Publication: Explotación de aceleradores y hardware gráfico de forma amigable
Loading...
Official URL
Full text at PDC
Publication Date
2014
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Los algoritmos de flujo óptico tienen un alto coste computacional, pero las operaciones que conllevan también muestran un alto grado de paralelismo. Estas dos cualidades convierten a este tipo de algoritmos en buenos candidatos para mejorar su rendimiento en aceleradores y hardware gráfico. El problema que lleva consigo el uso de estos aceleradores para los programadores es la necesidad de conocer su arquitectura, además de lenguajes de programación específicos; siendo muy costosa la tarea de migrar el código para su utilización en este tipo de hardware. La aparición reciente de nuevos paradigmas de programación basados en directivas como OpenMP y OpenACC resuelve dicho problema, ya que con un pequeño porcentaje de modificaciones en el código original (entorno al 5-7%) los algoritmos pueden ser acelerados; pudiéndose considerar un buen balance el obtenido entre el esfuerzo de codificación y rendimiento computacional. En este proyecto se estudiarán los beneficios antes comentados en una implementación del algoritmo de flujo óptico Lucas&Kanade. Para ello se paralelizará con OpenMP sobre una CPU multicore y posteriormente en GPU's mediante OpenACC.
Optical flow algorithms require a great amount of computational resources, but the operations involving also show high degree of parallelism. Both features make this type of algorithms suitable candidates to improve their performance in accelerators and graphic hardware. For programmers the problem associated with the use of these accelerators is the necessity of knowing their architecture, as well as specific programming languages; being very hard the task of migrating code to its usage on this type of hardware. The recent emergence of new programming paradigms, based on directives as OpenMP and OpenACC solves the problem, because with a small percentage of changes in the original code (with around 5-7%) algorithms can be accelerated; being a good balance the one obtained between the coding effort and computational performance. In this project we will study the previously mentioned benefits in an implementation of the Lucas & Kanade optical flow algorithm. So, it will be parallelized with OpenMP on a multicore CPU, and then on GPU’s by Open ACC.
Optical flow algorithms require a great amount of computational resources, but the operations involving also show high degree of parallelism. Both features make this type of algorithms suitable candidates to improve their performance in accelerators and graphic hardware. For programmers the problem associated with the use of these accelerators is the necessity of knowing their architecture, as well as specific programming languages; being very hard the task of migrating code to its usage on this type of hardware. The recent emergence of new programming paradigms, based on directives as OpenMP and OpenACC solves the problem, because with a small percentage of changes in the original code (with around 5-7%) algorithms can be accelerated; being a good balance the one obtained between the coding effort and computational performance. In this project we will study the previously mentioned benefits in an implementation of the Lucas & Kanade optical flow algorithm. So, it will be parallelized with OpenMP on a multicore CPU, and then on GPU’s by Open ACC.
Description
Trabajo de Fin de Grado en Ingeniería de Computadores (Universidad Complutense, Facultad de Informática, curso 2013/2014)