Les travaux de recherche détaillés dans ce document portent sur la conception et l’implantation d’un composant matériel jouant le rôle du sous-système cryptographique d’une radio logicielle sécurisée.A partir du début des années 90, les systèmes radios ont peu à peu évolué de la radio classique vers la radio logicielle. Le développement de la radio logicielle a permis l’intégration d’un nombre toujours plus grand de standards de communication sur une même plateforme matérielle. La réalisation concrète d’une radio logicielle sécurisée amène son concepteur à faire face à de nombreuses problématiques qui peuvent se résumer par la question suivante : Comment implanter un maximum de standards de communication sur une même plateforme matérielle et logicielle ? Ce document s’intéresse plus particulièrement à l’implantation des standards cryptographiques destinés à protéger les radiocommunications.Idéalement, la solution apportée à ce problème repose exclusivement sur l’utilisation de processeurs numériques. Cependant, les algorithmes cryptographiques nécessitent le plus souvent une puissance de calcul telle que leur implantation sous forme logicielle n’est pas envisageable. Il s’ensuit qu’une radio logicielle doit parfois intégrer des composants matériels dédiés dont l'utilisation entre en conflit avec la propriété de flexibilité propre aux radios logicielles.Or depuis quelques années, le développement de la technologie FPGA a changé la donne. En effet, les derniers FPGA embarquent un nombre de ressources logiques suffisant à l’implantation des fonctions numériques complexes utilisées par la radio logicielle. Plus précisément, la possibilité offerte par les FPGA d'être reconfiguré dans leur totalité (voir même partiellement pour les derniers d’entre eux) fait d’eux des candidats idéaux à l’implantation de composants matériels flexibles et évolutifs dans le temps. À la suite de ces constatations, des travaux de recherche ont été menés au sein de l’équipe Conception des Systèmes Numériques du Laboratoire IMS. Ces travaux ont d’abord débouché sur la publication d’une architecture de sous-système cryptographique pour la radio logicielle sécurisée telle qu’elle est définie par la Software Communication Architecture. Puis, ils se sont poursuivis par la conception et l’implantation d’un cryptoprocesseur multi-cœur dynamiquement reconfigurable sur FPGA. / The research detailed in this document deal with the design and implementation of a hardware integrated circuit intended to be used as a cryptographic sub-system in secure software defined radios.Since the early 90’s, radio systems have gradually evolved from traditional radio to software defined radio. Improvement of the software defined radio has enabled the integration of an increasing number of communication standards on a single radio device. The designer of a software defined radio faces many problems that can be summarized by the following question: How to implement a maximum of communication standards into a single radio device? Specifically, this work focuses on the implementation of cryptographic standards aimed to protect radio communications.Ideally, the solution to this problem is based exclusively on the use of digital processors. However, cryptographic algorithms usually require a large amount of computing power which makes their software implementation inefficient. Therefore, a secure software defined radio needs to incorporate dedicated hardware even if this usage is conflicting with the property of flexibility specific to software defined radios.Yet, in recent years, the improvement of FPGA circuits has changed the deal. Indeed, the latest FPGAs embed a number of logic gates which is sufficient to meet the needs of the complex digital functions used by software defined radios. The possibility offered by FPGAs to be reconfigured in their entirety (or even partially for the last of them) makes them ideal candidates for implementation of hardware components which have to be flexible and scalable over time.Following these observations, research was conducted within the Conception des Systèmes Numériques team of the IMS laboratory. These works led first to the publication of an architecture of cryptographic subsystem compliant with the security supplement of the Software Communication Architecture. Then, they continued with the design and implementation of a partially reconfigurable multi-core cryptoprocessor intended to be used in the latest FPGAs.

In this Ph.D. thesis, a new methodology for the fault tolerant system state synchronization design and its recovery from faults is presented. A state synchronization method designed by means of the proposed methodology allows to repair the state of sequential logic elements implemented in the FPGA application logic, which cannot be repaired by the partial dynamic reconfiguration. The proposed methodology describes possible state synchronization design methods with respect to TMR granularity, dependence of the system function on its previous states and the system architecture. The methodology focuses on coarse-grained TMR architectures and state synchronization in the systems controlled by means of finite state machines or a processor. The use of the methodology is demonstrated on the CAN bus control system and the microcontroller NEO430, for which specific synchronization methods were designed. The systems reliability and new ability of the systems for recovery from faults were verified in the presence of simulated SEU faults. The experimental results and the contribution of this thesis are discussed in the conclusion.

