Particle methods parallel implementations by GP-GPU strategies

This paper outlines the problems found in the parallelization of SPH (Smoothed Particle Hydrodynamics) algorithms using Graphics Processing Units. Different results of some parallel GPU implementations in terms of the speed-up and the scalability compared to the CPU sequential codes are shown. The most problematic stage in the GPU-SPH algorithms is the one responsible for locating neighboring particles and building the vectors where this information is stored, since these specific algorithms raise many dificulties for a data-level parallelization. Because of the fact that the neighbor location using linked lists does not show enough data-level parallelism, two new approaches have been pro- posed to minimize bank conflicts in the writing and subsequent reading of the neighbor lists. The first strategy proposes an efficient coordination between CPU-GPU, using GPU algorithms for those stages that allow a straight forward parallelization, and sequential CPU algorithms for those instructions that involve some kind of vector reduction. This coordination provides a relatively orderly reading of the neighbor lists in the interactions stage, achieving a speed-up factor of x47 in this stage. However, since the construction of the neighbor lists is quite expensive, it is achieved an overall speed-up of x41. The second strategy seeks to maximize the use of the GPU in the neighbor's location process by executing a specific vector sorting algorithm that allows some data-level parallelism. Al- though this strategy has succeeded in improving the speed-up on the stage of neighboring location, the global speed-up on the interactions stage falls, due to inefficient reading of the neighbor vectors. Some changes to these strategies are proposed, aimed at maximizing the computational load of the GPU and using the GPU texture-units, in order to reach the maximum speed-up for such codes. Different practical applications have been added to the mentioned GPU codes. First, the classical dam-break problem is studied. Second, the wave impact of the sloshing fluid contained in LNG vessel tanks is also simulated as a practical example of particle methods

