This is an OptiX/CUDA code sample showing how to quickly build ray-tracing acceleration structures for dynamic subdivision surfaces using the Cluster API introduced in OptiX 9.0. The scene is re-tessellated and a new acceleration structure is built on every frame. You can load your own scenes from OBJ files or JSON scene files.
It shares some common origins with the RTXMG DX12 code sample.
- Adaptive tessellation of Catmull-Clark limit surfaces using CUDA and OSD-Lite
- Animated scenes
- Optional displacement maps
- Fast BVH build with OptiX 9.0
- Real-time raytracing with OptiX. OpenGL is only used for image display, UI, and wireframe lines.
- DLSS-RR
IMPORTANT We do not include scene assets in this repo, due to their size. You will need to do one of two things to get them:
- Manually clone the RTXMG-Sample-Assets into a directory or symlink named "assets" inside the top of the project (the name is special and must be exactly as described).
- Or, set the CMake variable
ENABLE_ASSET_DOWNLOAD=1and reconfigure to automatically clone into "assets"
You can also download or create your own OBJ files, or use a simple json format for multi-object scenes with transforms. Please refer to the RTXMG documentation for scene file syntax.
Animation is supported via a series of OBJ files with vertex positions. See the Barbarian sample scene.
DLSS-RR uses AI to reconstruct high-quality images from noisy path-traced samples and motion vectors. The path tracing in this sample looks best with DLSS-RR enabled. The DLSS SDK is automatically fetched during CMake configuration, so no manual download is required.
If you prefer to use a local copy of the SDK, set DLSS_ROOT to its location and CMake will
use that instead of fetching. To disable DLSS entirely, set ENABLE_DLSS=OFF.
DLSS-RR exposes several quality modes in the UI: DLAA, Quality, Balanced, Performance, and Ultra Performance. See denoiserDlss.cu for implementation details.
- NVIDIA RTX GPU (Ampere or later with 8+ GB VRAM recommended)
- CMake version 3.25.3+ (3.27.7+ on Windows)
- Visual Studio 2022 or GCC 11.5+ (older versions may work but are untested)
- CUDA Toolkit, version 12+
-
Clone the repository
git clone https://github.com/NVIDIA/optix-subd.git -
Use CMake to configure the build and generate the project files or Makefiles. The first time will take a little longer, as OptiX headers, OSD-Lite, DLSS, GLFW, etc., are automatically fetched.
You can run CMake using the gui or via command line:
cmake CMakeLists.txt -B ./build -DCMAKE_BUILD_TYPE=Release -
Build the solution generated by CMake in the
./build/folder.For example, if using Visual Studio or vscode, open the generated solution
build/OptixSubd.sln, select the Release build type, and build all.You should prefer Release builds, since OptiX shaders built in Debug are significantly slower to execute.
-
(Optional) Clone the scene assets repo as described above and link or copy it to the
optix-subd/assetsdirectory. For example, to clone directly:git clone https://github.com/NVIDIA-RTX/RTXMG-Sample-Assets.git assets -
Select and run the
optixSubdproject. Binaries get built to thebuild/binfolder.
The first time you run it, the app will show a simple default subdivision surface and will populate the scene selection menu with files found in the assets directory. DLSS-RR is enabled by default; use the DLSS controls in the UI to toggle it or select a quality mode.
The camera is in Orbit mode by default (similar to Maya); use mouse buttons to orbit about a pivot point. Control-click on the scene to change the pivot. You can also switch to Roam mode in the UI and use W/S/A/D keys.
Hover the mouse over widgets to read tool-tips.
The window layout and camera mode are saved between runs to an imgui config file in the build dir. The Profiler window can always be toggled from the button next to the FPS counter.
-iloads a specific .scene.json or OBJ file-hprints a help message with other flags
Please report any issues directly through the GitHub issue tracker, or post general questions to the OptiX forum.
OpenSubdiv
OptiX:
DX:
Vulkan
- vk_animated_clusters :
VK_NV_cluster_acceleration_structurefor animated content. - vk_tessellated_clusters :
adaptive triangle tessellation and displacement using
VK_NV_cluster_acceleration_structure - vk_lod_clusters :
cluster-lod system and streaming using
VK_NV_cluster_acceleration_structure - vk_partitioned_tlas
update the TLAS of large dynamic scenes with
VK_NV_partitioned_acceleration_structure
Tools
- meshoptimizer.org - meshoptimizer provides ray tracing friendly clusterization through
meshopt_buildMeshletsSpatial. If you are interested in building a continuous level of detail system, have a look at its clusterlod.h demo, it shows the necessary steps
-
First a caveat: this started as a research project at NVIDIA, not as production code.
-
We have an old, modified version of optix-toolkit, inlined into our source tree.
-
Our memory management is basic: we grow geometry buffers on demand and never shrink them.
It is possible to create enough clusters and triangles to exceed GPU memory, and then the app will just crash. You may hit this for example by increasing the adaptive tess slider in the UI and then moving the camera very close to a surface. -
Since we don't constrain exactly how many triangles will be generated on any given frame, we read back intermediate values from the tessellation kernels and use them to resize our geometry buffers, and also to report stats in the API. This readback could be avoided if we were willing to allocate a large, fixed-size buffer on scene load.
-
We re-tessellate the scene every frame, even if nothing changed. This makes it easier to look at performance, but a real app may wish to track scene changes more carefully.