-e (2.5) where o' is the scatterer reflectivity and w (rj) is the antenna pattern in azimuth. The parameter TJ in the antenna pattern is the beam center offset time and represents the effect of the squint angle. The time that the target passes through the beam center is offset from the time of closest approach by r | c as can be seen in Figure 2.2. If the squint angle is zero then TJ is also zero. The basic intent of SAR processing is to recover a single point target response from the returned signal, which is represented by equation 2.5. This consists of a two-dimensional space variant convolution that compresses the target response of equation 2.5 to a single point. Usually the range and azimuth directions can be looked at separately during processing. In each dimension, the data needs to be convolved with a matched filter. In the range direction, a matched filter compresses the received chirp to a single point response. In azimuth, the situation is complicated by the fact that the returned signal from a single point target appears at a number of ranges during the time that it is illuminated by the radar. A l l the energy corre­sponding to a target needs to be collected and then compressed with another matched filter. This filter oper­ates by matching the azimuth variation in phase of the received signal for a point target. This variation is expressed by the appearance of the factor R (T) - T| 0, r 0 ) in the exponent in equation 2.5. Thus, the phase variation can be predicted from the geometry of the situation. The general theory of SAR processing will not be discussed further in this thesis. The next sections will focus on describing the equations that are used in defining the requirements for a SAR processor. 2.1.2 Azimuth Frequency A key step in analyzing the SAR processing problem is to examine the frequency content of the received signal in the azimuth direction. If the derivative of the azimuth phase term in the equation for the received signal (Equation 2.5) is taken, an expression for the instantaneous azimuth-frequency can be obtained. By also considering the geometry of the situation, the expression can be written as follows [28]. 2 - v / 0 - sin(9 (T I -T I 0 ) ) 2 - v sin(6 (T1-T10)) fn = c ° r A = X ' (2-6) 2. Overview of SAR Processing 7 where 8^  (T| - T| 0) is defined to be the instantaneous squint angle. At the beam center 6^  equals 6, the nominal squint angle. The signal energy is centered in azimuth frequency at the Doppler centroid: 2 v / 0 s i n ( e ) 2 - v s i n ( 8 ) The azimuth bandwidth for small squint angles is BD = 2-f. (2.8) The amount of Doppler bandwidth (BD) that is utilized by the SAR processor is called the processed band­width (Bp). For small squint angles, the azimuth frequency to azimuth time relationship can also be expressed as: where Ka is the azimuth F M rate as given by the following expression: *a = - ! - r - (2-10) The values of the Doppler centroid and the azimuth F M rate are needed by all SAR processing algorithms. Various methods of calculating and updating these parameters exist for different processing situations. 2.1.3 Resolution The resolution in the range and azimuth directions are fundamental parameters of every SAR system. In range, the spatial resolution in the slant range plane is ir = ^ m . (2.11, This shows that the range resolution is inversely proportional to the chirp bandwidth. In azimuth, the spatial resolution on the surface is 8x = ^- m, (2.12) P 2. Overview of SAR Processing 8 which comes from the fact that the azimuth resolution is inversely proportional to the processed azimuth bandwidth. An equivalent formulation (for rectilinear geometry) is: 8JC = | m. (2.13) This is a key result of SAR processing theory. The azimuth resolution is independent of the distance between antenna and target, and actually improves with a smaller antenna! 2.1.4 Range Cell Migration The expression for the range of the target from the radar (Equation 2.1) is often approximated by a second order Taylor series approximation [27]: R(T])~rc (T | - r , c ) ^ . i - O l - T l c ) 2 (2.14) or v 2 2 R(y\) = r c - v sine- (r\-i\c) + — • (r\-r\c) . (2.15) c The linear term of this approximation is referred to as range walk while the quadratic term is called range curvature. For zero squint, 9 = Oand TJ = 0, a simplified equation for the range to the target applies: In this case, the range walk component is zero. The amounts of range walk and range cell migration are critical in choosing a processing algorithm and designing a processor. The maximum amount of R C M to be expected in a processing scenario can be determined from the equations given for R (r\) above. The maximum amount of range walk is X-f.fax-B \ . f a X - B NRW = S ^ P s a m P l e s o r NRW = n i 4-T E m - (2-17) tin C a ~ -a One key decision in SAR processor design is whether or not it is necessary to correct for range walk. The amount of range walk is largely determined by the size of the squint angle. Correcting for range walk can 2. Overview of SAR Processing 9 imply an algorithm modification such as secondary range compression (SRC) in the range-Doppler algo­rithm. The criteria for needing SRC in range-Doppler is ^ - < 1, (2.18) N where TB = Bp- %c is the azimuth time bandwidth product and %c is the coherent integration time [27]. The maximum amount of range curvature expected is \-fs-B2 XB2 NDn = — t - samples or NDn = ^r- m . (2.19) R C 8 c K m m R C 1 6 - / T n a a By assuming Bp = BD and substituting Equations 2.8 and 2.10 into Equation 2.19, the amount of range curvature can also be written as: X rQ NRC = j m . (2.20) 8 • L The amount of range curvature is another factor that determines the type of algorithm that can be used. If range curvature is too large then the processing algorithm must correct for it. The rule of thumb for choos­ing an algorithm that cannot correct for range curvature (like SPECAN) is that the range curvature be less than one cell. Setting to one in Equation 2.19 results in the following criteria for not requiring range curvature correction: J s It can be seen from Equations 2.18 and 2.21 that the azimuth time bandwidth product is a key factor in determining which algorithm can be used. The time bandwidth product is a fundamental characteristic of a SAR system. Low time bandwidth product systems are typical of high frequency (e.g. X-band) SARs. 2.1.5 Correlation Processing When the data is processed, the correlation replicas in both range and azimuth have the form: 2 - T W . f - , 2 Sr(t) = e (2.22) 2. Overview of SAR Processing 10 The length of the range reference function is simply the length in samples of the transmitted chirp: When using fast convolution to convolve the SAR data with the correlation replica, Lr, output samples have to be discarded from each output block due to wrap-around. Thus a range processing efficiency factor can be defined as N -L S r = ^ W ~ 1 - (2-24) The range reference function does not have to be updated very often since the properties of the transmitted chirp usually do not vary greatly. The length of the azimuth reference function is given by X-rn-PRF 5 B • PRF D a Laz is proportional to the azimuth beamwidth, which is proportional to range. This means that a processor may have to use different azimuth matched filter lengths at different ranges. Some convenient vector length, Naz, needs to be chosen for azimuth processing that is long enough to pro­vide reasonable efficiency and not so long as to require unrealistic amounts of memory. Figure 2.3 illus­trates the splitting of the SAR data into azimuth blocks and the overlap regions that are required between blocks. The overlap between blocks must be adjusted according to the azimuth shift of each block relative to the next one. This azimuth shift is Ns = Waz-Laj'f- (2-26) a An azimuth processing efficiency factor can be written as N -L -N Sa = ^ N ° Z (2-27) az where Ns is taken to be the maximum block to block shift. In addition to considering the processing efficiency, the azimuth block length must also take into consider­ation the requirements for updating the azimuth reference function in the along-track direction. The update 2. Overview of SAR Processing 11 Overlap regions A Range \ \ \ \ \ \ k > Azimuth • (Direction of travel) Figure 2.3: Block processing of SAR data interval depends on the Doppler drift rates, the rates of change of Ka and / , and the amount of error that is allowable. The reference function update interval sets an upper limit on the azimuth block size Naz, although this is very rarely a restriction in practice. The azimuth reference function must also be Updated in the range direction. The update interval in range is driven by the slant range dependence of Ka. In order to keep errors resulting from incorrect Ka from caus­ing problems, the number of azimuth lines between updates is usually on the order of one to eight depend­ing on the error specification. 2.1.6 Data Set Sizes The choice of the data set size in the range direction is relatively straightforward. The range extent of the received signal is determined by the antenna pattern in range. The 3 dB beamwidth of the antenna is 8V = X/H radians, where H is the height of the antenna. If Rm is the range at mid-swath, then the extent of the radar beam in slant range is Slant range, SR, is related to ground range, GR, by SR = m. (2.29) sin y 2. Overview of SAR Processing 12 The conversion between slant range in meters, SR, and number of range samples, N, is 2'fs N = • SR samples. (2.30) c The size of the data array in range is also bounded by the pulse repetition period T = 1/ (PRF) . If the sampling rate in range is fs, then the maximum number of range samples is T • fs. The range sampling rate is chosen to be large enough to unambiguously sample the signal in range which has a bandwidth equal to the transmitted pulse bandwidth of A / x = %p • Kr. Thus the constraints on Nr, when processing the full range swath, are 1 2 - / r,rR • -—-<N <T f . (2.31) H m C r p Js v ' Typical values of Nr are 2000 to 10000 samples. Swath widths smaller than the full range swath can also be processed. This simply results in an image that covers a smaller extent in range with no loss of resolu­tion. In azimuth, the situation is significantly more complicated. The antenna 3 dB beamwidth in azimuth is 9 y = X/L radians where L is the length of the antenna. This means that the exposure time for small squint angles is X-rQ At) = - T — - sec. (2.32) L • v The size of the data array that spans the full exposure time is X-rQ Naz = AT| • PRF = -j—^ • PRF samples. (2.33) There are many other constraints that affect the choice of Naz. As mentioned in the previous section, the reference function update interval can impose an upper bound on Naz. The desired processing efficiency (Equation 2.27) imposes a lower limit on A ' f l z . If a value of Naz smaller than that called for in Equation 2.33 is used, the processed azimuth bandwidth is reduced and the azimuth resolution is degraded. If a larger value of N is used, the azimuth resolution does not improve significantly. A ' a z is often chosen to be a power of two to accommodate fast Fourier transform (FFT) algorithms. Some typical values are 2048, 4096 and 8192. 2. Overview of SAR Processing 13 2.1.7 Summary The preceding sections have given a brief summary of some basic equations related to SAR processing. This material gives the necessary background for the comparison of SAR algorithms and the derivation of their processing requirements in later sections. Additional information on the topics touched on here can be found in [27]. The next section gives some examples of SAR sensors, and uses the parameters of those sensors to illustrate the use of the equations given in this section. 2.2 SAR Sensors SAR sensors fall into two general categories: airborne and spaceborne. Although the theory for both types of sensors is similar, there are some differences in the data processing. This section lists some of the key characteristics of selected air- and spaceborne SAR sensors. This information was drawn together and pre­sented here since it is quite difficult to find detailed parameter lists of SAR sensors in the literature. A number of SAR sensors have been flown in aircraft by several countries. Table 2.1 gives a brief sum­mary of some of the parameters for a few airborne SARs [27] [67]. Table 2.1: Parameters of selected airborne radars System NASA/JPL CCRS/MDA China/MDA IRIS X2C Aircraft DC-8 CV-580 Lear-35A Frequency (GHz) 0.44,1.25,5.3 5.3,9.25 9.6 Polarization Quad Dual (like, cross) H Az resolution (m) / Looks 8/4 6/7 3 / 4 or 6 / 8 Range bandwidth (MHz) 20,40 35 60 Swath width (km) 10-18 18-65 10,20 Elevation angle y (deg) 10-65 0-87 not available Quantization (bits) 8 6 8 Airborne SARs typically offer higher resolution than satellite SARs but cover a smaller area per unit time. In airborne systems, the processing is simplified by the fact that R C M is not as great as for space-borne SARs, but is complicated by other issues such as motion compensation and greater variation in range of processing parameters. The first SAR equipped satellite was Seasat which was launched in 1978. Although its mission lasted only about 100 days, it collected a vast amount of data that was analyzed for many years. The success of Seasat lead to several SAR missions on the space shuttle (SIR-A, B, C series). Over ten years passed before another SAR satellite was launched. However, now several SAR satellites are in orbit and several more are 2. Overview of SAR Processing 14 planned. The next several years will see more SAR data becoming available than ever before. Table 2.2 summarizes the properties of several spaceborne SARs [27] [77] [72]. Table 2.2: Parameters of selected SAR satellites Satellite Seasat ERS-1/2 JERS-1 Radarsat ENVISAT Primary Parameters (approx.) Frequency Band (GHz) L(1.3) C (5.3) L(1.3) C (5.3) C (5.331) Wavelength X (m) 0.235 0.057 0.235 0.057 0.05627 Polarization HH V V HH HH HH, VV, HV.orVH Altitude (km) 800 785 568 793-821 786-814 Effective Velocity v (m/s) 7100 6928 7500 7071 6800 Nominal slant range (km) 850 850 690 844-1277 830-1130 Antenna Size L X H (m x m) 10.7x2.2 10.0 x 1.0 11.9x2.2 15.0 x 1.5 10.0 x 1.3 Transmitted pulse length %p (us) 34 37.1 35 42 11.6-41.3 Range freq. rate Kr (MHz/usec) 0.53 0.42 0.43 0.27,0.41, 0.71 0.03-1.4 Range sampling rate / (MHz) 45.52 (real) 18.96 17.076 12.9-32.2 19.208 PRF (Hz) 1647 1650 1506-1606 1270-1390 1580-2150 Slant range swath (km) 40 39 43 -25-250 -25-250 Ground range swath (km) 100 100 75 50-500 50-500 Elevation angle y (deg) 23 23 35 20-50 15-45 Expected squint 0 (deg) 4 0 4 4 0 Quantization (bits) analog 5 3 4 (BFP) 4 (BFP) Downlink rate (Mbps) analog 105 60 105 or 85 -100 Derived Parameters (approx.) Azimuth resolution (m) / Looks 23/4 25/3 30/4 28/4 30/4 Range Bandwidth (MHz) 19 13.5 15 11.6,17.3,30 0.4-16 Azimuth FM Rate KQ (Hz/s) 500 2340 690 1350-2100 1450-2000 Doppler Bandwidth BD (Hz) 1370 1385 1260 940 1350 Range matched filter length Lr 774 705 542-1352 223-793 Az matched filter length L f l z 4330 976 2930 622-970 1450-1900 Min. value of N needed to 4330 1154 2918 954 1800 process full exposure time Max. amount of range curvature (m) 51 3.5 34 2.3 4.5 2. Overview of SAR Processing 15 As can be seen from the table, the satellites have similar SAR system parameters in many cases. A major difference arises between the L-band satellites (Seasat and JERS-1) and the C-band satellites (ERS-1, Radarsat, and ENVISAT). Larger processor data arrays in azimuth direction are required to process the full target exposure time for the L-band satellites. The L-band systems also exhibit much larger amounts of range curvature, which must be properly handled by the processing algorithm in order to achieve high accuracy. The newer SAR satellites, like Radarsat and ENVISAT, have a much wider range of operating modes and a larger number of variable parameters. Radarsat has a steerable antenna beam that allows a range of swaths to be imaged. ENVISAT will add an alternating polarization mode to these capabilities. As a result of these multiple modes of operation, the values shown in Table 2.2 do not capture all the possibilities and should be seen as representative values. 2.3 SAR Processors Until about 1978, all satellite SAR processors used optical techniques. In the following three years, digital processing techniques almost completely replaced the optical ones. Digital processing offers major advan­tages in terms of accuracy and flexibility. However even today, there are occasions when size, weight or power constraints are crucial for which optical processing is useful. Modern optical processors are quite different from the ones used in the 1970s [45]. It is difficult to generalize a set of requirements for SAR processors since they have such a wide variety of uses. Table 2.3 lists some typical types of SAR processors and some typical attributes for each type [30]. Table 2.3: Attributes of some typical SAR processors Type Example SAR Processor Resolution Processing rate Required Processor Speed Primary Design Constraints Image Analy­sis/Research Radarsat Preci­sion Processor (RPP) med. to very high low varies (typically that achievable on a single workstation) flexibility Satellite Quicklook Radarsat Fasts-can low e.g. 100-180 m high(RT) 40 MOP/s (at least) speed Airborne MDA IRIS X2C high e.g. 1 to 5 m high (RT) 500 MOP/s (at least) speed, size, accu­racy Satellite Radarsat SAR Processor (RSARP) high e.g. 30 m high (RT) (or some fraction of) 3-4 GOP/s (or some fraction of) accuracy, speed 2. Overview of SAR Processing 16 Image analysis or research processors are typically used by researchers or scientists in labs for testing algo­rithm modifications or for verifying the results of production processors. Quicklook processors are typi­cally used for real-time processing of SAR data from a satellite as it is received in order to obtain a quick estimate of image quality and the overall correct operation of the system. Airborne processors are often mounted in the aircraft carrying the SAR sensor, and are used to obtain high quality images in real-time. Satellite SAR processors typically have the most demanding requirements. These processors are usually located at the satellite ground stations and are required to process all the data that the satellite generates at high levels of accuracy. Due to the extremely high processor speed required, most satellite SAR processors are built to achieve some fraction of real-time, for example, one tenth real-time. This is appropriate since the satellite is only in view of the ground station for a portion of the orbit, and full real-time processing is not required to process all the data acquired from the satellite. This limitation is also inherent in the design of the satellite which usually only has enough power to scan for a portion of its orbit. There are other types of SAR processors that have not been considered here. For example, spaceborne processors have not yet been built but are likely to appear in the future. Such systems will have extreme size, power, weight, and radiation immunity constraints. 2.4 Common SAR Algorithms This section briefly surveys a representative set of SAR processing algorithms. The purpose of this section is to introduce each algorithm's sequence of operations in preparation for the computational analysis in Section 2.5. The following common SAR processing algorithms or classes of algorithms are considered: • Time Domain algorithm • Spectral Analysis (SPECAN) algorithm • Range Doppler algorithm • 2-D Frequency Domain algorithms • Chirp Scaling algorithm These were chosen because they are either in common use or because they have been shown to be useful for Radarsat processing [65]. The main steps in each of these algorithms are summarized in Figure 2.4. This section characterizes each algorithm in terms of its major processing steps. Some indication of the accuracy of each algorithm is given but is not considered in detail. There are many other SAR algorithms that are not considered here. Many of these are simply variations of the ones discussed in this section. The algorithm descriptions in this section are based on [27], [65], [8], and [28]. 2. Overview of SAR Processing 17 10 Q <0 DC c t u_ •<3 t • cc — • o — • u_ Q CM Q CM Interi Q CM 1^ E o O to Q g> -in Figure 2.4: Main processing steps in some common SAR algorithms 2. Overview of SAR Processing 18 2.4.1 Time Domain Algorithm The time domain algorithm is conceptually the simplest SAR algorithm. It directly implements the two-dimensional correlation that is at the heart of SAR processing. The steps in the algorithm are as follows: Range Compression The radar signal data is convolved in the range direction with the range matched filter. Trajectory Extraction A curved trajectory corresponding to the path of a target through signal memory is extracted by interpo­lation for each output point. Azimuth Compression The data along each trajectory extracted in the previous step is convolved with the azimuth matched fil­ter. The main advantage of the time domain algorithm is that it can be the most accurate of all algorithms. Parameter variations, as well as any degree of antenna squint, can be handled. Another advantage is that any output sample grid can be chosen, which eliminates the need for later resampling. The main disadvan­tage of the time domain algorithm is that it is extremely computationally expensive. Time domain convolu­tions are required in both directions which require large amounts of computations for longer filter lengths. Additionally, for each output point, a complete trajectory must be extracted by interpolation from the data array before being compressed in azimuth. The huge number of operations required by this algorithm are difficult to achieve in a practical processor. The time domain algorithm is easier to use in applications that do not require R C M C . As a result, time domain approaches have been used in several airborne SAR processors. The development of extremely fast digital filter integrated circuits (ICs) has opened up the possibility of designing hardware implementa­tions of time domain algorithms with reasonable performance. This approach to SAR processor implemen­tation will be examined in a later section. 2.4.2 S P E C A N Algorithm S P E C A N is based on the idea of demodulating the returned signals with a linear F M signal which sepa­rates the returns from different targets to different frequencies. The targets can then be extracted using a single FFT. S P E C A N is more fully described in [63], [79], and [47]. Since S P E C A N is essentially a pulse compression technique, it can be used for either range compression, azimuth compression, or both. This section concentrates on the use of S P E C A N for azimuth compression. 2. Overview of SAR Processing 19 The steps in the algorithm are as follows: Range Compression (RC) and Linear RCMC Range compression is performed in any convenient manner, for example, either by fast convolution or by S P E C A N . Then a skew or interpolation is done in the range direction to compensate for range walk. Reference Function Multiply The data is multiplied by a linear F M waveform with an F M rate equal to the negative of the azimuth F M rate, Ka, in the azimuth direction. The reference function is updated as a function of cross-track position to track the Doppler parameter variation. FFT and Data Selection Short-length FFT operations are then applied to the data. This has the effect of compressing the returns from a target, which were confined to a single frequency by the previous step, into one bin. The output points of the FFT that correspond to valid data are selected, and the location of the next FFT is chosen. Radiometric Correction Radiometric corrections are neededbecause the look positions depend on azimuth time. This is usually a vector multiply operation. Deskew Deskewing is necessary because of the shift in range from the initial linear R C M C . This consists of an interpolation operation. Azimuth Sample Rate Adjustment Sample rate adjustment is needed to achieve a uniform output grid if different F M rates are used across the azimuth block. This is a resampling from the range-Doppler grid (fan-shaped) to an orthogonal grid. It consists of an interpolation operation. The output data after the FFT is not valid for all positions within the processed block. Values near the edges of the block are degraded and must be discarded. This decreases the processing efficiency. The frac­tion of data that is discarded increases as the required resolutions increases. Improved efficiency can be obtained by shortening the FFT length, but this comes with the penalty of decreased resolution. Some of the advantages of S P E C A N are listed below: • S P E C A N is the most efficient of all algorithms for satellite SAR in terms of both arithmetic and mem­ory. 2. Overview of SAR Processing 20 • S P E C A N can efficiently handle short vector lengths which can be important for special scanning modes like Radarsat's ScanSAR mode. The disadvantages of S P E C A N are as follows: • S P E C A N can only achieve about two-thirds of full resolution and still be computationally efficient. • Quadratic R C M C cannot easily be accommodated. • Extra processing is required to correct for radiometric scalloping that is present in the output image. • The output sample spacing is variable and depends on the F M rate of the azimuth matched filter, the PRF, and the FFT length. This complicates processing and necessitates a resampling operation. 2.4.3 Range-Doppler Algorithm The range-Doppler algorithm is probably the most widely used SAR algorithm. It was first developed by MacDonald Dettwiler and Associates (MDA) and the let Propulsion Lab (JPL) in 1979 for the processing of Seasat data [26] [93]. Since then, numerous variations and improvements have been developed, the most important of which is the addition of SRC [53]. Range-Doppler was designed to offer block processing efficiency in both the range and azimuth directions. Range curvature is corrected in the range-time / azi­muth-frequency domain which allows for block processing. SRC is a modification applied to the range compression matched filter to correct for range walk and other effects. The core steps in the range-Doppler algorithm are as follows: Range FFT An FFT is performed on the data in the range direction. Range Compression (RC) with SRC Range compression and SRC are performed by means of a vector multiply in the range direction. Range IFFT An IFFT is performed in the range direction. Azimuth FFT An FFT is performed in the azimuth direction. RCMC R C M C is performed by a shift and interpolation operation that lines up the target trajectories in mem­ory. 2. Overview of SAR Processing 21 Azimuth Compression (AC) Azimuth compression is accomplished with a vector multiply in the azimuth direction. Azimuth IFFT An IFFT in the azimuth direction completes the formation of the image. Some of the advantages of the range-Doppler algorithm are as follows: • Block processing efficiency is achieved for both R C M C and azimuth correlation. • Range variations of parameters can be accommodated with ease. Some of the disadvantages of range-Doppler are listed below: • There are problems with handling squint that are partially but not completely addressed by SRC. • Small errors are left by the fact the SRC is usually not made fully range and azimuth variant. • Interpolation is required to perform R C M C . This is complicated to implement correctly, and usually simple interpolators are used which introduce small errors. 2.4.4 Two Dimensional Frequency Domain Algorithms This is a class of algorithms that uses 2-D FFTs in their processing. Included in this class are the wave equation (WE) algorithm and its variations [21]. The steps in a typical algorithm of this class are as follows: 2-D FFT The data is transformed into the two dimensional frequency domain by a 2-D FFT. 2-D Reference Function Multiply (RFM) The data is then multiplied by a 2-D reference function. Change of variables or Stolt mapping This is an interpolation operation that accounts for the space variance of the correlation. 2-D IFFT The image formation process is completed by a 2-D IFFT. Some advantages of 2-D frequency domain algorithms are as follows: • Very accurate processing is possible for certain geometries. • Squinted SARs can be handled. 2. Overview of SAR Processing 22 Some disadvantages of 2-D frequency domain algorithms are as follows: • The algorithms usually require interpolations. • Large variations in the Doppler centroid are difficult to accommodate. 2.4.5 Chirp Scaling Algorithm The chirp-scaling algorithm is a relatively new algorithm that attempts to combine the advantages of the range-Doppler and wave-equation algorithms [78]. The main features of the algorithm are its high accu­racy, the lack of any interpolation operations, and its ability (with some modifications) to handle high squint [28]. The steps in the algorithm are as follows: Azimuth FFT The data is transformed into the range-Doppler domain with an FFT in the azimuth direction. Range Perturbation The data is then multiplied by a function in the range direction. This is the chirp scaling phase multiply that serves to equalize the range migration term of each scatterer's phase function with one at a certain reference range. Range FFT The data is transformed into the two dimensional frequency domain with an FFT in the range direction. RC, SRC, and RCMC RC, SRC, and R C M C is accomplished with a vector multiply in the range direction. Range IFFT An IFFT in the range direction is then applied to the data. Azimuth Compression (AC) and Phase Compensation Azimuth compression and phase compensation is performed with a vector multiply in the azimuth direction. Azimuth IFFT The image formation is completed with an IFFT in the azimuth direction. Some of the advantages of the chirp scaling algorithm are as follows: • No interpolations are required. 2. Overview of SAR Processing 23 • High squint can be handled. Some disadvantages of the algorithm are listed below: • Reference function calculations are more involved than in other algorithms. • It is difficult to handle large variations in the Doppler centroid. • Chirp scaling is not as well established as other older algorithms. 2.4.6 Algorithm Variations There are many variations and options that can be applied to the basic SAR image formation algorithms described in the previous sections. The processing of multiple looks is a common algorithmic enhance­ment. Multiple looks are used to reduce the amount of speckle that is a characteristic of SAR images in exchange for a loss in resolution. Multi-look processing consists of splitting the azimuth bandwidth and processing multiple independent views of the same area and then incoherently summing them. The algorithm descriptions given so far have only covered the basic image formation operations. There are many post-processing operations that can be applied to the outputs of any of the basic algorithms. These post-processing operations can form a significant portion of the computational requirements of a SAR processor. The exact nature of the post-processing operations applied to SAR images depends heavily on the final application. Some operations that may be performed are radiometric corrections and geometric corrections like resampling and geocoding. Radiometric and geometric calibrations are described in Sec­tions 7 and 8 of [27]. A sample sequence of post-processing operations is shown in Figure 2.5. Complex Image V Azimuth Resampling Corner Turn Range Resampling Detection V Detected and Resampled Image Figure 2.5: Sample post-processing sequence 2. Overview of SAR Processing 24 2.5 Computational Analysis of SAR Algorithms This section analyzes the computational requirements of the basic SAR processing algorithms presented in the last section. The results presented here extend those previously reported to a wider range of algorithms. A method of counting computing operations based on fundamental vector operations is given (Section 2.5.1). Expressions for the number of operations required by each step in each algorithm are derived (Sec­tion 2.5.2 and 2.5.3). Sample results are plotted for common processing situations. Finally a more detailed computational model of the range-Doppler algorithm is presented as a basis for architectural mappings in later sections (Section 2.5.4). 2.5.1 Operation Types This section describes the types of operations typically needed in SAR processing and the number of oper­ations (OPs) required for each operation on a vector of length N . To calculate computation amounts, a sin­gle operation (OP) on real operands is used as the base unit. An OP is defined as either an addition, subtraction, multiplication, or division. If floating point arithmetic is used in the processor, then the OPs can be interpreted as floating point operations (FLOPs). One dimensional FFT/IFFT The usual radix two algorithm is assumed with AV2 • \0g2N butterflies, 10 OPs per butterfly (four multiplies and six adds), and a total of 10 • N/2 • l og2^ OPs. Complex vector, element-by-element multiplication Each complex multiplication consists of four real multiplications and two adds, or 6 • N OPs. Complex vector / real vector, element-by-element multiplication Each complex/real multiplication consists of two real multiplications, or 2 • N OPs. Real vector scalar operations Each sample requires one operation, or N OPs. Filtering with Complex Data and Complex Coefficients Each filtered sample requires 4 • A j , / / e r OPs of multiplication and 4 • N^ilter - 2 OPs of addition for a total of 8 • Nfilter - 2 OPs. Filtering with Complex Data and Real Coefficients Each filtered sample requires 2 • N^iUer OPs of multiplication and 2 • (,N^il(er - 1) OPs of addition for a total of 2 • (2 • A ^ l 7 f e r - 1) OPs. This type of filtering commonly occurs in SAR processing as an interpolation operation. 