Field Programmable Gate Array (FPGA) devices offer a suitable platform for survivable hardware architectures in mission-critical systems. In this dissertation, active dynamic redundancy-based fault-handling techniques are proposed which exploit the dynamic partial reconfiguration capability of SRAM-based FPGAs. Self-adaptation is realized by employing reconfiguration in detection, diagnosis, and recovery phases. To extend these concepts to semiconductor aging and process variation in the deep submicron era, resilient adaptable processing systems are sought to maintain quality and throughput requirements despite the vulnerabilities of the underlying computational devices. A new approach to autonomous fault-handling which addresses these goals is developed using only a uniplex hardware arrangement. It operates by observing a health metric to achieve Fault Demotion using Recon- figurable Slack (FaDReS). Here an autonomous fault isolation scheme is employed which neither requires test vectors nor suspends the computational throughput, but instead observes the value of a health metric based on runtime input. The deterministic flow of the fault isolation scheme guarantees success in a bounded number of reconfigurations of the FPGA fabric. FaDReS is then extended to the Priority Using Resource Escalation (PURE) online redundancy scheme which considers fault-isolation latency and throughput trade-offs under a dynamic spare arrangement. While deep-submicron designs introduce new challenges, use of adaptive techniques are seen to provide several promising avenues for improving resilience. The scheme developed is demonstrated by hardware design of various signal processing circuits and their implementation on a Xilinx Virtex-4 FPGA device. These include a Discrete Cosine Transform (DCT) core, Motion Estimation (ME) engine, Finite Impulse Response (FIR) Filter, Support Vector Machine (SVM), and Advanced Encryption Standard (AES) blocks in addition to MCNC benchmark circuits. A iii significant reduction in power consumption is achieved ranging from 83% for low motion-activity scenes to 12.5% for high motion activity video scenes in a novel ME engine configuration. For a typical benchmark video sequence, PURE is shown to maintain a PSNR baseline near 32dB. The diagnosability, reconfiguration latency, and resource overhead of each approach is analyzed. Compared to previous alternatives, PURE maintains a PSNR within a difference of 4.02dB to 6.67dB from the fault-free baseline by escalating healthy resources to higher-priority signal processing functions. The results indicate the benefits of priority-aware resiliency over conventional redundancy approaches in terms of fault-recovery, power consumption, and resource-area requirements. Together, these provide a broad range of strategies to achieve autonomous recovery of reconfigurable logic devices under a variety of constraints, operating conditions, and optimization criteria.

FPGA based systems have been heavily used to prototype and test Application Specic Integrated Circuit (ASIC) designs with much lower costs and development time compared to hardwired prototypes. In recentyears, thanks to both the latest technology nodes and a change in the architecture of reconfigurable integrated circuits (from traditional Complex Programmable Logic Device (CPLD) to full-CMOS FPGA), FPGAs have become more popular in embedded systems, both as main computation resources and as hardware accelerators. A new era is beginning for FPGA based systems: the partial run-time reconguration of a FPGA is a feature now available in products already on the market and hardware designers and software developers have to exploit this capability. Previous works show that, when designed properly, a system can improve both its power efficiency and its performance taking advantage of a partial run-time reconfigurable architecture. Unfortunately, taking advantage of run-time reconfigurable hardware is very challenging and there are several problems to face: the reconfiguration overhead is not negligible compared to nowadays CPUs performance,the reconfiguration time is not easily predictable, and the software has to be re-though to work with a time-evolving platform. This thesis project aims to investigate the performance of a modern run-time reconfigurable SoC (a Xilinx Zynq 7020), focusing on the reconfiguration overhead and its predictability, on the achievable speedup, and the trade-off and limits of this kind of platform. Since it is not always obvious when an application (especially a real-time one) is really able to use at its own advantage a partial run-time reconfigurable platform, the data collected during this project could be a valid help for hardware designers that use reconfigurable computing. / FPGA-baserade system har tidigare främst använts för snabb och kostnadseffektiv konstruktion av prototyper vid framtagandet av applikationsspecika integrerade kretsar (ASIC). På senare år har användandet av FPGA:er i inbyggda system för implementation av hårdvaruacceleratorers