oneAPI Deep Neural Network Library (oneDNN)  1.95.0
Performance library for Deep Learning
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

Prepare the Build Space

mkdir -p build && cd build

Generate makefile

  • Native compilation:
    cmake .. <extra build options>
  • Cross compilation (AArch64 target on Intel 64 host)
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 and Install the Library

  • Build the library:
    make -j
  • Build the documentation:
    make doc
  • Install the library, headers, and documentation:
    make install

Windows

  • Generate a Microsoft Visual Studio solution:
    mkdir build && cd build && cmake -G "Visual Studio 15 2017 Win64" ..

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

cmake -G "Visual Studio 15 2017 Win64" -T "Intel C++ Compiler 19.0" ..
  • Build the library:
    cmake --build .

You can also use the msbuild command-line tool directly (here /p:Configuration selects the build configuration which can be different from the one specified in CMAKE_BUILD_TYPE, and /m enables a parallel build):

msbuild "oneDNN.sln" /p:Configuration=Release /m
  • Build the documentation
    cmake --build . --target DOC
  • Install the library, headers, and documentation:
    cmake --build . --target INSTALL

Building with DPCPP runtime

DPCPP runtime requires Intel oneAPI DPC++ Compiler. You can explicitly specify the path to Intel oneAPI DPC++ Compiler installation using -DDPCPPROOT CMake option.

C and C++ compilers need to be set to point to Intel oneAPI DPC++ Compilers.

Linux

# Set Intel oneAPI DPC++ Compiler environment
# <..>/setvars.sh
# Set C and C++ compilers
export CC=clang
export CXX=clang++
mkdir build
cd build
cmake -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP ..
cmake --build .

Windows

:: Set Intel oneAPI DPC++ Compiler environment
:: <..>\setvars.bat

Ninja

:: Set C and C++ compilers
set CC=clang
set CXX=clang++
mkdir build
cd build
cmake -G Ninja -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP ..
cmake --build .

Visual Studio

mkdir build
cd build
cmake -G "Visual Studio 16 2019" -A x64 -T "Intel(R) oneAPI DPC++ Compiler" -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP ..
msbuild "DNNL.sln"

Validate the Build

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

ctest