2. Overview of SAR Processing 25 Filtering with Real Data and Real Coefficients Each filtered sample requires Njilter OPs of multiplication and (Nfilter - 1) OPs of addition for a total of 2-Nfnter-l OPs. Rounding functions Rounding functions are assumed to require 5 OPs. Root functions Square root functions are assumed to require 10 OPs. The operation types and their operation counts are summarized in Table 2.4. Table 2.4: Operation types Operation Type Symbol Number of Operations One dimensional FFT/IFFT FFTld(N) 10 • N/2 • log2N Complex vector element -by-element multiplication CVM (AT) 6 -N Complex vector / real vector element-by-element multi­plication CVRVM(N) 2-N Real vector scalar operations RVS(N) N Filtering with complex data and complex coefficients FilterCC (NfUter) *-Nfilter-2 Filtering with complex data and real coefficients FilterCR(Nfiher) 2 - "filter-U Filtering with real data and real coefficients FilterRR (Nfiher) 2 -Nfllt„-\ Rounding Round 5 Roots Root 10 2.5.2 Computational Requirements of SAR Algorithms This section analyzes the computational requirements of the algorithms introduced in Section 2.4. The operation types and counts described in the previous section are used to quantify the analysis. The SAR algorithms are used to process an array of data to form a single look complex image. The only exception is the S P E C A N algorithm which is modelled as processing a four look image. Both range and azimuth compression are included. This analysis is an extension of that in Section 9 of [27]. This section covers a more complete range of algorithms and counts the number of operations for an entire image array instead of per input sample. Only azimuth compression was included in the comparison in [27] and thus that comparison was able to show the trade-offs between the azimuth compression algorithms alone. Range compression is included in the analysis of all algorithms in this section since, in some algorithms, 2. Overview of SAR Processing 26 range compression cannot be separated as a distinct step. In some cases, range compression can done by a number of methods and is not restricted to the same method that is used for azimuth compression. In this section, it is assumed that range compression is performed by the same method as azimuth compression. This means that in the analysis of the S P E C A N algorithm, range compression is performed using SPE­C A N . In range-Doppler, range compression is performed by fast convolution. The key parameters in the analysis are Nr and N , the sizes of the data array in range and azimuth, and Lr and Laz, the lengths of the reference functions in range and azimuth. The data set size is varied during the analysis. The variation in the computational requirements as a function of the reference function lengths is examined. The reference function lengths are allowed to increase by powers of two from 2° to 2 1 4 . It is assumed that the lengths of the range and azimuth reference functions are normally equal. In order to keep the scenarios a little more realistic, Lr is not allowed to grow any larger than 1024. Range refer­ence functions are not usually longer than this. The data set sizes are kept consistent with the reference function lengths by always setting the dimensions of the data array to be twice the azimuth reference func­tion length: i.e. Nr = 2 • Laz and Naz = 2 • Laz. This arrangement results in processing situations that are somewhat artificial since in practice reference function lengths and array sizes cannot be chosen arbitrarily. However, the arrangement is useful to show how the amount of computation grows as the problem size increases. It is assumed all data values are complex numbers. Some of the other definitions and assumptions used in the algorithm models are defined below. • The data array sizes are assumed to be constant throughout the processing. This ignores changes in the data array dimensions due to filter throw-away regions, which can have a quite significant effect on data array sizes and, therefore, on the amount of computations. This assumption also implies that the azimuth processing step operates on the full amount of data output by the range processing step. Often only a portion of this data will need to be compressed in azimuth since the size of the data array was probably padded in the range direction in preparation for range compression. Since the purpose of this section is only to get rough estimates and compare the algorithms, these issues are not a big concern. • The analysis adds the updating of the azimuth reference function to the algorithm steps shown in Fig­ure 2.4. The azimuth reference function is assumed to be updated every nu azimuth lines. nu is taken to be four in this section. The number of operations required to calculate each value in an azimuth matched filter is OPref which is taken to be equal to 10 operations. 2. Overview of SAR Processing 27 • Any operations that are not part of the core SAR algorithm or have much smaller computation requirements than the major steps are ignored. This includes any preprocessing operations, calcula­tions of range reference functions and SRC coefficients, Doppler centroid calculations, calculation of R C M coefficients, and any post-processing operations like resampling or detection. • The overlap between azimuth processing blocks is taken into account with the azimuth processing efficiency factor ga. ga is given by equation 2.27 with Ns taken to be zero. It is assumed that range processing is done in a single block so the range processing efficiency is one. The effect of multi-block range processing is taken up in Section 2.6.4. • The length of the filters used in the interpolation operations required by several of the algorithms is represented by NjiUer and is taken to be equal to four. The PRF is assumed to be 1500 Hz. The analysis of each algorithm except S P E C A N assumes a single look product. This is justified because processing a single look product requires more computation than a multi-look product. The difference in the amount of computation in the case of the range-Doppler algorithm arises out of the fact that the multi-look case has multiple shorter azimuth FFTs and requires look summation. The single and multi-look cases can be compared by looking at the amount of computation required by the inverse azimuth FFT and look summation. In the single look case, the azimuth IFFT requires FFTld (N ) operations. In the multi-look case the number of operations is N l o o f F F T l d (N \ az_ \NlookJ + Naz (2.34) which can be written as: FFTld (NAZ) -NAZ- ^5 • log- 2 (Af / o o j t ) - l j . (2.35) This shows that if the number of looks is greater than two, the amount of computation is smaller for the multi-look case. This means that the single look case should be used when deriving worst case computa­tion requirements. Multi-look processing may, however, require more memory. For each algorithm, bar graphs are shown that depict the relative numbers of operations that are performed in each step of the algorithm and the relative number of operations that are due to each of the operation types of Section 2.5.1. These results are given for a reference function length of 2 1 0 : i.e. Laz = 1024 and L. = 1024. 2. Overview of SAR Processing 28 Time Domain Algorithm The number of operations required by each step in the time domain algorithm is shown in Table 2.5. Table 2.5: Computational analysis of time domain algorithm No. Step in Time Domain Algorithm Number of Operations 1 Range Compression Nr • Naz • FilterCC (Lr) 2 RCMC (interpolate) 3 Azimuth Compression Nr-Naz-FilterCC (Laz) ref Reference function generation Nr — • (OP , - L ) nu v ref a z > Figure 2.6 shows a graph of the relative number of operations in each step of the algorithm. The adjoining graph shows the number of operations that are due to each of the basic operation types. Computational Breakdown Operation Usage 2 3 Algorithm Step Ref. 100 80 60 40 20 FFT CVM CVRVM RVS Filter Round Root Scalar Operation Type Figure 2.6: Computational analysis of time-domain algorithm SPECAN Algorithm Table 2.6 shows the operations required by each step in the S P E C A N algorithm. S P E C A N is handled somewhat differently than the other algorithms. This is due to the fact that S P E C A N cannot be used to process apertures as long as the other algorithms. S P E C A N is modelled as processing an aperture that is only 1/4 as long. In order to make the data sizes handled by all the algorithms the same, it is assumed that S P E C A N is used to process four azimuth looks. This is a realistic assumption since S P E C A N becomes more efficient as the resolution decreases. In Table 2.6 the steps involved in range compression and the look summation step were added to the basic algorithm steps in Figure 2.4. A much more complete discus­sion of the computational requirements of S P E C A N can be found in [79]. 2. Overview of SAR Processing 29 Table 2.6: Computational analysis of SPECAN algorithm No. Step in SPECAN algorithm Number of Operations la Range reference function multiply ^looks ^az-CVM{Nr) lb Linear RCMC (interpolate) *Jlooks NrNazFilterCR(hrflUJ lc Range FFT N l o o k s - N a z - F F T l d ^ r ) 2 Azimuth reference function multiply Niooks^/CVM(Naz) 3 Azimuth FFT and Data Select Niooks-lP^-Nr-FFTldiNJ lylooks 1 3a Look Summation N -N -2 r az 4 Radiometric Correction (multiply) Nr-CVM(Naz) 5 Deskew (interpolate) Nr • Naz • FilterCR {NfUter) 6 Azimuth SRA (interpolate) 2-Nr.Naz.FilterCR(Nfnter) ref Calculate reference function Nr ~ 0 P r e f N a Z Figure 2.7 shows the relative number of operations required by each step in the algorithm and by each operation type. The amount of computation for S P E C A N is somewhat overestimated in comparison with the other algorithms. Some radiometric and geometric calibration operations that are required by S P E C A N are included in the analysis. These types of operations are not included in the analyses of the other algo­rithms even though the overall system requirements may make them necessary for those algorithms as well. Computational Breakdown Operation Usage %of Total Ops 40 35 30 25 20 15 10 5 0 38.9 3S.7 9.1 I 3 9 I | , 9 | 2 3 3a 4 Algorithm Step ref 80 70 60 %of 50 Total 40 Ops 30 20 10 0 CVM CVRVM RVS Filter Round Root Scalar Operation Type Figure 2.7: Computational analysis of SPECAN algorithm 2. Overview of SAR Processing 30 Range-Doppler Algorithm The number of operations required by each step of the standard range-Doppler algorithm is shown in Table 2.7. The azimuth processing steps are affected by the azimuth processing efficiency ga. Table 2.7: Computational analysis of range-Doppler algorithm No. Step in Range-Doppler Algorithm Number of Operations 1 Range FFT NazFFTld(Nr) 2 RC with SRC (multiply) Naz-CVM(Nr) 3 Range IFFT Naz-FFTld(Nr) 4 Azimuth FFT — N -FFTld(N) 6 a r 5 RCMC (interpolation) 7 • Nr • " a z • FUterCR (Nfi[tJ °a 6 AC (multiply) l.Nr.CVM(Naz) °a 7 Azimuth IFFT — N FFTld(N) o<3 r ref Azimuth reference function generation ~ ^ ( O P r e f L a z + FFTld(Naz)) The relative number of operations required by each step in the algorithm and by each operation type is shown in Figure 2.8. Computational Breakdown Operation Usage %of Total Ops 28 24 20 16 12 8 4 0 13.6 13.5 I 1.5 I 6.9 3 4 5 Algorithm Step ref 100 80 % of 6 o Total Ops 40 20 0 6.9 | FFT CVM CVRVM RVS Filter Round Root Scalar Operation Type Figure 2.8: Computational analysis of range-Doppler algorithm 2. Overview of SAR Processing 31 Wave Equation Algorithm Table 2.8 shows the number of operations required by each step of the wave equation algorithm. Table 2.8: Computational analysis of wave equation algorithm No. Step in Wave Equation Algorithm Number of Operations 1 2D FFT Naz • FFTU(Nr) + — -N- FFTld(Naz) &a r 2 2DRFM — • N • CVM (Nr) 8 a az 3 Stolt Mapping (interpolation) ±-N -Nn.FilterCRWfllter) °a r 4 2D IFFT Nn7 • FFTld(N ) + — • N • FFT\d{Nn.) °a r ref Reference Function Generation v\vOPrefLaz + FFT^d<<NaZ) + N^FFTld (N,)] The relative number of operations required by each step in the algorithm and by each operation type is shown in Figure 2.9. Computational Breakdown Operation Usage 36 32 28 % o . 2 4 Total 20 Ops 1 6 12 8 4 0 I 2.4 5.5 2 3 Algorithm Step ref 100 80 % of 6o Total Ops 40 20 I 5.5 FFT CVM CVRVM RVS Filter Round Root Scalar Operation Type Figure 2.9: Computational analysis of wave equation algorithm Chirp Scaling Algorithm The number of operations required by each step in the chirp scaling algorithm is shown in Table 2.9. 2. Overview of SAR Processing 32 Table 2.9: Computational analysis of chirp scaling algorithm No. Step in Chirp Scaling Algorithm Number of Operations 1 Azimuth FFT Nr-FFTld (NJ 2 Range Peturbation (multiply) Naz-CVM(Nr) 3 Range FFT Naz-FFTld(Nr) 4 RC, SRC, and RCMC (multiply) Naz-CVM(Nr) 5 Range IFFT Naz- FFTld (AT) 6 AC and Phase Compensation (multiply) Nr-CVM(Naz) 7 Azimuth IFFT Nr-FFTld (Naz) ref Azimuth Reference Function Generation ~ (OPrefLaz + FFTld(NJ) Note: All expressions are multiplied by — to obtain the actual number of operations The relative number of operations required by each step in chirp scaling and by each operation type is shown in Figure 2.10. Computational Breakdown Operation Usage %of Total Ops 24 20 16 12 8 4 0 2.4 3 4 5 Algorithm Step ref 100 80 % of 6 0 Total Ops 40 20 0 FFT CVM CVRVM RVS Filter Round Root Scalar Operation Type Figure 2.10: Computational analysis of chirp scaling algorithm The wave equation and chirp scaling algorithms are made to look more computationally expensive than other algorithms in this model since the efficiency factor is applied to both range and azimuth processing while it is only applied to azimuth processing in the other algorithms. In practice, the other algorithms will also use multiple range blocks and will , therefore, suffer an efficiency penalty in the range direction. 2. Overview of SAR Processing 33 Comparison of Algorithms The total computation rates for each of the algorithms studied in this section are collected and shown as a function of the problem size in Figure 2.11. The vertical axis shows computation rate in GOP/s. This is cal­culated assuming a PRF of 1500 Hz. The horizontal axis shows the problem size in terms of n where Laz = 2". A somewhat different perspective is obtained by plotting the vertical axis on a log scale as shown in Figure 2.12. Computation Rate (GOP/s) 1 1 1 1 '• ' Wave Equation 1 \ ^ Time Domain ' Chirp Scaling^ / ' / — 1 1 1 + / 1 1 J" / / + / ^ S P E C A N / / / ^Range Doppler y 1 « i. ••li.---*rr.T&-- 1*""^  I i 10 12 14 Figure 2.11: Linear scale plot of computation rate as a function of L, 'az Computation 1 1 0 Rate (MOP/s) 1000 Domain Wave Equation Scaling 12 14 Figure 2.12: Log scale plot of computation rate as a function of La 2. Overview of SAR Processing 34 The figures show that the time domain algorithm has requirements comparable to the other algorithms only for very short filter lengths. Once Laz increases past about 32, the time domain approach becomes expo­nentially more expensive than the other algorithms. The range-Doppler, wave equation and chirp scaling algorithms have computation rates that are similar for all values of Laz. As expected, S P E C A N is the most efficient of all the algorithms. The azimuth reference function length is about 2 1 2 for Seasat and about 2 1 0 for ERS-1 and Radarsat. This information allows approximate computation rates to be predicted for real-time processing of data from these satellites (see Table 2.10). Since the algorithm models are too simplistic to produce accurate compu­tation rates, only relative rates are given in the table. Table 2.10: Approximate computational requirements for real-time processing Seasat ERS-l/Radarsat (computation rate (computation rate Algorithm relative to SPECAN) relative to SPECAN) SPECAN 1 1 Wave equation, range-Doppler, chirp scaling 3.5 3 Time domain 500 200 FFTs form a major portion of all algorithms with the exception of the time domain algorithm. Figure 2.13 is a plot of the fraction of the total number of operations required by each algorithm that is due to FFTs as a function of the problem size. Figure 2.14 shows a similar plot for operations due to interpolations. (Laz=2n) Figure 2.13: Plot of fraction of operations that are due to FFTs as a function of L, 2. Overview of SAR Processing 35 Fraction of Total Operations 0 6 ^ ^ — i p - u u t " r u Y l_l l|J L J " ( Time Domain -V ^ ^ - S P E C A N -.^ / R a n 9 e Doppler Wave Equation- ~° ~ v -I I i i * ^ 4 0 2 4 6 8 10 12 14 " (Laz=2n) Figure 2.14: Plot of fraction of operations that are due to interpolations as a function of Laz 2.5.3 Model of Post-Processing Operations Post-processing typically consists of multiplications and additions for radiometric corrections, resampling for geometric corrections, and squares and square roots for detection. These operations are common to all algorithms except the time domain algorithm. Resampling is not required in post-processing for the time domain algorithm since the resampling can be incorporated into the main processing operations. This sec­tion examines the computational requirements of the post-processing steps shown in Figure 2.5. Table 2.11 gives the number of operations required for the post-processing steps. A resampling ratio of 1.5 is assumed in both directions. This results in output arrays that are 1.5 times as large as the inputs to the post-processing functions. Table 2.11: Computational analysis of typical post-processing operations No. Step in Post-Processing Number of Operations 1 Radiometric Correction Naz- CVRVM (AT) 2 Azimuth Resampling W r • FinalNaz • FilterCR {NfiUer) 3 Range Resampling FinalNr • FinalNaz • FilterCR (Nfilter) 4 Detection FinalNr • FinalNaz • 3 Note: FinalNaz = Int (1.5 • NK) and FinalNr = Int {1.5 • Nr) 2. Overview of SAR Processing 36 Figure 2.15 shows plots of the relative number of operations required by each step in the algorithm and by each operation type. Computational Breakdown Operation Usage 2 3 Algorithm Step 100 %ot Total Ops 40 20 FFT CVM CVRVM RVS Filter Round Root Scalar Operation Type Figure 2.15: Computational analysis of post-processing operations Figure 2.16 shows how the amount of computation in the post-processing steps grows as the data set size is increased. It can be seen that the amount of computation grows very quickly with increasing array sizes. Post-processing can form a significant portion of the total load on a SAR processor. 2 1.5 Computation Rate (GOP/s) i 0.5 0 2 4 6 8 10 12 14 " M l Figure 2.16: Post-processing computation requirements for varying problem sizes The examination of post-processing in this section has only looked at a very simple scenario. More com­plex situations involving geocoding to a map projection and geocoding to a digital elevation model (DEM) are given in Section 9.4 of [27]. However, even in the more complex cases, the basic operations remain the same. More sophisticated post-processors make more passes over the data and use more complex address-2. Overview of SAR Processing 37 ing schemes for the resampling operations. Other types of post-processing operations like image compres­sion are not considered here. 2.5.4 Model of Range-Doppler Algorithm The previous section gave very high level models of five representative SAR processing algorithms. This section focuses on one algorithm, range-Doppler, and performs a more careful analysis of the computa­tional and memory requirements. A more detailed model is necessary for the development of processor architectures in a later section. The range-Doppler algorithm was chosen since it is the most widely used algorithm for space-borne SAR processing. The previous section showed that the range-Doppler algorithm has computational requirements that are similar to the other high accuracy algorithms. In addition, range-Doppler is representative since it requires all of the operation types that appear in the other algorithms. The model in this section adds more careful bookkeeping of array sizes and some post-processing steps to the previous analysis. A more detailed model of azimuth matched filter generation based on that given in [64] is used. The steps in the detailed range-Doppler algorithm model are shown in Figure 2.17. The values of the processing parameters are given in Table 2.12 and are based on typical Radarsat parameters. Table 2.12: Range-Doppler model parameters Parameter Symbol Value Pulse Repetition Frequency PRF 1345 Hz Range swath length Nr 6000 Range matched filter length Lr 600 Range compression FFT length Nrfft 8192 Azimuth matched filter length LaZ 600 Azimuth Compression FFT length NaZ 4096 RCMC interpolator filter length N rcmc 8 Resampling filter length Nres 8 Oversampling ratio for resampling OVSM 1.4 Azimuth filter update interval nu 1 Computation word length "bits 32 2. Overview of SAR Processing 38 raw radar data in Range Processing Corner Turn Azimuth Processing Corner Turn Range Resampling and Detection 1.0 l/Q Balance 2.0 Range FFT 3.0 Range Reference Function Multiply 4.0 Range IFFT 5.0 Azimuth FFT 6.0 RCMC Shift and Interpolation 7.0 Azimuth Reference Function Multiply 8.0 Azimuth IFFT 12.0 Calculate Range Filter and SRC 13.0 Calculate Doppler Centroid + 14.0 Calculate RCM Shifts and Indices 31 15.0 Calculate Ambiguity ; + ' 16X5 Calculate Azimuth Filter 9.0 Azimuth Resampling 9.1 Apply Interpolator 9.2 Generate Indices 10.0 Range Resampling 1 . 11.0 Detection 10.1 Apply Interpolator 10.2 Generate Indices f 14T1 A Calculate RCM 1^2 N Round for Integer Shift 14.3 > Calculate Interpolator y Indices / 16.1 Calculate Filter in Time Domain 16.2 FFT of Filter 16.3 Weight Filter 16.4 Phase Compensation detected image out Figure 2.17: Flow diagram of range-Doppler algorithm 2. Overview of SAR Processing 39 The range-Doppler algorithm was modelled using a Mathcad document. The key portions are shown in Appendix A . The amounts of computation and memory required by each bubble in the flow diagram of Figure 2.17 was calculated. The number of operations required by each of the major steps in the algorithm is shown in the first graph in Figure 2.18. The adjoining graph shows the portion of the total computation that is due to each of the basic operation types. Computational Breakdown Operation Usage %of Total Ops [ M l ll ll irmi 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Algorithm Step 70 60 50 40 30 20 10 0 22.1 I 3.8 r~4^~il FFT CVM CVRVM RVS Filter Round Root Scalar Operation Type Figure 2.18: Computational analysis of detailed range-Doppler algorithm Figure 2.18 shows that the more detailed model has a similar computational breakdown as the simpler model whose results are shown in Figure 2.8. One of the main differences is the additional processing that is required by the azimuth and range resampling operations which account for about 17% of the total. The model presented here is in some ways a worst case model since nu was set to one. This means that a new azimuth reference function is calculated for each azimuth line. The memory requirements for the algorithm vary during the course of execution from a minimum of 140 M B at the output to 280 M B before detection. It is possible to reduce the maximum value by using shorter FFT lengths, by combining some algorithm steps, or by using a shorter wordlength. 2.6 Partitioning of SAR Processing The high computational requirements for SAR processors result in a need to partition the SAR algorithm to run on multiple processors. This section examines the partitioning of SAR algorithms and provides a more general treatment of this topic than is available in the literature. Section 2.6.1 discusses some characteris­tics of SAR processors that affect partitioning. Section 2.6.2 looks at the issue of granularity. Section 2.6.3 presents a classification of coarse grain partitioning approaches. 2. Overview of SAR Processing 40 2.6.1 Considerations for Parallelism Some of the key design issues that need to be addressed when partitioning a SAR algorithm for parallel processing are listed below. • granularity of parallelism: Is parallelism exploited at a fine grain level, at a coarse grain level, or both? • choice of horizontal and/or vertical parallelism at both the fine and coarse grain levels: What combination of horizontal and vertical parallelism is used? • data partitioning scheme: How is the data divided between the processors? • amount of parallelism: How many processors are used? What is the size of the data partitions and how many of them are there? Some of the options in these areas are discussed from an algorithmic perspective in the sections that fol­low. Specific examples of these characteristics are given when actual processor architectures are studied in Section 9. Some of the characteristics of an algorithm that affect the partitioning in a parallel SAR processor are listed below. • structure of the processing flow diagram: This determines the number and type of operations, data dependencies, as well as the scheduling and synchronization requirements. Most SAR algorithms decompose naturally into large sequential steps with fairly simple data dependencies and synchroni­zation requirements. • data array sizes in range and azimuth directions: Large arrays have higher computational require­ments but allow for more parallelism. • reference function lengths: The long reference function lengths required for high accuracy process­ing and the resulting overlaps required between data partitions have an influence on efficiency. • reference function update intervals: The filter update intervals can place upper bounds on data par­tition sizes. In SAR, this applies especially to the partitioning of azimuth compression since reference function updates are much more frequent during azimuth processing than during range processing. 2. Overview of SAR Processing 41 2.6.2 Granularity of Parallelism The granularity of a parallel system is the size of the units by which work is allocated to processors. For SAR processing, two general levels of granularity can be established. The bulk of SAR processing is made up of DSP operations like FFTs and vector multiplies. Fine grain parallelism cm be considered to be par­allelism that subdivides these fundamental operations. Coarse grain parallelism allocates each fundamen­tal operation to a single processor. Coarse grain parallelism in SAR processing offers the advantages of simpler scheduling, fewer synchroni­zation problems, and the ability to use commercial system components. In contrast, fine grain parallelism offers the potential of more parallelism and, therefore, higher speed-ups, and reduces load balancing prob­lems. However, fine grain parallelism requires low communication times between processing elements (PEs), and the lack of suitable commercial components usually implies custom VLSI designs. Both approaches have potential application in SAR processing. Some of the characteristics of SAR processing that make it suitable for coarse grain parallelism are as follows: • SAR requires large data sets that can easily be partitioned among PEs. • SAR algorithms are composed of distinct sequential large grain blocks that can be separated and pipe­lined. • Computationally expensive algorithm steps consist of standard DSP operations for which standard components exist. SAR is also suited for fine grain parallelism: • Most algorithm steps consist of standard DSP operations for which parallel structures are well known. • The primary data flow in SAR processing has no low level recursive structures like HR filters. The loops that do occur are at the highest level of the algorithm and involve less computationally expen­sive operations. This facilitates fine grained parallelism since it results in fairly simple data dependen­cies. As a result, scheduling and synchronization are greatly simplified. • The vector lengths are generally quite large so very long fine grained pipelining can be used with little overhead. Fine grain parallelism is potentially a very attractive approach with the promise of large speed-ups. As a result, the area has attracted a great deal of research. Some areas, like parallel FFTs, have been studied in great depth (see for example [43]). However, due to the high cost of implementing an algorithm in a fine 2. Overview of SAR Processing 42 grained parallel fashion, this approach is more suited to custom systems with special requirements. The high cost is mainly due to the lack of existing hardware and software components for implementing such systems. Most of the developments in the computer industry have focused on developing systems based on microprocessors which are coarse grain components. The lack of mass-produced fine grained parallel processors, industry standards, and software infrastructure have led to fine grained systems being limited to custom designs or research projects. Since one of the attributes of a good SAR processor architecture is low development cost, much of this thesis will focus on coarse grain approaches. However, some consider­ation will also be given to certain fine grained architectures. 2.6.3 Classification of Approaches to Partitioning This section gives a classification of common coarse grain partitioning approaches for SAR. The classifi­cation is based on the scheme given in Section 5.2 of [15]. There are many characteristics that need to be considered when classifying parallelizations of algorithms, and no one set covers all situations. The parti­tioning approaches in this section should be taken as starting points for describing a specific design. The approaches considered are: • vertical partitioning, • horizontal and horizontal-vertical partitioning, and • vertical-horizontal partitioning. These approaches are described in the paragraphs that follow. The partitioning approaches are also illus­trated in Figures 2.20 through 2.22. The meanings of the symbols that are used in these figures are shown in Figure 2.19. Vertical Partitioning The first approach to partitioning SAR processing is vertical partitioning, which is also known as pipelin­ing or temporal parallelism. Figure 2.20 shows an example of vertical partitioning. The granularity of the partitioning can be varied by changing the number of pipeline stages. This form of partitioning increases the throughput of the processor but increases its latency. Pipelining is well suited to SAR processing since most SAR algorithms consist of a number of sequential steps. 2. Overview of SAR Processing 43 Algorithm steps: (T) Data Transformation (IT) Memory Buffer A generic algorithm: ^ - K ^ T V + r T data partition overlap between partitions Image data block: £ " N CO $ * $ r ^ v v $ $ $ $ s s s s range B The generic algorithm mapped to range-Doppler: V V K R F F T W R M Y > ( ^ h-KAFFTl-W AM H>(AIFFTh B Res) Figure 2.19: Symbols used to depict approaches to parallelism No pipelining )-*(T)-•(B)-Coarse-grain pipelining Finer-grain pipelining *(T) Figure 2.20: Vertical partitioning Horizontal Partitioning The next approach is horizontal partitioning or data parallelism as shown in Figure 2.21. Horizontal parti­tioning divides the data set among the processors. Each processor performs essentially identical operations on subsets of the data. There are a number of options that can be applied to horizontal partitioning: • the use and granularity of pipelining in each of the parallel processors: • none (strictly horizontal partitioning) • pipelining with some amount of granularity (horizontal-vertical partitioning) • the orientation of the data partitioning: • azimuth subswaths or strips • range subswaths or strips 2. Overview of SAR Processing 44 • submatrices the size of the data partitions: Smaller partitions will allow for more parallelism but will result in decreased efficiency. V v *(BY >(TV \ B \ - +TB)-(A) Azimuth strips Data Partitioning Options (B) Range strips (C) Submatrices 3 E N CO N N N range -t 3 E N CO range > s ^ 5 5 range • Figure 2.21: Horizontal partitioning The data partitioning can be static or dynamic. In the static case, each PE operates on the same data set throughout. In the dynamic case, some or all of the data is exchanged between PEs during processing. Static partitioning makes use of distributed corner turns. Each processor has enough data to perform processing in both range and azimuth directions. This places a lower limit on the size of the data partitions since the partitions must be at least as long as the reference functions in each direction. In a dynamic parti­tioning, the option exists to perform corner turn operations in which PEs exchange data until each has a data set suitable for processing in the orthogonal direction. Vertical-Horizontal Partitioning Vertical-horizontal partitioning is a combination of the previous two partitioning methods. The processing is partitioned vertically through the use of pipelining and horizontally by dividing the data among the proc­essors. This approach can be used to take advantage of the fact that in many SAR algorithms, range processing is performed first, then the data is corner turned, and then azimuth processing is performed. The corner turn buffer is a natural synchronization point for the parallel processors. Vertical-horizontal parti­tioning and some of its data partitioning options are shown in Figure 2.22. Some of the options that can be chosen are as follows: • the use and granularity of pipelining in the individual processing sections 2. Overview of SAR Processing 45 Parallel Stage #1 Corner Parallel Stage #2 Turn Memory Data Partitioning Options Parallel Stage #1 Parallel Stage #2 Range strips (A) (B) (C) E CO range Azimuth strips range Range strips £ 'a range -Azimuth strips range-Azimuth strips range -Range strips E 'a range -Figure 2.22: Vertical-horizontal partitioning data partitioning: • range-azimuth • azimuth-azimuth • range-range • submatrices or other options the size of the data partitions 2. Overview of SAR Processing 2.6.4 An Example of Horizontal Data Partitioning: Range-Doppler Algorithm The range-Doppler algorithm described in Section 2.5.4 is partitioned into range subswaths and analyzed in this section. Each range line is split into sublines with enough overlap to account for the matched filter length and R C M C overlap. The range FFT size is chosen to be the next largest power of two. The range subswath widths and range FFT sizes are shown in Figure 2.23. The operation rate required for each subs-wath processor is plotted in Figure 2.24. The graph also shows the total amount of computation which is the sum of the computation done by each processor. The adjoining graph plots the efficiency as a function of the number of processors. It can be seen from the graphs that increasing the number of processors within regions where the range FFT length is constant decreases the efficiency significantly because the workload per processor does not decrease very much. The overhead imposed by the overlapped regions imposes a limit on the parallelization of the algorithm (see equation 2.24 for the range processing efficiency gr). 1 T 10000 8000 F number of range 6°oo samples 4000 2000 h ~l -Range FFT length 'Range subswath width _L _L 5 10 number of processors 15 Figure 2.23: Range block size operation 5 rate . (GOP/s) -1 1 1 - , -total operation rate —\ operation rate -per processor — -i i -5 10 number of processors 15 5 10 number of processors Figure 2.24: Parallel operation rates and efficiency 2. Overview of SAR Processing 47 The memory requirements of a processing system also increase as the number of processors is increased. A plot of the required size of the input buffer memory is plotted against the number of processors in Figure 2.25. It can be seen that the total amount of memory required rises linearly with the number of processors. 