såväl som huvudsaklig beräkningsenhet ökat. Denna ökning har möjliggjorts mycket tack vare den utveckling som har skett av rekonfigurerbara integrerade kretsar: från de mer traditionella Complex Programmable Logic Devices (CPLD) till helt CMOS-baserade FPGA:er. Nu inleds en ny era för FPGA-baserade system tack vare möjligheten att under körning rekonfigurera delar av FPGA:n genom så kallad partial run-time reconguration(RTR) - en teknik som redan idag finns tillgänglig i produkter på marknaden. Tidigare forskning visar att användandet av en RTR-baserad hårdvaruarkitektur kan ha en positiv effekt med avseende på prestanda såväl som strömförbrukning. Att använda RTR-baserad hårdvara innebär dock flera utmaningar: En ej försumbar rekonfigurationstid måste tas i beaktning, så även den icke-deterministiska exekveringstiden som en rekonfiguration kan innebära. Vidare måste anpassningar av mjukvaran göras för att fungera med en hårdvaruplattform som förändras över tid. Denna uppsats syftar till att undersöka prestandan hos ett modernt RTRbaserat SoC (Xilinx Zynq 7020) med fokus på rekonfigurationstider och dess förutsägbarhet, prestanda ökning, begränsningar samt nödvändiga kompromisser som denna arkitektur innebär. Huruvida en applikation kan dra nytta av en RTR-baserad arkitektur eller inte kan vara svårt att avgöra. Den insamlade datan som presenteras i denna rapport kan dock fungera som stöd för hårdvarukonstruktörer som önskar använda en RTR-baserad plattform.

Partial dynamic reconfiguration of FPGAs has attracted high attention from both academia and industry in recent years. With this technique, the functionality of the programmable devices can be adapted at runtime to changing requirements. The approach allows designers to use FPGAs more efficiently: E. g. FPGA resources can be time-shared between different functions and the functions itself can be adapted to changing workloads at runtime. Thus partial dynamic reconfiguration enables a unique combination of software-like flexibility and hardware-like performance. Still there exists no common understanding on how to assess the overhead introduced by partial dynamic reconfiguration. This dissertation presents a new cost model for both the runtime and the memory overhead that results from partial dynamic reconfiguration. It is shown how the model can be incorporated into all stages of the design optimization for reconfigurable hardware. In particular digital circuits can be mapped onto FPGAs such that only small fractions of the hardware must be reconfigured at runtime, which saves time, memory, and energy. The design optimization is most efficient if it is applied during high level synthesis. This book describes how the cost model has been integrated into a new high level synthesis tool. The tool allows the designer to trade-off FPGA resource use versus reconfiguration overhead. It is shown that partial reconfiguration causes only small overhead if the design is optimized with regard to reconfiguration cost. A wide range of experimental results is provided that demonstrates the benefits of the applied method. / Partielle dynamische Rekonfiguration von FPGAs hat in den letzten Jahren große Aufmerksamkeit von Wissenschaft und Industrie auf sich gezogen. Die Technik erlaubt es, die Funktionalität von progammierbaren Bausteinen zur Laufzeit an veränderte Anforderungen anzupassen. Dynamische Rekonfiguration erlaubt es Entwicklern, FPGAs effizienter einzusetzen: z.B. können Ressourcen für verschiedene Funktionen wiederverwendet werden und die Funktionen selbst können zur Laufzeit an veränderte Verarbeitungsschritte angepasst werden. Insgesamt erlaubt partielle dynamische Rekonfiguration eine einzigartige Kombination von software-artiger Flexibilität und hardware-artiger Leistungsfähigkeit. Bis heute gibt es keine Übereinkunft darüber, wie der zusätzliche Aufwand, der durch partielle dynamische Rekonfiguration verursacht wird, zu bewerten ist. Diese Dissertation führt ein neues Kostenmodell für Laufzeit und Speicherbedarf ein, welche durch partielle dynamische Rekonfiguration verursacht wird. Es wird aufgezeigt, wie das Modell in alle Ebenen der Entwurfsoptimierung für rekonfigurierbare Hardware einbezogen werden kann. Insbesondere wird gezeigt, wie digitale Schaltungen derart auf FPGAs abgebildet werden können, sodass nur wenig Ressourcen der Hardware zur Laufzeit rekonfiguriert werden müssen. Dadurch kann Zeit, Speicher und Energie eingespart werden. Die Entwurfsoptimierung ist am effektivsten, wenn sie auf der Ebene der High-Level-Synthese angewendet wird. Diese Arbeit beschreibt, wie das Kostenmodell in ein neuartiges Werkzeug für die High-Level-Synthese integriert wurde. Das Werkzeug erlaubt es, beim Entwurf die Nutzung von FPGA-Ressourcen gegen den Rekonfigurationsaufwand abzuwägen. Es wird gezeigt, dass partielle Rekonfiguration nur wenig Kosten verursacht, wenn der Entwurf bezüglich Rekonfigurationskosten optimiert wird. Eine Anzahl von Beispielen und experimentellen Ergebnissen belegt die Vorteile der angewendeten Methodik.

