Sparse volumetric data, like 3D grids where most cells are empty, are used in simulations and graphics. Older systems could only read such grids on GPUs, not update them, limiting their use in dynamic processes.
This work introduces two advances for the NanoVDB data structure on GPUs: a set of operators that can quickly change the grid's structure (dilation, merging, pruning, refinement), and a new way to access data by streaming active cells in a linear order, avoiding slow tree searches.
The main finding is that these changes allow dynamic topological updates of grids with hundreds of millions of active cells in just a few milliseconds. This makes GPU-based sparse grids viable for simulations and geometry manipulation, removing a previous read-only restriction.