4501 1 size of input buffers (MB) 400 350 300 250 200 150 100 50 0 T T * \ total size of input buffers size of input buffer / for each processor _L J_ 5 10 number of processors 15 Figure 2.25: Input buffer memory size The restraints on parallelization imposed by the matched filter lengths in a static horizontal partitioning lead us to look for ways around this limitation. Some possible ideas and their drawbacks are: • use dynamic data partitioning to add a global corner turn step: The amount of inter-processor commu­nication will increase. • use shorter filter lengths: The accuracy of the processor will suffer. • use vertical-horizontal partitioning: The amount of inter-processor communication will increase and a centralized corner turn memory may become a bottleneck. • use an algorithm like S P E C A N that does not require overlaps: S P E C A N has precision limitations and has a more complicated data partitioning scheme. • use fine grain parallelism: The system design may be more difficult. 2. Overview of SAR Processing 48 3 SAR Processor System Design The SAR processing problem was outlined in the previous section. This section begins the process of designing a processor that provides a solution for that problem. SAR processors are large complex systems and, therefore, there are many issues that need to be considered during the design process. This section looks at some of these general design issues in preparation for a more specific look at SAR processor designs in the chapters that follow. General overviews are given of system design methodology (Section 3.1), system requirements (Section 3.2), selected design considerations (Section 3.3), and performance prediction (Section 3.4). Finally, the architecture selection criteria used in this thesis is presented (Section 3.5). The intent is not to cover these topics in depth but rather to identify them as important considerations in SAR processor design. A SAR processor is part of a very large system that in its entirety includes everything from the SAR sensor to the display, storage, and dissemination systems for the fully processed products. This thesis only looks at the signal processing subsystem. None of the other system components are considered. These compo­nents can include the communications system, frame synchronizer, tape drives, disks, display systems, backup and archival systems, higher level control functions, user interfaces, and other external interfaces. 3.1 Design Methodology Since SAR systems tend to be large and complex, methodologies used for designing other complex sys­tems are applicable. Methodology is a large topic and the discussion here is just a brief look at the steps involved in designing a SAR processor. More detailed information on signal processor design methodolo­gies can be found in [13] [15]. The general process of designing a SAR processor can be summarized as follows: • define the processor requirements, • select algorithm and evaluate image quality and computational loading, • survey candidate signal processor architectures and the available technology, • design and evaluate several candidate logical architectures, and • select an architecture and perform detailed design. Throughout the design process, a constant evaluation of design trade-offs needs to take place. The evalua­tion is based on cost/performance trade-offs that consider both implementation and maintenance costs. 3. SAR Processor System Design 49 The ever-shrinking design times have created a need for improvements to the design process. The concepts of rapid prototyping and hardware/software codesign attempt to provide solutions to this problem [62]. For concepts like this to be useful, they need to be supported by good tools. Many tools exist that aid in the design of algorithms (e.g. Matlab). Some tools exist that support limited versions of codesign (e.g. Ptolemy [56]). Both software design and detailed hardware design are well supported by existing tools. Fewer tools exist to aid in the development of architectures. An ongoing program exists in the United States that will have an impact on how embedded SAR process­ing systems are designed in the future. The US Department of Defence Advanced Research Projects Agency (ARPA) program for Rapid Prototyping of Application Specific Signal Processors (RASSP) aims to significantly improve the process by which embedded digital signal processors are developed [81]. Some of the main directions being taken by the RASSP program are given below. • executable requirements [3] • virtual prototyping: software models of hardware • co-development • standard hardware interfaces • reuse libraries The RASSP program is specifically aimed at improving the design process of systems like high-perform­ance SAR processors. The directions taken by this program are likely to be representative of methods that will be used in future SAR system designs. 3.2 System Requirements As for any system, a SAR processing system needs a complete set of requirements. Some basic algorithm related requirements for a SAR processor and their effects on system design are as follows: • radiometric accuracy: affects algorithm, number representation, post-processing corrections, etc. • resolution requirement: affects algorithm, the amount of data, computation, memory, etc. • choice of SAR algorithm: choice of processing parameters, data set sizes, filter lengths, number of looks, etc. These areas are covered in Section 9 of [27]. 3. SAR Processor System Design 50 The system requirements also need to address overall system functionality, external interfaces, perform­ance, cost, and other operational factors like reliability, maintainability, manability, supportability, and economic feasibility [13]. The goal in defining the system requirements is to ensure that the final system meets the needs of the customer. 3.3 System Design Considerations This section summarizes some of the general design considerations that must be addressed by every SAR processor design irrespective of the requirements of the specific system. A l l of the issues discussed here present the designer with decisions and trade-offs. Many of the decisions are interrelated. There is no cookbook recipe for coming up with an optimal design. Numerous high-level design iterations are neces­sary before an optimal solution can be found. A reference on systems engineering should be consulted for a broader look at all the issues of designing a complex system [13]. One of the most fundamental trade-offs in signal processor design is the one between performance and generality/flexibility. The more general the system the lower its performance. It is always possible to achieve higher performance if the architecture is tuned to one algorithm and associated set of processing parameters. Performance decreases if the architecture is made more flexibility to algorithm and processing parameter changes. This is not an easy trade-off to make since it involves many implementation and lifecy-cle issues. Some implementation issues in each of the three areas, algorithm, architecture, and technology, are listed below: Algorithm • fast algorithms: A multitude of options exist for choosing fast ways of implementing the basic DSP operations. For example, there are at least a dozen ways of implementing FFTs [12]. Each of these is suited for different situations and makes a unique set of trade-offs. For each of DSP operations in the SAR algorithm, an algorithm needs to be chosen or a new one developed. Different number systems or other number theoretic techniques may be used to speed up the computations. • arithmetic: Many options also exist for the method of performing arithmetic in the system. Some options are parallel, bit-serial, distributed arithmetic or some combination of these. 3. SAR Processor System Design 51 • number representation: The way numbers are represented in the system is a major factor in system performance and memory sizes. A fixed, floating point or block floating point representation with some word length needs to be chosen. A careful analysis of the number representation's effect on accuracy, truncation errors, memory sizes, and arithmetic unit complexity must be made. Architecture • data transfer rates: Both external and internal transfer rates need to be determined and carefully allocated. • amount and type of parallelism: The optimal level of parallelism must be decided upon based on a cost/performance analysis. • processor selection/design: The selection or design of the processing elements will determine the basic system building blocks. • memory organization and capacities: The types, sizes, and interconnection of all the memory types in the system present a wide range of design options. • general design principles: Principles like modularity, flexibility, and the others mentioned in Section 3.2 must be considered during the design of the architecture. Technology • use of COTS vs. custom hardware: In general, users prefer not to buy custom hardware if their requirements can be filled by standard computer system components. • implementation technology independence: It is advantageous, if future technology enhancements can be used to enhance the system rather than make it obsolete. • the technology freeze date: This sets the date by which all system components must be available; newer components are excluded from consideration. • tolerable level of technology risk: It is always riskier to use new cutting edge technologies or prod­ucts in a system even if they offer higher performance than older more established products. 3.4 Performance Prediction One of the key tasks in SAR processor design is predicting the performance of candidate architectures. Unfortunately, performance prediction can be very difficult. This is especially true for general purpose computers. Usually the more general the system, the more difficult performance prediction becomes. Per­formance analysis is covered in detail in [52]. 3. SAR Processor System Design 52 One of the most common methods of predicting performance is the use of standard benchmark programs. A large number of benchmarks exist. However, most are aimed at general purpose computers and not at DSPs. There are currently some efforts under way to set up a standard set of DSP benchmarks (see for example [94]) but few performance results are available. Most general benchmarks are not very accurate at predicting actual performance for DSP applications. The only truly accurate performance estimates are obtained when actual SAR code is executed on a processor. One of the few DSP benchmarks that has gained any widespread use is the time for a l k complex FFT. Given SAR's dependence on FFTs and the lack of other benchmarks, the l k CFFT time will be used exten­sively in the remainder of this thesis. The l k CFFT time is an important measure of the performance that can be expected from a given processor when running a SAR application since a large portion of the run time of most SAR algorithms consists of FFTs. However, it is dangerous to put too much importance on the l k FFT times. Certain processors may be optimized for FFTs and, therefore, have excellent FFT times but show very poor performance on other operations like filtering. Also, the FFT times for a range of vec­tor sizes should be looked at since the execution time may be strongly influenced by the data set size. This is usually due to the effects of caching. Difficulties arise when comparing machines that have different levels of generality. Performance measures must be looked at in context of the overall system capabilities. Thus, it is not appropriate to compare a gen­eral purpose machine to a special purpose or dedicated processor on only one benchmark result. 3.5 Architecture Selection Criteria In the sections that follow, many possible processor design approaches are presented. Some criteria are necessary to select the ones worthy of further study. This is best accomplished by establishing require­ments that must be met by the SAR processor. The four primary requirements are as follows. Performance The processor shall be capable of processing full resolution satellite SAR data at 1/10 of the real-time rate under the processing scenario defined in Section 2.5.4. Alternatively, it shall be possible to process data at a lower resolution at the full real-time rate. Scalability It shall be possible to scale the processor up to handle real-time full resolution processing with an approximately linear increase in cost. 3. SAR Processor System Design 53 Flexibility The processor shall be flexible enough to accommodate a number of possible SAR processing situa­tions. Processors based on frequency domain processing shall be capable of handling any of the four frequency domain based algorithms described in Section 2.4. Processors based on either time or fre­quency domain approaches shall be capable of handling changes in any of the processing parameters. Development cost The development cost shall be minimized by making use of existing hardware and software products wherever possible. Repeat cost The repeat cost (the cost of additional SAR processors) shall be minimized. One tenth real-time processing is chosen as the primary performance requirement since this level of per­formance is much more attainable by current technology than full real-time operation and is, therefore, much more cost effective. The lower cost of such processors also implies a larger potential market. In some cases, minimizing the development cost and the repeat cost are conflicting requirements. The intent here is to find a reasonable balance between these costs while meeting the other requirements. Some examples of other criteria that are not considered to be of primary importance in this work are power consumption, weight, and volume. 3. SAR Processor System Design 54 4 Architectural Approaches This section describes some general categories of SAR processor architectures. There is no one classifica­tion that covers all possible architectures. However, the one given here has been adapted from existing classifications to handle the most common SAR processing situations. The following list gives the classifi­cations as well as their relation to Flynn's taxonomy of parallel computers [37]. • single processor (SP) - SISD • common node processor (CN) - M I M D • multiprocessor (MP) - M I M D • pipeline processor (PL) - M I M D or pseudo-MISD • multicomputer (MC) - M I M D • SIMD processor - SDVID • Hardwired (HW) Many SAR processors make use of a mixture of these approaches. One architecture may describe the high level organization of the system, while another the lower level components, which are themselves parallel computers. More detailed information on the large variety of parallel architectures can be found in [29]. In all cases, a wide range of performance levels can be attained by each of the architectural approaches. Per­formance depends largely on the number of PEs, the choice/design of the individual PEs and the amount of memory associated with each. 4.1 Single Processor Single processor architectures offer the simplest and cheapest SAR processing solution. However, due to the large computational requirements, uniprocessors are only suited for low performance applications. As microprocessor performance increases, uniprocessors are becoming useful for more and more SAR appli­cations. An example of such a system is the running of SAR processing software on a modern high per­formance workstation with a high speed RISC processor and a large main memory and disk capacity. 4.2 Common Node Processor Like a uniprocessor, a common node processor has a single PE through which all data must pass at various points during execution. However, multiple processors are available to assist the common PE in a proces­sor farm configuration. A typical common node architecture is shown in Figure 4.1 Common node proces-4. Architectural Approaches 55 sors differentiate themselves from multiprocessors and multicomputers in that they have a single PE that distributes and collects the SAR data from the other PEs. Common node architectures have in the past been the most common way of building medium performance SAR processors. Mass Storage Operator Interaction •4 • Input/Output Interface i SAR Data Figure 4.1: Typical common node architecture Some advantages of common node architectures are listed below. • Common node architectures are flexible in handling algorithmic changes as well as different algo­rithms. • Most or all of the system components are usually available off-the-shelf. • Common node systems are usually not much more difficult to program than uniprocessors. • There is much existing experience in building such systems. Some disadvantages of common node systems are given below. • Scalability is a problem. The common PE can be a bottleneck that prevents the performance of the system from being increased. • Performance prediction can be difficult. An example of a common node processor is a minicomputer with an attached accelerator like an array processor. The accelerator may also be a parallel processor and may be designed using a different architec­ture (e.g. multicomputer). Another example of a common node processor is a network of workstations with a single master workstation. More examples can be found in Section 9.3 of [27]. Host/Control CPU PEs/Memories Common ^ w Node J T_N PEs/Memories PEs/Memories 4. Architectural Approaches 56 4.3 Multiprocessor A multiprocessor architecture consists of multiple processors that access a shared memory. Caches are typ­ically used to reduce the number of accesses to the shared memory. A typical multiprocessor architecture is shown in Figure 4.2. Mass Storage Operator Interaction • Input/Output Interface ^ SAR Data Figure 4.2: Typical multiprocessor architecture Some advantages of the multiprocessor approach are listed below. • Multiprocessors are flexible. They can handle algorithmic changes as well as different algorithms. • Either a complete off-the-shelf processor or off-the-shelf PEs can be used. • Multiprocessors use the same familiar global memory programming model as uniprocessors. Some disadvantages of multicomputers are given below. • Scalability is a problem. The shared memory can be a bottleneck that prevents the performance of the system from being increased. • Performance prediction can be difficult. Multiprocessors are likely the most common type of parallel machine being built today. Terms like tightly-coupled, shared-everything, or symmetric multiprocessing (SMP) are used to refer to multiprocessor vari­ants. Examples include the multiprocessor workstations available from vendors like DEC, HP, I B M , SGI and SUN. 4. Architectural Approaches 57 4.4 Pipeline Processor High performance SAR processors have traditionally been constructed as pipeline processors. Figure 4.3 shows a high-level diagram of a pipeline processor. I/O Input/Output Interface PEs/Memories/l/O PEs/Memories/l/O 1 PEs/Memories/l /O Input/Output Interface I/O Host/Control CPU Figure 4.3: Typical pipeline architecture Pipelining is often combined with other architecture approaches in the design of SAR processors. For example each pipeline stage can be designed using a multicomputer architecture. In general, each stage of the pipeline can itself be a parallel processor that is designed using any of the architectures in this section. A key characteristic of pipeline processors is the amount of flexibility designed into the system. Hardwired or fixed function pipelines have PEs and interconnections that are fixed and dedicated to a single algo­rithm. Programmable pipelines have programmable PEs and a more general interconnection scheme that can be adapted to a number of algorithms. Hardwired pipelines tend to be faster while programmable pipe­lines are more general. Some of the advantages of the pipeline approach are listed below. • A pipeline exploits the sequential nature of the steps in the algorithm. • Off-the-shelf processing elements (PEs) can be used. • The performance of a pipeline processor is relatively easy to predict. • There is a great deal of prior experience in building pipeline SAR processors. The following is a list of some of the potential problems with pipeline SAR processors. • The processor as a whole will most likely not be available off-the- shelf. • A specialized pipeline may not be flexible enough to handle changes in the algorithm or other aspects of SAR processing. Buffers are needed between all the stages. 4. Architectural Approaches 58 • Expandability can be expensive since all the stages need to be expanded together. • Fault tolerance may be difficult to achieve. In a simple pipeline, each PE or buffer is a critical link. • Load and I/O rate balancing can be difficult. The N A S A JPL Advanced Digital SAR Processor (ADSP) and its derivatives are hardwired pipelines. These processors exhibit extremely high performance but only for a specific algorithm. VASP based SAR processors like the M D A IRIS X2C can be described as programmable pipelines in which each pipeline stage is a multicomputer composed of Motorola 96000 DSPs. 4.5 Multicomputer The most flexible type of parallel architecture is the multicomputer, which is composed of a distributed collection of PEs with local memories. Multicomputers are gaining in popularity due to their flexibility and ability to use off-the-shelf components. Figure 4.4 is high-level diagram of a multicomputer SAR proces­sor. A key component of multicomputers is the interconnection network. The design of the interconnection needs to make a careful balance of complexity against cost since the choice of network is crucial to the per­formance of the system. 1 Mass Storage PEs/Memories/l/O 1 Input/Output Interface SAR Data T Figure 4.4: Typical multicomputer architecture Some of the advantages of the multicomputer approach are listed below. • Multicomputers are flexible. They can handle algorithmic changes as well as different algorithms. • Either a complete off-the-shelf processor or off-the-shelf PEs can be used. • Multicomputers are modular and easily expandable. Operator Interaction <—H 4. Architectural Approaches 59 • Fault tolerance can be achieved. The following is a list of some of the potential problems with multicomputers. • The cost of the interconnection network can be significant. • The inter-processor communication costs can severely hamper performance. • Load balancing can be a difficult problem. • Programming is more difficult than for shared memory machines. Some examples of multicomputer SAR processors are the RSARP processor that is based on Mercury multi-i860 processor boards, the SAR processors based on Meiko Transputer based multicomputers, and the M D A IRIS X2C VASP based airborne processor which is a multicomputer-pipeline combination. 4.6 SIMD Processor In a SIMD processor, a control unit issues the same instruction to a synchronous set of processors which each operate on a subset of the data. Figure 4.5 is a high-level diagram of a SIMD processor. One of the key design issues in SIMD architectures is the complexity of each PE. SIMD PEs can be simpler than those in multicomputers since they do not need to perform any program control functions. Many SIMD machines built to date have had large numbers of very simple PEs. This approach has not been well suited for long word length or floating point operations. The design of the interconnection network is another key issue and the same considerations described for multicomputers apply here. 7 = •• Mass Storage Operator Interaction •4 • I Input/Output L Interface i SAR Data Figure 4.5: Typical SIMD architecture The regular nature of the data processing operations in SAR point to SIMD processors as a possible solu-Host/Control CPU PEs/Memories/l/O Control Unit 4. Architectural Approaches 60 tion. SIMD machines are well suited to FFTs and vector multiplies; however, interpolation operations can pose efficiency problems. Since SIMD machines are best suited to vector operations, a way of efficiently performing the algorithm's scalar computations needs to be found. Some of the advantages of the SIMD approach are listed below. • SIMD processors are generally more efficient for vector and FFT calculations than M I M D processors. • Either a complete off-the-shelf processor or off-the-shelf PEs can be used. • SIMD processors are modular and easily expandable. • Fault tolerance can be achieved. The following is a list of some of the potential problems with SIMD processors. • It can be difficult to map an algorithm onto the SIMD architecture. • Pipelining is not possible in a pure SIMD processor. • SIMD processors can be inefficient for scalar operations, non-regular portions of SAR processing, and post-processing operations. • The simple processing elements in SIMD machines often lack floating point support. • Load balancing can be a problem. It may be difficult to keep all the PEs busy if the problem size changes. Machines that use SIMD architectures have traditionally not been used for high performance SAR process­ing. However, numerous studies have been done to examine the feasibility of using SIMD machines for SAR. An example using a MasPar machine is described in [11] and another using a Thinking Machines CM-2 in [60]. 4.7 Hardwired Hardwired architectures are a general category for processor architectures that use custom application-spe­cific designs and, therefore, do not fit into the other categories. Hardwired machines typically do not use regular von-Neumann processors. Rather, they use logic that is designed specifically for the application. An arbitrary amount of parallelism can be used. The hardware can also be tailored to the operations and wordlengths called for in the algorithm. Hardwired processors may be limited to a specific algorithm or they may be reconfigurable to some extent. Examples of this kind of architectures are processors imple­mented in discrete logic, custom VLSI , or programmable logic. 4. Architectural Approaches 61 5 Computing Technology This section gives a very brief summary of the key enabling technologies for SAR processing. A survey is made of the present state of the art and some indication is given of future directions. The technology areas covered are general and special purpose processors, programmable logic, memory, and interconnect. The information summarized here is drawn from a wide range of sources. The contribution of this section is to pull together and organize the information that allows these technologies to be compared on the basis of their suitability for SAR processing. In the sections that deal with processing devices, l k CFFT times are used as a method of comparison. FFT times are not a good general benchmark: much better ones are available for general purpose microproces­sors. However, this section looks at more than just general purpose processors. In order to process SAR data almost all processors will have to perform FFTs. As a result, FFTs are a reasonable common bench­mark. The other performance measure that is used is OP/s. OP/s has become a relatively meaningless measure given the diversity in the definition of an operation. However in this section, the definition of OP is the one given in Section 2.5.1 and will usually be based on FFT or FIR times. The purpose of these per­formance measures is not to allow detailed comparisons between individual devices but rather to allow order of magnitude comparisons between different types of devices. 5.1 General Purpose Microprocessors This section looks at some of the developments in general purpose microprocessors of both the RISC and CISC varieties. Table 5.1 gives approximate FFT times for some currently available high performance microprocessors. There is a wealth of information in the literature describing these chips. Much more information can be found on the internet [24]. It should be noted that many of the FFT times given in the table are for benchmarks coded in a high level language. Often significantly higher performance can be achieved using hand optimized assembly lan­guage. This should be kept in mind when comparing these processors to DSP chips, where benchmark results are almost always given for optimized assembly code. Although many types of high performance microprocessors are available, some common trends have emerged. One such trend is the increased emphasis on floating point performance. Until a few years ago, floating point performance was a weak point with microprocessors, which often required a separate chip to provide hardware floating point support. The most recent chips have integrated floating point support that 5. Computing Technology 62 Table 5.1: High performance microprocessors and their FFT performance lk CFFT SPEC Benchmarks Manufacturer Chip Time (msec) MOP/s Clock (MHz) Lang­uage int92 fp92 Clock (MHz) DEC Alpha 21064 0.48 107 200 133 200 200 HP PA-RISC 7100 1.1 47 99 C 80 150 100 IBM/Apple/Motorola PowerPC 601 4.0 13 C 40 60 50 Intel 486DX 60 0.9 33 C 28 13 50 Intel Pentium 1.25 41 133 assembly 67 64 66 Intel i860 0.74 70 40 assembly MIPS R4000 5.4 9 100 C 59 61 100 MIPS R4400 3.3 16 150 C 88 97 150 Sun SuperSPARC 0.95 54 50 assembly 89 103 60 Sun UltraSPARC 0.3 171 167 assembly 275 305 167 sometimes surpasses their integer performance. This development has made it easier to implement general DSP applications on these chips. The general purpose nature of these processors leads to some problems when they are used for DSP pur­poses. One of the problems with general purpose microprocessors in DSP applications is their extensive use of on-chip caching to attain high performance. The caching systems are designed for general applica­tions like operating systems and are not appropriate for many DSP applications. Microprocessors also lack the built-in support for common DSP operations like FIR filters and FFTs that DSP chips have. However, recent advances in microprocessors have helped to minimize these differences. Now the FFT performance numbers attained by microprocessors are almost as good as those of the fastest general purpose DSP chips [82]. This can be seen from Table 5.1 where assembly language optimized FFT routines on the latest crop of RISC processors, like the DEC Alpha and the Sun UltraSPARC, show extremely high performance. The most important difference between the use of microprocessors and DSPs in DSP applications is, in many cases, no longer performance. The major differences are now found in other areas like power consumption, on-chip peripherals, inter-processor communications support, and cost. One device that deserves special mention is the Intel i860. It was not used as a general purpose C P U but was targeted at signal processing type applications. This can be seen from its excellent FFT times. It found a great deal of use in multiple processor signal processing systems. Many different multi-i860 boards are available from a number of vendors. However, the DSP industry is looking at other chips for future growth since Intel does not plan any more developments with the i860 family. 5. Computing Technology 63 Systems that make use of multiple standard microprocessors are becoming the most common kind of par­allel machine. At the high end, supercomputer manufacturers are using large numbers of microprocessors in their systems. For more moderate performance, multiprocessor workstations are becoming very com­monplace. There are also many board-level and subsystem products available that can be used in parallel systems. 5.2 General Purpose DSPs One of the main drivers behind the rapid growth of the DSP market has been the availability of powerful low-cost DSP chips. Table 5.2 summarizes the currently available high-end DSP chips [58][19]. DSP chips can be divided into two types: floating point and fixed point. Floating point devices are easier to program but are more expensive. Fixed point devices are more difficult to program since truncation error needs to be carefully monitored by the programmer but they are less expensive and are the devices of choice in high-volume applications. Table 5.2: High performance DSPs and their FFT performance Manufacturer Chip Type Time (msec) lk CFFT MOP/s Clock (MHz) Notes Analog Devices ADSP-21020 32-bit floating 0.58 88 33 Analog Devices ADSP-2106x (SHARC) 32-bit floating 0.46 111 40 multiprocessing support AT&T DSP32C 32-bit floating 2.8 18 50 Motorola DSP56002 24-bit fixed 0.83 62 80 Motorola DSP96002 32-bit floating 1.05 49 40 multiprocessing support Texas Instruments TMS320C40 32-bit floating 1.55 33 50 multiprocessing support Texas Instruments TMS320C80 MVP 32-bit fixed/float­ing multiprocessor 1 floating pt. & 4 fixed pt. processors Zoran ZR38001 20-bit fixed 0.67 76 33 Currently the TMS320C40 and the ADSP-21060 S H A R C are the most popular devices for high-end DSP applications. This is partly due to these chips' built-in support for multiprocessing. Both the C40 and the SHARC have six 8-bit communication ports that are intended to facilitate inter-processor communication. A number of manufacturers offer scalable board products with multiple C40s or SHARCs. 5. Computing Technology 64 One of the most impressive chips to become available is the TMS320C80 M V P from Texas Instruments. This chip incorporates four integer processors and one floating point processor on one die. The peak per­formance is about 500 MOP/s. The M V P is primarily aimed at video signal processing applications like M P E G processing. This use of multiple processors on a chip can be seen as a sign of things to come. While general purpose RISC microprocessors have caught up to DSPs in terms of arithmetic performance, DSPs still have a number of advantages. In contrast to RISC chips, DSPs often have separate instruction and data buses, multiple data buses, numerous on-chip peripherals like D M A controllers, inter-processor communications support, lower power consumption, and lower cost. As always, the l k CFFT performance figures give an incomplete indication of actual performance. As an example, Figure 5.1 shows the operation rate achieved on FFTs of various lengths by the C40, SHARC, and i860 processors. A sharp drop-off in performance is evident at the point where the FFT length exceeds the size of the on-chip memory for the C40 and i860. Operation Rate (MOP/s) 6 8 10 12 " (FFT length = 2°) Figure 5.1: Performance for FFTs of various lengths Many board level products that incorporate DSP chips are available. The boards typically contain from one to eight DSP chips and various I/O and memory configurations. A great deal of support software is also available, ranging from compilers to DSP operating systems. 5.3 Special Purpose DSPs This section summarizes some of the currently available special purpose DSP processors. These devices are distinguished from the general processors described in the previous section by their promise of higher 5. Computing Technology 65 performance for specific algorithms. Section 5.3.1 describes some accelerated FFT chips, and Section 5.3.2 lists some recent digital filter chips. 