Run-Time Scalable Hardware for Reconfigurable Systems

La optimización de parámetros tales como el consumo de potencia, la cantidad de recursos lógicos empleados o la ocupación de memoria ha sido siempre una de las preocupaciones principales a la hora de diseñar sistemas embebidos. Esto es debido a que se trata de sistemas dotados de una cantidad de recursos limitados, y que han sido tradicionalmente empleados para un propósito específico, que permanece invariable a lo largo de toda la vida útil del sistema. Sin embargo, el uso de sistemas embebidos se ha extendido a áreas de aplicación fuera de su ámbito tradicional, caracterizadas por una mayor demanda computacional. Así, por ejemplo, algunos de estos sistemas deben llevar a cabo un intenso procesado de señales multimedia o la transmisión de datos mediante sistemas de comunicaciones de alta capacidad. Por otra parte, las condiciones de operación del sistema pueden variar en tiempo real. Esto sucede, por ejemplo, si su funcionamiento depende de datos medidos por el propio sistema o recibidos a través de la red, de las demandas del usuario en cada momento, o de condiciones internas del propio dispositivo, tales como la duración de la batería. Como consecuencia de la existencia de requisitos de operación dinámicos es necesario ir hacia una gestión dinámica de los recursos del sistema. Si bien el software es inherentemente flexible, no ofrece una potencia computacional tan alta como el hardware. Por lo tanto, el hardware reconfigurable aparece como una solución adecuada para tratar con mayor flexibilidad los requisitos variables dinámicamente en sistemas con alta demanda computacional. La flexibilidad y adaptabilidad del hardware requieren de dispositivos reconfigurables que permitan la modificación de su funcionalidad bajo demanda. En esta tesis se han seleccionado las FPGAs (Field Programmable Gate Arrays) como los dispositivos más apropiados, hoy en día, para implementar sistemas basados en hardware reconfigurable De entre todas las posibilidades existentes para explotar la capacidad de reconfiguración de las FPGAs comerciales, se ha seleccionado la reconfiguración dinámica y parcial. Esta técnica consiste en substituir una parte de la lógica del dispositivo, mientras el resto continúa en funcionamiento. La capacidad de reconfiguración dinámica y parcial de las FPGAs es empleada en esta tesis para tratar con los requisitos de flexibilidad y de capacidad computacional que demandan los dispositivos embebidos. La propuesta principal de esta tesis doctoral es el uso de arquitecturas de procesamiento escalables espacialmente, que son capaces de adaptar su funcionalidad y rendimiento en tiempo real, estableciendo un compromiso entre dichos parámetros y la cantidad de lógica que ocupan en el dispositivo. A esto nos referimos con arquitecturas con huellas escalables. En particular, se propone el uso de arquitecturas altamente paralelas, modulares, regulares y con una alta localidad en sus comunicaciones, para este propósito. El tamaño de dichas arquitecturas puede ser modificado mediante la adición o eliminación de algunos de los módulos que las componen, tanto en una dimensión como en dos. Esta estrategia permite implementar soluciones escalables, sin tener que contar con una versión de las mismas para cada uno de los tamaños posibles de la arquitectura. De esta manera se reduce significativamente el tiempo necesario para modificar su tamaño, así como la cantidad de memoria necesaria para almacenar todos los archivos de configuración. En lugar de proponer arquitecturas para aplicaciones específicas, se ha optado por patrones de procesamiento genéricos, que pueden ser ajustados para solucionar distintos problemas en el estado del arte. A este respecto, se proponen patrones basados en esquemas sistólicos, así como de tipo wavefront. Con el objeto de poder ofrecer una solución integral, se han tratado otros aspectos relacionados con el diseño y el funcionamiento de las arquitecturas, tales como el control del proceso de reconfiguración de la FPGA, la integración de las arquitecturas en el resto del sistema, así como las técnicas necesarias para su implementación. Por lo que respecta a la implementación, se han tratado distintos aspectos de bajo nivel dependientes del dispositivo. Algunas de las propuestas realizadas a este respecto en la presente tesis doctoral son un router que es capaz de garantizar el correcto rutado de los módulos reconfigurables dentro del área destinada para ellos, así como una estrategia para la comunicación entre módulos que no introduce ningún retardo ni necesita emplear recursos configurables del dispositivo. El flujo de diseño propuesto se ha automatizado mediante una herramienta denominada DREAMS. La herramienta se encarga de la modificación de las netlists correspondientes a cada uno de los módulos reconfigurables del sistema, y que han sido generadas previamente mediante herramientas comerciales. Por lo tanto, el flujo propuesto se entiende como una etapa de post-procesamiento, que adapta esas netlists a los requisitos de la reconfiguración dinámica y parcial. Dicha modificación la lleva a cabo la herramienta de una forma completamente automática, por lo que la productividad del proceso de diseño aumenta de forma evidente. Para facilitar dicho proceso, se ha dotado a la herramienta de una interfaz gráfica. El flujo de diseño propuesto, y la herramienta que lo soporta, tienen características específicas para abordar el diseño de las arquitecturas dinámicamente escalables propuestas en esta tesis. Entre ellas está el soporte para el realojamiento de módulos reconfigurables en posiciones del dispositivo distintas a donde el módulo es originalmente implementado, así como la generación de estructuras de comunicación compatibles con la simetría de la arquitectura. El router has sido empleado también en esta tesis para obtener un rutado simétrico entre nets equivalentes. Dicha posibilidad ha sido explotada para aumentar la protección de circuitos con altos requisitos de seguridad, frente a ataques de canal lateral, mediante la implantación de lógica complementaria con rutado idéntico. Para controlar el proceso de reconfiguración de la FPGA, se propone en esta tesis un motor de reconfiguración especialmente adaptado a los requisitos de las arquitecturas dinámicamente escalables. Además de controlar el puerto de reconfiguración, el motor de reconfiguración ha sido dotado de la capacidad de realojar módulos reconfigurables en posiciones arbitrarias del dispositivo, en tiempo real. De esta forma, basta con generar un único bitstream por cada módulo reconfigurable del sistema, independientemente de la posición donde va a ser finalmente reconfigurado. La estrategia seguida para implementar el proceso de realojamiento de módulos es diferente de las propuestas existentes en el estado del arte, pues consiste en la composición de los archivos de configuración en tiempo real. De esta forma se consigue aumentar la velocidad del proceso, mientras que se reduce la longitud de los archivos de configuración parciales a almacenar en el sistema. El motor de reconfiguración soporta módulos reconfigurables con una altura menor que la altura de una región de reloj del dispositivo. Internamente, el motor se encarga de la combinación de los frames que describen el nuevo módulo, con la configuración existente en el dispositivo previamente. El escalado de las arquitecturas de procesamiento propuestas en esta tesis también se puede beneficiar