Partial dynamic reconfiguration of FPGAs has attracted high attention from both academia and industry in recent years. With this technique, the functionality of the programmable devices can be adapted at runtime to changing requirements. The approach allows designers to use FPGAs more efficiently: E. g. FPGA resources can be time-shared between different functions and the functions itself can be adapted to changing workloads at runtime. Thus partial dynamic reconfiguration enables a unique combination of software-like flexibility and hardware-like performance. Still there exists no common understanding on how to assess the overhead introduced by partial dynamic reconfiguration. This dissertation presents a new cost model for both the runtime and the memory overhead that results from partial dynamic reconfiguration. It is shown how the model can be incorporated into all stages of the design optimization for reconfigurable hardware. In particular digital circuits can be mapped onto FPGAs such that only small fractions of the hardware must be reconfigured at runtime, which saves time, memory, and energy. The design optimization is most efficient if it is applied during high level synthesis. This book describes how the cost model has been integrated into a new high level synthesis tool. The tool allows the designer to trade-off FPGA resource use versus reconfiguration overhead. It is shown that partial reconfiguration causes only small overhead if the design is optimized with regard to reconfiguration cost. A wide range of experimental results is provided that demonstrates the benefits of the applied method.:1 Introduction 1 1.1 Reconfigurable Computing . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1.1 Reconfigurable System on a Chip (RSOC) . . . . . . . . . . . . 4 1.1.2 Anatomy of an Application . . . . . . . . . . . . . . . . . . . . . . 6 1.1.3 RSOC Design Characteristics and Trade-offs . . . . . . . . . . . 7 1.2 Classification of Reconfigurable Architectures . . . . . . . . . . . . . . . 10 1.2.1 Partial Reconfiguration . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2.2 Runtime Reconfiguration (RTR) . . . . . . . . . . . . . . . . . . . 10 1.2.3 Multi-Context Configuration . . . . . . . . . . . . . . . . . . . . . 11 1.2.4 Fine-Grain Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.2.5 Coarse-Grain Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.3 Reconfigurable Computing Specific Design Issues . . . . . . . . . . . . 12 1.4 Overview of this Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . 14 2 Reconfigurable Computing Systems – Background 17 2.1 Examples for RSOCs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2 Partially Reconfigurable FPGAs: Xilinx Virtex Device Family . . . . . . 20 2.2.1 Virtex-II/Virtex-II Pro Logic Architecture . . . . . . . . . . . . . 20 2.2.2 Reconfiguration Architecture and Reconfiguration Control . . 21 2.3 Methods for Design Entry . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.3.1 Behavioural Design Entry . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.2 Design Entry at Register-Transfer Level (RTL) . . . . . . . . . . 25 2.3.3 Xilinx Early Access Partial Reconfiguration Design Flow . . . . 26 2.4 Task Management in Reconfigurable Computing . . . . . . . . . . . . . 27 2.4.1 Online and Offline Task Management . . . . . . . . . . . . . . . 28 2.4.2 Task Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.4.3 Task Placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.4.4 Reconfiguration Runtime Overhead . . . . . . . . . . . . . . . . 31 2.5 Configuration Data Compression . . . . . . . . . . . . . . . . . . . . . . . 32 2.6 Evaluation of Reconfigurable Systems . . . . . . . . . . . . . . . . . . . . 35 2.6.1 Energy Efficiency Models . . . . . . . . . . . . . . . . . . . . . . . 35 2.6.2 Area Efficiency Models . . . . . . . . . . . . . . . . . . . . . . . . 37 2.6.3 Runtime Efficiency Models . . . . . . . . . . . . . . . . . . . . . . 37 2.7 Similarity Based Reduction of Reconfiguration Overhead . . . . . . . . 38 2.7.1 Configuration Data Generation Methods . . . . . . . . . . . . . 39 2.7.2 Device Mapping Methods . . . . . . . . . . . . . . . . . . . . . . . 40 2.7.3 Circuit Design Methods . . . . . . . . . . . . . . . . . . . . . . . . 