Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time

OWL: A Node Graph "Wrapper" Library for OptiX 7

What is OWL?

OWL is a convenience/productivity-oriented library on top of OptiX 7.x, and aims at making it easier to write OptiX programs by taking some of the more arcane arts (like knowing what a Shader Binding Table is, and how to actually build it), and doing that for the user. For example, assuming the node graph (ie, the programs, geometries, and acceleration structures) have already been built, the shader binding table (SBT) can be built and properly populated by a single call owlBuildSBT(context).

In addition, OWL also allows for somewhat higher-level abstractions than native OptiX+CUDA for operations such as creating device buffers, uploading data, building shader programs and pipelines, building acceleration structures, etc.

Who is OWL designed/intended for?

OWL is particularly targetted at two groups of users: First, those that do want to use GPU Ray Tracing and RTX hardware acceleration, and that are comfortable with typical GPU concepts such as GPU memory vs device memory, ray tracing pipeline, shader programs, and some CUDA programming - but that are not "Ninja" OptiX/Vulkan/DirectX users, and might not be 100% sure about the most nitty-bitty grits of details on SBT data layout and order, or on just how exactly to do the BVH compaction, how exactly to deal with async launches or refitting, etc.

Second, it targets those that do know all these concepts, but would rather spent their time on the actual shader programs and functionality of the program, rather than on doing and all the low-level steps themselves; ie, those that are willing to trade a bit of low-level control (and maybe some tiny amount of performance) for higher developing productivity.

Simple Example

As an example of how easy it is to use OWL to build OptiX data strucutres, the following example code snippet takes a host-based triangle mesh and:

  • uploads the index and vertex buffer to the active GPU(s)
  • creates a triangle mesh geometry with a sample 'color' SBT entry
  • puts this mesh into a triangle bottom-level accel structure (BLAS)
  • builds that acceleration structure, including BVH compaction
  • creates an instance with an instance transform, and finally
  • builds and returns an instance acceleration structure over that.

Note how this little example will do these step: including data upload, set-up of build inputs, BVH construction, BVH compaction, and everything else that's required for this. Though still a relatively benign example, doing the same in low-level CUDA and OptiX code would result in significantly more code that the user would have to write, debug, and maintain.

/* simple sample of setting up a full geometry, BLAS and
   IAS for a simple single-triangle mesh model */
OWLGroup buildBlasAndIas(max3x4f             &instXfm,
                         std::vector<float3> &vtx,
                         std::vector<int3>   &idx,
						 float3               color)
   /* upload the buffers */
   OWLBuffer vtxBuffer
      = owlDeviceBufferCreate(ctx,OWL_FLOAT3,
   OWLBuffer idxBuffer
      = owlDeviceBufferCreate(ctx,OWL_INT3,

   /* create triangle mesh geometry */
   OWLGeom mesh = owlGeomCreate(ctx,myMeshGT);

   /* create and build triangle BLAS */
   OWLGroup blas = owlTrianglesGroupCreate(ctx,1,&mesh);

   /* create and build instance accel struct (IAS) */
   OWLGroup ias = owlInstanceGroupCreate(ctx,1,
       /* instantiated BLASes */&blas,
       /* instance IDs:       */nullptr,
	   /* instance transforms */&instXfm);
   return blas; // that's it!

Of course, even with OWL there's still much more that needs to be done for a full renderer: For example, in this code we assumed that a context (ctx) and a geometry type for this mesh (myMeshGT) have already been created; the user also still has to set up the programs, create frame buffer and launch data, build the programs (owlBuildPrograms()), the pipline (owlBuildPipeline()), and the SBT (owlBuildSBT(ctx)), etc.

What about Advanced Users?

As stated above, OWL explicitly aims for helping entry-level or casual RTX users get started, and get working productively with OptiX and RTX without having to first become an OptiX "Ninja".

