2 unstable releases

0.2.0 Aug 27, 2022
0.1.0 Jan 17, 2022

#787 in Data structures

32 downloads per month

MIT/Apache

62KB
1.5K SLoC

grid-tree

Crates.io Docs.rs

Pixel quadtrees and voxel octrees.

Store any type in an OctreeI32, OctreeU32, QuadtreeI32, or QuadtreeU32, all of which are specific instances of the generic Tree. A Tree represents a map from (Level, Integer Coordinates) to T. Thus it is useful for storing pixel or voxel data with level-of-detail. The tree also requires that if a node slot is occupied (has data), then all ancestor slots are also filled.

Design Advantages

  • Since a Tree has its own internal allocators, any pointers are completely local to the data structure. In principle, this makes it easy to clone the tree for e.g. uploading to a GPU (although we haven't tried it for ourselves).
  • The level 0 allocator does not store pointers, only values. Pointer overhead at higher levels can be amortized using chunked data, i.e. [T; CHUNK_SIZE]. The alternative "pointerless" octrees take up less memory, but are also more complex to edit and traverse.
  • By using a hash map of root nodes, the addressable space is not limited by the height of the tree, and it is not necessary to "translate" the octree as it follows a focal point.
  • By having a very simple data layout, access using a NodePtr is simply an array lookup.
  • The NodeEntry and Tree::child_entry APIs allow for very simple code that fills entire trees with a single visitor closure.
  • By implementing VectorKey for a custom key type, the addressable range can be extended to coordinates of arbitrary precision.

Performance

This structure is optimized for iteration speed and spatial queries that benefit from a bounding volume hierarchy (like raycasting). Finding a single node by NodeKey starting from the root should be minimized as much as possible, so you might find it useful to cache NodePtrs or amortize the search with a full tree traversal. Memory usage is decent given the simplicity of the implementation, and the pointer overhead is easily amortized by using dense chunk values.

  • random access with NodeKey: O(depth)
  • random access with NodePtr: O(1)
  • iteration: O(nodes)
  • memory usage per node:
    • level 0: size_of::<T>() bytes
    • level N > 0: size_of::<T>() + CHILDREN * 4 bytes
    • where CHILDREN=4 for a quadtree and CHILDREN=8 for an octree

License: MIT OR Apache-2.0

Dependencies

~4MB
~102K SLoC