41 2.7.4 Model for Partial Configuration . . . . . . . . . . . . . . . . . . . 44 2.8 Contributions of this Work . . . . . . . . . . . . . . . . . . . . . . . . . . 44 3 Runtime Reconfiguration Cost and Optimization Methods 47 3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.2 Reconfiguration State Graph . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.2.1 Reconfiguration Time Overhead . . . . . . . . . . . . . . . . . . 52 3.2.2 Dynamic Configuration Data Overhead . . . . . . . . . . . . . . 52 3.3 Configuration Cost at Bitstream Level . . . . . . . . . . . . . . . . . . . . 54 3.4 Configuration Cost at Structural Level . . . . . . . . . . . . . . . . . . . 56 3.4.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 3.4.2 Virtual Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 62 3.4.3 Reconfiguration Costs in the VA Context . . . . . . . . . . . . . 65 3.5 Allocation Functions with Minimal Reconfiguration Costs . . . . . . . 67 3.5.1 Allocation of Node Pairs . . . . . . . . . . . . . . . . . . . . . . . 68 3.5.2 Direct Allocation of Nodes . . . . . . . . . . . . . . . . . . . . . . 76 3.5.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 4 Implementation Tools for Reconfigurable Computing 95 4.1 Mapping of Netlists to FPGA Resources . . . . . . . . . . . . . . . . . . . 96 4.1.1 Mapping to Device Resources . . . . . . . . . . . . . . . . . . . . 96 4.1.2 Connectivity Transformations . . . . . . . . . . . . . . . . . . . . 99 4.1.3 Mapping Variants and Reconfiguration Costs . . . . . . . . . . . 100 4.1.4 Mapping of Circuit Macros . . . . . . . . . . . . . . . . . . . . . . 101 4.1.5 Global Interconnect . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4.1.6 Netlist Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 4.2 Mapping Aware Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . 103 4.2.1 Generalized Node Mapping . . . . . . . . . . . . . . . . . . . . . 104 4.2.2 Successive Node Allocation . . . . . . . . . . . . . . . . . . . . . 105 4.2.3 Node Allocation with Ant Colony Optimization . . . . . . . . . 107 4.2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 4.3 Netlist Mapping with Minimized Reconfiguration Cost . . . . . . . . . 110 4.3.1 Mapping Database . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 4.3.2 Mapping and Packing of Elements into Logic Blocks . . . . . . 112 4.3.3 Logic Element Selection . . . . . . . . . . . . . . . . . . . . . . . 114 4.3.4 Logic Element Selection for Min. Routing Reconfiguration . . 115 4.3.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 5 High-Level Synthesis for Reconfigurable Computing 125 5.1 Introduction to HLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.1.1 HLS Tool Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.1.2 Realization of the Hardware Tasks . . . . . . . . . . . . . . . . . 128 5.2 New Concepts for Task-based Reconfiguration . . . . . . . . . . . . . . 131 5.2.1 Multiple Hardware Tasks in one Reconfigurable Module . . . . 132 5.2.2 Multi-Level Reconfiguration . . . . . . . . . . . . . . . . . . . . . 133 5.2.3 Resource Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 5.3 Datapath Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 5.3.1 Task Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 5.3.2 Resource Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 5.3.3 Resource Binding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 5.3.4 Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 5.3.5 Constraints for Scheduling and Resource Binding . . . . . . . . 151 5.4 Reconfiguration Optimized Datapath Implementation . . . . . . . . . . 153 5.4.1 Effects of Scheduling and Binding on Reconfiguration Costs . 153 5.4.2 Strategies for Resource Type Binding . . . . . . . . . . . . . . . 154 5.4.3 Strategies for Resource Instance Binding . . . . . . . . . . . . . 157 5.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 5.5.1 Summary of Binding Methods and Tool Setup . . . . . . . . . . 163 5.5.2 Cost Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 5.5.3 Implementation Scenarios . . . . . . . . . . . . . . . . . . . . . . 166 5.5.4 Benchmark Characteristics . . . . . . . . . . . . . . . . . . . . . . 