Conditional parallelization of nonStrict independence. Procedures and assessmen

This paper presents a conditional parallelization process for and-parallelism based on the notion of non-strict independence, a more relaxed notion than the traditional of strict independence. By using this notion, a parallelism annotator can extract more parallelism from programs. On the other hand, the intrinsic complexity of non-strict independence poses new challenges to this task. We report here on the implementation we have accomplished of an annotator for non-strict independence, capable of producing both static and dynamic execution graphs. This implementation, along with the also implemented independence checker and their integration in our system, have resulted what is, to the best of our knowledge, the first parallelizing compiler based on nonstrict independence which produces dynamic execution graphs. The paper also presents a preliminary assessment of the implemented tools, comparing them with the existing ones for strict independence, which shows encouraging results.

Automatic compile-time parallelization of CLP programs by analysis and transformation to a concurrent constraint language.

The concept of independence has been recently generalized to the constraint logic programming (CLP) paradigm. Also, several abstract domains specifically designed for CLP languages, and whose information can be used to detect the generalized independence conditions, have been recently defined. As a result we are now in a position where automatic parallelization of CLP programs is feasible. In this paper we study the task of automatically parallelizing CLP programs based on such analyses, by transforming them to explicitly concurrent programs in our parallel CC platform (CIAO) as well as to AKL. We describe the analysis and transformation process, and study its efficiency, accuracy, and effectiveness in program parallelization. The information gathered by the analyzers is evaluated not only in terms of its accuracy, i.e. its ability to determine the actual dependencies among the program variables, but also of its effectiveness, measured in terms of code reduction in the resulting parallelized programs. Given that only a few abstract domains have been already defined for CLP, and that none of them were specifically designed for dependency detection, the aim of the evaluation is not only to asses the effectiveness of the available domains, but also to study what additional information it would be desirable to infer, and what domains would be appropriate for further improving the parallelization process.

Effectiveness of combined sharing and freeness analysis using abstract interpretation

This paper presents improved unification algorithms, an implementation, and an analysis of the effectiveness of an abstract interpreter based on the sharing + freeness domain presented in a previous paper, which was designed to accurately and concisely represent combined freeness and sharing information for program variables. We first briefly review this domain and the unification algorithms previously proposed. We then improve these algorithms and correct them to deal with some cases which were not well analyzed previously, illustrating the improvement with an example. We then present the implementation of the improved algorithm and evaluate its performance by comparing the effectiveness of the information inferred to that of other interpreters available to us for an application (program parallelization) that is common to all these interpreters. All these systems have been embedded in a real parallelizing compiler. Effectiveness of the analysis is measured in terms of actual final performance of the system: i.e. in terms of the actual speedups obtained. The results show good performance for the combined domain in that it improves the accuracy of both types of information and also in that the analyzer using the combined domain is more effective in the application than any of the other analyzers it is compared to.

Transformation�based implementation and optimization of programs exploiting the basic Andorra model.

The characteristics of CC and CLP systems are in principle very dierent However a recent trend towards convergence in the implementation techniques for these systems can be observed While CLP and Prolog systems have been incorporating capabilities to deal with userdened suspension and coroutining CC compilers have been trying to coalesce negrained tasks into coarsergrained sequential threads This convergence of techniques opens up the possibility of having a general purpose kernel language and abstract machine to serve as a compilation target for a variety of userlevel languages We propose a transformation technique directed towards such an objective In particular we report on techniques to support the Andorra computational model essentially emulating the AndorraI system via program transformation into a sequential language with delay primitives The system is automatic comprising an optional program analyzer and a basic transformer to the kernel language It turns out that a simple parallel CLP or Prolog system with dynamic scheduling is sucient as a kernel language for this purpose The preliminary results are quite encouraging performance of the resulting system is comparable to the current AndorraI implementation.

