Build from Source

Download the Source Code

Download oneDNN source code or clone the repository.

git clone https://github.com/oneapi-src/oneDNN.git

Build the Library

Ensure that all software dependencies are in place and have at least the minimal supported version.

The oneDNN build system is based on CMake. Use

  • CMAKE_INSTALL_PREFIX to control the library installation location,

  • CMAKE_BUILD_TYPE to select between build type (Release, Debug, RelWithDebInfo).

  • CMAKE_PREFIX_PATH to specify directories to be searched for the dependencies located at non-standard locations.

See Build Options for detailed description of build-time configuration options.

Linux/macOS

GCC, Clang, or Intel oneAPI DPC++/C++ Compiler

  • Set up the environment for the compiler

  • Configure CMake and generate makefiles

    mkdir -p build
    cd build
    
    # Uncomment the following lines to build with Clang
    # export CC=clang
    # export CXX=clang++
    
    # Uncomment the following lines to build with Intel oneAPI DPC++/C++ Compiler
    # export CC=icx
    # export CXX=icpx
    cmake .. <extra build options>
  • Build the library

    make -j

Intel oneAPI DPC++/C++ Compiler with SYCL runtime

  • Set up the environment for Intel oneAPI DPC++/C++ Compiler using the setvars.sh script. The command below assumes you installed to the default folder. If you customized the installation folder, setvars.sh (Linux/macOS) is in your custom folder:

    source /opt/intel/oneapi/setvars.sh
  • Configure CMake and generate makefiles

    mkdir -p build
    cd build
    
    export CC=icx
    export CXX=icpx
    
    cmake .. \
              -DDNNL_CPU_RUNTIME=SYCL \
              -DDNNL_GPU_RUNTIME=SYCL \
              <extra build options>

Note

Open-source version of oneAPI DPC++ Compiler does not have the icx driver, use clang/clang++ instead. Open-source version of oneAPI DPC++ Compiler may not contain OpenCL runtime. In this case, you can use OPENCLROOT CMake option or environment variable of the same name to specify path to the OpenCL runtime if it is installed in a custom location.

  • Build the library

    make -j

GCC targeting AArch64 on x64 host

  • Set up the environment for the compiler

  • Configure CMake and generate makefiles

    export CC=aarch64-linux-gnu-gcc
    export CXX=aarch64-linux-gnu-g++
    cmake .. \
              -DCMAKE_SYSTEM_NAME=Linux \
              -DCMAKE_SYSTEM_PROCESSOR=AARCH64 \
              -DCMAKE_LIBRARY_PATH=/usr/aarch64-linux-gnu/lib \
              <extra build options>
  • Build the library

    make -j

GCC with Arm Compute Library (ACL) on AArch64 host

  • Set up the environment for the compiler

  • Configure CMake and generate makefiles

    export ACL_ROOT_DIR=<path/to/Compute Library>
    cmake .. \
              -DDNNL_AARCH64_USE_ACL=ON \
              <extra build options>
  • Build the library

    make -j

Windows

Microsoft Visual C++ Compiler or Intel oneAPI DPC++/C++ Compiler

  • Generate a Microsoft Visual Studio solution

    mkdir build
    cd build
    cmake -G "Visual Studio 16 2019" ..

    For the solution to use the Intel C++ Compiler, select the corresponding toolchain using the cmake -T switch:

    cmake -G "Visual Studio 16 2019" -T "Intel C++ Compiler 2022" ..
  • Build the library

    cmake --build . --config=Release

Warning

Intel oneAPI DPC++/C++ Compiler on Windows requires CMake v3.21 or later.

Note

CMake’s Microsoft Visual Studio generator does not respect CMAKE_BUILD_TYPE option. Solution file supports both Debug and Release builds with Debug being the default. You can choose specific build type with --config option.

Note

You can also open oneDNN.sln to build the project from the Microsoft Visual Studio IDE.

Intel oneAPI DPC++/C++ Compiler with SYCL Runtime

  • Set up the environment for Intel oneAPI DPC++/C++ Compiler using the setvars.bat script. The command below assumes you installed to the default folder. If you customized the installation folder, setvars.bat is in your custom folder:

    "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"

    or open Intel oneAPI Commmand Prompt instead.

  • Generate a Microsoft Visual Studio solution

    cmake .. -G "Visual Studio 16 2019" ^
             -T "Intel C++ Compiler 2022" ^
             -DDNNL_CPU_RUNTIME=SYCL ^
             -DDNNL_GPU_RUNTIME=SYCL ^
             <extra build options>
  • Alternatively, you can use CMake’s Ninja generator

    mkdir build
    cd build
    
    :: Set C and C++ compilers
    set CC=icx
    set CXX=icx
    cmake .. -G Ninja -DDNNL_CPU_RUNTIME=SYCL ^
                      -DDNNL_GPU_RUNTIME=SYCL ^
                      <extra build options>

Warning

Intel oneAPI DPC++/C++ Compiler on Windows requires CMake v3.21 or later.

Note

CMake’s Microsoft Visual Studio generator does not respect CMAKE_BUILD_TYPE option. Solution file supports both Debug and Release builds with Debug being the default. You can choose specific build type with --config option.

Note

Open-source version of oneAPI DPC++ Compiler does not have the icx driver, use clang/clang++ instead. Open-source version of oneAPI DPC++ Compiler may not contain OpenCL runtime. In this case, you can use OPENCLROOT CMake option or environment variable of the same name to specify path to the OpenCL runtime if it is installed in a custom location.

  • Build the library

    cmake --build .

Validate the Build

If the library is built for the host system, you can run unit tests using:

ctest

Build documentation

  • Install the requirements

    conda env create -f ../doc/environment.yml
    conda activate onednn-doc
  • Build the documentation

    cmake --build . --target doc

Install library

Install the library, headers, and documentation

cmake --build . --target install

The install directory is specified by the CMAKE_INSTALL_PREFIX cmake variable. When installing in the default directory, the above command needs to be run with administrative privileges using sudo on Linux/Mac or a command prompt run as administrator on Windows.