5.3.1 Accelerated F F T Chips Since the FFT is such a commonly used DSP algorithm, a variety of special purpose chips have been developed to accelerate its execution. Table 5.3 lists some of the currently available products. Table 5.3: FFT processor ICs lk CFFT Manufacturer Chip Time (usee) MOP/s Clock (MHz) Data Format Max FFT Size Notes Array Microsystems A66111 and A66211 131 391 40 16-bit floating 64k supports 16 func­tions (but no FIRs) GEC Plessey PDSP16510 96 533 40 16-bit block floating 1024 FFTs only LSI Logic L64280 256 200 40 24-bit floating 2048 FFTs and MACs only Raytheon/ TRW TMC2310 514 100 20 16-bit fixed 1024 supports 16 functions Sharp/Butterfly LH9124 and LH9320 81 632 40 24-bit block floating unlimited supports 26 functions Texas Memory Systems TM66 swiFFT 76 674 40 32-bit IEEE floating unlimited supports basic com­plex arithmetic The chips listed in the table have lk CFFT execution times that are up to an order of magnitude faster than those of the fastest general purpose DSPs. In addition to execution time, a number of important features differentiate these chips. One of the most important of these is the generality of the device. Some perform only FFTs and nothing else. Others are more general purpose and support a set of common DSP opera­tions. Some other characteristics are the number of chips required for a basic system, the maximum sup­ported FFT length, the number format, and the wordlength. It is usually possible to use multiple chips to increase performance. The most popular of the latest crop of these chips is the Sharp/Butterfly LH9124/LH9320 chip set. This chip set is near the lead in the race for the fastest single chip FFT time. It also supports a wide range of other arithmetic functions which is important for SAR processing since FFTs are not the only computation­ally expensive operation. The possibility of using it as the basis for a SAR processor architecture is exam­ined in Section 9.3. The newer TM66 chip is similar to the LH9124 but is slightly faster and supports IEEE floating point arithmetic. 5. Computing Technology 66 A number of manufacturers have board level products available that make use of accelerated FFT chips. Currently, the most common products are ISA or V M E bus boards that contain single or multiple LH9124 chips. 5.3.2 Digital Filter Chips Since digital filtering is the most basic operation in DSP, quite a few special purpose chips have been developed to perform fast filtering. Some of more recent ones are listed in Table 5.4. These chips are typi­cally designed for applications that require extremely high-speed real-time filtering like video signal processing. These chips usually work with fixed point numbers with fairly short wordlengths and do not usually directly accommodate complex numbers. They do, however, achieve extremely high speeds in terms of numbers of operations per second. Table 5.4: High speed filter ICs Manufacturer Chip Data (bits) Coef­ficient (bits) Max.# of Taps Approx Max. MOP/s Max sample rate for N-tap FIR (MHz) Max sample rate for 50-tap FIR (MHz) GEC/Plessey PDSP16256A 16 16 128 800 25/(N/16) 8 GrayChip GC2011 24 14 32 4500 70/(N/32) 44 Harris HSP43124 24 32 256 180 45/(N/2 + 1) 1.7 Raytheon/TRW TMC2243 10 10 3 120 20/(N/3) 1.2 Zoran ZR33288 8 10 288 8234 14.32 for all filter lengths 14.32 Motorola DSP56200 16 24 256 19 10.25/(12+N) 0.165 Even though these chips have limitations, they do have potential applications in SAR processing. Section 9.5 looks at some possible applications of digital filter chips in SAR processing. Not many off-the-shelf products are available that make use of these devices since they are usually designed into custom systems. 5.4 Custom and Semi-Custom VLSI The highest possible performance for a DSP algorithm that today's technology will allow can only be achieved by designing custom VLSI circuits to implement the algorithm. Custom VLSI chips are usually only used for extremely high volume applications or applications where specific power, space, or perform­ance requirements dictate their use. VLSI design costs have been decreasing with the increased use of sophisticated C A D tools. The use of hardware description languages (HDLs) and synthesis tools has made it easier and much faster to develop complex VLSI designs. These advances have made VLSI and ASIC design an option for a much larger number of designers than in the past. 5. Computing Technology 67 Custom VLSI design has not been a cost effective approach for general purpose SAR processors since these are only built in very low volumes. This is not likely to change in the near future despite the advances in VLSI design tools. One of the reasons custom designs are not as attractive for S A R processing is that the majority of the computation in SAR is made up of standard operations like FFTs for which standard parts are available. An exception to this may arise for SAR processors intended for airborne or spaceborne applications where stringent requirements rule out the use of general purpose parts. In addition to fully custom VLSI , several types of semi-custom devices are available that allow the designer to customize only several mask layers of an otherwise standard chip. The lower density end of the ASIC market is being taken over by user programmable devices such as FPGAs. 5.5 Field Programmable Logic Field programmable gate arrays (FPGAs) are user programmable integrated circuits that consist of uncom­mitted logic elements that can be interconnected in a general way. FPGAs have become extremely popular and, as a result, their prices have decreased rapidly. Great advances in density have also been made. Higher capacity FPGAs are announced almost weekly and 100 000 gate capacity FPGAs are expected to be avail­able within the next two years [38]. Until recently, FPGAs lacked the density to support meaningful DSP applications. As F P G A densities increase, the interest in implementing DSP algorithms in FPGAs grows. F P G A based computers are dis­cussed in Section 6.6. 5.6 Memory The capacity of D R A M chips has been increasing dramatically since their introduction. The most common D R A M sizes available today are 1, 4 and 16 Mbits. It is expected that 64 Mbit chips will soon be com­monly available. 256 Mbit chips are in development. Figure 5.2 plots D R A M size against the year of peak production for various D R A M chip sizes [76]. The ever increasing size of memory chips is very important to the design of SAR processors due to the large image sizes typically involved in SAR processing. For example, a 4k x 8k complex image, where each pixel is represented by two 32-bit complex numbers, requires 256 Mbytes of memory. Table 5.5 lists the number of memory chips needed to hold such an image. Unfortunately, while D R A M capacities have increased greatly, access speeds have not increased by the same amount. Currently, memory speed poses the most serious bottleneck to increased processor perform-5. Computing Technology 68 chip capacity 1 0 0 0 (kbits) 100 10 1 1 1 1 64Mb - 16Mb -/ * 4 M b 1 Mb 256 kb X l 6 k b | 1 I I 1980 1985 1990 1995 2000 2005 Year of maximum production Figure 5.2: Memory chip sizes and their year of peak production Table 5.5: Effect of memory chip size on chip and module count Chip size (Mbits) Number of chips for a 256 MB memory Number of modules containing 16 chips for a 256 MB memory 1 2048 128 4 512 32 16 128 8 64 32 2 256 8 1 ance [14]. At present no absolute solution to this problem exists. Partial solutions are provided by caching and memory interleaving. Caching is the most commonly used scheme. In the current generation of micro­processors, elaborate caching schemes are used in an attempt to hide the slow speed of the DRAMs. New developments in the D R A M devices themselves are also appearing to help alleviate the access time problem. These schemes usually allow the data to be accessed much more quickly if the data is accessed by successive memory locations. Some of the more common examples of these devices are listed in Table 5.6 in order of their date of availability [20]. D R A M s with fast page mode access are standard today. EDO D R A M s are expected to become commodity products in the near future. In several years, synchronous D R A M s are predicted to become the most common type of D R A M . It should be stressed that even though certain access sequences are greatly accelerated by these new D R A M s their random access time is still essentially unchanged. The new types of D R A M s such as EDO D R A M s are potentially useful in SAR processing because much of the data accesses occur to sequential memory accesses. However, a problem arises when the data needs 5. Computing Technology 69 Table 5.6: Bandwidth of some DRAM types Type Approximate Bandwidth (MHz) Regular DRAMs 10 Fast page mode DRAMs 33 Extended Data Out (EDO) DRAMs 50 Burst EDO DRAMs 66 Synchronous DRAMs 100 to be accessed in a different direction, like during a corner turn in SAR processing. If these types of devices are to be used to increase performance, the implementation must take the characteristics of the devices in account when writing and reading values from memory. 5.7 Interconnect In addition to large amounts of computing power, high performance processors require sufficient commu­nications bandwidth. This bandwidth is provided by interconnection technologies. Like memory, currently available interconnections have not kept pace with the rapid growth in computing power. However, the pri­mary reasons have not been technological. Rather, the problems have been to do with standards and market acceptance. Demand in the marketplace for interconnect has been concentrated on well accepted standards. This has resulted in interconnection technologies like the ISA bus remaining common long after higher performance buses became available. In recent years, there have been many developments in interconnection networks. There has been a prolif­eration of standards and pseudo-standards. It is difficult to classify these interconnection schemes. The tra­ditional boundaries between buses and I/O interfaces have become blurred. Serial interfaces are presently being used where previously parallel interfaces or backplane buses were used. This section presents one simple classification and gives examples of currently available interconnections. More information on modern interconnections can be found in [71], [70], and [51]. • Personal computer and workstation buses are interconnections that are commonly used in PCs and workstations as backplane buses. Table 5.7 summarizes the most common currently available buses. • High speed buses and related interfaces are interconnections that are either backplane buses or addi­tions to existing backplane buses. These interconnections are typically used in embedded or control applications. The most common category is the VMEbus and its derivatives. Table 5.8 lists the more common members of this class. 5. Computing Technology 70 Table 5.7: Personal computer and workstation buses Bus Clock Rate (MHz) Maximum Throughput (MB/s) Typical Throughput (MB/s) Max # Slots Bus Width Comment/ Applications Industry Standard Architecture (ISA) 8.33 8.33 1-2 no logical limit, practically 3-8 from electrical characteristics 16-bit similar to x86 bus, used in IBM PC com­patibles Extended Industry Standard Architec­ture (EISA) 8.33 33 in burst mode 8.33 logical limit 15 32-bit used in IBM PC compatibles Micro Channel Architecture (MCA) 10-20 157 in burst mode 76 sustained 30 15 16 & 32-bit used in IBM PS/2 VESA Local Bus (VL) up to 66 160 at 32 bits 267 at 64 bits (50 MHz) 67 3 32 or 64 bit used in PC graphics/video applications Peripheral Compo­nent Interconnect (PCI) up to 33 132 at 32 bits 264 at 64 bits 65 10 loads, typi­cally 2-3 cards 32 or 64-bit used in PCs NuBus 10 40 10 up to 8 typi­cally 32-bit developed by Apple MBus 40 320 in burst mode 200 sus­tained 6 theoretically, 4 in practice 64-bit developed by Sparc vendors SBus 20-25 80 at 32 bits 168 at 64 bits limited by elec­trical character­istics 32 or 64 bit developed by Sun Microsys­tems, Turbo Channel 12.5-25 50 to 98 2-6 32 bit developed by DEC • The other high speed interfaces category encompass new interconnection schemes used for interproc-essor connections as well as processor to peripheral connections. Some of the high speed interfaces are listed in Table 5.9. The numerous new standards for lower speed interfaces are not listed. The categories and tables listed above do not cover L A N standards or standards aimed primarily at periph­eral interfaces like SCSI and its variants. Existing high performance SAR processors have usually used some of the standardized interfaces in com­bination with a higher speed proprietary interface. Examples include the VASP signal bus from M D A or the interconnection network from Mercury. As with all areas of computing, there is a great desire to use standardized interfaces in the future. Some of the new high speed interfaces listed in the tables are aimed directly at high performance multiprocessor systems that can be used for applications like SAR processing. Examples include V M E additions like Autobahn and RACEway. 5. Computing Technology 71 Table 5.8: High speed buses and related interfaces Interface Parallel/ Serial Topology Maximum Throughput (MB/s) Maximum # Slots Bus Width Comment/ Applications Multibus I parallel bus 12 15 16 bits Multibus II parallel bus 80 in burst 64 sustained at 20 MHz 21 32 bits VME32 parallel bus 40 21 32 bits VME Subsystem Bus (VSB) parallel bus 40 21 32 bits sidebus to VME, uses P2 connector VME64 parallel bus 80 21 64 bits extension to VME, uses address lines to transfer data Futurebus+ parallel bus 3.2 GB/s at 256 bits 14 32 to 256 bits AutoBahn serial user defined 200 >8 2 serial pins extension to VME, developed by PEP Computer Heterogeneous Interconnect (HIC) IEEE P1355 serial user defined 160 QuickRing serial ring 200 16 (in one ring) 6 bits developed by National Semicon­ductor RACEway parallel crossbar 320 16 32 bits extension to VME, developed by Mer­cury SkyChannel parallel bus or crossbar 320 64 bit developed by SKY computer Table 5.9: Other high speed interfaces Interconnect Parallel /Serial Topology Throughput Number of Nodes Addressing/ Data Paths Comments/ Applications Scalable Coherent Interface (SCI) ANSI/IEEE 1596 parallel point to point 1 GB/s theoretical 125 MB/s with CMOS chipset 500 MB/s with GaAs chipset. 64 k 64 bit address, up to 256 data bits high perf. par­allel systems Fibre Channel ANSI X3T11 serial point to point up to 100 MB/s 16 mil­lion 24 bit ID fields mass storage interfaces High Performance Parallel Interface (fflPPI) parallel point to point 100 MB/s per cable millions 32 bit data per cable OPTOBUS serial point to point 20 MB/s per fiber developed by Motorola 5. Computing Technology 72 6 Architecture Implementation Alternatives This section looks at some alternatives for implementing SAR processors based on the architectures given in Section 4 and the technologies surveyed in Section 5. Only implementations that show promise for high performance processing are considered. The alternatives discussed here were chosen on the basis of being the most cost-effective ones given the state of technology now and in the near future. The implementation approaches considered in this section are listed below in approximate order of gener­ality. The architectures (as defined in Section 4) that each approach encompasses is shown in brackets. • Workstation (SP, MP) • Accelerated general purpose computer (CN) • Supercomputer (any) • Network of workstations (MC) * • DSP uniprocessor and multicomputer (SP, MC) * • Reconfigurable computing machine (HW) * • Custom algorithm specific processor (HW, PL or MC) Each of these alternatives is discussed in the sections that follow. The options marked with asterisks are identified for more in-depth study in Section 9 based on the selection criteria presented in Section 3.5. The discussion in this section extends and updates previous work to encompass a wider range of alternatives. No broad ranging overview of the suitability of architectures for SAR processing, like the one presented here, has been found in the literature. 6.1 Workstation The rapidly increasing performance of RISC based workstations has made them suitable for a wide variety of tasks that previously required large computers with special accelerators. For cases where a single proc­essor is not sufficient, workstation vendors provide an upgrade path by allowing extra processors to be added. These parallel systems are multiprocessors (the processors access a shared memory). The large market for workstations gives them an advantage over other high performance systems, in that they are usually quick to incorporate new technologies and software advances. Although workstations can be used for DSP purposes, they are really optimized for running applications under general purpose operating sys-6. Architecture Implementation Alternatives 73 terns like UNIX. This leads to a number of potential problems when they are applied to demanding DSP applications: • Workstations make extensive use of caching. Depending on how the caching is handled, it is not nec­essarily useful and may even be detrimental for SAR processing (see Section 7.4.2). • The shared memory and bus can be a major bottleneck in multiprocessor systems, especially for cor­ner turn operations. • Performance is difficult to predict in comparison with more specialized architectures. • Workstations are slower at FFTs by a factor of 10 in comparison with the fastest FFT processor chips. • Workstations are not suited for embedded applications like airborne or spaceborne processors. More and more applications will move to general purpose workstations as their computational require­ments are overtaken by the ever increasing capabilities of workstations. The flexibility, ease of use, and economies of scale that characterize workstations and PCs can more than make up for any extra costs asso­ciated with using a general purpose computer for an application. Unfortunately, the SAR processing requirements considered in this thesis are too demanding for cost effective implementation on single and multiprocessor workstations. SAR processing implementations on general purpose workstations are not looked at further in this thesis. 6.2 Accelerated General Purpose Computer Accelerated general purpose computers have been a very common implementation architecture for SAR processors. These systems are a type of common node architecture and the comments made in Section 4.2 about common node architectures also apply here. Typically, a powerful superminicomputer like an Alliant or a DEC V A X is used as the common node. It is supported by an attached accelerator which can be com­posed of an array processor from a vendor like STAR or Sky, or custom accelerator boards. Today the accelerator is likely to be a multicomputer composed of DSP chips. An advantage of accelerated general purpose computers is that most of the components are likely to be available off-the-shelf. Most of the algo­rithm is usually implemented in software on the central computer so the processor is relatively easy to develop and modify. The main disadvantages of these systems are their cost and the performance limita­tions due to the bottleneck posed by the common node. Some analysis and some examples of accelerated general purpose computers are given in Section 9.3.3 of [27]. Since this approach does not scale well to higher performance levels, it will not be examined further in this thesis. 6. Architecture Implementation Alternatives 74 6.3 Supercomputer The traditional solution to large scale scientific computation has been the use of supercomputers. A wide variety of architectures have been used to build supercomputer-class machines. Some examples are vector supercomputers like the Cray series, M I M D machines like the Thinking Machines CM-5 and the Kendall Square Research series, and SIMD machines like the Thinking Machines CM-2 and the Maspar series. The execution of S A R algorithms on supercomputers has been examined in quite a few instances [60] [11] [66]. Traditionally, however, supercomputers have not presented a cost effective solution for SAR process­ing. This is primarily due to their high cost. Also most high performance SAR processors are production processors that execute an application that is limited to a number of common DSP algorithms. Supercom­puters attempt to provide a high degree of generality. This is not needed for most SAR processors but con­tributes greatly to the system cost. Recently traditional supercomputers have become an even less attractive option. Currently there is a trend away from such systems. This trend has been demonstrated in the recent bankruptcies of many supercom­puter manufactures like Cray Computer, Thinking Machines, and Kendall Square Research. At one point, massively parallel processor (MPP) machines were thought to be the wave of the future. However, these have also not done as well as expected. This was due to high prices, long and costly development times, and the difficulty of software development. Traditional supercomputers are not examined in any more detail in this work. However, some of the lower cost supercomputer replacements that many researchers are investigating for some applications are consid­ered. Some examples are networks of workstations, multiprocessor workstations, and reconfigurable com­puting. 6.4 Network of Workstations A new option for computationally intensive problems is the use of networked groups of the many powerful and inexpensive workstations that are available today. This idea is rapidly growing in popularity and is attracting much research interest [5]. Some of the factors that have contributed to usefulness of networks of workstations (NOW) are listed below. • New technologies tend to appear in workstations faster than in larger systems like supercomputers. • Workstations provide more computing power at a lower price than most other computing systems. 6. Architecture Implementation Alternatives 75 • Excellent software is available for workstations since the high cost of software development can be spread over large sales volumes. • An opportunity exists to use the unused capacity of existing workstations for distributed tasks. This essentially provides computing power at no additional cost. • Unused R A M capacity of existing workstations can be put to use for applications like distributed disk caches. Several inter-process communication standards for high performance computing on workstation clusters have been established. The two most popular ones both have freely available software distributions: paral­lel virtual machine (PVM) [42], and message passing interface (MPI) [31]. Both of these systems provide software that facilitate message passing interfaces between programs running on different workstations [46]. The performance figures that are currently available show that relatively high overheads can be expected. For example, in an MPI performance evaluation paper, an FDDI network capable of 100 Mbps was used, but MPI could only reach a maximum throughput of under 20 Mbps [73]. As the software matures, it will no doubt be tuned to lower some of the overheads. For SAR processing applications, workstation clusters have all the advantages and disadvantages of using single workstations (Section 6.1). NOW attempt to circumvent as many of the disadvantages as possible by using more and more workstations in parallel. An examination of the literature shows that while these sys­tems are being used for a large number of applications, these applications are mostly ones that were previ­ously implemented on supercomputers. Little experience exists in predicting the usefulness of NOW for high performance SAR applications. (A P V M based system using older technology is described in [25]). The main disadvantages of existing NOW systems are the severe performance limitations due to the high latency and low bandwidth of their communication networks. Huge overheads imposed by the existing software packages that support NOW have magnified this problem. In order for N O W implementations to be successful, both a high bandwidth network and low overhead communications are necessary. NOW-like supercomputers, such as the I B M SP and the Cray T3D/E, overcome some of these limitations but at increased cost. The development of the Scalable Coherent Interface (SCI) [44] also attempts to address these issues. The key to successful algorithm implementation on workstation clusters is to come up with a partitioning of the problem that has a computation to communication mix that matches that of the workstations and their interconnection network. Since SAR processing can be split into fairly large chunks of computation it 6. Architecture Implementation Alternatives 76 appears to be a promising application for workstation clusters. This option is examined further in Section 9.1. 6.5 DSP Uniprocessor and Multicomputer Currently the most common approach to building high performance DSP systems is to use DSP ICs as the building blocks. General purpose, special purpose, or even custom DSP chips are used in single or multiple processor configurations. The key characteristic of this approach is that the architectures use components optimized for DSP operations but are flexible enough to handle a wide range of algorithms. The multicomputer architecture with general purpose DSPs as the processing elements combines the advantages of multicomputers with those of DSP chips. Modem DSP chips like the TMS320C40, TMS320C80, DSP96002, and ADSP-21060 have on-chip support for multiprocessing. This usually con­sists of high bandwidth parallel data ports with associated controllers that allow interprocessor communi­cation to occur with minimal overhead to the main CPUs. Depending on how the processors are interconnected, these systems can be quite scalable. Hundreds of DSP chips can used in parallel. Many commercial products are available in the area of DSP based systems. A great deal of software is available to support development. This includes compilers, debuggers and operating systems. However, the software support for parallel systems is not yet as mature as for uniprocessor systems. A major disad­vantage of current DSP based systems is that there are not many industry wide standards for hardware or software. Each chip, board, and software vendor has a different set of standards. This makes development more difficult and severely hampers portability. It is no surprise that DSPs are being used for many of the SAR processors designed recently. Some exam­ples of SAR processors based on general purpose DSP based multicomputers include M D A processors based on the MDA/Spectrum VASP multi-96000 boards, and on Mercury Computer's multi-i860 boards. The RASSP program in the United States has defined a series of benchmark projects that will allow the development process for application specific signal processors to be studied, measured, and refined. The first two of these benchmarks involve the design of a real-time airborne SAR processor [95]. Two contrac­tors designed systems to meet the requirements set by the RASSP program. Both teams chose a multiple processor DSP chip based approach [2]. One team designed custom boards and used a specialized DSP chip: the LH9124. The other team used commercial boards that feature either the ADSP-21060 or the i860. The choice of these processors is hardly surprising since each of these chips has the fastest l k FFT time for the processors currently available in its class as shown in Tables 5.1, 5.2, and 5.3. 6. Architecture Implementation Alternatives 77 The DSP chip approach to building SAR processors is studied in much more detail in Section 9. Section 9.2 looks at processors based on general purpose DSPs. Vector DSP based architectures are covered in Section 9.3. Section 9.4 discusses optimized DSP ICs. Processor designs based on digital filter chips are examined in Section 9.5. Fine grained parallel approaches based on DSPs are discussed in Section 9.6. Finally, a sample architecture that combines vector and scalar DSPs is described in Section 9.7. 6.6 Reconfigurable Computing Machine Reconfigurable computing machines use reprogrammable building blocks to implement application spe­cific processors. Such computing devices have been given a number of different names: reconfigurable computers, transformable computers, virtual computers, or flexible processors. A l l of these names refer to the concept of using reconfigurable or programmable hardware to implement processing devices. These processors attempt to use hardware to circumvent the one-instruction-at-a-time bottleneck inherent in Von Neumann computers. Reconfigurable computing is a rapidly growing research area [54][17][7][23][55]. FPGAs are usually used as the reconfigurable building blocks in such machines. The reprogrammability of FPGAs allows algorithm implementations to be debugged and modified like software but still run at the speed of dedicated hardware. The major problems with FPGAs in high performance processing applica­tions are the low speed and density of the existing F P G A devices. F P G A gate density is typically an order of magnitude below that of microprocessors. As a result, unless the algorithms to be implemented are very simple, multiple F P G A devices are needed. Some examples of reconfigurable computing machines based on multiple FPGAs are the Programmable Active Memory (PAM) developed by the DEC Paris Research Lab (PRL) [10], the Anyboard system developed at North Carolina State University [88], the S P L A S H system from the Supercomputing Research Center [40] [89], and the processors developed by Virtual Computer Corp. [54] [22]. An up-to-date and extensive list of FPGA based computing machines is available on the internet [74]. The key fea­ture of all of these machines is the ability to repeatedly prototype and optimize designs quickly and cheaply. The growing popularity of these machines has led Xilinx, the leading F P G A vendor, to recently announce a new family of FPGAs, the XC6000 family, that is optimized for reconfigurable architectures. Reconfigurable computers have been used for a variety of applications but until recently little work was done in applying them to DSP. This is changing quickly and a flood of F P G A applications in DSP is begin­ning to appear [6] [48] [49] [39]. In instances where FPGA based computers were used for DSP, the algo­rithms have typically been fairly simple and involved short wordlengths. 6. Architecture Implementation Alternatives 78 A comparison of F P G A custom computers and conventional processors in DSP applications is given in [9]. It is concluded that, at present, reconfigurable architectures do not give any cost/performance improve­ments over conventional uniprocessor and parallel processor systems for DSP applications. This is mainly because commercial chips have dedicated hardware support for arithmetic. Custom computers offer the best speed-ups when they can implement in hardware what sequential computers have to do in software. The extensive on-chip support for fixed and floating point arithmetic in modern DSPs and general purpose CPUs makes them difficult to surpass using programmable logic. Two major challenges are evident in all the FPGA based reconfigurable systems developed to date. The first is the relatively low density of the F P G A devices. This requires that large arrays of FPGAs be used. The limited number and low speed of the off-chip connections make such designs difficult and time-con­suming. The second challenge is the need for a new set of tools to deal with the design of reconfigurable machines. At the higher level, a new type of compiler is necessary to specify programmable hardware designs. At the lower level, better synthesis and place-and-route tools are required. Since there is an overhead associated with reprogrammability, FPGAs will always be a step behind single purpose VLSI devices for fixed function DSP applications. For example, it will always be possible to design an FFT processor in a custom VLSI chip that is faster than a version implemented in FPGAs. How­ever, since FPGAs have the advantage that one device can be used to implement more than one algorithm, they may be the more cost-effective solution for some applications. The applicability of reconfigurable computing to SAR processing is examined in more depth in Section 9.8. 6.7 Custom Algorithm Specific Processor Custom processors are the traditional approach to building very high throughput DSP systems. Custom processors are defined to be processors that are hardwired for one algorithm and cannot easily be used for another algorithm that is significantly different. Custom processors make use of either custom or commer­cial VLSI chips on custom circuit boards. This type of processor can utilize an arbitrary amount of parallelism and can achieve extremely high per­formance. Custom SAR processors have usually used some form of coarse grain pipelining. For example, the N A S A Advanced Digital S A R Processor (ADSP) built by JPL implements the range-Doppler algo­rithm in a hardwired pipeline [27]. It achieves a computation rate of over 6 GFLOPs at a cost of 73 boards. A simplified and more modern design is the N A S A Alaska SAR Processor (ASP). It achieves about 3 6. Architecture Implementation Alternatives 79 GFLOPs at a cost of 35 boards. The IRIS airborne SAR processor from M D A is also a custom hardwired pipeline processor. The cost-effectiveness of custom SAR processors needs to be examined careful before embarking on a custom processor design. The major disadvantages of custom processors are high development cost, poor flexibility, and potentially low reliability. These factors rule out the use of custom processors in many applications. In addition, the speed of commercial more general purpose architectures has increased greatly, and these architectures are now suitable for applications that previously required a custom approach. As a result, there is a trend today away from custom processors. Custom processors are not con­sidered further in this thesis since one of the goals of this work is to find high-performance architectures that are also flexible. 6. Architecture Implementation Alternatives 80 7 Memory System Design SAR processing requires extremely large amounts of memory and high data transfer rates. It is important to emphasize memory system design right from the beginning of the design process since slow D R A M access times or memory access conflicts can easily make memory latency the main bottleneck in the sys­tem. SAR system designers have an advantage over general purpose computer system designers since the memory access patterns in SAR can be largely predicted. This predictability must be exploited in the design of high performance but cost effective memory systems. Memory requirements also set SAR proc­essors apart from many other high performance DSP systems. Typical high throughput DSP systems are designed to perform a single set of operations on a stream of data. They do not have the memory support to perform repeated operations on a large data array that is stored in memory and accessed in various direc­tions. This section examines the various issues involved in memory systems design for SAR processing. There is no reference that explains the design process for such systems. The contribution of this section is to draw from the wealth of well-known design ideas related to memory system design, choose the most important ones, and apply them to SAR processing. High performance data path design is introduced in Section 7.1. Section 7.2 looks at the memory access patterns in SAR. The problem of corner turning is discussed in Section 7.3. The most common ways of overcoming the problem of high memory latency are covered in Section 7.4. 