high performance computing on graphics processing units: hgpu.org

Programs developed under the Compute Unified Device Architecture (CUDA) obtain the highest performance rate, when the exploitation of hardware resources on a Graphics Processing Unit (GPU) is maximized. In order to achieve this purpose, load balancing among threads and a high value of processor occupancy, i.e. the ratio of active threads, are indispensable. However, in certain applications, an optimally balanced implementation may limit the occupancy, due to a greater need of registers and shared memory. This is the case of the Fast Generalized Hough Transform (Fast GHT), an image processing technique for localizing an object within an image. In this work, we present two parallelization alternatives for the Fast GHT, one that optimizes the load balancing and another that maximizes the occupancy. We have compared them using a large amount of real images to test their strong and weak points and we have drawn several conclusions about under which conditions it is better to use one or another. We have also tackled several parallelization problems related to sparse data distribution, divergent execution paths and irregular memory access patterns in updating operations by proposing a set of generic techniques as compacting, sorting and memory storage replication.