high performance computing on graphics processing units: hgpu.org

In this paper we study two parallelization strategies (loop-level parallelism and domain decomposition), and we investigate their impact in terms of performance and scalability on two different parallel architectures. As a test application, we use the Canny Edge Detector due to its wide range of parallelization opportunities, and its frequent use in computer vision applications. Different parallel implementations of the Canny Edge Detector are run on two distinct hardware platforms, namely a multi-core CPU, and a many-core GPU. Our experiments uncover design rules that, depending on a set of applications and platform factors (parallel features, data size, and architecture), indicate which parallelization scheme is more suitable.