7.1 High Performance Data Path Design The general goal of high performance data path design is to minimize latency and maximize throughput. Many different factors can cause these areas to miss their performance goals. Latency problems can be caused by slow R A M access times, cache misses, and memory or bus contention. Throughput problems can stem from slow memories, slow clock speeds, narrow data path widths, and interference from other tasks or processors. These problems can also cause the latency and throughput to be non-deterministic, which is a problem in real-time systems. In a real-time processor, it should be possible to exactly predict all data transfer times. Some general data path design techniques that can help to achieve high performance are listed below. • Use multiple data buses to increase throughput and minimize contention. • Dedicate each data bus to a small number of precisely defined uses. 7. Memory System Design 81 • Use as wide a word width as possible on each data path. • Use high data transfer clock rates. • Operate the data buses synchronously where possible to avoid synchronization delays. • Use a fast, deterministic bus arbitration method. • Isolate data paths with double buffers or dual port memories. • Use one of the techniques from Section 7.4 to reduce average memory latency to the required level. An appropriate balance between performance and cost must be found when choosing which of these tech­niques to apply in a system. 7.2 Memory Access in SAR Processing In SAR processing, memory is generally accessed in a regular fashion along the rows and columns of the data array. This suggests that R A M s which provide faster access for consecutive addresses would be use­ful. Once a row or column has been read from main memory, it is typically used a number of times before the next row or column is required. This indicates that caching of data rows or columns may be used to advantage. The main exception to the rule of regular accesses is the corner turn operation, where memory accesses switch between row-wise and column-wise. This point in SAR algorithms was a major bottleneck in the days when the data was stored on disk. One would expect that once the entire data array is stored in R A M , the need for corner turning would disappear. However, the relatively slow random access time of D R A M devices and the characteristics of the methods that are used to overcome this D R A M latency problem have made this untrue. 7.3 Corner turns 7.3.1 Singe Processor Corner Turns Whether or not a comer turn is necessary depends on several factors: • Caches The presence and organization of caches will often be the deciding factor in the decision to perform comer turns. Caches are discussed in more detail in Section 7.4.2. 7. Memory System Design 82 • R A M access speeds The difference in R A M cycle times between random access and sequential access will help decide if a separate comer turn is worthwhile. • Memory access patterns In a system with R A M s that have faster sequential access, the number of times that memory is accessed in the orthogonal direction between implied comer turns will help determine if a separate comer turn operation is worthwhile. The effect of the last two factors is examined in more detail in Section 7.4.4. The simplest way of performing a corner turn is to simply read one row and column, and then exchange them while writing them back to memory. This requires temporary storage of the size of one row or col­umn. In this method, one half of the memory accesses will be in the non-preferred direction. A number of methods exist for performing comer turns in cases when memory can only be read in one direction [32] [33]. Most of these methods were developed for data arrays that were stored on disk or tape. They require multiple passes through the data matrix. In general, the size of the temporary storage area is traded-off against the number of passes that are needed through the data array. Whether or not these meth­ods offer a performance improvement in the case of D R A M memory depends on the difference in access times for random and sequential access. If it is decided not to perform a comer turn, it is still possible to perform 2D processing with only single direction accesses. One method for performing the 2D FFT with accesses only in one direction is described in [4]. This approach is potentially useful for range-Doppler SAR processing since it simplifies the han­dling of R C M C . 7.3.2 Multiple Processor Corner Turns A comer turn in a parallel system involves an exchange of data between PEs such that each PE obtains a strip of data that is orthogonal to the one it had originally. The comer turn is troublesome in parallel sys­tems since it involves only communication between memory and processors and no computation. This stresses the interconnection bandwidth in the parallel system. Unfortunately in many parallel systems, communication is much slower than computation. The communication-only nature of comer turns can expose bottlenecks in parallel systems. For example, on a system with a shared memory or a shared bus, the parallel speed-up can be very low during corner 7. Memory System Design 83 turns. This is due to the fact that all PEs attempt to access the shared resource at the same time. For corner turning to be truly parallelized, distributed memories and parallel communication paths are necessary. In a multicomputer with a horizontal partitioning of the SAR algorithm, the need for corner turns depends on the way the data is partitioned. If the data is suitably overlapped, each processor only needs to operate on one portion of the data (static partitioning) and a global corner turn operation can be avoided. In this case, each PE may not need to communicate with the other PEs during processing. However, this approach leads to inefficiencies and inaccuracies due to the required data overlaps. The decision whether or not to perform global corner turns during processing depends on the relative costs of the corner turn and the decrease in efficiency that results if no corner turn is done. The decrease in effi­ciency in the no corner turn case arises from data overlaps due to filter lengths. The cost of a corner turn depends on the speed of interprocessor communication and the topology of the interconnection network. In a vertically partitioned (pipelined) system, the corner turn memory is usually used as the memory buffer between pipeline stages. The corner turn buffer can be single or double buffered. Concurrent access to both rows and columns from two pipeline stages can be achieved in both cases. The simplest situation is the use of full double buffered memories. This has the advantage of having sim­pler control sequences and the ability to access any part of the entire image. However, it has twice the memory cost of the single buffered scheme. A single buffered corner turn memory is less expensive but requires a somewhat more complicated control and addressing mechanism. The control logic has to ensure that no access conflicts occur. Also the output data from a single buffered corner turn memory is only available in a staggered format. 7.4 Memory Latency Although SAR processing has different characteristics than many other processing problems, the ways around the memory latency problem are the standard methods of • wide data path widths, • caches, • interleaving, and • faster RAMs . 7. Memory System Design 84 7.4.1 Data Path Width SAR processing typically operates on fairly wide data words. The most common arrangement is to use complex 32-bit numbers for a total of 64 bits. The transfer of one complex word would take eight cycles on an 8-bit data bus. Obviously it is desirable to use a data bus that can accommodate the full data word width. 7.4.2 Caches Caches are extremely widely used in general purpose computers. Caches typically operate by fetching a whole data line from main memory when one address is requested. This is very inefficient i f the memory accesses are orthogonal to the direction of the cache lines. This is exactly what can happen during a comer turn operation. In the worst case, each memory access will result in a cache miss and a whole new cache line will need to be fetched from memory. As a result, on many modem general purpose computers, an explicit corner turn operation is required to make memory accesses efficient in both directions. Some other considerations on the use of caches in SAR systems are given below. • Data caches can also cause problems if they are not large enough to hold the entire quantum of data being processed: a range or azimuth line in the case of SAR. If a cache is too small, a data line may have to be fetched from main memory multiple times instead of just once. • While normal data caching schemes can be a problem in SAR processing, program code caching can be quite useful in SAR since signal processing code is often quite small. The compute intensive por­tions often consist of several tight loops which can be stored in a cache. • Caches are especially problematic in a real-time system since caching can introduce non-deterministic delays. • If possible, it can be advantageous to disable the caching mechanism during some parts of processing. Most of SAR processing operates on one range or azimuth line at a time. This indicates that a cache that holds one or several lines would be useful. A data line can be fetched from main memory and stored in a high speed cache memory. If the cache is made double buffered, one of the buffers can be written back to memory and refilled with the next data line while the other one is being processed. Provided the memory system can keep up with the processor, memory latency and data transfer times can be completely hidden in this way. The access time requirements of the cache memories can be reduced by using separate caches for the input and output to the processor. Figure 7.1 shows a sample architecture that makes use of these techniques. 7. Memory System Design 85 Controller Figure 7.1: Sample double buffered caching scheme 7.4.3 Interleaving Interleaved memory banks can be used to improve the average memory access time. Memory locations are spread out among a number of parallel memory banks. When an address is activated, a location in each bank is accessed simultaneously. Provided all these locations need to be accessed by the processor, the effective memory access time is reduced to t a c c e s s / N , where N is the number of memory banks. Memory interleaving also has some disadvantages. Interleaving requires a more complicated memory controller. It also requires that memory be upgraded in larger increments since the size of each memory bank must be increased together. Like the other schemes for speeding up memory access, interleaving relies on the fact that most memory accesses are to sequential locations. Thus it should be possible to design an effective memory interleaving scheme for SAR processing since the memory access patterns can be predicted. In SAR, fast sequential access is required to both the rows and columns of the data array. The techniques developed for parallel memory access in array processors [18] [57] are directly applicable here. In these methods, there is a gen­eral trade-off between the complexity of the memory system and its potential performance. More sophisti­cated schemes require more elaborate memory controllers and alignment networks. However, since memory access methods in SAR processing aire confined to row and column accesses, a straightforward skewed storage approach can be used to provide fast access in both the row and column directions. More complex methods, such as prime memories, are not needed. Figure 7.2 illustrates the storage of a data array with no interleaving, skewed interleaving, and a simpler interleaving scheme. With four memory banks, skewed storage gives a four times improvement in memory access time. However, skewed storage requires a fairly complicated memory control system. The second 7. Memory System Design 86 interleaving scheme shown in Figure 7.2 only provides a two times improvement in memory access. How­ever, it is much simpler since it relies on the fact that the array sizes are a power of two and several of the address lines can be used to rearrange the data in memory. Through the use of a programmable memory controller, various array sizes can be accommodated. If the array dimensions are not powers of two, there will be some wasted storage space since this scheme will need to use the next largest power of two. No interleaving 0, 1. 2, ... 8192. 8193. 8194.... 16384, 16385, 16386, ... 32768, 32769, 32770, ... Four memory banks with skewed interleaving (8,=1, 82=1) 0,4, ... 8195, .. 1,5, ... 8192, ... 16387, . 32770, . 2, 6, ... 8193, ... 3, 7,... 8194,... 16385, . 32768,, Four memory banks with simple interleaving 0. 2, 4,... 1,3, 5. ... 16384, 16386, ... 16385, 16387, ... J > a13=0 8192, 8194, ... 8193, 8195, ... \ 24576, 24578, ... 24577, 24579,... J > a13=1 At lease one dimension of the data array is equal to 8192 in this figure Figure 7.2: Some interleaved memory schemes 7.4.4 Fast DRAMs Most new families of fast D R A M devices provide accelerated access to sequential memory locations but do not improve the random access time. In SAR processing, the memory must be accessed sequentially in orthogonal directions. This causes problems with the use of the new fast types of D R A M S like burst, col­umn, nibble, EDO, and synchronous. A l l of these devices allow fast access only in one direction. In order to reap the benefits of this fast access, the data would have to be explicitly corner turned during execution. This would allow all accesses for processing to occur at the fast speed. However, during corner turning, either some of the accesses have to occur at the slow random access speed of the D R A M or several passes through the data must be made. Figure 7.3 shows a graph of the average access times for an 8k x 8k data array for regular D R A M s and fast D R A M s as a function of the ratio of number of accesses to number of corner turns. In the case of the regu­lar DRAMs, no corner turns are performed and all the accesses occur at the slower random access speed. In the case of the fast DRAMs, two corner turns are performed and all other accesses occur at the fast speed. It is assumed that the random access cycle time is 110 ns for both types of D R A M . The cycle time for the fast D R A M is assumed to be 40 ns in fast direction. 7. Memory System Design 87 1 1 1 I 1 1 1 1 — fast DRAMs with comer turn * . normal DRAMs without comer turn / I ' l l I 1 I I I 1 2 3 4 5 6 7 8 9 10 Number of accesses per corner turn Figure 7.3: Average data array access time in slow direction when using fast RAMs The figure shows that the choice of memory configurations depends on the number of times the data is accessed between comer turns. In SAR this depends on both the algorithm and the architecture and can range between two and ten. The graph shows that the crossover point for this scenario is about four accesses per comer turn. This can be reduced by using a more sophisticated comer turning scheme like the one described in [32] at the cost of a larger temporary storage area. The crossover point can be reduced to two or three accesses if a temporary storage array large enough to hold 32 rows is used. Other types of fast D R A M s incorporate on-chip caches with extremely fast cache to memory transfers. These chips do not necessarily offer good performance for SAR because of the fixed mapping and rela­tively small size of the caches. This includes the new DRAMs that adhere to the RAMbus standard. These chips have extremely fast peak transfer rates (up to 500 MB/s) but have the dual disadvantages of fixed caches and a narrow data path (8 bits). Since SAR processors use large amounts of memory, it is desirable to use readily available inexpensive D R A M s as opposed to more expensive SRAMs. At present, this means fast page mode D R A M s and in the future it may mean EDO or synchronous RAMs . In any case, the overall design of the SAR system should attempt to take advantage of the characteristics of the D R A M devices used in the memory. ZD 20 Average time to 1 5 access complete 8k x 8k data array (sec) 1 0 7. Memory System Design 88 8 Interconnection Networks Interconnection networks are a key architectural component of most parallel systems. This section looks at some of the issues related to interconnection networks for modem parallel SAR systems. The focus is on those networks that are the most cost-effective with today's technology. The networks are analyzed with respect to the requirements of SAR processing. In particular, an original analysis based on comer turn and matrix transpose performance is presented. There are two facets to interconnections in parallel systems: • system I/O and • inter-processor communication. These two areas are examined in the sections that follow. 8.1 System I/O System I/O is comprised of the data transfers necessary to move the data into and out of the SAR proces­sor. Since these transfers often occur over a global system bus, they are a potential bottleneck. The delay due to system I/O can be removed by double buffering the input and output at the cost of additional mem­ory. The cost of this option is often prohibitive in SAR systems due to the large image sizes. System I/O interconnection should offer • high bandwidth, • independent I/O channels, • efficient interfaces to mass storage devices, • minimal requirements for C P U intervention, and • double buffering where necessary. Since the design of the system I/O interconnection is usually driven by the requirements of the specific sys­tem, it is not explored in more detail here. 8.2 Inter-processor Communication The inter-processor connection is responsible for carrying the communication that occurs between the PEs. Some of the key attributes required of interconnection networks for SAR processors are a high data trans-8. Interconnection Networks 89 fer rate, scalability, and efficient support for corner turning. The greatest test of the interconnection net­work in a SAR processor arises during parallel corner turn operations. The discussion presented here focuses on the network topology. As with all aspects of the interconnection, the performance and cost of a given topology depends largely on the type of inter-processor communica­tion support built into the PEs. There are many possible interconnection networks topologies [35]. Some common topologies that are practical for use with current DSP chips are the bus, mesh, and crossbar. These topologies are examined in the sections that follow. 8.2.1 Bus The bus is perhaps the simplest interconnection scheme as can be seen from Figure 8.1. It can be imple­mented with almost any DSP chip. DSP chips that have multiple independent buses have a definite advan­tage for bus interconnections since they can separate the local and global memory accesses onto distinct buses. The TMS320C40 and the DSP96000 are examples of DSP chips with independent buses. PE PE PE PE PE PE PE PE Figure 8.1: Bus interconnection Two problems with buses are poor scalability and slow comer turns. Some of these disadvantages can be overcome by using multiple buses within the system. This is the approach taken by the VASP boards from Spectrum Signal Processing. These boards use multiple independent high bandwidth buses in conjunction with two-port memories to provide multiple data paths. During a global comer turn in a distributed memory system, each PE must communicate with all other PEs. This is problematic on a bus since all data transfers must be serialized. The time required to perform a comer turn on a system with a single bus can be expressed by pe pe where • N is the number of processors, • Nr is the dimension of the data array in range, 8. Interconnection Networks 90 • N is the dimension of the data array in azimuth, and • BW is the bandwidth of the bus in MB/s. Complex 32-bit values are assumed. Figure 8.2 shows a plot of the time required to corner turn an 8k x 4k image on a system with a bus bandwidth of 150 MB/s against the number of PEs. 2| Comer turn time (sec) 1 60 70 20 30 40 50 number of processors Figure 8.2: 8k x 4k image corner turning time on various interconnection networks 8.2.2 Mesh Meshes are the simplest two dimensional interconnection scheme. Figure 8.3 shows a sample of a mesh interconnection. PE PE PE PE PE PE PE PE PE PE PE PE PE PE PE PE Figure 8.3: Mesh interconnection Meshes are typically implemented using DSP chips that have on-chip I/O ports designed to support multi­processor communications. Both the TMS320C40 and ADSP21060 have six such communication ports. 8. Interconnection Networks 91 The on-chip ports allow for very low chip count solutions. For example, the ADSP21060 with its 512 kB of on-chip memory can form a single chip PE. Board products that support meshes of DSPs are available from Ariel, Spectrum, Integrated Computing Engines, and numerous others. Meshes have the advantage of being quite efficient at corner turns. The time required for a corner turn on a mesh of processors can be expressed as: where the symbols have the same meanings as in Equation 8.1. A toroidal mesh is assumed. This implies a maximum data shift distance of | ~ ^ / V p e J / 2 ~ J hops. This expression is plotted in Figure 8.2 for an 8k x 4k image with a 40 MB/s transfer rate between PEs. Some disadvantages of meshes are as follows: • The greater number of inter-PE connections makes them more difficult to build. • The lack of direct access to each PE can make external I/O and broadcasts inefficient. Hypercubes are a related type of interconnect. They are not as commonly used since they require more connections per node and since they do not scale as easily. Hypercubes exhibit similar corner turning per­formance to meshes. The corner turning time can be found from equation 8.2 by replacing the factor Crossbar interconnections offer the highest connectivity and also the highest cost. Crossbars have typically not been used in large multicomputers since full crossbar interconnections become very expensive. A com­mercial partial crossbar mechanism that circumvents this problem is the RACEway interconnect from Mercury Computer Systems [50]. Mercury's RACEway is a scalable crossbar scheme that has been stand­ardized. RACEway uses custom six port crossbar switches that are capable of two simultaneous 150 MB/s transfers at a time. RACEway offers good scalability because interconnect performance and cost is not fixed like on a bus, but scales with the number of processes. Figure 8.4 shows some topologies based on RACEway-style crossbar switches. Crossbar interconnects are often used for CPUs that do not have dedicated ports for interprocessor commu­nication like the i860. Crossbar based board products using the i860 are available from Mercury, Sky, and pe (8.2) 8.2.3 Crossbar CSPI. 8. Interconnection Networks 92 PE PE PE PE Crossbar PE PE PE PE PE PE PE PE Crossbar Crossbar Crossbar PE PE PE PE PE PE PE PE Crossbar Crossbar PE PE PE PE PE PE PE PE Crossbar Crossbar Crossbar Figure 8.4: Crossbar interconnection The time required to perform a comer turn on RACEway-style crossbar interconnections like the ones shown in Figure 8.4 can be expressed as (\log<{Npe)-\ k \ Y N - 3 - 4 Z-i pe k-l Pe 2 • N N • N i i -id"-2-4-aw- <83» £ _ i pes pe pe where the symbols have the same meanings as in Equation 8.1. Since the R A C E way crossbar chips support six ports, extra crossbars can be added to increase bandwidth and to provide fault tolerance. The model given here assumes no extra crossbars are used, which means at least one port is left unused in each cross­bar chip. The comer turn time for an 8k x 4k complex data array is plotted in Figure 8.2. Some of the issues that affect crossbar based systems are listed below: • The cost of the crossbar chips is a major factor in these systems. • The crossbar chips are often single sourced custom chips that increase the chip and pin count of the processing boards. 8.2.4 Conclusion This section has looked at three common topologies. The mesh topology was found to have the most scal­able comer turn performance. Many other possibilities exist that may also offer good performance. The comer turning performance of the topologies has been emphasized. This may not always be an important consideration. If the data is partitioned differently or if a global memory buffer is used, comer turning may not be necessary at all. 8. Interconnection Networks 93 9 Examination of Architectures This section brings together the trio of algorithm, architecture and technology that have been looked at separately in previous sections. The following design approaches are considered. • network of workstations • general purpose DSP architecture • vector DSP architecture • optimized DSP architecture • digital filter chip architecture • fine grained parallel architecture • a sample vector/scalar architecture • F P G A computing machine The goal is to find architectures that meet the requirements set out in Section 3.5. This entails the use of the best currently available technology in cost effective architectures that are tuned to the needs of SAR processing but are still flexible enough to handle a range of algorithms. The selection and analysis of the architectures in this section is new work. The results and conclusions pre­sented here are new contributions to the study of SAR processor architectures. 9.1 Network of Workstations 9.1.1 NOW Model This section presents a simple model that will be used to test the potential usefulness of NOW for SAR processing. The intent is to obtain an idea of the performance that could be expected and the factors that affect performance. Static horizontal partitioning is used as the partitioning approach since it involves the least inter-processor communication of all the partitioning strategies. The details of the model are given below. • A network of ten workstations is assumed, where each workstation has either 25, 50, 75, or 100 MOP/s of sustained processing speed for SAR operations. These are reasonable values for current and near-future workstations. A 1 ms l k CFFT time is typical for many modem workstations, which cor-9. Examination of Architectures 94 responds to 50 MOP/s. Ten 50 MOP/s workstations would be capable of approximately 1/10 real-time processing. As an example, a DEC Alpha workstation has a peak performance of 200 MOP/s and achieves about 100 MOP/s on a lk CFFT. • The processing scenario is that of the range-Doppler algorithm modelled in Section 2.5.4. The real­time computation rate is about 4.5 GOP/s. The image size is 160 M B at the beginning of execution and 140 M B at the end. • A l l data is assumed to originate from a single node in the network and the output data is collected at the same node (see Figure 9.1). This is a realistic situation since a single frame synchronizer or tape drive is usually the source of the data. The network bandwidth figures used in this section refer to the bandwidth required at the connection(s) of the I/O node to the network. No other assumptions about the network topology are made. Figure 9.1: NOW SAR processor • The data is partitioned between the processors at the beginning of the computation and is collected at the end. Thus, the only inter-processor communication is the transferring of the complete image data array twice: once at the beginning and once at the end. • It is assumed that communication can be completely overlapped with computation. This implies that the execution time will be determined by the longer of the computation time and the communication time. • Algorithm overheads due to things like data overlaps are ignored in this section. This allows the effect of network communications to be studied in isolation. It should be remembered that the actual speed­up may be considerably less depending on the algorithm parameters. 9. Examination of Architectures 95 speed-up 200 R/C ratio (OPs/byte) 200 \\ \ 1 1 1 . \ • • \ \ \ . 75 -\ .50 V \ — \25 • . \ - -- - . — i r — — - — i — 0 50 100 150 200 network throughput (Mbps) network throughput (Mbps) (numbers on graphs indicate speed of workstations in MOP/s) Figure 9.2: Speed-up for ten workstation network vs network throughput and computation/ communication ratio vs network throughput 9.1.2 Results The plot of speed-up versus network throughput in Figure 9.2 shows that the network bandwidth must be very large before the full computational power of the workstations can be realized. The speed-up is defined as the execution time on one workstation divided by the execution time on ten workstations of the same type. For example, an attempt to process a full size SAR image on a network of workstations connected by a conventional 10 Mbps Ethernet network would be a disaster: the speed-up would be less than one. The relationship between computational performance and communications bandwidth can be described by the computation to communication ratio (R/C ratio). For example, a workstation capable of 100 MOP/s and a network capable of 10 MB/s have an R/C ratio of 10 OPs of computation per byte of communication. The SAR processing scenario used in this section has an R/C ratio of about 38 OPs/byte. If the R/C ratio of the architecture is lower than that of the algorithm, then the processors will be the bottleneck. If it is higher, then communication will be the bottleneck. R/C ratios are plotted as a function of network bandwidth in the second graph in Figure 9.2. A horizontal line indicates the ratio that characterizes the hypothetical SAR problem. The intersections between the curves and the line correspond to the communications throughputs at which the ten workstation network achieves linear speed-up. Since the R/C ratio of the S A R problem is known, the network bandwidth required for a given amount of processing power can be predicted. This is shown in Figure 9.3, which plots the required communications bandwidth against the number of workstations. It can be seen that extremely high network bandwidths are required to achieve linear speed-up. 9. Examination of Architectures 96 400 h 300 h communication bandwidth required for linear speed-up 2 0 0 (Mbps) 100 H numbers on graph indicate speed of workstations in MOP/s 10 15 number of processors Figure 9.3: Network throughput required vs. number of workstations If the more realistic assumption is made that only 50% of the computation and communication can be overlapped in time, then it becomes much more difficult to achieve linear speed-up. This can be seen from Figure 9.4 which is a plot of the speed-up versus network bandwidth for this new assumption. 10 | speed-up 1 1 s o . - • / . - ' 7 5 „ — / ' ^ i " 1 0 0 , - ' _ / - ' / - ' ' ^ / - s • / • ' / - ' / ' ' ' / / -L • y.' _ • ' / / f i i i 50 100 150 network throughput (Mbps) 200 Figure 9.4: Speed-up for ten workstation network vs. network bandwidth with partially overlapped communications 9.1.3 N O W Conclusions A performance analysis of the range-Doppler algorithm on a NOW architecture has shown that very large network bandwidths are necessary for practical implementation of SAR processing on NOW. For example, in a ten workstation network where each workstation has a performance of 100 MOP/s, an effective net­work bandwidth of over 200 Mbps is required. This is beyond most commonly available networking tech­nologies. Extremely low overhead communications software is also a necessity. It should be stressed that 9. Examination of Architectures 97 the bandwidth numbers shown in this section are effective numbers after all overheads. Protocol and oper­ating system overheads can easily add 100% to the bandwidth required. In the analysis of the NOW architecture, the horizontal partitioning approach was used since it requires the minimum amount of inter-processor communication. However, as was seen in Section 2.6.4, this approach also suffers serious efficiency problems due to data overlap requirements. An approach that includes a glo­bal corner turn could potentially solve this problem but it would increase interprocessor communication significantly. NOW become more appropriate for SAR processing if the algorithm requires a significantly higher amount of computation per image pixel than the standard image formation algorithms like range-Doppler. A possi­ble candidate application is a SAR processor that incorporates large amounts of computation in post­processing functions such as those required for image analysis and understanding. 9.2 General Purpose DSP Architecture The most general purpose architecture that is still DSP-specific is one that uses general purpose DSP chips as building blocks. This section discusses the processor, memory, and interconnection subsystems for gen­eral purpose DSP based SAR processors. The performance of a sample architecture for SAR processing is also examined. 9.2.1 Processor Some desirable features of a general purpose DSP for a SAR processing system are listed below. • floating point number support • especially fast FFT execution time • high data transfer bandwidth • multiple independent data buses • inter-processor communication support • large on-chip memory Currently the chips that come closest to meeting this ideal are the Analog Devices ADSP-21060 SHARC [1] [16], the TITMS320C40 [84], or the TITMS320C80 [86]. These processors each have strengths and weaknesses. Before choosing any one processor, a trade-off analysis must be performed. 9. Examination of Architectures 98 9.2.2 Memory General issues related to memory system design were covered in Section 7. A few of the issues that are particularly important in DSP based designs are given below. • the distribution of the total memory into local (to each DSP) and global sections • the scalability of the memories over a large range of sizes • the need for distributed memories for high bandwidth portions of the memory system • the potential bottleneck posed by shared memory areas • the careful use of caches or buffers • efficient support for D M A transfers 9.2.3 Interconnect The properties of several of the most common interconnection networks were discussed in Section 8. Some issues specific to DSP based systems are listed below. • The choice of interconnection is affected by the inter-processor communications support provided by the DSP chip. • The use of standardized interfaces makes the system more general and more easily expandable. 