de este mecanismo. Se ha incorporado también un acceso directo a una memoria externa donde se pueden almacenar bitstreams parciales. Para acelerar el proceso de reconfiguración se ha hecho funcionar el ICAP por encima de la máxima frecuencia de reloj aconsejada por el fabricante. Así, en el caso de Virtex-5, aunque la máxima frecuencia del reloj deberían ser 100 MHz, se ha conseguido hacer funcionar el puerto de reconfiguración a frecuencias de operación de hasta 250 MHz, incluyendo el proceso de realojamiento en tiempo real. Se ha previsto la posibilidad de portar el motor de reconfiguración a futuras familias de FPGAs. Por otro lado, el motor de reconfiguración se puede emplear para inyectar fallos en el propio dispositivo hardware, y así ser capaces de evaluar la tolerancia ante los mismos que ofrecen las arquitecturas reconfigurables. Los fallos son emulados mediante la generación de archivos de configuración a los que intencionadamente se les ha introducido un error, de forma que se modifica su funcionalidad. Con el objetivo de comprobar la validez y los beneficios de las arquitecturas propuestas en esta tesis, se han seguido dos líneas principales de aplicación. En primer lugar, se propone su uso como parte de una plataforma adaptativa basada en hardware evolutivo, con capacidad de escalabilidad, adaptabilidad y recuperación ante fallos. En segundo lugar, se ha desarrollado un deblocking filter escalable, adaptado a la codificación de vídeo escalable, como ejemplo de aplicación de las arquitecturas de tipo wavefront propuestas. El hardware evolutivo consiste en el uso de algoritmos evolutivos para diseñar hardware de forma autónoma, explotando la flexibilidad que ofrecen los dispositivos reconfigurables. En este caso, los elementos de procesamiento que componen la arquitectura son seleccionados de una biblioteca de elementos presintetizados, de acuerdo con las decisiones tomadas por el algoritmo evolutivo, en lugar de definir la configuración de las mismas en tiempo de diseño. De esta manera, la configuración del core puede cambiar cuando lo hacen las condiciones del entorno, en tiempo real, por lo que se consigue un control autónomo del proceso de reconfiguración dinámico. Así, el sistema es capaz de optimizar, de forma autónoma, su propia configuración. El hardware evolutivo tiene una capacidad inherente de auto-reparación. Se ha probado que las arquitecturas evolutivas propuestas en esta tesis son tolerantes ante fallos, tanto transitorios, como permanentes y acumulativos. La plataforma evolutiva se ha empleado para implementar filtros de eliminación de ruido. La escalabilidad también ha sido aprovechada en esta aplicación. Las arquitecturas evolutivas escalables permiten la adaptación autónoma de los cores de procesamiento ante fluctuaciones en la cantidad de recursos disponibles en el sistema. Por lo tanto, constituyen un ejemplo de escalabilidad dinámica para conseguir un determinado nivel de calidad, que puede variar en tiempo real. Se han propuesto dos variantes de sistemas escalables evolutivos. El primero consiste en un único core de procesamiento evolutivo, mientras que el segundo está formado por un número variable de arrays de procesamiento. La codificación de vídeo escalable, a diferencia de los codecs no escalables, permite la decodificación de secuencias de vídeo con diferentes niveles de calidad, de resolución temporal o de resolución espacial, descartando la información no deseada. Existen distintos algoritmos que soportan esta característica. En particular, se va a emplear el estándar Scalable Video Coding (SVC), que ha sido propuesto como una extensión de H.264/AVC, ya que este último es ampliamente utilizado tanto en la industria, como a nivel de investigación. Para poder explotar toda la flexibilidad que ofrece el estándar, hay que permitir la adaptación de las características del decodificador en tiempo real. El uso de las arquitecturas dinámicamente escalables es propuesto en esta tesis con este objetivo. El deblocking filter es un algoritmo que tiene como objetivo la mejora de la percepción visual de la imagen reconstruida, mediante el suavizado de los "artefactos" de bloque generados en el lazo del codificador. Se trata de una de las tareas más intensivas en procesamiento de datos de H.264/AVC y de SVC, y además, su carga computacional es altamente dependiente del nivel de escalabilidad seleccionado en el decodificador. Por lo tanto, el deblocking filter ha sido seleccionado como prueba de concepto de la aplicación de las arquitecturas dinámicamente escalables para la compresión de video. La arquitectura propuesta permite añadir o eliminar unidades de computación, siguiendo un esquema de tipo wavefront. La arquitectura ha sido propuesta conjuntamente con un esquema de procesamiento en paralelo del deblocking filter a nivel de macrobloque, de tal forma que cuando se varía del tamaño de la arquitectura, el orden de filtrado de los macrobloques varia de la misma manera. El patrón propuesto se basa en la división del procesamiento de cada macrobloque en dos etapas independientes, que se corresponden con el filtrado horizontal y vertical de los bloques dentro del macrobloque. Las principales contribuciones originales de esta tesis son las siguientes: - El uso de arquitecturas altamente regulares, modulares, paralelas y con una intensa localidad en sus comunicaciones, para implementar cores de procesamiento dinámicamente reconfigurables. - El uso de arquitecturas bidimensionales, en forma de malla, para construir arquitecturas dinámicamente escalables, con una huella escalable. De esta forma, las arquitecturas permiten establecer un compromiso entre el área que ocupan en el dispositivo, y las prestaciones que ofrecen en cada momento. Se proponen plantillas de procesamiento genéricas, de tipo sistólico o wavefront, que pueden ser adaptadas a distintos problemas de procesamiento. - Un flujo de diseño y una herramienta que lo soporta, para el diseño de sistemas reconfigurables dinámicamente, centradas en el diseño de las arquitecturas altamente paralelas, modulares y regulares propuestas en esta tesis. - Un esquema de comunicaciones entre módulos reconfigurables que no introduce ningún retardo ni requiere el uso de recursos lógicos propios. - Un router flexible, capaz de resolver los conflictos de rutado asociados con el diseño de sistemas reconfigurables dinámicamente. - Un algoritmo de optimización para sistemas formados por múltiples cores escalables que optimice, mediante un algoritmo genético, los parámetros de dicho sistema. Se basa en un modelo conocido como el problema de la mochila. - Un motor de reconfiguración adaptado a los requisitos de las arquitecturas altamente regulares y modulares. Combina una alta velocidad de reconfiguración, con la capacidad de realojar módulos en tiempo real, incluyendo el soporte para la reconfiguración de regiones que ocupan menos que una región de reloj, así como la réplica de un módulo reconfigurable en múltiples posiciones del dispositivo. - Un mecanismo de inyección de fallos que, empleando el motor de reconfiguración del sistema, permite evaluar los efectos de fallos permanentes y transitorios en arquitecturas reconfigurables. - La demostración de las posibilidades de las arquitecturas propuestas en esta tesis para la implementación de sistemas de hardware evolutivos, con una alta capacidad de procesamiento de datos. - La implementación de sistemas de hardware evolutivo escalables, que son capaces de tratar