168 5.5.5 Benchmark Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 5.5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 6 Summary and Outlook 185 Bibliography 189 A Simulated Annealing 201 / Partielle dynamische Rekonfiguration von FPGAs hat in den letzten Jahren große Aufmerksamkeit von Wissenschaft und Industrie auf sich gezogen. Die Technik erlaubt es, die Funktionalität von progammierbaren Bausteinen zur Laufzeit an veränderte Anforderungen anzupassen. Dynamische Rekonfiguration erlaubt es Entwicklern, FPGAs effizienter einzusetzen: z.B. können Ressourcen für verschiedene Funktionen wiederverwendet werden und die Funktionen selbst können zur Laufzeit an veränderte Verarbeitungsschritte angepasst werden. Insgesamt erlaubt partielle dynamische Rekonfiguration eine einzigartige Kombination von software-artiger Flexibilität und hardware-artiger Leistungsfähigkeit. Bis heute gibt es keine Übereinkunft darüber, wie der zusätzliche Aufwand, der durch partielle dynamische Rekonfiguration verursacht wird, zu bewerten ist. Diese Dissertation führt ein neues Kostenmodell für Laufzeit und Speicherbedarf ein, welche durch partielle dynamische Rekonfiguration verursacht wird. Es wird aufgezeigt, wie das Modell in alle Ebenen der Entwurfsoptimierung für rekonfigurierbare Hardware einbezogen werden kann. Insbesondere wird gezeigt, wie digitale Schaltungen derart auf FPGAs abgebildet werden können, sodass nur wenig Ressourcen der Hardware zur Laufzeit rekonfiguriert werden müssen. Dadurch kann Zeit, Speicher und Energie eingespart werden. Die Entwurfsoptimierung ist am effektivsten, wenn sie auf der Ebene der High-Level-Synthese angewendet wird. Diese Arbeit beschreibt, wie das Kostenmodell in ein neuartiges Werkzeug für die High-Level-Synthese integriert wurde. Das Werkzeug erlaubt es, beim Entwurf die Nutzung von FPGA-Ressourcen gegen den Rekonfigurationsaufwand abzuwägen. Es wird gezeigt, dass partielle Rekonfiguration nur wenig Kosten verursacht, wenn der Entwurf bezüglich Rekonfigurationskosten optimiert wird. Eine Anzahl von Beispielen und experimentellen Ergebnissen belegt die Vorteile der angewendeten Methodik.:1 Introduction 1 1.1 Reconfigurable Computing . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1.1 Reconfigurable System on a Chip (RSOC) . . . . . . . . . . . . 4 1.1.2 Anatomy of an Application . . . . . . . . . . . . . . . . . . . . . . 6 1.1.3 RSOC Design Characteristics and Trade-offs . . . . . . . . . . . 7 1.2 Classification of Reconfigurable Architectures . . . . . . . . . . . . . . . 10 1.2.1 Partial Reconfiguration . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2.2 Runtime Reconfiguration (RTR) . . . . . . . . . . . . . . . . . . . 10 1.2.3 Multi-Context Configuration . . . . . . . . . . . . . . . . . . . . . 11 1.2.4 Fine-Grain Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.2.5 Coarse-Grain Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.3 Reconfigurable Computing Specific Design Issues . . . . . . . . . . . . 12 1.4 Overview of this Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . 14 2 Reconfigurable Computing Systems – Background 17 2.1 Examples for RSOCs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2 Partially Reconfigurable FPGAs: Xilinx Virtex Device Family . . . . . . 20 2.2.1 Virtex-II/Virtex-II Pro Logic Architecture . . . . . . . . . . . . . 20 2.2.2 Reconfiguration Architecture and Reconfiguration Control . . 21 2.3 Methods for Design Entry . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.3.1 Behavioural Design Entry . . . . . . . . . . . . . . . . . . . . . . . 25 2.3.2 Design Entry at Register-Transfer Level (RTL) . . . . . . . . . . 25 2.3.3 Xilinx Early Access Partial Reconfiguration Design Flow . . . . 26 2.4 Task Management in Reconfigurable Computing . . . . . . . . . . . . . 27 2.4.1 Online and Offline Task Management . . . . . . . . . . . . . . . 28 2.4.2 Task Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.4.3 Task Placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.4.4 Reconfiguration Runtime Overhead . . . . . . . . . . . . . . . . 31 2.5 Configuration Data Compression . . . . . . . . . . . . . . . . . . . . . . . 32 2.6 Evaluation of Reconfigurable Systems . . . . . . . . . . . . . . . . . . . . 35 2.6.1 Energy Efficiency Models . . . . . . . . . . . . . . . . . . . . . . . 35 2.6.2 Area Efficiency Models . . . . . . . . . . . . . . . . . . . . . . . . 37 2.6.3 Runtime Efficiency Models . . . . . . . . . . . . . . . . . . . . . . 