9.2.4 Sample Architecture An example of a general purpose DSP based architecture is presented in this section. The goal is to design the most cost effective architecture that is capable of 1/10 real-time satellite SAR processing. The AD21060 S H A R C is selected as the DSP for the architecture. Some of the strengths of this chip are given below. • large high bandwidth dual-ported S R A M (512kB) • flexible D M A structure • six 40 MB/s link ports • high FFT performance • IEEE floating point number format • potentially a single chip PE for multiprocessor systems 9. Examination of Architectures 99 The S H A R C also has some weaknesses: • only one external bus • on-chip R A M may be too small for some applications If the TMS320C40 were used instead of the SHARC, some design differences would result. • The lower FFT performance would mean that more PEs are required. • Each PE would require more external memory. • The dual external buses would provide more design flexibility. The additional bus could be used as a global high bandwidth communication channel. A toroidal mesh is chosen as the interconnection network for the processor. A block diagram of the proces­sor is given in Figure 9.5. The mesh is supported by the S H A R C with no need for extra chips. Two of the SHARC's link ports are left unused by the inter-processor connections. On certain PEs, one of these ports is used for system I/O. memory host bus H interface Figure 9.5: General purpose DSP based mesh processor If it is assumed that the SHARCs can sustain 25% of the performance they exhibit on FFTs, then about 16 processors would be needed for 1/10 real-time satellite SAR processing. Since the SHARCs ' internal memories are not big enough to hold the entire image, external memories are required at each PE. To proc­ess an 8k x 4k image, 16 M B of memory is required at each node. 9. Examination of Architectures 100 A l l inter-processor communication occurs over the link ports. One of the link ports of each of the PEs at one edge of the mesh is connected to an additional SHARC. These four 40 MB/s links provide a 160 MB/s system I/O link. If more I/O capacity is required, an additional S H A R C can be added and attached to the extra link ports of other PEs in the array. The other unused link ports can be used for additional I/O or for fault tolerance purposes. A simple model was created to predict the performance of this architecture for the range-Doppler process­ing scenario of Section 2.5.4. The model shows that this architecture is capable of about 20% of real-time operation. The model takes into account the time to load data into and out of the mesh, all core SAR processing calculations as well as resampling, and the time required for two corner turns. Theoretical SHARC execution times are used. The time required by overhead functions executed by the DSPs and any delays due to external memory access have not been taken into account. Also, coefficient and filter calcula­tion times have not been considered. These functions are fairly easy to accommodate in this architecture since they can be performed by any PE. However, they will lead to additional computation and communi­cation. Since it is reasonable to plan for about a 100% overhead to the theoretical model, a 16 processor architec­ture seems well suited to achieve 1/10 real-time operation. This verifies the earlier prediction based on FFT performance alone. In an actual design, it is often desirable to have a safety margin. If a 50% safety margin is used, then 24 processors are required. 9.2.5 General Purpose DSP Conclusions General purpose DSP based architectures pose a good solution for low to medium speed SAR processing. They strike a reasonable balance between performance and flexibility. Compared to architectures based on microprocessors, they offer a lower chip count and lower cost solution. However, they are more special­ized and require more design effort than SAR processors based on standard microprocessor systems. For processing speeds that approach real-time, very large numbers of processors are required compared to designs based on more specialized chips. The design of general purpose DSP systems is also somewhat riskier than designs of more specialized architectures since performance prediction is much more difficult. 9.3 Vector DSP Architecture This section examines the design of a SAR processor based on the Sharp/Butterfly LH9124 vector DSP. This chip was chosen for a number of reasons: extremely fast FFT time, support for other operations besides FFTs, availability of boards and software, and adequate numerical precision (24-bit block floating 9. Examination of Architectures 101 point). The time spent looking at LH9124 designs in this section is justifiable since the chip has a well thought-out architecture that is a good example of high performance design. The results obtained are also applicable to architectures based on other chips. An overview of the LH9124 is given in Section 9.3.1. Standard single chip architectures are presented in Section 9.3.2. A mapping of the range-Doppler algorithm to an LH9124 based processor and some per­formance predictions are given in Section 9.3.3. Sections 9.3.4 and 9.3.5 give some multi-chip architec­tures and range-Doppler performance predictions. The performance of the LH9124 on some other SAR algorithms is discussed in Section 9.3.6. Section 9.3.7 gives some conclusions regarding the use of the LH9124 in SAR processing. 9.3.1 Introduction to the LH9124 The LH9124 is a high-performance fixed-point DSP optimized for vector oriented operations [80]. An overview of the chip is shown in Figure 9.6. Some of its key features are given below: • The data word width can be 8 to 24 bit real or complex with support for block floating point. • 25, 40 and 50 M H z versions are available and 66 and 80 M H z versions are planned. • Data I/O occurs via four independent, bi-directional 24-bit complex buses. • Built in operations are available for 26 functions like real and complex FIRs, multiplications, and radix-2, -4, and -16 butterflies. • The heavily pipelined architecture is very efficient for long vector lengths. • Programming is relatively straight-forward due to the high level of the instructions provided. • Performance prediction is also quite easy. Some shortcomings of the LH9124 are listed below: • There is no on-chip R A M or address generator. Address patterns can be generated by a companion address generator chip: the LH9320. • It is not suited for scalar computations. • There are no comparative, branching, or flow control capabilities. • The block floating point format is incompatible with floating point formats used by most general pur­pose DSPs. • FIR filtering speed is much slower than the speed of specialized digital filter chips. 9. Examination of Architectures 102 • A relatively large number of support chips, like address generators and SRAMs, are required. • The interface to a larger memory system most likely requires a custom design. ACQUISITION P O R T R E A L A 4 IMAGINARY 24-7 D A T A P O R T A - R E A l I 2 4 * ' F U N C T I O N C O D E D A T A F L O W 3 ' Z4t D A T A P O R T A IMAGINARY Q R AR F C Ql an LH9124 Al C R 24' 24'. ' D A T A P O R T B - R E A L 24; D A T A P O R T B IMAGINARY R E A L T T IMAGINARY C O E F F I C I E N T S * D A T A WIDTH S E L E C T A B L E F R O M 8 T O 24 Figure 9.6: LH9124 digital signal processor A list of execution times for the basic DSP operations is given in Table 9.1. FFTs are performed in passes composed of radix 2, 4, or 16 stages. For example, the most efficient way of computing a l k FFT is to decompose it into two radix-16 stages and one radix-4 stage. The data vector is transferred through the chip for each pass. The highest processing efficiency is achieved with radix-16 butterflies. Table 9.1: LH9124 execution times Operation Time Time at 40 MHz (tcycle=25 ns) (usee) Filtering (Nfilter' N + 18) ' {cycle f o r e a c h fi,tered point -Multiplication "" fcycle -512 CFFT (3 -512 + 2- 68+ 18) • tcycU 42.3 lkCFFT (3-1024 + 2-68 + 18) • tcycle 80.7 2k CFFT (4-2048 + 2-68 + 2-18) • tcycle 209.1 4k CFFT (3-4096 + 3-68) • tcyc[e 312.3 8k CFFT (4-8192 + 3-68 + 18) • tcycle 824.8 Board products that incorporate the LH9124 are available from Catalina, Interactive Circuits and Systems, Pacific Cyber/Metrix, Valley Technologies, and others. 9. Examination of Architectures 103 9.3.2 Single LH9124 Architectures A sample single LH9124 configuration is shown in Figure 9.7. A double buffer is used on the A , B and C ports and a single buffer on the Q port. Any other buffering configuration is possible. Not shown in the fig­ure are • the scheduler that provides the overall control functions, • the main memory, its controller, and its interconnection with the buffers, and • a general purpose C P U for scalar computation. a R A M ! t Imag. Real _ Imag Q Real Real LH9124 D A B Imag Imag c Real Imag TT 3" 1 T • Real LH9320 Address Generator Real Imag. — ^ Imag. TT Figure 9.7: LH9124 single chip architecture The C P U used for scalar computation must have access to the data being processed. This C P U could be a general purpose DSP, and will most likely be a floating point processor since it wil l perform more compli­cated higher level functions. For the algorithms that are modelled in the sections that follow, it is assumed that all the other elements of the system are fast enough to allow the LH9124 to execute at its maximum rate. This is not unreasonable since most of the computation will be done by the LH9124. It will be shown that the data transfer rates are also not unreasonable provided the data path widths are wide enough to handle the wide word widths demanded by the LH9124. The task of keeping the LH9124 busy can be achieved by using: • a main memory that is fast enough to fill and empty the buffers on time, • double buffering to hide the time taken by data transfers, and 9. Examination of Architectures 104 • a sufficiently fast scalar C P U . Some of the design issues for the memory subsystem are • the number of memories and memory buses, • the nature of the connection between the LH9124 buffers and the main memory, • the design and complexity of the control structure, and • the connection point for external I/O. Some sample interconnections between LH9124 based PEs and a main memory are shown in Figure 9.8. As always, a trade-off exists in these designs between complexity and performance. A LH9124 processing element is represented by the simplified block diagram shown at right. Only the LH9124 buffers that are on an I/O path are shown. double buffer Single port, single bus, single memory memory -*\ controller double buffer LH9124 LH9124 Multi-port, multi-bus, single memory memory double buffer memory controller Multi-port, single bus, multi-memory double buffer LH9124 double buffer memory controller Figure 9.8: Some sample LH9124 to memory interconnections The interconnection with scalar CPUs has not yet been addressed. There are many ways to do this and the method selected will depend on the specific application. This is explored in more depth in Section 9.7, where a design that combines the LH9124 with several scalar DSPs is presented. 9. Examination of Architectures 105 9.3.3 Range-Doppler Algorithm on a Single LH9124 Architecture An implementation of the range-Doppler algorithm described in Section 2.5.4 on an LH9124 is modelled in this section. The vector operations in the algorithm are assigned to the LH9124 and the scalar operations to the general purpose C P U . Table 9.2 gives the mapping for each step in the algorithm. The table also shows the execution times and data transfer rates for each of the steps allocated to the LH9124. Some assumptions that were used when generating these values are listed below. • Complex vector multiplies are rolled into the first stage of the FFTs. • A single bus, single main memory architecture is assumed (see Figure 9.8). • The address generators are able to support all the operations required. • A double buffer is used between the LH9124 and main memory. A l l data transfers between main memory and the double buffer occur while the LH9124 is processing a vector in the other half of the double buffer. • The data is transferred back to main memory after the azimuth FFT and then fetched back in the range direction for R C M C . This essentially adds two corner turns, but the alternative of performing R C M C across multiple lines stored in the buffer is not supported by the LH9320. Table 9.2: Range-Doppler algorithm implementation with a single LH9124 Algorithm step Processor Execution time (sec.) Data transfer rate between main memory and buffer (MCW/s) 1.01/Q Balance GPCPU - -2.0 Range FFT, 3.0 Range Multiply, 4.0 Range IFFT LH9124 5.8 9.9 5.0 Azimuth FFT LH9124 1.7 26.2 6.0 RCMC Shift and Interpolation LH9124 4.4 10.0 7.0 Azimuth multiply, 8.0 Azimuth IFFT LH9124 2.2 6.3 9.0 Azimuth Resampling LH9124 5.3 8.4 10.0 Range Resampling LH9124 7.4 -11.0 Detection GPCPU - -12.0 Calculate Range Filter and SRC GPCPU - -13.0 Calculate Doppler Centroid GPCPU - -14.0 Calculate RCM Shifts and Indices GPCPU - -15.0 Calculate Ambiguity GPCPU - -16.2 Azimuth filter FFT LH9124 1.7 -16.1, 16.3, 16.4 Calculate Azimuth Filter GPCPU - -9. Examination of Architectures 106 As can be seen from the table, the resampling operations take up 45% of the total execution time. Without the resampling operations, the architecture is able to achieve 17% of real-time. With the resampling, only 9.3% is obtained. To achieve 1/10 real-time operation with resampling, two LH9124s in a parallel configu­ration are needed. The data transfer rates between main memory and the buffer at the LH9124 input port are shown in the right-most column of Table 9.2. The highest data transfer rate between main memory and the buffer occurs for the azimuth FFT stage of the algorithm. This is due to the fact that the azimuth FFT stage performs the least amount of computation between data transfers. The maximum transfer rate is 26 Mwords/s or 150 MB/s. This is higher than that achievable with normal DRAMs. However, a memory that makes use of a combination of wide data paths, interleaving, and fast R A M s can reach this transfer rate. The computation requirements on the general purpose C P U are 36.3 MOP/s if it is to keep up with the LH9124 on the range-Doppler algorithm with resampling. This is close to the performance available on a single chip DSP. However, it is doubtful if this level of performance could be sustained on a single C P U if it is also responsible for higher level control functions. 9.3.4 Multi-LH9124 Processor Architectures The LH9124's data path oriented architecture makes it fairly well suited to parallel configurations. The many possible configurations can be grouped into two general categories. Fine grain Fine grain approaches use multiple LH9124s to speed up fundamental operations. Cascade or parallel configurations are used with interconnections that support pipelined operations and parallel FFTs. The advantage of this approach is higher throughput and near-linear speed-up. The main disadvantage is the potential lack of flexibility. For example, the speed-up of cascaded architectures depends on the FFT length. Thus, good speed-up may not be available for all algorithms. Coarse grain Coarse grain approaches use replicated LH9124 based processor-memory subsystems. These PEs ran in parallel on distinct data subswaths. This type of architecture is much more general, however, it may not be as efficient as a fine grain approach. A combination of coarse and fine grain approaches may also be used. Some possible interconnections for both coarse and fine grain parallelism are shown in simplified form in Figure 9.9. Each PE in the figure represents an LH9124 and its associated buffers and address generators. 9. Examination of Architectures 107 main memory Parallel main memory main memory PE \ • Bus-based parallel PE PE PE Cascade main memory PE PE PE Cascade/parallel main memory PE Hb \ i PE A PE r PF PF Figure 9.9: Some options for LH9124 parallelism main memory main memory It should be noted that if the number of LH9124s is increased, the number of general purpose CPUs in the system will have to increase by a similar amount. 9.3.5 Range Doppler Algorithm on Multi-LH9124 Architectures This section examines the parallel execution of the range-Doppler algorithm described in Section 2.5.4 on an architecture composed of multiple LH9124 processors. Each PE is assumed to be an LH9124 subsystem like the one shown in Figure 9.7. It is again assumed that the memory and interconnect can keep up with the PEs. The first method of parallelism that is examined is the coarse grain horizontal partitioning modelled in Sec­tion 2.6.4. The same allocation of tasks to the LH9124 as in the single processor model is used. The time taken by each processor is computed using the vector sizes from Section 2.5.4. Figure 9.10 shows a graph of the percentage of real-time achieved for varying numbers of LH9124s. This graph is similar to the ones obtained in Section 2.6.4, with the same discontinuities caused by the jumps in range FFT length. With resampling, real-time operation requires 15 LH9124s. Without resampling, real-time operation is attained with 10LH9124s. As before, the main problem with this approach is the inefficiency caused by the data overlap due to the relatively long range matched filter. Some possible solutions to this problem were mentioned in Section 2.6.4. If each PE were faster, then larger range blocks could be used. The PEs can be made faster by either using faster processors, like an LH9124 with a higher clock speed, or by using multiple LH9124s for each range subswath. 9. Examination of Architectures 108 number of processors Figure 9.10: Performance achieved for various numbers of LH9124 processors on the range-Doppler algorithm A method of partitioning that might improve the efficiency of the parallel architecture is to use dynamic horizontal partitioning. This method may require the use of a centralized comer turn memory. This would imply a system architecture like the one shown in the top half of Figure 9.11. Each PE is an LH9124 sub­system. This architecture was modelled and the results are shown in Figure 9.12. Horizontal partitioning Vertical-horizontal partitioning Figure 9.11: Simple architectures with corner turn memories Figure 9.12 shows that 11 LH9124s are required for real-time operation with resampling. Fewer processors are required than for the no comer turn approach. The penalty here is the cost of the comer turn memory. It has been assumed that the comer turn memory can supply data fast enough to keep up with the LH9124s. This may be hard to achieve in practice. 9. Examination of Architectures 109 2 4 6 8 10 number of processors Figure 9.12: Performance of horizontal partitioning with a corner turn on multiple LH9124 processors with the range-Doppler algorithm Another partitioning method that can be used is vertical-horizontal partitioning. This scheme also requires a centralized comer turn memory and allows range compression and azimuth compression to be pipelined. This implies that the comer turn memory be double buffered or dual-ported as is shown in the bottom half of Figure 9.11. This architecture was modelled and the results are shown in Figure 9.13. The percentage of real-time operation achieved is plotted against the number of processors in each pipeline stage. Separate graphs are shown for range and azimuth compression. Azimuth compression is taken to mean all the oper­ations after the comer turn in range-Doppler. 300 200 processing speed relative to real time rate (%) 1 1 1 / 1 range compression---^^^ . / / azimuth compression-^ - + - r i - + 1 1 2 4 6 8 10 number of processors Figure 9.13: Performance of vertical-horizontal partitioning with a corner turn on multiple LH9124 processors with the range-Doppler algorithm 9. Examination of Architectures 110 Figure 9.13 shows that for real-time operation, three LH9124s are required for range compression and nine for azimuth compression for a total of 12. The central corner turn memory is again a potential bottleneck in the architecture. At real-time operation, the transfer rate into the corner turn memory is about 48 MB/s or about 8 MW/s. This data rate probably cannot be sustained with a commercial memory but can be rela­tively easily accomplished in a custom design. It should be emphasized that the models used in this section are quite rudimentary. The purpose of the per­formance estimates is only to give an idea of the feasibility of building a real-time S A R processor using an LH9124 based approach. The conclusions reached are quite positive. With about 11-15 LH9124s and a similar number of general purpose CPUs, a real-time SAR processor for the range-Doppler algorithm for Radarsat can be constructed. This could conceivably fit in one 19" card cage. This is no small feat consid­ering the sustained computation is on the order of 4.5 GOP/s. 9.3.6 Other Algorithms The LH9124 DSP was also applied to models of some other SAR processing scenarios. The first is the algorithm used in the M D A IRIS X2C airborne SAR processor. A mapping of algorithm steps to either the LH9124 or a general purpose C P U was performed in a similar manner to the one shown in Table 9.2. A single LH9124 configuration is able to process a 4k x 2k frame in approximately 6.4 seconds. Since the frame time is 4.5 seconds, at least two LH9124s are needed to meet real-time operation. The operation rate required of the GP C P U is 7.3 MOP/s which is within the capabilities of a single chip C P U . Thus, a possi­ble system would consist of 2 LH9124s and at least 1 GP C P U (most likely about three or four) on a total of about 4 or 5 boards. This should be compared with the actual system that was built for this application using VASP boards based on the DSP96000 general purpose DSP. The system consisted of 18 General Sig­nal Processor boards with 4 processors each, four I/O Processor boards with one processor each, and four Two-Port Memory boards. This gives a total of 76 processors and 26 boards. The lack of interpolations makes the chirp scaling algorithm particularly well suited for implementation on the LH9124. Chirp scaling requires more vector multiplies, but these can usually be performed with no overhead when they are rolled into the first stage of an FFT. A chirp scaling implementation on a single LH9124 that uses the same parameters as the range:Doppler model of Section 2.5.4 can achieve almost 25% of real-time operation. This is without any post-processing resampling operations. Chirp scaling requires similar data transfer rates as range-Doppler. It also suffers from the fact that extremely high mem­ory to buffer data transfer rates are necessary for algorithm steps that involve small amounts of processing per vector. In chirp scaling, like in range-Doppler, this occurs for the azimuth FFT. 9. Examination of Architectures 111 9.3.7 Vector DSP Conclusions The use of the LH9124 DSP has been shown to be a promising implementation alternative for SAR processing. This chip provides a fairly straight-forward solution for high performance vector processing. Despite the high performance attained by the chip, it is not fast enough for real-time processing and paral­lel architectures are needed. The design of the memory system and control structure are left open by the architecture of the LH9124 and will be major challenges in the design of an LH9124 based processor. A practical design that incorporates the LH9124 is examined in Section 9.7. Some of the shortcomings of the LH9124 are discussed below. • Numerous additional chips are needed to support the LH9124. External R A M s and address generators are required at each port. If these functions were incorporated on the main chip the system design would be considerable simpler. Butterfly DSP is attempting to address this by building M C M s that combine the LH9124 with some support chips. • A l l four data ports are rarely necessary in a system due to the cost of all the support circuitry for each port. For most applications, two ports would be enough: one for data and one for coefficients. • The FIR filtering speed of the LH9124 is disappointing. For a fixed algorithm like FIRs, much higher performance should be possible. The LH9124 is no faster per filter tap than general purpose DSPs which can also usually handle one filter tap per cycle. The only performance advantage of the LH9124 is that it can handle complex M A C s in a single cycle. A possible work-around for this prob­lem is to use the LH9124 in conjunction with a fast digital filter chip. • The need for separate scalar CPUs complicates the design. • More flexible address generation circuitry is needed. The LH9230 cannot handle the more complex addressing patterns that are required for sample rate conversion. Also the LH9230 cannot handle out-of-the ordinary situations. For example, it would be convenient to be able to perform filtering across vectors stored in the buffers since this would speed up R C M C in range-Doppler. Currently the only way around this problem is to design a custom address generator. A more flexible solution would be to make the address generator programmable like a microsequencer. This way the programmer could choose any addressing sequence. The LH9124 is not the only chip in its class. In fact it seems that a number of new chips optimized for fast FFTs will become available shortly. Texas Memory Systems has announced the TM66 swiFFT. It has a slightly faster lk CFFT time than the LH9124 but uses full 32-bit IEEE floating point arithmetic. The TM66 architecture is quite different than the LH9124. It also makes extensive use of pipelining, but the 9. Examination of Architectures 112 data flows in only one direction, from two input ports to one output port. Butterfly DSP, who have taken over the support of the LH9124 from Sharp, claim to be working on the BDSP32000, which will feature a 32-bit fixed or floating point l k CFFT time of 10 ps. 9.4 Optimized DSP Architecture Based on the experience gained from studying SAR algorithms and existing DSPs, it is possible to charac­terise a DSP IC that has an architecture optimized to handle the requirements of SAR processing. This sec­tion examines some aspects of optimized DSP architectures. A natural question that arises is whether or not it is possible to build a single chip processor that can handle all the arithmetic needed for SAR. The requirements on such a processor can be taken from the range-Dop­pler model of Section 2.5.4. The amount of computation in the main processing flow (steps 2.0 to 8.0 in Figure 2.17) is on the order of 3 GOP/s. The data transfer rates are upwards of 50 MW/s. We are quite close to realizing this level of performance. One of the fastest FFT processors today is the LH9124 which achieves an execution rate of 15 OPs/cycle when executing FFTs. One of the fastest general purpose proc­essors available today is the DEC Alpha 21064 which runs at clock speeds in excess of 200 MHz. If these features were to be combined, one would obtain a computation rate of 3 GOP/s which is what is required for a single chip real-time SAR processor. Of course, the LH9124 only achieves this high operation rate for the fixed FFT algorithm. But the intent here is to show that the technology exists today for a single chip real-time SAR processor. A possible supporting example is the new chip that Butterfly DSP is claiming to be developing that can perform a l k CFFT in 10 ps. This translates into an operation rate of 5.1 GOP/s. The attributes of an ideal SAR processor chip can be based on solutions to the shortcomings of the LH9124 discussed in the previous section. The features listed below would be desirable in a SAR processing chip. • A high clock speed • A vector processor that supports extremely fast FFTs and other vector operations: Like the LH9124, this would be achieved using multiple arithmetic units and extensive pipelining. The vector processor should support complex IEEE floating point numbers. • S R A M double buffers large enough to hold the longest data vector expected • Programmable address generators for the inputs and output of the vector processor • Efficient support for filtering: It should be possible to process at least eight or more complex filter taps per clock cycle 9. Examination of Architectures 113 • A scalar processor: This could be similar to one of the existing general purpose floating point DSP chips. This processor would also be responsible for the control of the vector processor, eliminating the need for a separate scheduler. • Sufficient S R A M to minimize off-chip access for both program and data by the scalar processor It is not possible to achieve all this with today's technology. However, it is not unrealistic for the near future, especially considering TI's achievement in putting five DSP CPUs on one chip in the C80 MVP. Memory Processor double buffers l&Q bus Figure 9.14: Hypothetical single chip SAR processor As the speed of a single chip processor is increased, the requirements on the data path elements that route data to and from the processor also increase. A simple estimate of the overall transfer rate for a real-time SAR processor can be obtained by making the following assumptions: (see Figure 9.14). • One corner turn operation is required. • The complete data array is transferred between the external memory and the processor four times (once at the beginning, twice for the corner turn, and once at the end). • Parallel transfers of complex words are possible. (This requires a 64-bit wide data path for 32-bit data.) • Data transfers are fully overlapped with computation; i.e. the full processing time is available for data transfers. • The data volumes from Section 2.5.4 are used. These assumptions result in a data transfer rate of 31 MCW/s. This fairly low number results because we are assuming very few data transfers are necessary and that a large amount of computation happens between data transfers. This is attainable with existing D R A M s and an appropriate interleaving scheme. Much larger data transfer rates are needed between the vector buffers and the processing unit. This indi­cates that these buffers should reside on the same IC as the processing unit. 9. Examination of Architectures 114 Combining so much functionality makes for a very large IC. It may be more cost effective to use an alter­nate approach and use fine grained parallelism to spread out the functionality over multiple chips. This requires extremely fast inter-chip communication. No currently available chip supports this. This option is discussed further in Section 9.6. 9.5 Digital Filter Chip Architecture The appearance of extremely high performance digital filter chips has made the time domain approach to SAR processing more appealing. The highest performance digital filter chip encountered to date is the Zoran ZR33288 Video Rate Digital Filter. Some of its key characteristics are listed below. • 8 bit data input • 288 coefficients (10 bit) • 14.3 M H z sample rate for all filter lengths • real data only (Four chips plus adders are needed to handle complex data.) • Longer filter lengths can be handled by cascading a number of chips. The Zoran chip is capable of about 8 GOP/s which is the highest performance of any currently available chip surveyed. This makes it about 13 times as powerful as the fastest FFT chip which has a peak perform­ance of about 600 MOP/s (but with a larger word length). A potential application of the Zoran chip to a quick-look SAR processing situation is described in [68]. A board is described that makes use of two sets of four Zoran chips in a pipeline configuration. The design is able to achieve 2/3 real-time operation for high precision Radarsat processing (without quadratic R C M C ) on 2048-sample range lines. This section examines the ZR33288 based time domain approach to SAR processing. It is based on [68] but adds a performance analysis, some design additions, and an overall analysis. 9.5.1 Architecture A block diagram of a filter-chip based SAR processor is shown in Figure 9.15. The input FIFO acts as a buffer for the incoming data and serves to equalize the input data rate with the data rate supported by the digital filters. The high level design of a complex filter block is shown in Figure 9.16. The same block is used for both range and azimuth processing. It is made up of four ZR33288 digital filter ICs. Longer filters than the 288 taps supported by the filter chips can be accommodated by feeding the outputs back to the inputs through FIFOs that provide the appropriate amount of delay. Filter coefficients are stored in a coef-9. Examination of Architectures 115 Input Input FIFO Complex Filter (Range Compressor)! Corner Turn Memory Look Memory Complex Filter (Azimuth Compressor) Detector l2+Q° Look Processor Output Controller Figure 9.15: Time domain SAR processor block diagram ficient memory and are written to the filter chips by the control logic. The coefficient memory can be quite large since different filter coefficients are potentially needed for each vector. The comer turn memory can be either single or double buffered using the methods described in Section 7.3.2. Low resolution processing may call for each range line to be subsampled before it is stored in mem­ory. The size of the memory can be chosen to match the size of the image being processed. The data rate in and out of the memory is 14.3 MHz , which is equivalent to a cycle time of about 70 ns. Since the data needs to accessed at this speed in both row and column directions, existing D R A M s are too slow and SRAMs must be used. The look processor stores the values from the initial look in the look memory. Each subsequent value is added to the one in the look memory and the new value is stored in the look memory. Once all the looks have been summed, the data is read out of the look memory in the range direction, and a deskew filter is applied to remove the effects of the shifts applied during range processing. A useful refinement to the basic time domain processor design is the addition of buffers and more sophisti­cated address generation logic at the inputs and outputs of the complex filters. This would facilitate the implementation of sample-rate conversion. This is required if the image needs to be resampled. In the time domain approach, this can be combined with range and azimuth processing. A possible structure is shown in Figure 9.17. Such a filter based resampling structure might also be a useful addition to a fast convolution SAR proces­sor. It could be used to off-load the resampling operations from the main processors. This may be worth-9. Examination of Architectures 116 Coefficient Memory ZR33288 DIN DOUT CASIN CASOUT CADD CIN FIFO In Control Out ZR33288 DIN DOUT CASIN CASOUT CADD CIN FIFO In Out Control 18 ZR33288 DIN DOUT CASIN CASOUT CADD CIN FIFO In Out Control ZR33288 DIN DOUT CASIN CASOUT CADD CIN FIFO In Control Out Control control subtract Q Figure 9.16: Time domain complex filter block while since resampling takes up a large portion of the total number of operations and since fast FFT chips are usually not as efficient for filtering. 