con la fluctuación de la cantidad de recursos disponibles en el sistema, de una forma autónoma. - Una estrategia de procesamiento en paralelo para el deblocking filter compatible con los estándares H.264/AVC y SVC que reduce el número de ciclos de macrobloque necesarios para procesar un frame de video. - Una arquitectura dinámicamente escalable que permite la implementación de un nuevo deblocking filter, totalmente compatible con los estándares H.264/AVC y SVC, que explota el paralelismo a nivel de macrobloque. El presente documento se organiza en siete capítulos. En el primero se ofrece una introducción al marco tecnológico de esta tesis, especialmente centrado en la reconfiguración dinámica y parcial de FPGAs. También se motiva la necesidad de las arquitecturas dinámicamente escalables propuestas en esta tesis. En el capítulo 2 se describen las arquitecturas dinámicamente escalables. Dicha descripción incluye la mayor parte de las aportaciones a nivel arquitectural realizadas en esta tesis. Por su parte, el flujo de diseño adaptado a dichas arquitecturas se propone en el capítulo 3. El motor de reconfiguración se propone en el 4, mientras que el uso de dichas arquitecturas para implementar sistemas de hardware evolutivo se aborda en el 5. El deblocking filter escalable se describe en el 6, mientras que las conclusiones finales de esta tesis, así como la descripción del trabajo futuro, son abordadas en el capítulo 7. ABSTRACT The optimization of system parameters, such as power dissipation, the amount of hardware resources and the memory footprint, has been always a main concern when dealing with the design of resource-constrained embedded systems. This situation is even more demanding nowadays. Embedded systems cannot anymore be considered only as specific-purpose computers, designed for a particular functionality that remains unchanged during their lifetime. Differently, embedded systems are now required to deal with more demanding and complex functions, such as multimedia data processing and high-throughput connectivity. In addition, system operation may depend on external data, the user requirements or internal variables of the system, such as the battery life-time. All these conditions may vary at run-time, leading to adaptive scenarios. As a consequence of both the growing computational complexity and the existence of dynamic requirements, dynamic resource management techniques for embedded systems are needed. Software is inherently flexible, but it cannot meet the computing power offered by hardware solutions. Therefore, reconfigurable hardware emerges as a suitable technology to deal with the run-time variable requirements of complex embedded systems. Adaptive hardware requires the use of reconfigurable devices, where its functionality can be modified on demand. In this thesis, Field Programmable Gate Arrays (FPGAs) have been selected as the most appropriate commercial technology existing nowadays to implement adaptive hardware systems. There are different ways of exploiting reconfigurability in reconfigurable devices. Among them is dynamic and partial reconfiguration. This is a technique which consists in substituting part of the FPGA logic on demand, while the rest of the device continues working. The strategy followed in this thesis is to exploit the dynamic and partial reconfiguration of commercial FPGAs to deal with the flexibility and complexity demands of state-of-the-art embedded systems. The proposal of this thesis to deal with run-time variable system conditions is the use of spatially scalable processing hardware IP cores, which are able to adapt their functionality or performance at run-time, trading them off with the amount of logic resources they occupy in the device. This is referred to as a scalable footprint in the context of this thesis. The distinguishing characteristic of the proposed cores is that they rely on highly parallel, modular and regular architectures, arranged in one or two dimensions. These architectures can be scaled by means of the addition or removal of the composing blocks. This strategy avoids implementing a full version of the core for each possible size, with the corresponding benefits in terms of scaling and adaptation time, as well as bitstream storage memory requirements. Instead of providing specific-purpose architectures, generic architectural templates, which can be tuned to solve different problems, are proposed in this thesis. Architectures following both systolic and wavefront templates have been selected. Together with the proposed scalable architectural templates, other issues needed to ensure the proper design and operation of the scalable cores, such as the device reconfiguration control, the run-time management of the architecture and the implementation techniques have been also addressed in this thesis. With regard to the implementation of dynamically reconfigurable architectures, device dependent low-level details are addressed. Some of the aspects covered in this thesis are the area constrained routing for reconfigurable modules, or an inter-module communication strategy which does not introduce either extra delay or logic overhead. The system implementation, from the hardware description to the device configuration bitstream, has been fully automated by modifying the netlists corresponding to each of the system modules, which are previously generated using the vendor tools. This modification is therefore envisaged as a post-processing step. Based on these implementation proposals, a design tool called DREAMS (Dynamically Reconfigurable Embedded and Modular Systems) has been created, including a graphic user interface. The tool has specific features to cope with modular and regular architectures, including the support for module relocation and the inter-module communications scheme based on the symmetry of the architecture. The core of the tool is a custom router, which has been also exploited in this thesis to obtain symmetric routed nets, with the aim of enhancing the protection of critical reconfigurable circuits against side channel attacks. This is achieved by duplicating the logic with an exactly equal routing. In order to control the reconfiguration process of the FPGA, a Reconfiguration Engine suited to the specific requirements set by the proposed architectures was also proposed. Therefore, in addition to controlling the reconfiguration port, the Reconfiguration Engine has been enhanced with the online relocation ability, which allows employing a unique configuration bitstream for all the positions where the module may be placed in the device. Differently to the existing relocating solutions, which are based on bitstream parsers, the proposed approach is based on the online composition of bitstreams. This strategy allows increasing the speed of the process, while the length of partial bitstreams is also reduced. The height of the reconfigurable modules can be lower than the height of a clock region. The Reconfiguration Engine manages the merging process of the new and the existing configuration frames within each clock region. The process of scaling up and down the hardware cores also benefits from this technique. A direct link to an external memory where partial bitstreams can be stored has been also implemented. In order to accelerate the reconfiguration process, the ICAP has been overclocked over the speed reported by the manufacturer. In the case of Virtex-5, even though the maximum frequency of the ICAP