Skip to content

Find neighbors gpu

DMI-Cavelan Aurélien requested to merge findNeighbors_GPU into gpu-hackathlon

New iterative findNeighbors with good offloading performance.

Inspired by: https://devblogs.nvidia.com/thinking-parallel-part-ii-tree-traversal-gpu/

The tree is transformed into a 1D array just for the GPU walk (LinearOctree.hpp). The code tries to minimize execution divergence by keeping CUDA threads in sync.

A CUDA version of the code is also included.

Marking as draft for the following reasons:

  • Hard-coded constants in findNeighbors.hpp should be moved elsewhere
  • Can we compute them?

Merge request reports