However, that is not to mean that it is only useful for beginners. In fact, OWL currently supports lots of rather advanced features as well, including, for example:

  • multi-level instancing
  • accel structure refitting (compaction is always on)
  • multiple raygen programs and multiple ray types
  • motion blur, including instance motion blur
  • multi-GPU support, including proper handling of entities that might be different per GPU (such as buffers, traversables, and textures)
  • async launches
  • different buffer types including pinned and managed memory, and including buffers of buffers, buffers of traversables, and buffers of textures
  • textures, with different formats and filter modes
  • triangle mesh and user-defined geometry types (curves to be supported soon)
  • etc

In particular for advanced users, OWL is explicitly intended to allow advanced users to mix OWL code and data structures with other, manually written CUDA code if and whenever so desired. For example, OWL offers functions to easily query the CUDA device-addresses of buffers, OptixTraversableHandle's from groups, CUDA streams from launches, etc. As such, it is absolutely possible to mix OWL and CUDA code by, for example, having a multi-pass renderer in which CUDA does all the shading code and set-up of ray streams, and OWL doing the acceleration structure build and (RTX hardware-accelerated) tracing of these ray streams, even in multi-threaded and multi-GPU settings, with proper CUDA streams, etc (in fact, I do that in several of my own OWL applications).

Current State of Development

OWL was first publicly released early 2019, and has been used in several research/paper projects (see below). OWL initially targetted a much smaller scope of work - initially it was supposed to be only a "wrapper" around things like building acceleration structures (hence the name "OptiX Wrapper Library"), but the need for a higher abstraction level soon became evident, primarily due to the need to help users build and populate the SBT - which needs more "global" information than a single acceleration structure.

Despite these significant changes after the initial release, the current abstraction level and API have remained stable over roughly a year now, with only relatively minor additions such as buffers of buffers, refitting, textures, or motion blur. Some features will still need adding (e.g., curves, which got added to OptiX 7.1 but are not yet exposed in OWL); however, we consider the current release to be sufficiently stable to finally have given it the long-awaited "version 1.x".

Sample Use Cases

Some sample use projects/papers that recently used OWL:


Building OWL / Supported Platforms

General Requirements:

  • OptiX 7 SDK (version 7.0, 7.1, 7.2, 7.3, or 7.4; should work with either)
  • CUDA version 10 or 11
  • a C++11 capable compiler (regular gcc on CentOS, Ubuntu, or any other Linux should do; as should VS on Windows)
  • OpenGL

Per-OS Instructions:

  • Ubuntu 18, 19, and 20 (automatically tested on 18, mostly developed on 20)
    • Dependencies
      • cmake for building (sudo apt install cmake-curses-gui)
      • if you want to build the graphical examples: glfw (sudo apt-get install libglfw3-dev), or all the libraries to build it from included source code (sudo apt-get install x11-xserver-utils libxrandr-dev libxinerama-dev libxcb-xkb-dev libxcursor-dev libxcb-xinput-dev libxi-dev)
    • Build:
     mkdir build
     cd build
     cmake ..
  • CentOS 7:
    • Requires: sudo yum install cmake3
    • Build:
     mkdir build
     cd build
     cmake3 ..
    (mind to use cmake3, not cmake, using the wrong one will mess up the build directory)
  • Windows
    • Requires: Visual Studio (both 2017 and 2019 work), OptiX 7.0, cmake
    • Build: Use CMake-GUI to build Visual Studio project, then use VS to build
      • Specifics: source code path is ...Gitlab/owl, binaries ...Gitlab/owl/build, and after pushing the Configure button choose x64 for the optional platform.
      • You may need to Configure twice.
      • If you get "OptiX headers (optix.h and friends) not found." then define OptiX_INCLUDE manually in CMake-gui by setting it to C:/ProgramData/NVIDIA Corporation/OptiX SDK 7.0.0/include

Using OWL through CMake

Though you can of course use OWL without CMake, it is highly encouraged to use OWL as a git submodule, using CMake to configure and build this submodule. In particular, the suggested procedure is to first do a add_subdirectory with the owl submodules as such:

set(owl_dir ${PROJECT_SOURCE_DIR}/whereeverYourOWLSubmoduleIs)
add_subdirectory(${owl_dir} EXCLUDE_FROM_ALL)

(the EXCLUDE_FROM_ALL makes sure that your main project won't automatically build any owl samples or test cases unless you explicitly request so).