is reported to be 100 MHz, valid operations at 250 MHz have been achieved, including the online relocation process. Portability of the reconfiguration solution to today's and probably, future FPGAs, has been also considered. The reconfiguration engine can be also used to inject faults in real hardware devices, and this way being able to evaluate the fault tolerance offered by the reconfigurable architectures. Faults are emulated by introducing partial bitstreams intentionally modified to provide erroneous functionality. To prove the validity and the benefits offered by the proposed architectures, two demonstration application lines have been envisaged. First, scalable architectures have been employed to develop an evolvable hardware platform with adaptability, fault tolerance and scalability properties. Second, they have been used to implement a scalable deblocking filter suited to scalable video coding. Evolvable Hardware is the use of evolutionary algorithms to design hardware in an autonomous way, exploiting the flexibility offered by reconfigurable devices. In this case, processing elements composing the architecture are selected from a presynthesized library of processing elements, according to the decisions taken by the algorithm, instead of being decided at design time. This way, the configuration of the array may change as run-time environmental conditions do, achieving autonomous control of the dynamic reconfiguration process. Thus, the self-optimization property is added to the native self-configurability of the dynamically scalable architectures. In addition, evolvable hardware adaptability inherently offers self-healing features. The proposal has proved to be self-tolerant, since it is able to self-recover from both transient and cumulative permanent faults. The proposed evolvable architecture has been used to implement noise removal image filters. Scalability has been also exploited in this application. Scalable evolvable hardware architectures allow the autonomous adaptation of the processing cores to a fluctuating amount of resources available in the system. Thus, it constitutes an example of the dynamic quality scalability tackled in this thesis. Two variants have been proposed. The first one consists in a single dynamically scalable evolvable core, and the second one contains a variable number of processing cores. Scalable video is a flexible approach for video compression, which offers scalability at different levels. Differently to non-scalable codecs, a scalable video bitstream can be decoded with different levels of quality, spatial or temporal resolutions, by discarding the undesired information. The interest in this technology has been fostered by the development of the Scalable Video Coding (SVC) standard, as an extension of H.264/AVC. In order to exploit all the flexibility offered by the standard, it is necessary to adapt the characteristics of the decoder to the requirements of each client during run-time. The use of dynamically scalable architectures is proposed in this thesis with this aim. The deblocking filter algorithm is the responsible of improving the visual perception of a reconstructed image, by smoothing blocking artifacts generated in the encoding loop. This is one of the most computationally intensive tasks of the standard, and furthermore, it is highly dependent on the selected scalability level in the decoder. Therefore, the deblocking filter has been selected as a proof of concept of the implementation of dynamically scalable architectures for video compression. The proposed architecture allows the run-time addition or removal of computational units working in parallel to change its level of parallelism, following a wavefront computational pattern. Scalable architecture is offered together with a scalable parallelization strategy at the macroblock level, such that when the size of the architecture changes, the macroblock filtering order is modified accordingly. The proposed pattern is based on the division of the macroblock processing into two independent stages, corresponding to the horizontal and vertical filtering of the blocks within the macroblock. The main contributions of this thesis are: - The use of highly parallel, modular, regular and local architectures to implement dynamically reconfigurable processing IP cores, for data intensive applications with flexibility requirements. - The use of two-dimensional mesh-type arrays as architectural templates to build dynamically reconfigurable IP cores, with a scalable footprint. The proposal consists in generic architectural templates, which can be tuned to solve different computational problems. •A design flow and a tool targeting the design of DPR systems, focused on highly parallel, modular and local architectures. - An inter-module communication strategy, which does not introduce delay or area overhead, named Virtual Borders. - A custom and flexible router to solve the routing conflicts as well as the inter-module communication problems, appearing during the design of DPR systems. - An algorithm addressing the optimization of systems composed of multiple scalable cores, which size can be decided individually, to optimize the system parameters. It is based on a model known as the multi-dimensional multi-choice Knapsack problem. - A reconfiguration engine tailored to the requirements of highly regular and modular architectures. It combines a high reconfiguration throughput with run-time module relocation capabilities, including the support for sub-clock reconfigurable regions and the replication in multiple positions. - A fault injection mechanism which takes advantage of the system reconfiguration engine, as well as the modularity of the proposed reconfigurable architectures, to evaluate the effects of transient and permanent faults in these architectures. - The demonstration of the possibilities of the architectures proposed in this thesis to implement evolvable hardware systems, while keeping a high processing throughput. - The implementation of scalable evolvable hardware systems, which are able to adapt to the fluctuation of the amount of resources available in the system, in an autonomous way. - A parallelization strategy for the H.264/AVC and SVC deblocking filter, which reduces the number of macroblock cycles needed to process the whole frame. - A dynamically scalable architecture that permits the implementation of a novel deblocking filter module, fully compliant with the H.264/AVC and SVC standards, which exploits the macroblock level parallelism of the algorithm. This document is organized in seven chapters. In the first one, an introduction to the technology framework of this thesis, specially focused on dynamic and partial reconfiguration, is provided. The need for the dynamically scalable processing architectures proposed in this work is also motivated in this chapter. In chapter 2, dynamically scalable architectures are described. Description includes most of the architectural contributions of this work. The design flow tailored to the scalable architectures, together with the DREAMs tool provided to implement them, are described in chapter 3. The reconfiguration engine is described in chapter 4. The use of the proposed scalable archtieectures to implement evolvable hardware systems is described in chapter 5, while the scalable deblocking filter is described in chapter 6. Final conclusions of this thesis, and the description of future work, are addressed in chapter 7.