Herramienta de medida y visualización de prestaciones basada en eventos para especificaciones RVC-CAL

Este proyecto fin de grado presenta dos herramientas, Papify y Papify-Viewer, para medir y visualizar, respectivamente, las prestaciones a bajo nivel de especificaciones RVC-CAL basándose en eventos hardware. RVC-CAL es un lenguaje de flujo de datos estandarizado por MPEG y utilizado para definir herramientas relacionadas con la codificación de vídeo. La estructura de los programas descritos en RVC-CAL se basa en unidades funcionales llamadas actores, que a su vez se subdividen en funciones o procedimientos llamados acciones. ORCC (Open RVC-CAL Compiler) es un compilador de código abierto que utiliza como entrada descripciones RVC-CAL y genera a partir de ellas código fuente en un lenguaje dado, como por ejemplo C. Internamente, el compilador ORCC se divide en tres etapas distinguibles: front-end, middle-end y back-end. La implementación de Papify consiste en modificar la etapa del back-end del compilador, encargada de la generación de código, de modo tal que los actores, al ser traducidos a lenguaje C, queden instrumentados con PAPI (Performance Application Programing Interface), una herramienta utilizada como interfaz a los registros contadores de rendimiento (PMC) de los procesadores. Además, también se modifica el front-end para permitir identificar cierto tipo de anotaciones en las descripciones RVC-CAL, utilizadas para que el diseñador pueda indicar qué actores o acciones en particular se desean analizar. Los actores instrumentados, además de conservar su funcionalidad original, generan una serie de ficheros que contienen datos sobre los distintos eventos hardware que suceden a lo largo de su ejecución. Los eventos incluidos en estos ficheros son configurables dentro de las anotaciones previamente mencionadas. La segunda herramienta, Papify-Viewer, utiliza los datos generados por Papify y los procesa, obteniendo una representación visual de la información a dos niveles: por un lado, representa cronológicamente la ejecución de la aplicación, distinguiendo cada uno de los actores a lo largo de la misma. Por otro lado, genera estadísticas sobre la cantidad de eventos disparados por acción, actor o núcleo de ejecución y las representa mediante gráficos de barra. Ambas herramientas pueden ser utilizadas en conjunto para verificar el funcionamiento del programa, balancear la carga de los actores o la distribución por núcleos de los mismos, mejorar el rendimiento y diagnosticar problemas. ABSTRACT. This diploma project presents two tools, Papify and Papify-Viewer, used to measure and visualize the low level performance of RVC-CAL specifications based on hardware events. RVC-CAL is a dataflow language standardized by MPEG which is used to define video codec tools. The structure of the applications described in RVC-CAL is based on functional units called actors, which are in turn divided into smaller procedures called actions. ORCC (Open RVC-CAL Compiler) is an open-source compiler capable of transforming RVC-CAL descriptions into source code in a given language, such as C. Internally, the compiler is divided into three distinguishable stages: front-end, middle-end and back-end. Papify’s implementation consists of modifying the compiler’s back-end stage, which is responsible for generating the final source code, so that translated actors in C code are now instrumented with PAPI (Performance Application Programming Interface), a tool that provides an interface to the microprocessor’s performance monitoring counters (PMC). In addition, the front-end is also modified in such a way that allows identification of a certain type of annotations in the RVC-CAL descriptions, allowing the designer to set the actors or actions to be included in the measurement. Besides preserving their initial behavior, the instrumented actors will also generate a set of files containing data about the different events triggered throughout the program’s execution. The events included in these files can be configured inside the previously mentioned annotations. The second tool, Papify-Viewer, makes use of the files generated by Papify to process them and provide a visual representation of the information in two different ways: on one hand, a chronological representation of the application’s execution where each actor has its own timeline. On the other hand, statistical information is generated about the amount of triggered events per action, actor or core. Both tools can be used together to assert the normal functioning of the program, balance the load between actors or cores, improve performance and identify problems.