37 2.7 Similarity Based Reduction of Reconfiguration Overhead . . . . . . . . 38 2.7.1 Configuration Data Generation Methods . . . . . . . . . . . . . 39 2.7.2 Device Mapping Methods . . . . . . . . . . . . . . . . . . . . . . . 40 2.7.3 Circuit Design Methods . . . . . . . . . . . . . . . . . . . . . . . . 41 2.7.4 Model for Partial Configuration . . . . . . . . . . . . . . . . . . . 44 2.8 Contributions of this Work . . . . . . . . . . . . . . . . . . . . . . . . . . 44 3 Runtime Reconfiguration Cost and Optimization Methods 47 3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.2 Reconfiguration State Graph . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.2.1 Reconfiguration Time Overhead . . . . . . . . . . . . . . . . . . 52 3.2.2 Dynamic Configuration Data Overhead . . . . . . . . . . . . . . 52 3.3 Configuration Cost at Bitstream Level . . . . . . . . . . . . . . . . . . . . 54 3.4 Configuration Cost at Structural Level . . . . . . . . . . . . . . . . . . . 56 3.4.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 3.4.2 Virtual Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 62 3.4.3 Reconfiguration Costs in the VA Context . . . . . . . . . . . . . 65 3.5 Allocation Functions with Minimal Reconfiguration Costs . . . . . . . 67 3.5.1 Allocation of Node Pairs . . . . . . . . . . . . . . . . . . . . . . . 68 3.5.2 Direct Allocation of Nodes . . . . . . . . . . . . . . . . . . . . . . 76 3.5.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 4 Implementation Tools for Reconfigurable Computing 95 4.1 Mapping of Netlists to FPGA Resources . . . . . . . . . . . . . . . . . . . 96 4.1.1 Mapping to Device Resources . . . . . . . . . . . . . . . . . . . . 96 4.1.2 Connectivity Transformations . . . . . . . . . . . . . . . . . . . . 99 4.1.3 Mapping Variants and Reconfiguration Costs . . . . . . . . . . . 100 4.1.4 Mapping of Circuit Macros . . . . . . . . . . . . . . . . . . . . . . 101 4.1.5 Global Interconnect . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4.1.6 Netlist Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 4.2 Mapping Aware Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . 103 4.2.1 Generalized Node Mapping . . . . . . . . . . . . . . . . . . . . . 104 4.2.2 Successive Node Allocation . . . . . . . . . . . . . . . . . . . . . 105 4.2.3 Node Allocation with Ant Colony Optimization . . . . . . . . . 107 4.2.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 4.3 Netlist Mapping with Minimized Reconfiguration Cost . . . . . . . . . 110 4.3.1 Mapping Database . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 4.3.2 Mapping and Packing of Elements into Logic Blocks . . . . . . 112 4.3.3 Logic Element Selection . . . . . . . . . . . . . . . . . . . . . . . 114 4.3.4 Logic Element Selection for Min. Routing Reconfiguration . . 115 4.3.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 5 High-Level Synthesis for Reconfigurable Computing 125 5.1 Introduction to HLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.1.1 HLS Tool Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5.1.2 Realization of the Hardware Tasks . . . . . . . . . . . . . . . . . 128 5.2 New Concepts for Task-based Reconfiguration . . . . . . . . . . . . . . 131 5.2.1 Multiple Hardware Tasks in one Reconfigurable Module . . . . 132 5.2.2 Multi-Level Reconfiguration . . . . . . . . . . . . . . . . . . . . . 133 5.2.3 Resource Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 5.3 Datapath Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 5.3.1 Task Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 5.3.2 Resource Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 5.3.3 Resource Binding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 5.3.4 Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 5.3.5 Constraints for Scheduling and Resource Binding . . . . . . . . 151 5.4 Reconfiguration Optimized Datapath Implementation . . . . . . . . . . 153 5.4.1 Effects of Scheduling and Binding on Reconfiguration Costs . 153 5.4.2 Strategies for Resource Type Binding . . . . . . . . . . . . . . . 154 5.4.3 Strategies for Resource Instance Binding . . . . . . . . . . . . . 157 5.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 5.5.1 Summary of Binding Methods and Tool Setup . . . . . . . . . . 163 5.5.2 Cost Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 5.5.3 Implementation Scenarios . . . . . . . . . . . . . . . . . . . . . . 166 5.5.4 Benchmark Characteristics . . . . . . . . . . . . . . . . . . . . . . 168 5.5.5 Benchmark Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 5.5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 6 Summary and Outlook 185 Bibliography 189 A Simulated Annealing 201