9.5.2 Range and Azimuth Processing Range compression is performed using standard time domain convolution. Matched filter lengths of up to 288 taps can be accommodated in a single pass through the ZR33288. Longer filter lengths require multi-9. Examination of Architectures 117 Input Buffer Complex Filter Buffer Address Generator Output • Adder and Address Generator Figure 9.17: Modification to complex filter for sample-rate conversion pie passes. Linear R C M C is accommodated by using different filter coefficients for each range line. The coefficients implement a shift that removes range walk. This shift must be removed after azimuth process­ing-Azimuth compression is handled similarly to range compression. Quadratic R C M C is handled by process­ing multiple input lines for each output line using appropriate coefficients that extract the energy that belongs in the output line from each of the input lines. Multi-look processing is handled by processing each azimuth line once per look. 9.5.3 Performance The performance of the time domain processor described above was modelled and the results are shown in Table 9.3. Table 9.3: Time domain processor performance Scenario %ofRTforRC % of RT for AC high resolution processing 1 look 55% 59% ' 4 looks 55% 15% 4 looks with quadratic RCMC 55% 2.5% low resolution processing 1 look 180% 1000% 4 looks 180% 250% The SAR scenario from Table 2.12 is used for high resolution processing. Quadratic R C M C is assumed to require the processing of six adjoining azimuth lines. For the low resolution processing scenario the parameters from Table 2.12 are used again with the following exceptions: L r=100, La=60, Naz=\024, and the output range line length is 1024. 9. Examination of Architectures 118 The throughput of the time domain processor falls off rapidly for higher resolution processing. However, the processor is well suited to a low resolution four look processing situation. This situation is also suited for an implementation using the S P E C A N algorithm. In order to process the same data using S P E C A N , about 30 MOP/s of computation are required. This could be achieved using two or three general purpose DSPs. Some advantages of both approaches are given below. • general purpose DSPs • much greater flexibility • much lower hardware development cost • greater precision due to longer word length • digital filter processor • more predictable performance and scalability • able to use longer filter lengths and still achieve real-time speed • possibility of lower repeat costs 9.5.4 Analysis Some of the main advantages of the filter-chip based time domain approach to SAR processing are listed below. • Digital filter ICs offer the highest operation rate per chip available today. • Scalable architectures can be built with predictable performance. • A range of processing situations can potentially be accommodated: from low to high resolution. • Output image resampling can be handled more efficiently than in a frequency domain processor. Some potential disadvantages surrounding filter-chip based SAR processors are summarized below. • A great deal of custom hardware design is required. • The amount of custom hardware makes it likely that a time domain design intended for high perform­ance will exhibit low flexibility. • Many chips only handle short word lengths. This results in either lower precision processing or a higher chip count. The design is also complicated by the need to carefully track how the data is scaled. • Most filter chips do not directly handle complex values. This results in a higher chip count. • A design is likely to be very specific to a particular digital filter device. 9. Examination of Architectures 119 • The control requirements for the filtering structures can be relatively complex. • Control and coefficient generation needs to be handled by a separate C P U . • In a pipelined design, it can be difficult to balance the amount of processing between the range and azimuth stages. • The efficiency of the time domain approach can decrease rapidly for higher precision processing due to longer filter lengths, and due to the extra processing that is required once effects like R C M C need to be incorporated. • Multi-look processing requires additional processing, and as a result is harder to accommodate than in a frequency domain approach. The viability of the time domain approach is dependent on the matched filter length. If the matched filter is short and the vector length is relatively long, then a time domain approach can be efficient relative to a fre­quency domain approach. This applies primarily to low resolution processing where filter lengths are shorter. For example, with a filter length of 100 and vector length of 1024 the time domain approach only requires about eight times as much computation as fast convolution. However, for a filter length of 600 and a vector length of 8k, the time domain approach requires about 35 times as much computation. 9.5.5 Digital Filter Conclusions The digital-filter chip based time domain SAR processor has been shown to be most suitable to high speed low resolution (greater than about 50 m) processing. Specifically, the time domain processor is suited for applications that require: • relatively short matched filter lengths (less than 288 for the ZR33288) • no long wordlengths • a small number of looks • no quadratic R C M C For a processing problem that has these characteristics, the time domain processor can require significantly less hardware than an equivalent fast convolution based processor. For example, for medium resolution single look processing with no quadratic R C M C , two time domain processors in parallel would be able to achieve real-time operation for Radarsat. This might otherwise require upwards of 40 general purpose CPUs or six vector DSPs like the LH9124. 9. Examination of Architectures 120 Performance falls off rapidly with longer filter lengths, larger numbers of looks and quadratic R C M C cor­rection. As a result, in most high resolution cases there is no real performance advantage to using the time domain approach. Any speed advantage is outweighed by the lack of flexibility. The advantage presented by the high speed of the digital filter chips is also partially offset by the availability of efficient frequency domain processing algorithms, especially for low resolution processing. The dedicated filter chip approach to SAR processing would be significantly more attractive if some of the flexibility and fixed wordlengths problems could be overcome. This may be addressed through the use of FPGAs to build custom hardware filter devices. This option is addressed in Section 9.8. 9.6 Fine Grained Parallel Architecture Fine grained architectures have a number of advantages over the coarse grain architectures considered so far (as previously mentioned in Section 2.6.2). Fine grained DSP architectures • have more opportunities for parallelism, • are easier to program/schedule, and • are suitable for automatic programming tools. In order to build a fine grained SAR processor, the appropriate building blocks must be made available. Current microprocessors and DSPs do not have the level of support for high speed, low latency communi­cation that is necessary for fine grained processing. Chips that address this issue have been proposed but have not been produced commercially. An example is the SIMC (Special Instruction Set Multiple Chip Computer) proposed in [62], which is based on the TMS320C40, but which incorporates features that make it more suitable for fine grained parallelism. Fine grain approaches are useful with existing processors in the following situations: • It is not possible to use coarse grain partitioning to achieve the required performance level. • The partitioning allows all the data required for an operation to fit into on-chip memory. This can result in super linear speed ups. Examples of this for the case of splitting the ID FFT over multiple C40s is given in [85]. Due to the lack of building block components, no fine grained system level products are available. As a result of this lack of commercial products, not much experience exists in implementing large scale signal processing applications on fine grained machines. 9. Examination of Architectures 121 Fine grained signal processors make more intensive use of software tools. A typical programming design flow might entail the following four steps, the last three of which can be automated. • model the algorithm in terms of a signal flow graph • obtain a schedule • find a processor mapping • translate into machine code One area that will see the emergence of massively parallel fine grained architectures is the area of F P G A computing covered in Section 9.8. F P G A architectures may be one of the main beneficiaries of the research in fine grained parallelism. The potential advantages of fine grain architectures are mitigated somewhat by the fact that the large data set sizes used in SAR processing make coarse grain partitioning relatively easy. Fine grain approaches may be more attractive for extremely high performance applications where coarse grain architectures become less efficient. In summary, fine grained architectures are not found to be a viable alternative for most SAR processing applications for the following reasons. • Availability: Both the hardware and software components necessary for a fine grained design are not available commercially. • Cost: The implementation of a SAR processor on a fine grained processor would entail the design of a completely custom processor. • Risk: Fine grain processors are not a very mature technology. 9.7 A Heterogeneous Architecture This section describes the architecture and analysis of a proposed SAR processor design that combines some of the design approaches described in the previous sections. The design makes use of both vector and scalar DSP chips and is referred to as a vector/scalar architecture (VSA). The goal of the design is to achieve 1/10 real-time processing using the scenario of Section 2.5.4 on images up to 8k x 4k in size. 9. Examination of Architectures 122 9. Examination of Architectures 123 A block diagram of the architecture is shown in Figure 9.18. The design can be divided into three parts: • the vector processor, • the scalar processor, and • the image double buffer. Some of the key architectural features of the V S A are given below: • Vector processing is separated from scalar processing. • Vector and scalar processor data flows are decoupled via twin D R A M banks. • High bandwidth data flows are assigned their own buses. • The computational elements and the data paths are flexible enough to handle a range of processing sit­uations. • The performance goals are achieved while using commodity D R A M s for all large memories. 9.7.1 Vector Processor An LH9124 based design is used as the vector processor for the system. The double buffers attached to the data ports of the LH9124 are at least 48 kB for 24-bit data or 64 kB for 32-bit data for each half of the dou­ble buffer. An additional buffer is attached to the data port and is accessible from the I/O bus. This buffer allows the scalar DSPs to directly supply data to the vector processor. This vector access buffer is made somewhat larger to give the general purpose DSPs more time to access the data. A possible size is 128k x 48. At least 15 ns SRAMs are needed for a 40 M H z LH9124. The buffer attached to the coefficient port of the LH9124 is organized as multiple pages. The purpose of this is to reduce the otherwise very large bandwidth requirements on the coefficient bus. When a new set of coefficients is needed, a switch is simply made between pages. If 16 8k coefficient vectors are stored, the size of the buffer is 128k x 48 or 768 kB. The coefficient D R A M provides storage for coefficient vectors. If all the azimuth matched filters are pre-stored in the memory and a new filter is used for every four azimuth lines, the memory will need to be at least 48 M B in size. Fast page mode access can always be used for this memory since all accesses are made to consecutive vector addresses and no corner-turns are required. This memory is accessible from both the coefficient bus and the I/O bus. The intent is to allow either the coefficient controller DSP or the host com­puter to update the coefficients. 9. Examination of Architectures 124 9.7.2 Scalar Processor The scalar processing devices are general purpose DSPs like the C40 or the S H A R C . The design relies on their level of performance and inter-processor communication capabilities. In the remainder of this description, the scalar processor is taken to be a mesh of SHARC chips. The large on-chip memory of the S H A R C is particularly attractive in this design. Since the entire data array never has to be distributed among the DSPs, the on-chip memory is large enough to serve as a buffer and no external memory is nec­essary. If the C40 is used instead, each PE would require memory external to the processor. However, the C40's dual bus structure would greatly simplify the interfacing between buses in the system. Functionally, the scalar processor can be divided into two portions: fixed functions and non-fixed func­tions. The two fixed functions of scheduler and coefficient controller are fixed by the hardware design to two distinct DSPs. The non-fixed functions, like pre/post-processing and parameter generation, can be per­formed by any combination of the remaining DSPs. The scalar processor is very scalable since the number of general purpose DSPs can be changed relatively easily. 9.7.3 Image Double Buffer The image double buffer is composed of two identical banks of D R A M organized as 32 bit words. Each bank can be up to 256 M B in size. The most efficient way to implement this memory is to use 32 M B SIMMs. Eight SIMMs are required for each bank. Accesses to the image double buffer from the I/O bus are always to consecutive addresses allowing the fast page mode cycle time to be used. Accesses from the vector bus can be to either rows or columns of the memory. As a result, the slower random access cycle time is assumed for all accesses from this bus. The memory controller is responsible for managing the image double buffer. It implements a crossbar switch that allows independent access of either memory bank from either bus. It is also handles • D R A M access timing, parity (or ECC), and refresh, • the vector bus interface including address generation, and • the I/O bus interface. It is not necessary to use interleaving to meet the performance goals. However, the design of the memory controller should be able to accommodate the future addition of interleaving which may be necessitated by an increase in clock rate in either the vector or scalar processor. The memory controller would most likely be implemented in a PLD. 9. Examination of Architectures 125 9.7.4 Data Movement The V S A is designed to allow high speed data transfer with no bottlenecks. To achieve this, there are four main buses in the V S A : the I/O bus, the vector bus, the coefficient bus, and the control bus. The I/O bus is a general purpose bus that is used for many different data transfers in the system. None of these are high bandwidth transfers. In order to keep the latency for all the different users of this bus low, the utilization of the bus is kept low. The I/O bus has a high bandwidth; it runs at a speed of about 30 MHz. The clock speed is limited by the fast page mode cycle time of the DRAMs in the image double buffer. The I/O bus is also connected to the host bus interface, which can be a standard PCI or V M E interface chip. The vector bus connects the LH9124 to the image double buffer. It carries the high bandwidth traffic that fills and empties the LH9124 buffers. The vector bus utilization is high, but since it is point-to-point and its workload can be completely predicted, this is not a problem. The available bandwidth of the vector bus is limited by the need to accommodate random access to the image double buffer. As a result, its average transfer rate is about 10 MHz. The coefficient bus services the LH9124's coefficient buffer. The traffic is similar to that on the vector bus. The control bus is a special purpose bus that carries the control signals for the various chips in the design. It originates in control logic that is attached to the scheduler DSP. 9.7.5 Implementation The V S A should fit on 1.5 to 2 full-length PCI bus or 6U V M E bus boards depending on how aggressively it is packaged. The use of daughtercards may make it possible to put everything in one board slot. It is also possible to assemble an architecture similar to the V S A using mainly commercial boards. The fol­lowing products are needed: • an LH9124 board (like the P C / M VSP-9 or the Catalina CRV1M40) • a memory board to support the LH9124 board • a general purpose DSP board with multiple SHARCs or C40s (available from a number of vendors) The functionality of the image double buffer would be more difficult to find in a standard product. 9. Examination of Architectures 126 A V S A design based on COTS products would have to be carefully evaluated since it is hard to achieve optimal performance from a mix of standard products. It would be especially difficult to match the high data transfer rates achieved by the VSA's multiple bus design using commercial products. 9.7.6 Overall System The V S A can form a part of an overall SAR processing system like the one depicted in Figure 8.22. I/O Video Controller Host bus (eg. PCI or VME) I Host CPU Figure 9.19: VSA based SAR processing system A high bandwidth host bus is required as the backbone for the system. Either PCI or VME64 would be good choices for this bus since they offer sufficient bandwidth and the opportunity to use standard PC or V M E platforms. It is assumed that the SAR data is transferred over the host bus before and after it is proc­essed. While this places significant loads on the host bus, it allows standard mass storage interfaces to be used. No other time critical SAR processing related traffic is carried by the host bus. The following subunits are connected to the host bus: • VSAboard(s) • frame synchronizer (if necessary) • high speed mass storage interface (to disk and tape) • video controller to interface to a high resolution monitor (not shown): if necessary • host C P U interface (additional peripheral devices can be attached to the host C P U via a lower speed peripheral bus like ISA) QDl A tn mnnitnr VSA VSA (board 1) (board 2) Storage Controller! 9. Examination of Architectures 127 9.7.7 Range-Doppler Algorithm This section describes the mapping of the range-Doppler algorithm of Section 2.5.4 to the V S A . The amounts of computation and data set sizes are taken from the analysis of the algorithm in that section. Scalar Processor The raw SAR data is transferred over the host bus to general purpose DSPs which handle: • input data unpacking • I/Q balance (step 1.0 of Figure 2.17) • number format conversion • data transfer into the image memory buffer When both the scalar processor and the vector processor have completed processing the data in one of the image memory banks, the banks are switched. The scalar processor general purpose DSPs then perform: • format conversion: block floating point to floating point • azimuth resampling (step 9.1 of Figure 2.17) • range resampling (step 10.1 of Figure 2.17) • detection (step 11.0 of Figure 2.17) • radiometric correction • output format conversion • data transfer out of the V S A over the host bus Once all this is complete, the general purpose DSPs ingest a new frame of data and the process begins again. These operations are performed by the four general purpose DSPs that make up the pre/post-processor por­tion of the scalar processor. When predicting the performance of the general purpose DSPs, an operation rate of 30 MOP/s is assumed. This is 25% of the claimed peak rate of 120 MOP/s and about 38% of the claimed sustained performance figure of 80%. This a fairly safe assumption since these PEs are performing fairly fixed algorithms consisting of filtering and arithmetic functions with little variation. When calculat­ing data transfer times, a transfer rate of 30 MB/s is assumed for the I/O bus. It is assumed that there is no 9. Examination of Architectures 128 overlap between computation and data transfer. With these assumptions, the total time to perform all the functions listed is about 21 seconds, which is about 12% of real-time. The coefficient and filter generation functions are performed by the two parameter generator DSPs in the scalar processor. These DSPs perform steps 9.2, 10.2,12.0, 13.0, 14.0, 15.0,16.1, 16.3, and 16.4 of Figure 2.17. The azimuth matched filter is assumed to be updated every four lines. The computation rate required is 19 MOP/s for 1/10 real-time. This is easily achievable on two DSPs. Vector Processor The timings for the functions performed by the vector processor are shown in Table 9.4. Table 9.4: Times for vector processor operations Step in Algorithm | Time required (sec) 2.0 Range FFT, 3.0 Range multiply, 4.0 Range IFFT Processing 5.8 Data transfer 5.7 5.0 Azimuth FFT Processing 1.7 Data transfer 4.4 6.0 RCMC Processing 4.4 Data transfer 4.4 7.0 Azimuth multiply, 8.0 Azimuth IFFT Processing 2.2 Data transfer 4.4 16.2 FFT of azimuth matched filter Processing 0.4 Data transfer 0.1 Total: 19.5 Time allowed for 1/10 RT: 26 sec. The data is transferred between the image double buffer and the LH9124 buffers for each step in the table. With the exception of step 16.2, the data transfer times are based on a data transfer rate of 10 M H z from the image double buffer. The data transfer for step 16.2 is assumed to occur between the general purpose DSPs and the vector access buffer and so it occurs at a rate of 40 M H z . For each step, the larger of the computation time and the data transfer time is taken as the total time. 9. Examination of Architectures 129 The data transfer and computation times are well balanced for all steps except for the azimuth FFT and IFFT. This imbalance results because the data is transferred back to memory after the azimuth FFT so that it can be read out in the range direction for R C M C . The delay introduced by these data transfers means that the LH9124 is not fully utilized. Fast access modes of the D R A M s cannot be used to reduce this delay unless an explicit corner turn is performed. One way this delay could be reduced is by introducing inter­leaving in the image double buffer at the expense of a more complex memory controller. 9.7.8 Vector/Scalar Conclusions The analysis of the V S A has shown that it is capable of handling 1/10 real-time satellite SAR processing using the range-Doppler algorithm of Section 2.5.4. A similar analysis was carried out for the IRIS X2C airborne processing situation. A single V S A should able to handle the all IRIS X 2 C processing require­ments. It is interesting to compare the V S A with a general purpose DSP only design. The LH9124 could be replaced by about 8 to 10 general purpose DSPs. This would result in a more general, flexible, and stand­ard product that might be more cost effective for some applications. However, it would result in signifi­cantly higher chip and board counts. 9.8 FPGA Computing Machine F P G A based computers allow their hardware resources to be reconfigured to fit the job at hand. They offer the possibility of combining the speed of custom hardware with the reprogrammability of software. For example, an F P G A based SAR processor could be configured as an FFT processor for the fast convolution operations, and then reconfigured as a resampling processor for the geometric correction operations. The circuit used in each case would be optimized for the required algorithm. This section examines the feasibility of using FPGAs in SAR processing applications, and makes some predictions of the performance that can be expected from existing F P G A devices. 9.8.1 Standard Arithmetic This section analyzes the performance achieved by FPGAs on DSP operations based on results reported in the literature. In the analysis given in [9], the basic building block is a multiplier designed by Casselman. The 24-bit fixed point version of the multiplier requires 48 configurable logic blocks (CLB) in a Xilinx XC4000 series 9. Examination of Architectures 130 device and produces a result in 16 clock cycles at 16 M H z . This gives 1 MOP/sec per multiplier. Given that the Xilinx XC4010 device has 400 CLBs, and assuming 25% overhead, results in about 6 MOP/sec per chip. The single precision floating point version of the multiplier requires 60 CLBs and produces a result in 16 cycles at 16 MHz. This gives about 5 MOP/sec per XC4010 chip. Somewhat higher perform­ance could be obtained by using the largest available F P G A which is currently the XC4025 with 1024 CLBs. In the results given in [6] for the SPLASH-2 hardware platform which uses 16 XC4010 FPGAs as the processing units, a 2D FFT implementation can process two 512 x 512 images per second. The wordlength is 8 bits. This gives an operation rate of 47 MOP/s. Only five FPGAs are used for the actual butterfly com­putations. Both [9] and [6] give the idea that the order of magnitude of performance to be expected is about 5-10 MOP/s per 400-CLB chip for MAC-type operations with 8 to 24-bit wordlengths. This is a relatively low level of performance. A large number of chips is required to achieve performance levels higher than that available from general purpose DSPs. Current FPGAs do not have sufficient resources for the efficient implementation of standard arithmetic units like multipliers. It can be concluded that current FPGAs are not suitable for SAR processing if regular M A C structures are implemented. 9.8.2 Distributed Arithmetic An interesting result with much higher performance is contained in a 16-tap 8-bit FTR filter macro for the XC4000 series devices produced by Xilinx [41]. The filter mns at 5.44 Msamples/sec and uses 67 CLBs. This gives an operation rate of 169 MOP/sec per filter. If multiple filters are programmed into an XC4010 with 400 CLBs, a single FPGA is able to realize about 800 MOP/s if 25% overhead is assumed. This impressive level of performance is much higher than a general purpose DSP, but is still lower than a spe­cial purpose DSP device like the ZR33288. The Xilinx filter macro shows that it is possible to achieve very high performance with FPGAs for a spe­cific application. The key to success in the case of the filter is the use of distributed arithmetic [92] [69] to avoid multiplications, which are expensive to implement in FPGAs. The Xilinx macro has some serious limitations however: • The word length is only eight bits. • The coefficients are not variable. The F P G A must be reprogrammed to change them. 9. Examination of Architectures 131 • The filter relies on the symmetry of linear phase FTRs: actually only eight unique coefficients are han­dled. As a result, the design as it stands is not very practical. Table 9.5 shows how the filter scales as the design parameters are changed and how different approaches affect the scalability. Table 9.5: Scaling the distributed arithmetic filter Design Parameter Change Approach Result increase filter length in one filter section area increases exponentially partition into multiple parallel filter sections area increases linearly partition into multiple cascade filter sections area increases linearly (but slower than for parallel sections) speed decreases linearly increase wordlength in one filter section area increases linearly speed decreases linearly in duplicate sections area increases linearly The original Xilinx filter macro is able to achieve 800 MOP/s per chip with 8-bit real numbers. If the filter is adapted to handle 24-bit complex numbers, the performance per 400 C L B F P G A would be expected to decrease to a maximum of about 100 MOP/s. This is somewhat faster than most general purpose DSPs but is slower than the LH9124. When compared with commercial fixed-function filter chips like the ZR33288, the F P G A approach to time domain SAR processing exhibits • lower performance, • even more hardware design, • a much more flexible architecture, and • a range of wordlengths and number formats. So far, the good performance obtained from distributed arithmetic has only been applied to filtering. While time domain processing can be used for SAR, it has a number of disadvantages as seen in Section 9.5. The key to making more general application of FPGAs to SAR processing is an efficient implementation of FFTs. FFTs can be implemented using distributed arithmetic using the same principles as are used in filters [59] [75] [90] [61] [91] [83]. The multipliers are again replaced with look-up tables and adders. However, considerably more resources are required in this case since • support for complex arithmetic is needed, 9. Examination of Architectures 132 • some kind of floating point support is probably also needed, and • the number of twiddle factors required for longer FFT lengths make an external R O M necessary. It is predicted that a l k CFFT time of about 1 ms for 16-bit data is achievable in the largest F P G A currently available. This prediction is based on an analysis of reported FFT and filter implementations in FPGAs. This is equivalent to a processing rate of about 5 Mbutterflies/sec. This is in the same league as general purpose DSPs, but is slower than the fastest ones, and is only for a relatively short word length. Of course, higher performance can be achieved by using multiple FPGAs. 9.8.3 F P G A Conclusions Some of the advantages of FPGA based SAR processing approaches are listed below. • Reconfigurability: The hardware designs implemented in the FPGAs can be changed to adapt to dif­ferent processing requirements. • Flexibility: FPGA based designs can handle a wide range of algorithms, word lengths, number for­mats, etc. • Generality: It is possible to build relatively general computing machines that are useful for other algorithms or applications. • Parallelism: Much potential exists for fine grain parallelism. • Migration: A clear migration path to ASICs exists. Some disadvantages are: • Performance: Performance will be mediocre. • Cost: The development costs are likely to be high. • Risk: F P G A based DSP is an unproven technology. A possible solution to the problem of implementing arithmetic functions in FPGAs is to combine FPGAs with commercial arithmetic chips like multipliers and ALUs . In this case, the commercial chips would implement the arithmetic and the FPGAs the control logic. There are several potential problems with this approach: • Commercial arithmetic chips are not much faster than general purpose DSPs. 9. Examination of Architectures 133 • The greatest potential for speed-up in FPGAs is through the extensive use of parallelism. This leads to large chip counts if separate arithmetic chips are used. • A primary advantage of FPGAs is their flexibility. This begins to be lost if commercial chips are used. F P G A performance also depends on reprogramming time. Existing reconfigurable machines usually con­figure the FPGAs before processing starts. In DSP applications that make use of a number of different operations, it is desirable to reconfigure the FPGAs during processing. This requires the ability to reconfig­ure the FPGAs quickly. Until now FPGAs have had relatively long programming times, however, this is decreasing in more recent devices. In order to improve the cost effectiveness of FPGA based computers, the development costs need to be reduced significantly. This will be facilitated by the development of • standard F P G A based computing platforms, • better design tools, and • standard libraries. FPGAs will undoubtedly play a large part in any custom hardware design due to their usefulness in imple­menting control logic. However, their role as computational devices is limited. F P G A based signal processing is not yet a mature enough technology to be cost effective. Higher density FPGAs, hardware platforms that support extensive fine grain parallelism, and powerful software tools are needed for FPGAs to be an attractive alternative to commercial DSP chips. 9.9 Architecture Examination Conclusions This section summarizes the conclusions reached about each architecture and makes some comparisons between them. The degree to which each of the architectures meets the criteria set out in Section 3.5 is shown in Table 9.6. A l l of the architectures discussed in this section are listed in the table with the excep­tion of the fine grain architecture. A design for a fine grain processor would have to be specified before comparisons can be made. In the table, a scale of one to five is used, where a score of five is the most desirable and one the least. It should be emphasised that the scores are quite subjective and have no absolute meaning. The scores simply provide a means of comparing the architectures. The scores in Table 9.6 for each of the five criteria are dis­cussed in the paragraphs that follow. 9. Examination of Architectures 134 Table 9.6: Architecture trade-off Architecture Performance Scalability Flexibility Development Cost Repeat Cost NOW 2 3 5 5 1 GPDSP 4 4 4 4 4 Vector DSP 5 4 3 3 4 Optimized DSP 5 4 2 1 4 Filter 4 4 2 2 4 VSA 5 4 4 3 4 FPGA 3 4 3 2 3 Performance A l l of the architectures are capable of high performance. In the table, the scores for performance are based on the performance that can be expected per board. One of the fundamental trade-offs in these architectures is the one between the performance and the generality of the architecture. This is depicted in Figure 9.20. high Optimized DSP Vector DSP Filter Chip Performance for SAR Processing per Circuit Board GPDSP FPGA Workstation/NOW low low high Flexibility and Generality Figure 9.20: Performance/generality trade-off of processor architectures Scalability A l l architectures are easily scaled to higher performance levels. NOW is considered to be the least scal­able because of the limited interconnect bandwidth and the high cost of additional PEs. Flexibility As can be seen from Figure 9.20, there is a large range of flexibility in the processor architectures, rang­ing from the extremely flexible NOW to much less flexible architectures like the filter chip approach. 