Once your project has called add_subdirectory on owl, it only has to link the owl::owl target in order to bring in all includes, linked libraries, etc. to fully use it. This might look like:

target_link_libraries(myOwlApp PRIVATE owl::owl)

If your sample uses the owlViewer base class and/or ptx embedding, add those as well:

target_link_libraries(myOwlApp PRIVATE myOwlApp-ptx owl::owl owl_viewer)

OptiX will need to be in a place that can be found by CMake. Point CMake at your OptiX directory by adding it to CMAKE_PREFIX_PATH (where it works on all platforms similar to how LD_LIBRARY_PATH resolves runtime linking on Linux). Note that CMAKE_PREFIX_PATH can be specified as an environment variable or as a CMake variable when you run CMake on your project.

Latest Progress/Revision History

Latest additions, not yet in any release

  • build fix to automatically build glfw on linux if no system-glfw is installed (kudos srogatch)

  • now handling empty user-geometries gracefully (fixes #147)

v1.1 - Switched to "modern cmake" technology (kudos lpisha, and jda)


  • bugfix: supporting optix 7.4 now.

  • renamed all CUDA_ macros to OWL_CUDA_ to avoid naming conflicts with other projects

v.1.1.5: bugfix: various buffer types didn't properly release all memory.


  • added new (optional) EMBEDDED_SYMBOL_NAMES argument to embed_ptx() to permit overriding the C symbol names used (default is ${CU_FILE_NAME}_ptx)
    • see tests/t01-many-spheres as an example of using this feature
  • added new (optional) PTX_TARGET argument to embed_ptx() to allow specifying the target name used to compile .cu files to PTX
    • see tests/t02-group-rebuilds as an example of using this feature

v.1.1.3: bugfix: fixed TBB includes for windows, when sued as submodule


  • added owlBufferClear(), at least for copyable data

  • added owlLaunch3D(), as requested in feature request issue #123

  • added new cmdline sample s10-launch3D that demonstrates both 3D launches, and how to do 'interop' between CUDA and OWL (using owl for rendering, and CUDA to convert framebuffer from float3 to RGBA8.

1.1.1: various fixes for the 'modern cmake' version, with a pretty big re-vamp of the entire build system. projects using owl as a submodule need to update how they use cmake accordingly; pls consult the samples for how to cleanly do that.

1.1.0: first release with louis pisha's 'modern cmake' version of the build system. Also includes several new samples, including the voxel renderer from the "ray tracing gems2" article.

v1.0.x - First "considered to be complete" version

1.0.4: fixed issue #68 - now compiles with optix 7.2, and newewst intel tbb

1.0.3: bugfix: no longer fatally failing when memadvise optimization didn't work

1.0.2: (finally) fixed long-standing bug in owlViewer that caused samples to crash when forcing OWL to run on a GPU that's different from the GPU that held the OpenGL graphics context for the viewer. Fixed.

*1.0.1: bugfix for missing owlSet4{}() functions

v0.9.x - Elimination of LL layer, and support for motion blur

v0.9.1: added support for more texture formats, access to the raw texture objects

v0.9.0: initial motion blur, and inital elimination of ll layer

  • Major code re-org: eliminated most of ll layer, includign most of ll::Device and virutally all of ll::DeviceGroup; in new design Device will only contain device context, and all object-specific stuff will live in the respective api::Group, api::Geom etc classes. Device-specific data for a given object is handled by this object itself.