Space debris and space situational awareness (SSA) have become growing concerns for national security and the sustainability of space operations, where timely detection and tracking of space objects is critical in preventing collision events. Traditional computer-vision algorithms have been used extensively to solve detection and tracking problems in flight, but recently deep learning approaches have seen widespread adoption in non-space related applications for their high accuracy. The performanceper-watt and flexibility of reconfigurable Field-Programmable Gate Arrays (FPGAs) make them a good candidate for deep learning model deployment in space, supporting in-flight updates and maintenance. However, the FPGA design costs of custom accelerators for complex algorithms remains high. The research focus of the thesis relies on novel high-level synthesis (HLS) workflows that allow the developer to raise the level of abstraction and lower design costs for deep learning accelerators, particularly for space-representative applications. To this end, four different hardware accelerators of convolutional neural network models for spacebased debris detection are implemented (ResNet, SqueezeNet, DenseNet, TinyCNN), using the open-source HLS tool NNgen. The obtained hardware accelerators are deployed to a reconfigurable module of the Zynq Ultrascale+ MPSoC programmable logic, and compared in terms of inference performance, resource utilization and latency. The tests on the target hardware show a detection accuracy over 95% for ResNet, DenseNet and SqueezeNet, and a localization intersection-over-union over 0.5 for the deep models, and over 0.7 for TinyCNN, for space debris objects at a range between 1km and 100km for a diameter of 1cm, or between 100km and 1000km for a diameter of 10cm. The obtained speed-ups with respect to software-only implementations lay between 3x and 32x for the different hardware accelerators. / Rymdskrot och rymdsituationstänksamhet (SSA) har blivit växande oro för nationell säkerhet och hållbarheten för rymdoperationer, där snabb upptäckt och spårning av rymdobjekt är avgörande för att förhindra kollisioner. Traditionella datorseendealgoritmer har använts omfattande för att lösa problem med upptäckt och spårning i flygning, men på senare tid har djupinlärningsmetoder fått stor användning inom icke rymdrelaterade applikationer på grund av sin höga noggrannhet. Prestandaper-watt och flexibiliteten hos omkonfigurerbara Field-Programmable Gate Arrays (FPGAs) gör dem till en bra kandidat för distribution av djupinlärningsmodeller i rymden, med stöd för uppdateringar och underhåll under flygning. Men FPGAdesignkostnaderna för anpassade acceleratorer för komplexa algoritmer är fortfarande höga. Forskningsfokus för avhandlingen ligger på nya högnivåsyntes (HLS) arbetsflöden som gör det möjligt för utvecklaren att höja abstraktionsnivån och sänka designkostnaderna för acceleratorer för djupinlärning, särskilt för tillämpningar i rymden. För detta har fyra olika hårdvaruacceleratorer för modeller av konvolutionsnätverk för upptäckt av rymdbaserat skrot implementerats (ResNet, SqueezeNet, DenseNet, TinyCNN), med hjälp av öppen källkod HLS-verktyget NNgen. De erhållna hårdvaruacceleratorerna distribueras till en omkonfigurerbar modul av Zynq Ultrascale+ MPSoC-programmerbar logik och jämförs med avseende på inferensprestanda, resursutnyttjande och latens. Testerna på målhardwaren visar en upptäktnoggrannhet på över 95% för ResNet, DenseNet och SqueezeNet, och en lokaliserings-intersektion-över-union på över 0,5 för de djupa modellerna och över 0,7 för TinyCNN för rymdskrotobjekt på en avstånd mellan 1 km och 100 km för en diameter på 1 cm eller mellan 100 km och 1000 km för en diameter på 10 cm. De erhållna hastighetsökningarna i förhållande till endast programvara ligger mellan 3x och 32x för de olika hårdvaruacceleratorerna.