Parallel Computer Vision Algorithms for Graphics Processing Units

La evolución de los teléfonos móviles inteligentes, dotados de cámaras digitales, está provocando una creciente demanda de aplicaciones cada vez más complejas que necesitan algoritmos de visión artificial en tiempo real; puesto que el tamaño de las señales de vídeo no hace sino aumentar y en cambio el rendimiento de los procesadores de un solo núcleo se ha estancado, los nuevos algoritmos que se diseñen para visión artificial han de ser paralelos para poder ejecutarse en múltiples procesadores y ser computacionalmente escalables. Una de las clases de procesadores más interesantes en la actualidad se encuentra en las tarjetas gráficas (GPU), que son dispositivos que ofrecen un alto grado de paralelismo, un excelente rendimiento numérico y una creciente versatilidad, lo que los hace interesantes para llevar a cabo computación científica. En esta tesis se exploran dos aplicaciones de visión artificial que revisten una gran complejidad computacional y no pueden ser ejecutadas en tiempo real empleando procesadores tradicionales. En cambio, como se demuestra en esta tesis, la paralelización de las distintas subtareas y su implementación sobre una GPU arrojan los resultados deseados de ejecución con tasas de refresco interactivas. Asimismo, se propone una técnica para la evaluación rápida de funciones de complejidad arbitraria especialmente indicada para su uso en una GPU. En primer lugar se estudia la aplicación de técnicas de síntesis de imágenes virtuales a partir de únicamente dos cámaras lejanas y no paralelas—en contraste con la configuración habitual en TV 3D de cámaras cercanas y paralelas—con información de color y profundidad. Empleando filtros de mediana modificados para la elaboración de un mapa de profundidad virtual y proyecciones inversas, se comprueba que estas técnicas son adecuadas para una libre elección del punto de vista. Además, se demuestra que la codificación de la información de profundidad con respecto a un sistema de referencia global es sumamente perjudicial y debería ser evitada. Por otro lado se propone un sistema de detección de objetos móviles basado en técnicas de estimación de densidad con funciones locales. Este tipo de técnicas es muy adecuada para el modelado de escenas complejas con fondos multimodales, pero ha recibido poco uso debido a su gran complejidad computacional. El sistema propuesto, implementado en tiempo real sobre una GPU, incluye propuestas para la estimación dinámica de los anchos de banda de las funciones locales, actualización selectiva del modelo de fondo, actualización de la posición de las muestras de referencia del modelo de primer plano empleando un filtro de partículas multirregión y selección automática de regiones de interés para reducir el coste computacional. Los resultados, evaluados sobre diversas bases de datos y comparados con otros algoritmos del estado del arte, demuestran la gran versatilidad y calidad de la propuesta. Finalmente se propone un método para la aproximación de funciones arbitrarias empleando funciones continuas lineales a tramos, especialmente indicada para su implementación en una GPU mediante el uso de las unidades de filtraje de texturas, normalmente no utilizadas para cómputo numérico. La propuesta incluye un riguroso análisis matemático del error cometido en la aproximación en función del número de muestras empleadas, así como un método para la obtención de una partición cuasióptima del dominio de la función para minimizar el error. ABSTRACT The evolution of smartphones, all equipped with digital cameras, is driving a growing demand for ever more complex applications that need to rely on real-time computer vision algorithms. However, video signals are only increasing in size, whereas the performance of single-core processors has somewhat stagnated in the past few years. Consequently, new computer vision algorithms will need to be parallel to run on multiple processors and be computationally scalable. One of the most promising classes of processors nowadays can be found in graphics processing units (GPU). These are devices offering a high parallelism degree, excellent numerical performance and increasing versatility, which makes them interesting to run scientific computations. In this thesis, we explore two computer vision applications with a high computational complexity that precludes them from running in real time on traditional uniprocessors. However, we show that by parallelizing subtasks and implementing them on a GPU, both applications attain their goals of running at interactive frame rates. In addition, we propose a technique for fast evaluation of arbitrarily complex functions, specially designed for GPU implementation. First, we explore the application of depth-image–based rendering techniques to the unusual configuration of two convergent, wide baseline cameras, in contrast to the usual configuration used in 3D TV, which are narrow baseline, parallel cameras. By using a backward mapping approach with a depth inpainting scheme based on median filters, we show that these techniques are adequate for free viewpoint video applications. In addition, we show that referring depth information to a global reference system is ill-advised and should be avoided. Then, we propose a background subtraction system based on kernel density estimation techniques. These techniques are very adequate for modelling complex scenes featuring multimodal backgrounds, but have not been so popular due to their huge computational and memory complexity. The proposed system, implemented in real time on a GPU, features novel proposals for dynamic kernel bandwidth estimation for the background model, selective update of the background model, update of the position of reference samples of the foreground model using a multi-region particle filter, and automatic selection of regions of interest to reduce computational cost. The results, evaluated on several databases and compared to other state-of-the-art algorithms, demonstrate the high quality and versatility of our proposal. Finally, we propose a general method for the approximation of arbitrarily complex functions using continuous piecewise linear functions, specially formulated for GPU implementation by leveraging their texture filtering units, normally unused for numerical computation. Our proposal features a rigorous mathematical analysis of the approximation error in function of the number of samples, as well as a method to obtain a suboptimal partition of the domain of the function to minimize approximation error.