Abstract: The rapid changes in the present computer architectures throw questions concerning the future of the hardware-oriented programming based on the usage of heterogeneous parallel computing resources, consisting of nodes of interconnected, but separate CPUs and GPUs. However, it is believed that the sustainability of the programming efforts nowadays is to be searched in the fact that the principle source of the performance of the present GPUs as well as the emerging APUs is the hardware acceleration of vector operations. Therefore, this contribution delivers an assessment of a hardware-accelerated parallel implementation of a principal algorithm for a vertically integrated finite difference scheme for free surface flows including non-linear treatment of wetting and drying. The chosen approach is to expose the fine-grained parallelism of the scheme and execute the whole computational kernel of the code on a presently available streaming vector processor – a state-of-the-art GPU. The reached speedups compared to a single CPU core are in the order 20 or 30 for the double or single precision, respectively, which opens new possibilities for high resolution flood modelling. However, in order to reach this speedup level the coding must be radically adapted for the vectorised execution.
The preprint text can be obtained locally. Unpublished.
Reference: J.A. Jankowski (2011) Assessment of a parallel implementation of a two-dimensional scheme for free surface flows using a GPU. Unpublished.