Build from source

System requirements

  • C++14 compiler:

    • Ubuntu 18.04+: GCC 5+, Clang 7+

    • macOS 10.14+: XCode 8.0+

    • Windows 10 (64-bit): Visual Studio 2019+

  • CMake: 3.18+

    • Ubuntu (18.04 / 20.04):

      • Install with apt-get: see official APT repository

      • Install with snap: sudo snap install cmake --classic

      • Install with pip (run inside a Python virtualenv): pip install cmake

    • macOS: Install with Homebrew: brew install cmake

    • Windows: Download from: CMake download page

  • CUDA (optional): Open3D supports GPU acceleration of an increasing number of operations through CUDA on Linux. We recommend using CUDA 11.0 for the best compatibility with recent GPUs and optional external dependencies such as Tensorflow or PyTorch.

    Please see the official documentation to install the CUDA toolkit from Nvidia.

Cloning Open3D

Make sure to use the --recursive flag when cloning Open3D.

git clone --recursive https://github.com/intel-isl/Open3D

# You can also update the submodule manually
git submodule update --init --recursive

Ubuntu/macOS

1. Install dependencies

# On Ubuntu
util/install_deps_ubuntu.sh

# On macOS: skip this step

2. Setup Python environments

Activate the python virtualenv or Conda environment. Check which python to ensure that it shows the desired Python executable. Alternatively, set the CMake flag -DPYTHON_EXECUTABLE=/path/to/python to specify the python executable.

If Python binding is not needed, you can turn it off by -DBUILD_PYTHON_MODULE=OFF.

3. Config

mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=<open3d_install_directory> ..

The CMAKE_INSTALL_PREFIX argument is optional and can be used to install Open3D to a user location. In the absence of this argument Open3D will be installed to a system location where sudo is required) For more options of the build, see Compilation options.

4. Build

# On Ubuntu
make -j$(nproc)

# On macOS
make -j$(sysctl -n hw.physicalcpu)

5. Install

To install Open3D C++ library:

make install

To link a C++ project against the Open3D C++ library, please refer to Link Open3D in C++ projects.

To install Open3D Python library, build one of the following options:

# Activate the virtualenv first
# Install pip package in the current python environment
make install-pip-package

# Create Python package in build/lib
make python-package

# Create pip wheel in build/lib
# This creates a .whl file that you can install manually.
make pip-package

# Create conda package in build/lib
# This creates a .tar.bz2 file that you can install manually.
make conda-package

Finally, verify the python installation with:

python -c "import open3d"

Windows

1. Setup Python binding environments

Most steps are the steps for Ubuntu: 2. Setup Python environments. Instead of which, check the Python path with where python.

2. Config

mkdir build
cd build

:: Specify the generator based on your Visual Studio version
:: If CMAKE_INSTALL_PREFIX is a system folder, admin access is needed for installation
cmake -G "Visual Studio 16 2019" -A x64 -DCMAKE_INSTALL_PREFIX="<open3d_install_directory>" ..

3. Build

cmake --build . --config Release --target ALL_BUILD

Alternatively, you can open the Open3D.sln project with Visual Studio and build the same target.

4. Install

To install Open3D C++ library, build the INSTALL target in terminal or in Visual Studio.

cmake --build . --config Release --target INSTALL

To link a C++ project against the Open3D C++ library, please refer to Link Open3D in C++ projects.

To install Open3D Python library, build the corresponding python installation targets in terminal or Visual Studio.

:: Activate the virtualenv first
:: Install pip package in the current python environment
cmake --build . --config Release --target install-pip-package

:: Create Python package in build/lib
cmake --build . --config Release --target python-package

:: Create pip package in build/lib
:: This creates a .whl file that you can install manually.
cmake --build . --config Release --target pip-package

:: Create conda package in build/lib
:: This creates a .tar.bz2 file that you can install manually.
cmake --build . --config Release --target conda-package

Finally, verify the Python installation with:

python -c "import open3d; print(open3d)"

Compilation options

OpenMP

We automatically detect if the C++ compiler supports OpenMP and compile Open3D with it if the compilation option WITH_OPENMP is ON. OpenMP can greatly accelerate computation on a multi-core CPU.

The default LLVM compiler on OS X does not support OpenMP. A workaround is to install a C++ compiler with OpenMP support, such as gcc, then use it to compile Open3D. For example, starting from a clean build directory, run

brew install gcc --without-multilib
cmake -DCMAKE_C_COMPILER=gcc-6 -DCMAKE_CXX_COMPILER=g++-6 ..
make -j

Note

This workaround has some compatibility issues with the source code of GLFW included in 3rdparty. Make sure Open3D is linked against GLFW installed on the OS.