  • initial support for motion blur on triangle meshes, by specifying two vertex arrays

  • initial support of motion blur on instances, by specifyign two sets of transforms.

  • new api fct owlMotionBlurEnable() to enable support for motion blur.

  • groups and geoms now have methods to compute their world-space bounding boxes, as required for instance motion blur. These will get called/evaluated/used only hwne motion blur is enabled.

  • moved some files from .cpp to .cu; to allow calling device kernels for computing bboxes.

  • initial support for optix 7.1; code will automatically detect version and compile to proper version where they differ.

  • owlLaunch2D now synchronous, async version explicitly owlLaunhc2DAsync

  • added owlMissProgSet(context,rayType,missProg) to set miss program for a given ray type

v0.8.x - Revamped build system, owl viewer, interative samples, and textures

v0.8.3: fixes, github issues, and naming

  • added OWL_CHAR and OWL_UCHAR types

  • renamed: owlParamsLaunch2D -> owlLaunch2D (added to owl_deprecated.h, and also axed lloLaunch function for cleanups)

  • renamed owlLaunchParams -> owlParams (create, set, vetvariable etc)


  • added owlXyzSetPointer()

  • variables can now have type OWL_BUFFER (not just BUFPTR), and will write a owl::device::Buffer type (with size, type, and data members)

  • device buffers can now be created over OWL_BUFFER and OWL_TEXTURE types

  • added int12-buffer-of-objects sample that shows/tests buffers of buffers, and buffers of textures (by creating a buffer of buffers of textures)

v0.8.2: double types, interactive sample

  • added OWL_DOUBLE type for variables, and al owl3d, setVariable, etc

  • added int11-rotationCubes sample that has NxMxK roating textured cubes

v0.8.1: first light of textures

  • added basic infrastructure for textures: OWLTexture type, OWLTextureFormat and OWLTextureFilterMode enums, OWL_TEXTURE variable types, owlVariableSetTexture, owlSetTexture(), etc. Textures currently only working for OWL_TEXEL_FORMAT_RGBA8, OWL_TEXEL_FORMAT_RGBA32F, OWL_TEXEL_FORMAT_R8, OWL_TEXEL_FORMAT_R32F and with OWL_TEXTURE_FILTER_LINEAR.

  • added int10-texturedTriangles that opens a window with a checkerboard-textured box.

v0.8.0: build system, glfw, and owl viewer

  • cmake build system now cleaner, and can use entire owl dir as add_subdirectory in other projects; main owl variables (OWL_INCLUDES, OWL_VIEWER_LIBRARIES, etc) now get exported to whoever includes, thus allowing includee to use same build flags, proper set of libraries an dincludes, etc.

  • existing (glut-based) viewerWidget got replaced with glfw based OWLViewer class. New class has cleaner setup code, and no longer requires installing glut binaries for windows build

  • build system picks up glfw where available, and otherwise builds glfw from source (full glfw source in samples/3rdParty)

  • owlViewer now handles frame buffer resize and display internally (no longer app's job to do that), and does so with cuda/gl resource sharing using managed mem frame buffer. App still writes render() method, but simply writes final pixels to viewer-handles frame buffer.

  • added first two interactive samples, using owlViewer

  • changed samples/advaned/optixCourse to use owlViewer - glut now completely eliminated from owl, and all samples use same viewer.

v0.7.x - Unifiction of ng and ll APIs into one single owl API

v0.7.4: major cleanups of "low-level" and "api" layer abstractions

  • 'll' and 'ng' layers mostly merged, at least from the API layer; led to significant reduction in duplicate code.

  • merged in PRs to enable TBB on windows, and to add cuda grphics resource buffer

v0.7.3: performance "guiding"

  • OWL no longer allows for tracing directly into BLASes ... this is highly discouraged in the driver, so better to just disallow it.

v0.7.3: bug hotfix

  • hotfix for bug introduced when auto-freeing of device memory, which accidentally freed instance BVH whiel still in use.

  • various windows fixes; in particular removing tbb by default (windows only)

v0.7.2: various feature extensions and bug fixes

  • lots of additional use in owl prime, m-owl-ana, distributed renderer, etc.

  • fixed memory leak in instance builder

  • added multi-device sample (s07-rtow-multiGPU), including samples/s07-rtow-multiGPU/ with brief notes on how to do multi-GPU in owl

  • added a lot more documentation to api functions (though much is still missing)

  • added ManagedMemory buffer type

  • added several sanity checks for group sizes, traversable graph depth, etc (checking w/ optix limits before trying to build)

  • added AnyHit shader support

v0.7.1: bugfix release.

  • added variable plumbing for missing unsigned and 64-bit types

  • fixed race condition in creating/using api handles that caused random crashes when setting launch params from different threads

v0.7.0: merged ng and ll APIs into one single API

  • now have a single owl library, a single header file, etc

  • eliminated all old ll/ samples (they only confused users)

  • sierpinski, rtow, and rtow-mixedGeom samples now in owl API

  • eliminated compaction in user geom and instance groups (doesn't help, anyway, and now have lower peak memory)

  • camke now defines (and all samples etc use) cmake variables for OWL_INCLUDES and OWL_LIBRARIES

  • fixes for TBB; TBB now gets detected more automatically, and used if found, with fallback to serial implementation if not

  • added array3D, various cleanups and exntensions to owl/common

  • various bugfixes and sanity/range checks throughout

v0.6.x - Buffer updates, launch params, first interactive example, ...

v0.6.1: cleanup/flesh-out of instance transform API

  • added creation of instance groups to ng/owl API

  • added ability to use different matrix memory layouts (row major, column major)

  • on ll layer: axed 'transform' parameter in InstanceGroupSetChild; matrices should now get passed to dedicated InstanceGroupSetTransform function.

  • fixed various bugs related to instancing

v0.6.0: Buffer updates, launch params, first interactive example, ...

  • Added OWL-port of siggraph course notes sample as a first 'advanced' sample that allows for interactive fly-throughs trhough a "real" model. See original course notes for instructions on how to use.

Process of adding this sample also required, among others, the following feature changes/additions

  • added support for resizing (owlBufferResize), destroying (owlBufferDestroy), and uploading data to (owlBufferUpload) buffers.

  • added concept of LaunchParams (with variables, similar to geoms), both for creating (owlLaunchParamsCreate) and launching with launch params (owlParamLaunch2D)

  • added support for asynchronous launches, where multiple launches can be in flight in parallel, using different streams. Wrote experimental "owl prime" project to debug, debug, and test this; project allow, for example, highly threaded cpu-side shading with owl-based, async GPU offload of the ray tracing (this project is not yet included).

  • worked on better interoperability between owl and CUDA-based host app; app can query buffer device pointers, add can add raw data (eg, cuda textures) to objects, app can query streams used for async launches (eg to add async cudamemcpys to that same stream, sync itself w/ owl, etc) and fixed includes to allow mixing cuda and optix code

  • added ability to create user-type variables (to pass, e.g., CUDA texture objects as parameters to meshes)

  • rewritten interface for getting and setting variables using macros, all object and variable types now supported

  • added support for more than one ray type (owlContextSetRayTypeCount)

  • clamped down on verbosity of the logging - most logs now visible only in debug mode

  • various minor bugfixes throughout the code

v0.5.x - First Public Release Cleanups

v0.5.4: First external windows-app

  • various changes to cmake scripts, library names, and in partciualr owl/common/viewerWidget to remove roadblocks for windows apps using that infrastructure

  • first external windows sandbox app (particle viewer) using owl/ng and owl/viewerWidget

v0.5.3: First serious node graph sample

  • ported ll05-rtow sample to node graph api

  • added bound program, user geom, user geom group, setprimcount and other missing functionality to node graph api

  • ng05-rtow ported, working, and passing tests

v0.5.2: First (partial) node graph sample

  • first working version of subset of node graph library (all that is required for 'firstTriangleMesh' example)

  • ng01-firstTriangleMesh working

  • significant renames and cleanups of owl/common (in particular, all 'gdt::' and 'gdt/' merged into owl::common and owl/common)

  • cleaned up owl/common/viewerWidget. Not used in owl itself (to avoid dependencies to glut etc), but now working successfully in first external test project

v0.5.1: First "c-api" version

  • added public c-linkage api (in include/owl/ll.h)

  • changed to build both static and dynamic/shared lib (tested working both linux and windows)

  • ported all samples to this new api

v0.5.0: First public release

  • first publicly accessible project on

  • major cleanups: "inlined" al the gdt submodule sources into owl/common to make owl external-dependency-fee. Feplaced gdt:: namespace with owl::common:: to match.

v0.4.x - Instances

v0.4.5: ll08-sierpinski now uses path tracing

v0.4.4: multi-level instancing

  • added new DeviceGroup::setMaxInstancingDepth that allows to set max instance depth and stack depth on pipeline.

  • added ll08-sierpinski example that allows for testing user-supplied number of instance levels with a sierpinski pyramid (Thx Nate!)

v0.4.3: new api fcts to set transforms and children for instance groups

  • added instanceGroupSetChild and instanceGroupSetTransform
  • extended ll07-groupOfGroups by two test cases that set transforms

v0.4.2: bugfix - all samples working in multi-device again

v0.4.1: example ll06-rtow-mixedGeometries.png working w/ manual sucessive traced into two different accels

v0.4.0: new way of building SBT now based on groups

  • api change: allocated geom groups now have their program size set in geomTypeCreate(), miss and raygen programs have it set in type rather than in sbt{raygen/miss}build (ie, program size now for all types set exactly once in type, then max size computed during sbt built)

  • can handle more than one group; for non-0 group has to query geomGroupGetSbtOffset() and pass that value to trace

  • new sbt structure no longer uses 'one entry per geom' (that unfortunately doesnt' work), but now builds sbt by iterating over all groups, and putting each groups' geom children in one block before putting next group. groups store the allcoated SBT offset for later use by instances

v0.3.x - User Geometries

v0.3.4: bugfix: adding bounds prog broke bounds buffer variant. fixed.

v0.3.4: first 'serious' example: RTOW-finalChapter on OWL

  • added s05-rtow example that runs Pete's "final chapter" example (iterative version) on top of OWL, with multi-device, different material, etc.

v0.3.3: major bugfix in bounds program for geoms w/ more than 128 prims.

v0.3.2: added two explicit examples for uesr geom - one with host-generation of bounds passed thrugh buffer, and one with bounds program

v0.3.1: First draft of device-side user prim bounds generation

  • added groupBuildPrimitiveBounds function that builds, for a user geom group, all the the primbounds required for the respective user geoms and prims in that group. The input for the user geoms' bounding bxo functions is generated using same callback mechanism as sbt writing.

v0.3.0: First example of user geometry working

  • can create user geometries through createUserGeom, and set type's isec program through setGeomTypeIntersect
  • supports passing of new userGeomSetBoundsBuffer fct to pass user geoms through a buffer
  • first example (8 sphere geometries, each with one sphere per geom) available as s03-userGeometry


v0.2.1: multiple triangle meshes working

  • multiple triangle meshes in same group debugged and working
  • added ll02-multipleTriangleGroups sample that generates 8 boxes

v0.2.0: first triangle mesh with trace and SBT data working

  • finalized llTest sample that ray traced image of one (tessellated) box


  • first version that does "some" sort of launch with mostly functional SBT


  • Ingo Wald
  • Nate Morrical
  • Eric Haines