Securing implementations of feedback-shift-register-based ciphers using compiler optimizations and co-processors

Los algoritmos basados en registros de desplazamiento con realimentación (en inglés FSR) se han utilizado como generadores de flujos pseudoaleatorios en aplicaciones con recursos limitados como los sistemas de apertura sin llave. Se considera canal primario a aquel que se utiliza para realizar una transmisión de información. La aparición de los ataques de canal auxiliar (en inglés SCA), que explotan información filtrada inintencionadamente a través de canales laterales como el consumo, las emisiones electromagnéticas o el tiempo empleado, supone una grave amenaza para estas aplicaciones, dado que los dispositivos son accesibles por un atacante. El objetivo de esta tesis es proporcionar un conjunto de protecciones que se puedan aplicar de forma automática y que utilicen recursos ya disponibles, evitando un incremento sustancial en los costes y alargando la vida útil de aplicaciones que puedan estar desplegadas. Explotamos el paralelismo existente en algoritmos FSR, ya que sólo hay 1 bit de diferencia entre estados de rondas consecutivas. Realizamos aportaciones en tres niveles: a nivel de sistema, utilizando un coprocesador reconfigurable, a través del compilador y a nivel de bit, aprovechando los recursos disponibles en el procesador. Proponemos un marco de trabajo que nos permite evaluar implementaciones de un algoritmo incluyendo los efectos introducidos por el compilador considerando que el atacante es experto. En el campo de los ataques, hemos propuesto un nuevo ataque diferencial que se adapta mejor a las condiciones de las implementaciones software de FSR, en las que el consumo entre rondas es muy similar. SORU2 es un co-procesador vectorial reconfigurable propuesto para reducir el consumo energético en aplicaciones con paralelismo y basadas en el uso de bucles. Proponemos el uso de SORU2, además, para ejecutar algoritmos basados en FSR de forma segura. Al ser reconfigurable, no supone un sobrecoste en recursos, ya que no está dedicado en exclusiva al algoritmo de cifrado. Proponemos una configuración que ejecuta múltiples algoritmos de cifrado similares de forma simultánea, con distintas implementaciones y claves. A partir de una implementación sin protecciones, que demostramos que es completamente vulnerable ante SCA, obtenemos una implementación segura a los ataques que hemos realizado. A nivel de compilador, proponemos un mecanismo para evaluar los efectos de las secuencias de optimización del compilador sobre una implementación. El número de posibles secuencias de optimizaciones de compilador es extremadamente alto. El marco de trabajo propuesto incluye un algoritmo para la selección de las secuencias de optimización a considerar. Debido a que las optimizaciones del compilador transforman las implementaciones, se pueden generar automáticamente implementaciones diferentes combinamos para incrementar la seguridad ante SCA. Proponemos 2 mecanismos de aplicación de estas contramedidas, que aumentan la seguridad de la implementación original sin poder considerarse seguras. Finalmente hemos propuesto la ejecución paralela a nivel de bit del algoritmo en un procesador. Utilizamos la forma algebraica normal del algoritmo, que automáticamente se paraleliza. La implementación sobre el algoritmo evaluado mejora en rendimiento y evita que se filtre información por una ejecución dependiente de datos. Sin embargo, es más vulnerable ante ataques diferenciales que la implementación original. Proponemos una modificación del algoritmo para obtener una implementación segura, descartando parcialmente ejecuciones del algoritmo, de forma aleatoria. Esta implementación no introduce una sobrecarga en rendimiento comparada con las implementaciones originales. En definitiva, hemos propuesto varios mecanismos originales a distintos niveles para introducir aleatoridad en implementaciones de algoritmos FSR sin incrementar sustancialmente los recursos necesarios. ABSTRACT Feedback Shift Registers (FSR) have been traditionally used to implement pseudorandom sequence generators. These generators are used in Stream ciphers in systems with tight resource constraints, such as Remote Keyless Entry. When communicating electronic devices, the primary channel is the one used to transmit the information. Side-Channel Attack (SCA) use additional information leaking from the actual implementation, including power consumption, electromagnetic emissions or timing information. Side-Channel Attacks (SCA) are a serious threat to FSR-based applications, as an attacker usually has physical access to the devices. The main objective of this Ph.D. thesis is to provide a set of countermeasures that can be applied automatically using the available resources, avoiding a significant cost overhead and extending the useful life of deployed systems. If possible, we propose to take advantage of the inherent parallelism of FSR-based algorithms, as the state of a FSR differs from previous values only in 1-bit. We have contributed in three different levels: architecture (using a reconfigurable co-processor), using compiler optimizations, and at bit level, making the most of the resources available at the processor. We have developed a framework to evaluate implementations of an algorithm including the effects introduced by the compiler. We consider the presence of an expert attacker with great knowledge on the application and the device. Regarding SCA, we have presented a new differential SCA that performs better than traditional SCA on software FSR-based algorithms, where the leaked values are similar between rounds. SORU2 is a reconfigurable vector co-processor. It has been developed to reduce energy consumption in loop-based applications with parallelism. In addition, we propose its use for secure implementations of FSR-based algorithms. The cost overhead is discarded as the co-processor is not exclusively dedicated to the encryption algorithm. We present a co-processor configuration that executes multiple simultaneous encryptions, using different implementations and keys. From a basic implementation, which is proved to be vulnerable to SCA, we obtain an implementation where the SCA applied were unsuccessful. At compiler level, we use the framework to evaluate the effect of sequences of compiler optimization passes on a software implementation. There are many optimization passes available. The optimization sequences are combinations of the available passes. The amount of sequences is extremely high. The framework includes an algorithm for the selection of interesting sequences that require detailed evaluation. As existing compiler optimizations transform the software implementation, using different optimization sequences we can automatically generate different implementations. We propose to randomly switch between the generated implementations to increase the resistance against SCA.We propose two countermeasures. The results show that, although they increase the resistance against SCA, the resulting implementations are not secure. At bit level, we propose to exploit bit level parallelism of FSR-based implementations using pseudo bitslice implementation in a wireless node processor. The bitslice implementation is automatically obtained from the Algebraic Normal Form of the algorithm. The results show a performance improvement, avoiding timing information leakage, but increasing the vulnerability against differential SCA.We provide a secure version of the algorithm by randomly discarding part of the data obtained. The overhead in performance is negligible when compared to the original implementations. To summarize, we have proposed a set of original countermeasures at different levels that introduce randomness in FSR-based algorithms avoiding a heavy overhead on the resources required.

