OpenACC-based GPU acceleration of an optical flow algorithm
Nelson Martin, Jorge Collado, Guillermo Botella, Carlos García, Manuel Prieto · 2015
Despite optical flow algorithms requiring a great amount of computational resources, they do show a high degree of data parallelism. Both features make this type of algorithms suitable candidates for having their performance improved on accelerators or Graphics Processing Units (GPUs). However, from a programmer's perspective, the use of accelerators entails a detailed knowledge of architecture as well as specific programming languages. The recent emergence of new programming paradigms, based on directives such as OpenMP and OpenACC mitigate this drawback because with a small percentage of source code modification a GPU executable version could be compiled. This paper addresses the fist OpenACC implementation of the well-known Lucas & Kanade algorithm. A successful tradeoff between coding effort (5-7% source lines hand-tuned) and speedup is observed, with speedups of upto 40× achieved.