Demonstration of octree construction on CPUs and GPUs for 3D point clouds. This is a simplified version of the code found here: Cornestone Octree. Includes neighbor searches as an example application for the constructed octrees.
- Octrees represented based on Space-Filling-Curves (SFCs). Here, we use 3D-Hilbert curves.
- Performance portable octree construction on CPUs and GPUs based on common building blocks such as radix sort and prefix sums as described in [1].
- Portable neighbor search implementation
- Warp-level optimized neighbor search for the GPU
octree-miniapp
├── CMakeLists.txt
├── findneighbors.hpp - CPU/GPU portable neighbor search implementation
├── findneighbors_warps.cuh - warp-level optimized neighbor search implementation
├── neighbor_search.cu - neighbor search mini-app
├── octree.cpp - octree build mini-app for the CPU
├── octree.cu - octree build mini-app for the GPU
├── sfc - Hilbert SFC implementation
│ ├── bitops.hpp
│ ├── box.hpp
│ └── hilbert.hpp
├── tree - octree construction implementation
│ ├── csarray_gpu.cuh - extension of csarray.hpp to GPUs
│ ├── csarray.hpp - octree leaf-cell array construction (Sec. 4 of [1])
│ ├── octree_gpu.cuh - extension of octree.hpp to GPUs
│ └── octree.hpp - internal (fully-linked) octree construction on top of leaf-cells
│ (Sec. 5 of [1])
├── util - common boiler-plate code
│ ├── accel_switch.hpp
│ ├── annotation.hpp
│ ├── array.hpp
│ ├── cuda_utils.hpp
│ ├── random.hpp
│ ├── stl.hpp
│ ├── timing.cuh
│ └── tuple.hpp
└── warps - warp-level functions required for the warp-aware GPU neighbor search
├── gpu_config.cuh
└── warpscan.cuh
- Major dependencies: Thrust (ships with the CUDA Toolkit)
- HIP support for AMD devices: yes, after hipifying the sources.
mkdir build
cd build
cmake -DCMAKE_CUDA_ARCHITECTURES=<60,70 or 80> <THIS_REPO_GIT_SOURCE_DIR>
make -j
# running the mini-apps
./neighbor_search
./octree_cpu
./octree_gpuAll executables are single-source, therefore you may also compile them directly on the command line, e.g.:
nvcc neighbor_search.cu -O3 -std=c++17 --gpu-architecture=compute_<60,70 or 80> neighbor_search.cu[1] Cornerstone: Octree Construction Algorithms for Scalable Particle Simulations, PASC 2023, https://doi.org/10.1145/3592979.3593417
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent