Awesome
3DGS.cpp
3DGS.cpp is a high performance, cross-platform implementation of Gaussian Splatting using the Vulkan API and compute pipelines.
Why Vulkan? We want to democratize the access to high-performance point-based radiance fields. Existing implementations of Gaussian Splatting are often limited to CUDA, which only runs on NVIDIA GPUs, and OpenGL is deprecated on Apple platforms. Additionally, Vulkan's compute capabilities are the closest to CUDA's with support for warp-level primitives (subgroups).
Downloads
- Windows
- macOS (arm64)
- Linux
- iOS, iPad OS, visionOS (as an iPad app) on TestFlight
Command Line Viewer
./3dgs_cpp_viewer {OPTIONS} [scene]
Vulkan Splatting
OPTIONS:
-h, --help Display this help menu
--validation Enable Vulkan validation layers
-v, --verbose Enable verbose logging
-d[physical-device],
--device=[physical-device] Select physical device by index
-i, --immediate-swapchain Set swapchain mode to immediate
(VK_PRESENT_MODE_IMMEDIATE_KHR)
-w[width], --width=[width] Set window width
-h[height], --height=[height] Set window height
--no-gui Disable GUI
scene Path to scene fil
Building
Linux
3DGS.cpp requires the following dependencies:
Vulkan headers, Vulkan validation layers, glslangValidator, glfw, glm
The easiest way to install the first three is through the LunarG Vulkan SDK.
Alternatively, you can install the corresponding packages from your distro. For Ubuntu, the packages to install
are vulkan-headers, vulkan-validationlayers, glslang-dev, libglfw3-dev, libglm-dev
.
git clone https://github.com/shg8/3DGS.cpp/
cd 3DGS.cpp
mkdir build
cmake -DCMAKE_BUILD_TYPE=Release -S ./ -B ./build
cmake --build ./build -j4
The command line viewer will be under build/apps/viewer/
.
Windows
After installing Vulkan SDK, set the VULKAN_SDK
environmental variable to the install path. Alternatively,
pass -DVULKAN_SDK=\INSTALL\LOCATION\OF\YOUR\SDK
to CMake when configuring.
macOS
After installing the Vulkan SDK, please proceed with CMake configuration and build steps.
iOS, iPadOS, visionOS
- Make sure that the Vulkan SDK is installed
- Run the
xcode_setup
CMake target - Download the prebuilt MoltenVK libraries or build it yourself. Unzip the downloaded artifact from a MoltenVK workflow. If you're not sure which one you want to use, pick the one from the latest tag
- Place the MoltenVK folder that contains the
dynamic
,include
, andstatic
folders underapps/apple
- Set your development team id in
project.xcconfig
- Profit
TODO
- Better controls and GUI on GLFW
- Apps for iOS and visionOS
- Fully immersive app on visionOS using the Compositor Service framework
- OpenXR support
- App for Android and Qualcomm Spaces
- Implement SOTA parallel radix sort
- Use Vulkan subgroups to batch Gaussian retrievals at the warp level
- Training
Please feel free to open an issue if you have any feature suggestions or are interested in contributing.
Benchmarking
There is room for performance improvements. See benchmarks on a 4090 at 3200x1400.
Contributing
If you are interested in integrating your Gaussian Splatting variant, please open an issue or a pull request. 3DGS.cpp's shaders follow the procedures outlined in the original paper, so it should be relatively easy to port your CUDA code. With cross-platform support, it's a great way to expand the reach and adoption of your research. If there are any questions, feel free to send me an email.
License
The main project is licensed under LGPL.
This project uses several third-party libraries. Here is a list of these libraries along with their respective licenses:
- GLM: MIT License.
- args.hxx: MIT License.
- spdlog: MIT License.
- ImGUI: MIT License.
- Vulkan Memory Allocator: MIT License.
- VkRadixSort: MIT License.
- implot: MIT License.
- glfw: zlib/libpng license.
- libenvpp: Apache License 2.0.