GPU-Aware AMR on Octree-Based Grids
- 282 Downloads
Algorithms for refinement/coarsening of octree-based grids entirely on GPU are proposed. Corresponding CUDA/OpenMP implementations demonstrate good performance results which are comparable with p4est library execution times. Proposed algorithms permit to perform all dynamic AMR procedures on octree-based grids entirely in GPU as well as solver kernels without exploiting CPU resourses and pci-e bus for grid data transfers.
KeywordsAMR CUDA OpenMP Octree
This research was supported by the Grant No 17-71-30014 from the Russian Science Foundation.
- 1.Beckingsale, D., Gaudin, W., Herdman, A., Jarvis, S.: Resident block-structured adaptive mesh refinement on thousands of graphics processing units. In: Parallel Processing (ICPP), 2015 44th International Conference on Parallel Processing, pp. 61–70. IEEE (2015)Google Scholar
- 5.Brown, A.: Towards achieving GPU-native adaptive mesh refinement. Oxford e-Research Centre (2017). https://www.oerc.ox.ac.uk/sites/default/files/uploads/ProjectFiles/CUDA//Presentations/2017/A_Brown_8th_March.pdf
- 7.Xinsheng, Q., Randall, L., Michael, R.M.: Accelerating wave-propagation algorithms with adaptive mesh refinement using the Graphics Processing Unit (GPU) (2018). https://arxiv.org/abs/1808.02638
- 8.Burstedde, C., et al.: Extreme-scale AMR. In: Proceedings of the 2010 ACM/IEEE International Conference for High Performance Computing, Networking, Storage and Analysis, pp. 1–12. IEEE Computer Society (2010)Google Scholar