fatlipp / cuda-tree

CUDA-based Tree builder

CUDA based Tree builder (QuadTree, Octree)

Octree	QuadTree

Functionality

• CUDA tree builder
• CPU kNN, Radius Search
• 2d, 3d visualization (controls: WSAD R F, Mouse). Additional functionality for QuadTree - click on a Tree for kNN and RadiusSearch
• Random Points Generator
• Configuration file
• An obj file loader

Some implementation details: Medium

---

Build and Run

* mkdir build && cd build
* cmake .. `[-DCUDA_ARCH=SET_YOUR_ARCH]`
* make -j8
* `./cuda_tree_app_TYPE ./app/config/AppConfig.json`
* memory check: `compute-sanitizer --tool memcheck  ./cuda_tree_app_TYPE ./app/config/AppConfig.json`

• cuda_tree_app_TYPE:
  • cuda_tree_app_rng - to use random points
  • cuda_tree_app_obj - to read .obj file from model_path (only vertices)

Config

• AppConfig.json - contains general information (points count, pathes to Tree and Render config files)
• TreeConfig.json:
  • type - 'Quad' or 'Oct'
  • size - width, height, depth (if Octree)
  • threadsPerBlock - min 128 for QuadTree and 256 for Octree (1 warp for each leaf)
• RenderConfig.json - all about rendering

Requirements

• CUDA 11.8+
• C++ 17+
• OpenGL, GLFW
• nlohmann, tinyobjloader (included)

Limitations

• Tested on up to 150 mln points
• Max stable depth: QuadTree = 7, Octree = 6
• I recommend to disable rendering (AppConfig.json) if more than 50 mln points are used

Further development

1. CUDA based kNN, Radius Search
2. Refactoring, CPU, GPU optimization
3. OpenGL interop for faster visualization
4. Raycasting
5. KDTree

---

Disclaimer

• Tested on Linux: Ubuntu 22.02, CUDA: 12.2, GPU: 3060 (laptop)
• Approach based on Nvidia's QuadTree sample (cdpQuadtree)

cuda-tree's Issues

getLastCudaError() CUDA error : KernelError : (500) named symbol not found.

Hello,

epic library, thanks for providing such a great resource.

Am trying to run the code, and run into this error:

/tmp/cuda-tree/cmake-build-debug/app/cuda_tree_app_obj /tmp/cuda-tree/app/config/AppConfig.json
Load Model()
WARN: Material file [ 12140_Skull_v3_L2.mtl ] not found.
Failed to load material file(s). Use default material.

# of vertices  = 40062
# of normals   = 40062
# of texcoords = 42682
# of shapes    = 4
bmin = -15.487700, 0.050200, -9.920400
bmax = 13.964300, 27.575899, 9.919900
RunAppOctree(): 240048
Initialize() capacity: 240048
 duration: 0
Reset()
Build()
/tmp/cuda-tree/lib/core/octree/OctreeBuilderCuda.cu(57) : getLastCudaError() CUDA error : KernelError : (500) named symbol not found.

Process finished with exit code 1

I noticed #include <thirdparty/tinyobjloader/tiny_obj_loader.h> was missing so am using this version https://github.com/tinyobjloader/tinyobjloader/releases/tag/v2.0-rc1

Any ideas?

Recursive calling of kernel functions

Is it correct to recursively call the OctreeKernel function in the OctreeKernel function?