Filament

The visualization module depends on the Filament rendering engine and, by default, Open3D uses a prebuilt version of it. You can also build Filament from source by setting BUILD_FILAMENT_FROM_SOURCE=ON.

Note

Whereas Open3D only requires a C++14 compiler, Filament needs a C++17 compiler and only supports Clang 7+, the most recent version of Xcode, and Visual Studio 2019, see their building instructions. Make sure to use one of these compiler if you build Open3D with BUILD_FILAMENT_FROM_SOURCE=ON.

ML Module

The ML module consists of primitives like operators and layers as well as high level code for models and pipelines. To build the operators and layers, set BUILD_PYTORCH_OPS=ON and/or BUILD_TENSORFLOW_OPS=ON. Don’t forget to also enable BUILD_CUDA_MODULE=ON for GPU support. To include the models and pipelines from Open3D-ML in the python package, set BUNDLE_OPEN3D_ML=ON and OPEN3D_ML_ROOT to the Open3D-ML repository. You can directly download Open3D-ML from GitHub during the build with OPEN3D_ML_ROOT=https://github.com/intel-isl/Open3D-ML.git.

Warning

Compiling PyTorch ops with CUDA 11 may have stability issues. See Open3D issue #3324 and PyTorch issue #52663 for more information on this problem.

We recommend to compile Pytorch from source with compile flags -Xcompiler -fno-gnu-unique or use the PyTorch wheels from Open3D. To reproduce the Open3D PyTorch wheels see the builder repository here.

The following example shows the command for building the ops with GPU support for all supported ML frameworks and bundling the high level Open3D-ML code.

# In the build directory
cmake -DBUILD_CUDA_MODULE=ON \
      -DBUILD_PYTORCH_OPS=ON \
      -DBUILD_TENSORFLOW_OPS=ON \
      -DBUNDLE_OPEN3D_ML=ON \
      -DOPEN3D_ML_ROOT=https://github.com/intel-isl/Open3D-ML.git \
      ..
# Install the python wheel with pip
make -j install-pip-package

Note

Importing Python libraries compiled with different CXX ABI may cause segfaults in regex. https://stackoverflow.com/q/51382355/1255535. By default, PyTorch and TensorFlow Python releases use the older CXX ABI; while when they are compiled from source, newer ABI is enabled by default.

When releasing Open3D as a Python package, we set -DGLIBCXX_USE_CXX11_ABI=OFF and compile all dependencies from source, in order to ensure compatibility with PyTorch and TensorFlow Python releases.

If you build PyTorch or TensorFlow from source or if you run into ABI compatibility issues with them, please:

  1. Check PyTorch and TensorFlow ABI with

    python -c "import torch; print(torch._C._GLIBCXX_USE_CXX11_ABI)"
    python -c "import tensorflow; print(tensorflow.__cxx11_abi_flag__)"
    
  2. Configure Open3D to compile all dependencies from source with the corresponding ABI version obtained from step 1.

After installation of the Python package, you can check Open3D ABI version with:

python -c "import open3d; print(open3d.pybind._GLIBCXX_USE_CXX11_ABI)"

To build Open3D with CUDA support, configure with:

cmake -DBUILD_CUDA_MODULE=ON -DCMAKE_INSTALL_PREFIX=<open3d_install_directory> ..

Please note that CUDA support is work in progress and experimental. For building Open3D with CUDA support, ensure that CUDA is properly installed by running following commands:

nvidia-smi      # Prints CUDA-enabled GPU information
nvcc -V         # Prints compiler version

If you see an output similar to command not found, you can install CUDA toolkit by following the official documentation.

WebRTC remote visualization

We provide pre-built binaries of the WebRTC library to build Open3D with remote visualization. Currently, Linux, macOS and Windows are supported for x86_64 architecture. If you wish to use a different version of WebRTC or build for a different configuration or platform, please see the official WebRTC documentation and the Open3D build scripts.

Linux and macOS

Please see the build script 3rdparty/webrtc/webrtc_build.sh. For Linux, you can also use the provided 3rdparty/webrtc/Dockerfile.webrtc for building.

Windows

We provide Windows MSVC static libraries built in Release and Debug mode built with the static Windows runtime. This corresponds to building with the /MT and /MTd options respectively. For the build procedure, please see .github/workflows/webrtc.yml. Other configrations are not supported.

Unit test

To build and run C++ unit tests:

cmake -DBUILD_UNIT_TESTS=ON ..
make -j$(nproc)
./bin/tests

To run Python unit tests:

# Activate virtualenv first
pip install pytest
make install-pip-package -j$(nproc)
pytest ../python/test