9. Examination of Architectures 135 Development cost and repeat cost There is also a great deal of variation in the costs associated with each processor architecture. N O W has the lowest development cost but the highest repeat cost. The more specialized DSP based systems have the highest development cost and the lowest repeat cost. The paragraphs that follow summarize the usefulness of each architecture for SAR processing focusing on their performance. Network of workstations NOW is a high performance solution best suited for situations where: • Very complex processing algorithms are needed. • A large amount of computation in addition to the basic SAR processing algorithm needs to be per­formed. This helps hide the problem of limited network bandwidth. • A large amount of flexibility is needed. An example is if the same hardware is to be used for other purposes besides SAR. General purpose DSP architecture General purpose DSP architectures are a good compromise between flexibility and performance. Vector DSP architecture Vector DSP architectures offer the highest performance that is available with commercial ICs. The pen­alty in comparison to general purpose DSPs is less flexibility and more design work. Optimized DSP A custom optimized DSP approach must be taken if higher performance or integration than that offered by vector DSPs is required. An optimized DSP potentially offers the ultimate in performance for the highest development cost. Digital filter chip architecture Digital filter chip architectures can offer a significant performance advantage for a specific class of low to medium resolution processing scenarios. However, their development cost is higher and their flexi­bility is lower than for general purpose DSPs. Fine grained architecture Fine grained architectures were not found to be necessary for SAR processing except for certain situa­tions where benefits can be realized by using existing coarse grain components in a fine grain manner. This will likely remain the case unless commercial products become available that give fine grain paral­lel architectures a cost advantage. 9. Examination of Architectures 136 Sample heterogeneous architecture A hybrid architecture that combines general purpose DSPs with a vector DSP was found to offer a per­formance advantage over general purpose DSPs at the expense of flexibility and development cost. FPGA computing machine FPGA based computers were found to offer no performance advantages for general SAR processing over other devices, and additionally carry the price of higher development cost. Currently, FPGAs are potentially useful for some specific portions of SAR processing where a commercial chip is not appro­priate for some reason. Future advances in F P G A devices and systems will expand their usefulness. 9. Examination of Architectures 137 10 Conclusions and Future Work 10.1 Conclusions This thesis has examined processor architectures for SAR by looking at the three components of system design: algorithm, technology, and architecture. Algorithm (Section 2) An overview of SAR processing theory was given with an emphasis on deriving general performance and memory requirements for SAR processors. This overview attempted to fill a gap in the SAR processing lit­erature by providing a concise summary of the equations that affect processor design. A selection of repre­sentative SAR algorithms was discussed and compared. The comparison is a more comprehensive extension of similar studies previously documented. A numerical analysis of their computational require­ments supported the previously reported computational characterizations of these algorithms. FFTs were found to account for between 75% and 92% of the arithmetic operations required by the non-time domain algorithms. Since range-Doppler is representative of most common SAR algorithms, a more detailed model of this algorithm was developed for use in analyzing potential architectures. The high computation requirements of most SAR processing situations necessitates the use of multiple processors. The options for partitioning SAR algorithms onto multiple processors were examined. The rel­ative advantages of coarse and fine grain partitioning approaches for SAR processing were explored. A taxonomy of partitioning approaches was presented. The alternatives include horizontal and vertical parti­tioning and their combinations with various data partitioning schemes. The partitioning of the range-Dop­pler algorithm was modelled. Pure horizontal partitioning was found to introduce large inefficiencies due to the overlaps required between data blocks. Technology (Section 5) A survey was made of the current state of technology in the key areas that underlie SAR processor design. The most important recent developments in processing devices include the increase in performance for DSP-type operations, the rise in on-chip integration, and the emergence of built-in support for multiproc­essing. In order to provide a quantitative measure for comparison, FFT performance figures for various types of processing devices were collected and compared. Modem RISC processors were found to be equally as fast as DSPs at DSP operations like FFTs. However, DSP chips have other advantages that make them easier to use in high performance DSP applications. The highest performance available from com­mercial chips for general SAR operations is offered by vector DSPs. These devices have approximately an 10. Conclusions and Future Work 138 order of magnitude faster FFT performance than general purpose DSPs. Despite the high performance fig­ures, no existing single processing device is fast enough to support 1/10 real-time satellite SAR processing. Any high performance SAR processor must make use of some form of parallel processing. Recent developments in R A M technology have much to offer to SAR processors. Increased memory capacity is an obvious advantage. The increase in D R A M speed for certain types of accesses brings the promise of higher processing speed, provided it is utilized properly. However, the lack of significant improvements in D R A M random access times leads to a need for careful memory systems design. Many options exist for the interconnection networks of SAR processors. The most important development in interconnections has been the emergence of standard interconnections aimed specifically at multiproc­essing DSP applications. Architecture (Section 3,4,6,7,8,9) As a starting point in the architecture design process, the importance of considering all aspects of the regu­lar systems engineering process was highlighted. The requirements of high performance, scalability, flexi­bility, and cost effectiveness were chosen for the architecture candidates. They were selected to represent a balance of real world requirements. The remainder of the work was concerned with finding and under­standing the architecture options for SAR processors that address these requirements. The classification of SAR processor architectures developed in this report is depicted in Figure 10.1. The high level taxonomy of architectures is represented by the first level of decomposition in Figure 10.1. It differs from standard taxonomies in that it has been adapted to encompass existing and anticipated SAR processor architectures. Possible architecture implementation alternatives were chosen. The relation of these alternatives to the high level taxonomy is shown in the second level of decomposition in Figure 10.1. These combine the best of existing technology with the most promising architectures. The suitability of each architecture imple­mentation for SAR processing was analyzed. Three were identified for further study: networks of worksta­tions, DSPs, and FPGAs. The other architectures were eliminated from further study. These architectures and the reasons they were ruled out are as follows: workstation (performance and scalability), accelerated general purpose computer (cost and scalability), supercomputer (cost), and custom algorithm specific proc­essor (flexibility). Before continuing with a more detailed look at SAR processor implementations, two essential ingredients of the architectures where examined by themselves: memory systems and interconnection networks. 10. Conclusions and Future Work 139 SAR Processor Single Processor Workstation Common Node Accelerated Computer Pipeline Processor Multiprocessor DSP Workstation SIMD Hardwired Reconrigurable FPGA Custom Algorithm Specific Multicomputer DSP Multicomputer Network of Workstations General Vector DSP Optimized Filter Chip Fine Grain Heterogeneous P u r P ° s e (Vector/Scalar) DSP Figure 10.1: Classification of SAR processor architectures Memory system design was identified as a major system design issue. The suitability of modem memory system design techniques to SAR was examined: large data path widths, caches, interleaving, and fast R A M devices. A l l were found to be useful if properly tuned to the memory access patterns of SAR. In par­ticular, caches can potentially be detrimental if they are not flexible enough to handle the vector nature of SAR. The most likely interconnection networks to be used in multicomputer SAR processors were identified and studied. The networks are the bus, mesh, and crossbar. Their comer turning performance was derived and compared. A l l the interconnection networks are suitable for SAR processors with relatively low numbers of PEs. For larger numbers of PEs, mesh and mesh-like interconnections offer the best performance. A detailed analysis was performed of SAR processors based on networks of workstations, general purpose DSPs, vector DSPs, a proposed optimized DSP, digital filter chips, and FPGAs. Networks of workstations suffer from the problem of requiring much higher interconnection bandwidths than those offered by con­ventional L A N products. General purpose and vector DSPs provide good solutions for SAR processing. Approximately 16 Analog Devices S H A R C general purpose DSPs would be necessary for 1/10 real time 10. Conclusions and Future Work 140 processing. Approximately two LH9124 vector DSPs and a similar number of general purpose processors would be required. An optimized DSP based on the existing vector DSPs was proposed. The availability of such a device would significantly reduce the number of ICs required for a SAR processor and would even­tually lead to the development of a single chip SAR processor. Digital filter chips and FPGAs where found to be less suitable for general purpose SAR processing. The analysis demonstrated the strengths and weaknesses of commercial products. Commercial chips that address subsets of the needs of SAR processing are available. However, in general, it is difficult to find both good system level solution support and high performance on one chip. This observation led to the consideration of a heterogeneous approach that combines the advantages of various types of DSP devices. A design that combines a vector DSP with scalar general purpose DSPs was developed and its perform­ance analyzed. It was found to offer a good balance between performance and flexibility. A design with one vector DSP (Sharp/Butterfly LH9124) and eight general purpose DSPs (Analog Devices SHARC) was found to be capable of handling 1/10 real-time processing. The provision of efficient memories and data transfer paths was of equal importance in the design as the selection of the processing units. Each of the processor architectures considered meet the established requirements to varying degrees. A l l offer the potential for good performance and scalability. The trade-off between performance and flexibility was very evident in these architectures. In general, DSP based multicomputers were found to offer the best overall architecture for the SAR processing requirements considered. Future technological developments are expected to improve this further. Within a decade, single chip DSPs should be available that are able to handle all the computation required for the 1/10 real-time SAR processing scenario. 10.2 Future Work This section identifies some possible areas for future work that extend the work presented in this thesis. Network of workstations Provided appropriate networking hardware and software is available, it would be interesting to develop an experimental NOW SAR processor system. Several powerful workstations with large memories and disks are required. These would probably already be on-hand. A high speed network with speeds at least as high as FDDI or A T M is necessary (i.e. with a bandwidth greater than 100 Mbps). The network hardware would most likely need to be procured, unless a fast network is already available due to some other related research. The SAR processing software could be an adaptation of existing SAR software that runs on a single workstation. The application should be made scalable so that it can use whatever number of workstations is available. An attempt should be made to bypass the overheads posed by the 10. Conclusions and Future Work 141 networking protocols but still be as independent as possible of the underlying hardware. The software interface used for inter-processor communication needs to be chosen. Since the amount of coupling between the tasks mnning on different machines will need to be kept small, full message passing inter­faces like P V M or MPI may not be necessary. However, if a portable interface like P V M or MPI is used, the disadvantage of higher overheads may be outweighed by the ability to more easily port the application to other platforms (e.g. to parallel machines like the SP-2). General purpose DSP The DSP multicomputer approach to building SAR processors can most easily be explored using a commercially available multi-DSP board that plugs into a standard low-cost computer platform like a PC or workstation. The board should contain multiple floating point DSPs, but can otherwise be a fairly low-end, inexpensive board. An example would be a PCI board that contains 2 to 4 SHARCs. Once again, the SAR processing software could be an adaptation of existing SAR software. The application would be split: the control and user interface would ran on the host, while the signal processing portion would run on the DSPs. The challenge of the design would be to achieve high performance while leav­ing the software as portable as possible. A key decision that affects this is whether or not to use any high level programming toolkits offered by the DSP board vendor. Optimized DSP The design of an optimized DSP can be explored by modelling such a device using a behavioural HDL. Various design options and trade-offs can be explored. Existing models of standard building blocks or DSP cores can be used to accelerate this work. Synthesis tools can be used to verify the feasibility of the designs. It is quite possible that the designs produced will not be feasible for fabrication today. Var­ious options for mapping the design to current technologies can be investigated (e.g. multiple chips, M C M s , etc.). Digital filter chip processor An experimental SAR processor based on digital filter chips can be designed and built based on the description given in Section 9.5. Vector / scalar architecture A version of V S A described in Section 9.7 can be designed and built. This project would have the allure of trying to put the maximum amount of performance for SAR on a single board. However, the development of such a board would be fairly expensive. A significant amount of support software would have to be written for both the V S A and the host computer. From the software side, this project is an extension of the general purpose DSP work described above. 10. Conclusions and Future Work 142 F P G A computing machine A large amount of research work remains to be done in the field of F P G A computing machines. The application of this field to SAR processing can be pursued at two levels. Designs for key portions of SAR processing can be described at a fairly high level and existing synthesis tools can be used to obtain implementations. Alternatively, new low-level microarchitectures that are suitable for FPGAs can be developed for the key operations in SAR. However, both of these approaches are really in a separate research area and will not offer competitive solutions for SAR processor designs for a number of years. 10. Conclusions and Future Work 143 References [I] Analog Devices, Inc. 7995 DSP/MSP Products Reference Manual, 1995. [2] A . H . Anderson, G.S. Downs, and G.A. Shaw. "RASSP benchmark-1 and -2: A preliminary assess­ment." Technical report, M.I .T Lincoln Laboratory, 1995. [3] Allan H . Anderson, Gary A. Shaw, and Chris T. Sung. " V H D L executable requirements." In 1st Annual RASSP Conference, 1994. [4] G. Leigh Anderson. " A stepwise approach to computing the multidimensional fast fourier transform of large arrays." IEEE Transactions on Acoustics, Speech, and Signal Processing, ASSP-28(3):280, June 1980. [5] Thomas E. Anderson, David E. Culler, David A . Patterson, and Others. " A case for N O W (networks of workstations)." Technical report, UC Berkeley, 1994. [6] Peter M . Athanas and A . Lynn Abbott. "Real-time image processing on a custom computing platform." IEEE Computer, pages 16-24, February 1995. [7] Stan Baker. "Reconfigurable computing." Electronic Engineering Times, Feb. 28 1994. [8] J.R. Bennett, I.G. Cumming, and R . M . Wedding. "Algorithms for preprocessing of satellite SAR data." In Proceedings ISPRS Commission II Symposium, 1982. [9] Neil W. Bergmann and J. Craig Mudge. "Comparing the performance of FPGA-based custom comput­ers with general-purpose computers for DSP applications." In 7994 Workshop on FPGAs for Custom Com­puting, 1994. [10] Patrice Bertin, Didier Roncin, and Jean Vuillemin. "Introduction to programmable active memories." Technical Report 3, Digital Paris Research Laboratory, June 1989. [II] P. Bj0rstad, J. Cook, H . Munthe-Kaas, and T. S0revik. "Implementation of a SAR processing algo­rithm on MasPar MP-1208." Technical Report 57, MasPar, 1991. [12] Richard E. Blahut. Fast Algorithms for Digital Signal Processing. Addison-Wesley, Reading, Massa­chusetts, 1985. [13] Benjamin S. Blanchard and Wolter J. Fabrycky. Systems Engineering and Analysis. Prentice Hall, 1990. [14] Keith Boland and Apostolos Dollas. "Predicting and precluding problems with memory latency." IEEE Micro, pages 59-67, August 1994. [15] B .A. Bowen and W.R. Brown. Systems Design Volume II of VLSI Systems Design for Digital Signal Processing. Prentice-Hall, 1985. [16] Joe E. Brewer, L . Gray Miller, Ira H . Gilbert, Joseph F. Melia, Douglas Garde, and James E. DeMaris. " A monolithic processing subsystem." IEEE Transactions on Components, Packaging, and Manufacturing Technology - Part B, 17(3):310-317, Aug. 1994. 144 [17] Chappell Brown. "The programmable-logic assault." Electronic Engineering Times, pages 33,36, Feb. 81993. [18] Paul Budnik and David J. Kuck. "The organization and use of parallel memories." IEEE Transactions on Computers, pages 1566-1569, December 1971. [19] Dave Bursky. "Improved DSP ICs eye new horizons." Electronic Design, pages 69-82, Nov. 11 1993. [20] Dave Bursky. "Advanced D R A M s deliver peak systems performance." Electronic Design, 43(16):42, Aug. 7 1995. [21] C. Cafforio, C. Prati, and F. Rocca. "SAR data focussing using seismic migration techniques." IEEE Transactions Aerospace Electronic Systems, 27:199-207, 1991. [22] Steven Casselman. "Virtual computing and the virtual computer." In IEEE Workshop on FPGAs for Custom Computing Machines. IEEE, 1993. [23] Steven Casselman and Mike Thomburg. "Transformable computers." In 8th International Parallel Processing Symposium. IEEE, 1994. [24] C P U Info Center. http://infopad.eecs.berkely.edu/CIC/. [25] Jinsong Chong and Hailiang Peng. " A distributed SAR processing system based on P V M frame­work." Technical report, Institute of Electronics, Chinese Academy of Sciences, 1995. [26] Ian G. Cumming and John R. Bennett. "Digital processing of SEAS AT SAR data." In Record of the IEEE 1979 International Conference on Acoustics, Speech and Signal Processing, Washington, D.C., April 2-4 1979. [27] J. C. Curlander and R. N . McDonough. Synthetic Aperture Radar - Systems and Signal Processing. Wiley, 1991. [28] Gordon Davidson. Image Formation From Squint Mode Synthetic Aperture Radar Data. PhD thesis, UBC, 1994. [29] Angel L . DeCegama. Parallel Processing Architectures and VLSI Hardware. Prentice Hall, 1989. [30] Marc A . D'lorio, Joseph Lam, and John Whitman. "An overview of Canadian processing systems for Radarsat." In IGARSS'93,1993. [31] Jack J. Dongarra, Steve W. Otto, Marc Snir, and David Walker. " A n introduction to the MPI stand­ard." to appear in Communications of the ACM, 1995. [32] J.O. Eklundh. " A fast computer method for matrix transposing." IEEE Transactions on Computers, pages 801-803, July 1972. [33] J.O. Eklundh. "Efficient matrix transposition." In T.S. Huang, editor, Two-Dimensional Digital Signal Processing II: Transforms and Median Filters. Springer-Verlag, 1981. [34] Charles Elachi, Tom Bicknell, Rolando L . Jordan, and Chialin Wu. "Spaceborne synthetic-aperture imaging radars: Applications, techniques, and technology." Proceedings of the IEEE, 70(10):1174-1209, Oct 1982. 145 [35] Tse-Yun Feng. " A survey of interconnection networks." IEEE Computer, pages 12-27, December 1981. [36] J. Patrick Fitch. Synthetic Aperture Radar. Springer-Verlag, New York, 1988. [37] M.J . Flynn. "Some computer organizations and their effectiveness." IEEE Transactions on Comput­ers, C-21(9):948-960, Sept. 1972. [38] John Gallant. "High-density PLDs." EDN, page 31, March 16 1995. [39] A . Giri, V. Visvanathan, S.K. Nandy, and S.K. Ghoshal. "High speed digital filtering on SRAM-based FPGAs." In 7th International Conference on VLSI Design, pages 229-232. IEEE, 1994. [40] Maya Gokhale, William Holmes, Andrew Kopser, Sara Lucas, Ronald Minnich, Douglas Sweely, and Daniel Lopresti. "Building and using a highly parallel programmable logic array." IEEE Computer, pages 81-89, Jan 1991. [41] G. Goslin. "Using Xilinx FPGAs to design custom digital signal processing devices." Technical report, Xilinx Publications, 1995. [42] Brian K . Grant and Anthony Skjellum. "The P V M systems: An in-depth analysis and documenting study - concise edition." Technical report, Lawrence Livermore National Laboratory, 1992. [43] Anshul Gupta and Vipin Kumar. "On the scalability of FFT on parallel computers." In Third Sympo­sium on the Frontiers of Massively Parallel Computation. Proceedings, pages 69-74, New York, NY, 1990. IEEE. [44] David B. Gustavson and Qiang L i . "Local-area multiprocessor: the scalable coherent interface." Tech­nical report, SCIzzL, Santa Clara University, 1995. [45] M.W. Haney. "Compact acousto-optic processor for synthetic aperture radar image formation." Pro­ceedings of the IEEE, 82(11): 1735-48, Nov. 1994. [46] Johnathan C. Hardwick. "Porting a vector library: a comparison of MPI, Paris, C M M D and P V M . " Technical report, School of Computer Science, Carnegie Mellon University, 1994. [47] Haleh Hobooti. "Radiometric correction in range-SPECAN SAR processing." Master's thesis, U B C , 1995. [48] Frederick R. Hood, III. "Reconfigurable hardware for digital signal processing algorithms." Master's thesis, University of Washington, 1993. [49] Jeffrey I. Hutchings and Kent L . Gilson. "High performance digital signal processing using reconfig­urable F P G A logic." In The Proceedings of the 5th International Conference on Signal Processing Appli­cations and Technology (ICSPAT'94), 1994. [50] Barry Isenstein. "Scaling I/O bandwidth with multiprocessors." Electronic Design, pages 128-138, June 13 1994. [51] Rich Jaenicke, Ken Linton, and Doug Williams. " V M E , performance growing, gains support." Elec­tronic Engineering Times, (847):54, May 8 1995. 146 [52] Raj Jain. The Art of Computer Systems Performance Analysis. John Wiley and Sons, 1991. [53] Michael Y. Jin and Chialin Wu. " A SAR correlation algorithm which accomodates large-range migra­tion." IEEE Transactions on Geoscience and Remote Sensing, GE-22(6):592-597, Nov 1984. [54] Colin R. Johnson and Ron Wilson. "The 'virtual computer' gets boost from Altera." Electronic Engi­neering Times, page 1, Dec. 13 1993. [55] R. Colin Johnson. "Computing 2000: The PC unbound." OEM Magazine, pages 88-9, March 1994. [56] Asawaree Kalavade and Edward A . Lee. " A hardware-software codesign methodology for DSP appli­cations." IEEE Design & Test of Computers, pages 16-28, September 1993. [57] Duncan H . Lawrie. "Access and alignment of data in an array processor." IEEE Transactions on Com­puters, pages 1145-1155, December 1975. [58] Markus Levy and James P. Leonard. "EDN's DSP-chip directory." EDN, pages 40-95, May 11 1995. [59] Bede Liu and Abraham Peled. " A new hardware realization of high-speed fast fourier transformers." IEEE Transactions on Acoustics, Speech, and Signal Processing, ASSP-23(6):543-547, Dec. 1975. [60] Y. Luo, I.G. Cumming, and M.J . Yedlin. "Benchmarking a massively parallel computer by a synthetic aperture radar processing algorithm." In C.A. Brebbia and H . Power, editors, Applications of Supercom­puters in Engineering III, pages 393-408. Computational Mechanics Publications and Elsevier Science Publishers, 1993. [61] LR. Mactaggart and M . A . Jack. " A single chip radix-2 FFT butterfly architecture using parallel data distributed arithmetic." IEEE Journal of Solid-State Circuits, SC-19(3):368-373, June 1984. [62] Vijay K. Madisetti. VLSI Digital Signal Processors - An Introduction to Rapid Prototyping and Design Synthesis. IEEE Press, 1995. [63] M D A . "Spectral analysis approach to the compression of linear F M signals." Technical Report ESTEC Contract 3998/79/NL/HP(SC), M D A , 1979. [64] M D A . "Design concepts for the RadarSat SAR processor architecture." Technical Report RT-TN-50-3295, M D A , 1991. [65] M D A . "Survey of SAR processing algorithms for the Radarsat SAR processor." Technical Report RT-RP-50-3264, M D A , 1991. [66] M D A . "RSARP computing platform evaluation." Technical Report RZ-TN-50-4403, M D A , 1992. [67] M D A . "IRIS X2C system segment design manual." Technical Report IR-MA-50-5064, M D A , 1993. [68] M D A . "Proposal for the development of the Radarsat Fastscan processor." Technical Report 01-1453, M D A , 1993. [69] Les Mintzer. "Mechanization of digital signal processors." In Douglas F. Elliott, editor, Handbook of Digital Signal Processing Engineering Applications. Academic Press, Inc., 1987. 147 [70] Richard Nass. " A designer's guide to computer buses." Electronic Design, 41(24): 129, November 22 1993. [71] National Semiconductor. QuickRing Competitive Comparison, 1995. [72] Yoshiaki Nemoto, Hideo Nishino, Makoto Ono, Hitoshi Mizutamari, Katsuhiko Nishikawa, and Kaoru Tanaka. "Japanese Earth Resources Satellite-1 synthetic aperture radar." Proceedings of the IEEE, 79(6):800-809, June 1991. [73] Natawut Nupairoj and Lionel M . N i . "Performance evaluation of some MPI implementations on workstation clusters." Technical report, Dept. of Computer Science, Michigan State University, 1994. [74] List of FPGA-Based Computing Devices, http://www.io.com/ guccione/HW_list.html. [75] Abraham Peled and Bede Liu. Digital Signal Processing: Theory, Design, and Implementation. John Wiley and Sons, 1976. [76] Betty Prince. Semiconductor Memories. Wiley, 1991. [77] R. Keith Raney, Anthony P. Luscombe, E. J. Langham, and Shabeer Ahmed. "RADARSAT." Pro­ceedings of the IEEE, 79(6):839-849, June 1991. [78] R.K. Raney, H . Runge, R. Bamler, I. Cumming, and F. Wong. "Precision SAR processing using chirp scaling." IEEE Transactions on Geoscience and Remote Sensing, 32(4):786-799, July 1994. [79] M . Sack, M.R. Ito, and I.G. Cumming. "Application of efficient linear F M matched filtering algo­rithms to synthetic aperture radar processing." IEE Proceedings, 132 Pt. F(l):45, February 1985. [80] Sharp Corp. LH9124 Digital Signal Processor Data Sheet, 1993. [81] G.A. Shaw, J.C. Anderson, and V.K. Madisetti. "Assessing and improving current practice in the design of application-specific signal processors." Technical report, RASSP, 1994. [82] Michael R. Smith. "How RISCy is DSP?" IEEE Micro, pages 10-23, December 1992. [83] S.G. Smith and P B . Denyer. "Efficient bit-serial complex multiplication and sum-of-products compu­tation using distributed arithmetic." In Proc. 1986 International Conference on Acoustics, Speech, and Sig­nal Processing (ICASSP), pages 2203-2206, 1986. [84] Texas Instruments, Inc. TMS320C4x User's Guide, 1991. [85] Texas Instruments, Inc. Parallel Processing with the TMS32C4x, 1994. [86] Texas Instruments, Inc. TMS320C8x (MVP) Online Reference CD-ROM, 1995. [87] Kiyo Tomiyasu. "Tutorial review of synthetic-aperture radar (SAR) with applications to imaging the ocean surface." Proceedings of the IEEE, 66(5):563-583, May 1978. [88] David E. Van Den Bout, Joseph N . Morris, Douglas Thomae, Scott Labrozzi, Scot Wingo, and Dean Hallman. "AnyBoard: An FPGA-based, reconfigurable system." IEEE Design and Test of Computers, pages 21-30, Sept. 1992. 148 [89] Thomas C. Waugh. "Field programmable gate array key to reconfigurable array outperforming super­computers." In IEEE 1991 Custom Integrated Circuits Conference. IEEE, 1991. [90] S.A. White. " A simple FFT butterfly arithmetic unit." IEEE Transactions on Circuits and Systems, CAS-28(4):352-355, April 1981. [91] S.A. White. "Results of a preliminary study of a novel IC arithmetic unit ofr an FFT processor." In Proc. 18th Asilomar Conference on Circuits, Systems, and Computers, pages 67-71, 1984. [92] Stanley A . White. "Applications of distributed arithmetic to digital signal processing: A tutorial review." IEEEASSP Magazine, pages 4-19, July 1989. [93] C. Wu, K.Y. Liu, and M . Jin. "Modeling and a correlation algorithm for spacebome SAR signals." IEEE Transactions on Aerospace and Electronic Systems, AES-18:563-575, 1982. [94] Vojin Zivojnovic. "Benchmarks benediction." Electronic Engineering Times, (819):39, October 17 1994. [95] B . Zuerndorfer and G.A. Shaw. "SAR processing for RASSP application." Technical report, MIT Lin­coln Laboratory, 1994. Publications Derived from this Thesis Peter G. Meisl, Mabo R. Ito, and Ian G. Cumming. "Parallel synthetic aperture radar processing on work­station networks." Proceedings of 10th International Parallel Processing Symposium, pp. 716-723, IEEE, Honolulu, Hawaii, April 15-19, 1996. Peter G. Meisl, Mabo R. Ito, and Ian G. Cumming. "Parallel Processors for Synthetic Aperture Radar Imaging." To appear in Proceedings of 1996 International Conference on Parallel Processing, Bloomingdale, IL, Aug. 12-16, 1996. 149 A Model of Range-Doppler Algorithm The range-Doppler algorithm described in Section 2.5.4 was modelled using a Mathcad document. The key portions are shown in Figures A . l , A.2, and A.3. The amount of computation required by each bubble in the flow diagram of Figure 2.17 is shown and then summed to find the total amount of computation. Before and after each step (or implicit step) that changes the size of the data array, the total amount of memory required is shown. A. Model of Range-Doppler Algorithm 150 Derived Parameters N - I _ „ az az Frame time: Ftime = Ftime = 2.599 sec PRF Output azimuth dimension FinalNAZ := floor[OVSM^N ^ - L ^ - N res)] FinalNAZ =4883 samples Output range dimension FinalNR := floor[ OVSM- (N r - L r - N r c m c - N r e s ) ] FinalNR = 7537 samples Data Volume on Input N U J K D a t a l = ( N a z - L a z ) N r 2 ^ - Data, = 160.034 MB 2 - 8 1.0 l/Q Balance R D 1 0 := 3 { N f f i - L j . R V S ( N ^ 2.0 Range FFT R D 2 0 - ( N a z - L ^ - F F T ^ N r f n ) 3.0 Range Multiply R D 3 0 : = ( N a z - L a z ) - C V M ( N r f f t ) 4.0 Range IFFT RD 4 0:= ( N a z - L ( B ) - F F n d ( N r f B ) R D , 0 = 6.293-10 OPs R D 2 Q = 1.862-109 OPs R D 3 0 = 1.718-108 OPs R D , n = 1.862-109 OPs Data Volume after Range Compression N U j . „ • ^ ^ " B - ^ i N r - ^ V 0 ^ = 144.031 MB 2 '8 Data Volume before Azimuth Compression N h i t s Data, := f N r - L r V N „ - 2 Data, = 168.75 MB 3 v r r/ az g j 5.0 Azimuth FFT R D s o : = ( N r - L r ) - F F r i d ( N M ) R D 5 0 = 1.327-10 OPs 6.0 R C M C Shift and Interpolation " D f f l ^ B - l N f - L f - N j - I n l e r l N j R D 6 0 =6.626-10 OPs 7.0 Azimuth reference function multiply R D 7 0 : = ( N r - L r - N r c m c ) C V M ( N a z ) R D 7 0 = 1.325-10 OPs 8.0 Azimuth IFFT R D 8 0 = ( N r - L r - N r c m c ) F F T l d ( N a z ) R D g 0 = 1.325-10 OPs Data Volume after Azimuth Compression N D a t a 4 : = ( N r - L r - N r c m c ) ( N a z - L a z ) - 2 ^ - Data4 = 143.817 MB 2 • 8 Figure A.l: Computational analysis of range-Doppler algorithm (Part 1 of 3) A. Model of Range-Doppler Algorithm 9.0 Azimuth Resampling 9.1 Apply Interpolators RD91 := (N r - L r - N r c m c ) -FinalNAZ-Inter(N , 9.2 Generate Indices RD921 := FinalNAZ-1 RD922 := FinalNAZ-Round RD92 := RD921 + RD922 Total: R D 9 Q := RD91 + RD92 RD91 =7.899-108 OPs RD921 =4883 RD92 =2.93-10 OPS RD922 = 2.442-10 OPs OPS OPS Data Volume after Azimuth Resampling Data5 := (N r - L r - N rcmc)-FinalNAZ-2-1 bits 22 0.8 Data5 =200.875 MB 10.0 Range Resampling 10.1 Apply Interpolators RD101 := FinalNRFinalNAZInter(N r e s ) RD101 = 1.104-109 OPs 10.2 Generate Indices RD1021 = 2-FinalNR-l RD1021 = 1.507-104 OPs RD1022 := 2FinalNRRound RD1022 = 7.537-lO4 OPs RD102 := RD1021 + RD1022 RD102= 9.044-104 OPs Total: R D 1 0 Q := RD101 + RD102 R D , 0 0 = 1.104-109 OPs Data Volume after Range Resampling N bits Data, : = FinalNR- FinalNAZ- 2 22U-8 Data6 =280.786 MB 11.0 Detection RD, , 0 := 3FinalNRFinalNAZ R D l l n = 1.104-108 OPs Data Volume Nbits Data, := FinalNR-FinalNAZ 22U-8 Data, = 140.393 MB 12.0 Calculate Range Filter and S R C R D 1 2 „ : = 0 R D 1 2 0 = 0 OPs 13.0 Calculate Doppler Centroid R D 1 3 0 : = O R D 1 3 0 = 0 OPs Figure A.2: Computational analysis of range-Doppler algorithm (Part 2 of 3) A. Model of Range-Doppler Algorithm 14.0 Calculate RCM Shifts and Indices 14.1 Calculate RCM RD141 := N az^N r - L r - N r c m c ) ' 6 14.2 Round for Integer Shift RD142 := (N r - L r - N r c m c ) -N ^ R o u n d 14.3 Calculate Interpolator indices RD1431 := (N f - L r - N r c m c ) N ^ R o u n d RD1432 := (N r - L r - N r c m c ) -RVS(N ^ RD143 = RD1431 + RD1432 RD141 =1.325-10 OPs RD142 = 1.104-108 OPs RD1431 =1.104-10 RD1432 = 2.209-107 RD143 = 1.325-108 OPs OPS OPS Total R D , 4 0 := RD141 + RD142 + RD143 R D 1 4 0 = 3.755-10 OPS 15.0 Calculate Ambiguity R D 1 3 0 = 0 OPs 16.0 Calculate Azimuth Filter 16.1 Calculate Azimuth Filter in Time 1 RD161:= — ( N ^ L r - N ^ - L ^ O 16.2 FFT RD162 ;= J - . ( N r - L r - N r c m c ) - F F T l d ( N ^ n u 16.3 Frequency Weight Filter RD163 := — •[ (N r - L r - N rcmc) C V R V M ( N ^ ] n u 16.4 Phase compensation RD1641 := - L . ( N r - L f - N r c m c ) N ^ 1 0 RD1642 : N r C m c ) - C V M ( N a z ) RD164 = RD1641 + RD1642 RD161 =6.47-10' OPs RD162 = 1.325-109 OPs RD163 = 4.417-107 OPs RD1641 =2.209-108 OPs RD1642 = 1.325-108 OPs RD164= 3.534-108 OPs Total R D , 6 0 = RD161 + RD162 + RD163 + RD164 R D 1 6 0 = 1.787-10 OPs Statistics: i:= 10,20.. 160 Grand Total RD, ' o - Z > D i R D 0 = 1.157-10 OPs Operation Rate: R D n MOps := Ft imelO 6 MOps = 4452.245 MOP /s Figure A.3: Computational analysis of range-Doppler algorithm (Part 3 of 3) A. Model of Range-Doppler Algorithm 


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items