CUDA Mesh Voxelization

1. Introduction

CUDA Mesh Voxelization is a C++/CUDA library and CLI for 3D mesh voxelization, CSG operations, and Signed Distance Field (SDF) computation on the GPU.
The project is designed for benchmarking, visualization, and manipulation of volumetric 3D data, with a particular focus on performance and extensibility.

Features

• Mesh Voxelization: Converts 3D triangle meshes into solid or surface voxel grids.
• CSG Operations: Supports union, intersection, and difference operations between voxelized meshes.
• SDF Calculation: Computes the signed distance field using the Jump Flooding Algorithm (JFA).
• CLI Application: Command-line interface for batch processing and benchmarking.
• Benchmarking: Comparative analysis between sequential, OpenMP, and CUDA implementations.
• Visualization: Exports voxel grids for analysis and plotting.

---

2. Papers, Libraries, and Requirements

Reference Papers and Algorithms

• Fast Parallel Surface and Solid Voxelization on GPUs by Michael Schwarz and Hans-Peter Seidel (2010)
• Out-of-Core Construction of Sparse Voxel Octrees
• Jump Flooding in GPU with Applications to Voronoi Diagram and Distance Transform

Libraries Used

• cxxopts for CLI parsing
• CMake as the build system

Software Requirements

• C++: >= C++23
• CUDA Toolkit: >= 12.0
• CMake: >= 3.20
• Compiler: GCC/Clang/MSVC with C++23 support

---

3. Installation and Usage

Installation

git clone https://github.com/<your-username>/cuda-mesh-voxelization.git
cd cuda-mesh-voxelization
git submodule update --init --recursive
cmake -S . -B build
cmake --build build --config Release

Basic Usage

Voxelize a mesh and export the result:

./build/Release/apps/cli/cli assets/bunny.obj -n 128 -e -t 2 -p 1 -s

Main options:

• -n number of voxels per grid side
• -e enable export
• -t algorithm type (0 = sequential, 1 = naive, 2 = tiled, 3 = OpenMP)
• -p CSG operation type (1 = union, 2 = intersection, 3 = difference)
• -s compute SDF

For all available options:

./build/Release/apps/cli/cli -h

---

4. Results

Stanford Armadillo 64 voxels/side	Stanford Dragon 256 voxels/side	Stanford Bunny 128 voxels/side

Armadillo section (128 voxels), colored by SDF value.	CSG section (Bimba ∪ Bunny, 64 voxels), colored by SDF value.

---

5. Benchmark and Performance Analysis

Voxelization Performance Charts

Bunny 3,510 faces	Bunny 56,172 faces	Bunny 168,516 faces
Bunny 337,032 faces	Bunny 674,064 faces	Bunny 1,348,128 faces

JFA Performance Charts

Implementations comparison

Stages comparison