Automatic compile-time parallelization of logic programs for restricted, goal-level, independent and-parallelism.

A framework for the automatic parallelization of (constraint) logic programs is proposed and proved correct. Intuitively, the parallelization process replaces conjunctions of literals with parallel expressions. Such expressions trigger at run-time the exploitation of restricted, goal-level, independent and-parallelism. The parallelization process performs two steps. The first one builds a conditional dependency graph (which can be implified using compile-time analysis information), while the second transforms the resulting graph into linear conditional expressions, the parallel expressions of the &-Prolog language. Several heuristic algorithms for the latter ("annotation") process are proposed and proved correct. Algorithms are also given which determine if there is any loss of parallelism in the linearization process with respect to a proposed notion of maximal parallelism. Finally, a system is presented which implements the proposed approach. The performance of the different annotation algorithms is compared experimentally in this system by studying the time spent in parallelization and the effectiveness of the results in terms of speedups.

High-level characteristics of or-and independent and-parallelism in prolog

Although studies of a number of parallel implementations of logic programming languages are now available, their results are difficult to interpret due to the multiplicity of factors involved, the effect of each of which is difficult to sepárate. In this paper we present the results of a high-level simulation study of or- and independent and-parallelism with a wide selection of Prolog programs that aims to determine the intrinsic amount of parallelism, independently of implementation factors, thus facilitating this separation. We expect this study will be instrumental in better understanding and comparing results from actual implementations, as shown by some examples provided in the paper. In addition, the paper examines some of the issues and tradeoffs associated with the combination of and- and or-parallelism and proposes reasonable solutions based on the simulation data obtained.

Towards high-level execution primitives for and-parallelism: preliminary results

Most implementations of parallel logic programming rely on complex low-level machinery which is arguably difflcult to implement and modify. We explore an alternative approach aimed at taming that complexity by raising core parts of the implementation to the source language level for the particular case of and-parallelism. Therefore, we handle a signiflcant portion of the parallel implementation mechanism at the Prolog level with the help of a comparatively small number of concurrency-related primitives which take care of lower-level tasks such as locking, thread management, stack set management, etc. The approach does not eliminate altogether modiflcations to the abstract machine, but it does greatly simplify them and it also facilitates experimenting with different alternatives. We show how this approach allows implementing both restricted and unrestricted (i.e., non fork-join) parallelism. Preliminary experiments show that the amount of performance sacriflced is reasonable, although granularity control is required in some cases. Also, we observe that the availability of unrestricted parallelism contributes to better observed speedups.

Towards a high-level implementation of flexible parallelism primitives for symbolic languages

An abstract is not available.

Exploration of RVC applications using an ARM multicore processor = Exploración de aplicaciones RVC empleando un procesador ARM multinúcleo

Single core capabilities have reached their maximum clock speed; new multicore architectures provide an alternative way to tackle this issue instead. The design of decoding applications running on top of these multicore platforms and their optimization to exploit all system computational power is crucial to obtain best results. Since the development at the integration level of printed circuit boards are increasingly difficult to optimize due to physical constraints and the inherent increase in power consumption, development of multiprocessor architectures is becoming the new Holy Grail. In this sense, it is crucial to develop applications that can run on the new multi-core architectures and find out distributions to maximize the potential use of the system. Today most of commercial electronic devices, available in the market, are composed of embedded systems. These devices incorporate recently multi-core processors. Task management onto multiple core/processors is not a trivial issue, and a good task/actor scheduling can yield to significant improvements in terms of efficiency gains and also processor power consumption. Scheduling of data flows between the actors that implement the applications aims to harness multi-core architectures to more types of applications, with an explicit expression of parallelism into the application. On the other hand, the recent development of the MPEG Reconfigurable Video Coding (RVC) standard allows the reconfiguration of the video decoders. RVC is a flexible standard compatible with MPEG developed codecs, making it the ideal tool to integrate into the new multimedia terminals to decode video sequences. With the new versions of the Open RVC-CAL Compiler (Orcc), a static mapping of the actors that implement the functionality of the application can be done once the application executable has been generated. This static mapping must be done for each of the different cores available on the working platform. It has been chosen an embedded system with a processor with two ARMv7 cores. This platform allows us to obtain the desired tests, get as much improvement results from the execution on a single core, and contrast both with a PC-based multiprocessor system. Las posibilidades ofrecidas por el aumento de la velocidad de la frecuencia de reloj de sistemas de un solo procesador están siendo agotadas. Las nuevas arquitecturas multiprocesador proporcionan una vía de desarrollo alternativa en este sentido. El diseño y optimización de aplicaciones de descodificación de video que se ejecuten sobre las nuevas arquitecturas permiten un mejor aprovechamiento y favorecen la obtención de mayores rendimientos. Hoy en día muchos de los dispositivos comerciales que se están lanzando al mercado están integrados por sistemas embebidos, que recientemente están basados en arquitecturas multinúcleo. El manejo de las tareas de ejecución sobre este tipo de arquitecturas no es una tarea trivial, y una buena planificación de los actores que implementan las funcionalidades puede proporcionar importantes mejoras en términos de eficiencia en el uso de la capacidad de los procesadores y, por ende, del consumo de energía. Por otro lado, el reciente desarrollo del estándar de Codificación de Video Reconfigurable (RVC), permite la reconfiguración de los descodificadores de video. RVC es un estándar flexible y compatible con anteriores codecs desarrollados por MPEG. Esto hace de RVC el estándar ideal para ser incorporado en los nuevos terminales multimedia que se están comercializando. Con el desarrollo de las nuevas versiones del compilador específico para el desarrollo de lenguaje RVC-CAL (Orcc), en el que se basa MPEG RVC, el mapeo estático, para entornos basados en multiprocesador, de los actores que integran un descodificador es posible. Se ha elegido un sistema embebido con un procesador con dos núcleos ARMv7. Esta plataforma nos permitirá llevar a cabo las pruebas de verificación y contraste de los conceptos estudiados en este trabajo, en el sentido del desarrollo de descodificadores de video basados en MPEG RVC y del estudio de la planificación y mapeo estático de los mismos.

Annotation algorithms for unrestricted independent and-parallelism in logic programs

We present two new algorithms which perform automatic parallelization via source-to-source transformations. The objective is to exploit goal-level, unrestricted independent and-parallelism. The proposed algorithms use as targets new parallel execution primitives which are simpler and more flexible than the well-known &/2 parallel operator. This makes it possible to genérate better parallel expressions by exposing more potential parallelism among the literals of a clause than is possible with &/2. The difference between the two algorithms stems from whether the order of the solutions obtained is preserved or not. We also report on a preliminary evaluation of an implementation of our approach. We compare the performance obtained to that of previous annotation algorithms and show that relevant improvements can be obtained.

Task granularity analysis in logic programs

While logic programming languages offer a great deal of scope for parallelism, there is usually some overhead associated with the execution of goals in parallel because of the work involved in task creation and scheduling. In practice, therefore, the "granularity" of a goal, i.e. an estimate of the work available under it, should be taken into account when deciding whether or not to execute a goal concurrently as a sepárate task. This paper describes a method for estimating the granularity of a goal at compile time. The runtime overhead associated with our approach is usually quite small, and the performance improvements resulting from the incorporation of grainsize control can be quite good. This is shown by means of experimental results.

On the correctness and efficiency of independent and-parallelism in logic programs

This paper presents and proves some fundamental results for independent and-parallelism (IAP). First, the paper treats the issues of correctness and efficiency: after defining strict and non-strict goal independence, it is proved that if strictly independent goals are executed in parallel the solutions obtained are the same as those produced by standard sequential execution. It is also shown that, in the absence of failure, the parallel proof procedure doesn't genérate any additional work (with respect to standard SLDresolution) while the actual execution time is reduced. The same results hold even if non-strictly independent goals are executed in parallel, provided a trivial rewriting of such goals is performed. In addition, and most importantly, treats the issue of compile-time generation of IAP by proposing conditions, to be written at compile-time, to efficiently check strict and non-strict goal independence at run-time and proving the sufficiency of such conditions. It is also shown how simpler conditions can be constructed if some information regarding the binding context of the goals to be executed in parallel is available to the compiler trough either local or program-level analysis. These results therefore provide a formal basis for the automatic compile-time generation of IAP. As a corollary of such results, the paper also proves that negative goals are always non-strictly independent, and that goals which share a first occurrence of an existential variable are never independent.

Constructs and evaluation strategies for intelligent speculative parallelism - armageddon revisited

This report addresses speculative parallelism (the assignment of spare processing resources to tasks which are not known to be strictly required for the successful completion of a computation) at the user and application level. At this level, the execution of a program is seen as a (dynamic) tree —a graph, in general. A solution for a problem is a traversal of this graph from the initial state to a node known to be the answer. Speculative parallelism then represents the assignment of resources to múltiple branches of this graph even if they are not positively known to be on the path to a solution. In highly non-deterministic programs the branching factor can be very high and a naive assignment will very soon use up all the resources. This report presents work assignment strategies other than the usual depth-first and breadth-first. Instead, best-first strategies are used. Since their definition is application-dependent, the application language contains primitives that allow the user (or application programmer) to a) indícate when intelligent OR-parallelism should be used; b) provide the functions that define "best," and c) indícate when to use them. An abstract architecture enables those primitives to perform the search in a "speculative" way, using several processors, synchronizing them, killing the siblings of the path leading to the answer, etc. The user is freed from worrying about these interactions. Several search strategies are proposed and their implementation issues are addressed. "Armageddon," a global pruning method, is introduced, together with both a software and a hardware implementation for it. The concepts exposed are applicable to áreas of Artificial Intelligence such as extensive expert systems, planning, game playing, and in general to large search problems. The proposed strategies, although showing promise, have not been evaluated by simulation or experimentation.

Efficient management of backtracking in and-parallelism

A backtracking algorithm for AND-Parallelism and its implementation at the Abstract Machine level are presented: first, a class of AND-Parallelism models based on goal independence is defined, and a generalized version of Restricted AND-Parallelism (RAP) introduced as characteristic of this class. A simple and efficient backtracking algorithm for R A P is then discussed. An implementation scheme is presented for this algorithm which offers minimum overhead, while retaining the performance and storage economy of sequent ial implementations and taking advantage of goal independence to avoid unnecessary backtracking ("restricted intelligent backtracking"). Finally, the implementation of backtracking in sequential and AND-Parallcl systems is explained through a number of examples.

The role of computational logic as a hinge paradigm among deduction, problem solving, programming, and parallelism

This paper presents some brief considerations on the role of Computational Logic in the construction of Artificial Intelligence systems and in programming in general. It does not address how the many problems in AI can be solved but, rather more modestly, tries to point out some advantages of Computational Logic as a tool for the AI scientist in his quest. It addresses the interaction between declarative and procedural views of programs (deduction and action), the impact of the intrinsic limitations of logic, the relationship with other apparently competing computational paradigms, and finally discusses implementation-related issues, such as the efficiency of current implementations and their capability for efficiently exploiting existing and future sequential and parallel hardware. The purpose of the discussion is in no way to present Computational Logic as the unique overall vehicle for the development of intelligent systems (in the firm belief that such a panacea is yet to be found) but rather to stress its strengths in providing reasonable solutions to several aspects of the task.

A simulation study on parallel backtracking with solution memoing for independent and-parallelism

Goal-level Independent and-parallelism (IAP) is exploited by scheduling for simultaneous execution two or more goals which will not interfere with each other at run time. This can be done safely even if such goals can produce multiple answers. The most successful IAP implementations to date have used recomputation of answers and sequentially ordered backtracking. While in principle simplifying the implementation, recomputation can be very inefficient if the granularity of the parallel goals is large enough and they produce several answers, while sequentially ordered backtracking limits parallelism. And, despite the expected simplification, the implementation of the classic schemes has proved to involve complex engineering, with the consequent difficulty for system maintenance and expansion, and still frequently run into the well-known trapped goal and garbage slot problems. This work presents ideas about an alternative parallel backtracking model for IAP and a simulation studio. The model features parallel out-of-order backtracking and relies on answer memoization to reuse and combine answers. Whenever a parallel goal backtracks, its siblings also perform backtracking, but after storing the bindings generated by previous answers. The bindings are then reinstalled when combining answers. In order not to unnecessarily penalize forward execution, non-speculative and-parallel goals which have not been executed yet take precedence over sibling goals which could be backtracked over. Using a simulator, we show that this approach can bring significant performance advantages over classical approaches.