Link Search Menu Expand Document

Build ONNX Runtime from source



Start: Baseline CPU


  • Checkout the source tree:
     git clone --recursive
     cd onnxruntime
  • Install cmake-3.13 or higher from

Build Instructions


Open Developer Command Prompt for Visual Studio version you are going to use. This will properly setup the environment including paths to your compiler, linker, utilities and header files.

.\build.bat --config RelWithDebInfo --build_shared_lib --parallel

The default Windows CMake Generator is Visual Studio 2017, but you can also use the newer Visual Studio 2019 by passing --cmake_generator "Visual Studio 16 2019" to .\build.bat

Linux/Mac OS X

./ --config RelWithDebInfo --build_shared_lib --parallel

By default, ORT is configured to be built for a minimum target Mac OS X version of 10.12. The shared library in the release Nuget(s) and the Python wheel may be installed on Mac OS X versions of 10.12+.


  • Please note that these instructions build the debug build, which may have performance tradeoffs
  • To build the version from each release (which include Windows, Linux, and Mac variants), see these .yml files for reference: CPU, GPU
  • The build script runs all unit tests by default (for native builds and skips tests by default for cross-compiled builds).
  • If you need to install protobuf 3.6.1 from source code (cmake/external/protobuf), please note:
    • CMake flag protobuf_BUILD_SHARED_LIBS must be turned OFF. After the installation, you should have the ‘protoc’ executable in your PATH. It is recommended to run ldconfig to make sure protobuf libraries are found.
    • If you installed your protobuf in a non standard location it would be helpful to set the following env var:export CMAKE_ARGS="-DONNX_CUSTOM_PROTOC_EXECUTABLE=full path to protoc" so the ONNX build can find it. Also run ldconfig <protobuf lib folder path> so the linker can find protobuf libraries.
  • If you’d like to install onnx from source code (cmake/external/onnx), use:
      export ONNX_ML=1
      python3 bdist_wheel
      pip3 install --upgrade dist/*.whl

Supported architectures and build environments


  x86_32 x86_64 ARM32v7 ARM64


OS Supports CPU Supports GPU Notes
Windows 10 YES YES VS2019 through the latest VS2015 are supported
Windows 10
Subsystem for Linux
Ubuntu 16.x YES YES Also supported on ARM32v7 (experimental)

GCC 4.x and below are not supported.

OS/Compiler Matrix:

OS/Compiler Supports VC Supports GCC Supports Clang
Windows 10 YES Not tested Not tested
Linux NO YES(gcc>=4.8) Not tested
Mac OS X NO Not tested YES (Minimum version required not ascertained)

Common Build Instructions

Description Command Additional details
Basic build build.bat (Windows)
./ (Linux)
Release build –config Release Release build. Other valid config values are RelWithDebInfo and Debug.
Use OpenMP –use_openmp OpenMP will parallelize some of the code for potential performance improvements. This is not recommended for running on single threads.
Build using parallel processing –parallel This is strongly recommended to speed up the build.
Build Shared Library –build_shared_lib  

APIs and Language Bindings

API Command Additional details
Python –build_wheel  
C# and C packages –build_csharp  
WindowsML –use_winml
WindowsML depends on DirectML and the OnnxRuntime shared library
Java –build_java Creates an onnxruntime4j.jar in the build directory, implies --build_shared_lib
Compiling the Java API requires gradle v6.1+ to be installed in addition to the usual requirements.
Node.js –build_nodejs Build Node.js binding. Implies --build_shared_lib

Build ONNX Runtime Server on Linux

Read more about ONNX Runtime Server here.

Build instructions are here

Execution Providers



  • Install CUDA and cuDNN
    • ONNX Runtime is built and tested with CUDA 10.1 and cuDNN 7.6 using the Visual Studio 2019 14.12 toolset (i.e. Visual Studio 2019 v16.5). ONNX Runtime can also be built with CUDA versions from 9.1 up to 10.1, and cuDNN versions from 7.1 up to 7.4.
    • The path to the CUDA installation must be provided via the CUDA_PATH environment variable, or the --cuda_home parameter
    • The path to the cuDNN installation (include the cuda folder in the path) must be provided via the cuDNN_PATH environment variable, or --cudnn_home parameter. The cuDNN path should contain bin, include and lib directories.
    • The path to the cuDNN bin directory must be added to the PATH environment variable so that cudnn64_7.dll is found.

Build Instructions

.\build.bat --use_cuda --cudnn_home <cudnn home path> --cuda_home <cuda home path>
./ --use_cuda --cudnn_home <cudnn home path> --cuda_home <cuda home path>

A Dockerfile is available here.


  • Depending on compatibility between the CUDA, cuDNN, and Visual Studio 2017 versions you are using, you may need to explicitly install an earlier version of the MSVC toolset.
  • CUDA 10.0 is known to work with toolsets from 14.11 up to 14.16 (Visual Studio 2017 15.9), and should continue to work with future Visual Studio versions
  • CUDA 9.2 is known to work with the 14.11 MSVC toolset (Visual Studio 15.3 and 15.4)
    • To install the 14.11 MSVC toolset, see this page.
    • To use the 14.11 toolset with a later version of Visual Studio 2017 you have two options:
      1. Setup the Visual Studio environment variables to point to the 14.11 toolset by running vcvarsall.bat, prior to running the build script. e.g. if you have VS2017 Enterprise, an x64 build would use the following command "C:\Program Files (x86)\Microsoft Visual Studio\2017\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" amd64 -vcvars_ver=14.11 For convenience, .\build.amd64.1411.bat will do this and can be used in the same way as .\build.bat. e.g. ` .\build.amd64.1411.bat –use_cuda`
    1. Alternatively, if you have CMake 3.13 or later you can specify the toolset version via the --msvc_toolset build script parameter. e.g. .\build.bat --msvc_toolset 14.11
  • If you have multiple versions of CUDA installed on a Windows machine and are building with Visual Studio, CMake will use the build files for the highest version of CUDA it finds in the BuildCustomization folder. e.g. C:\Program Files (x86)\Microsoft Visual Studio\2017\Enterprise\Common7\IDE\VC\VCTargets\BuildCustomizations. If you want to build with an earlier version, you must temporarily remove the ‘CUDA x.y.*’ files for later versions from this directory.


See more information on the TensorRT Execution Provider here.


  • Install CUDA and cuDNN
    • The TensorRT execution provider for ONNX Runtime is built and tested with CUDA 10.2 and cuDNN 7.6.5.
    • The path to the CUDA installation must be provided via the CUDA_PATH environment variable, or the --cuda_home parameter. The CUDA path should contain bin, include and lib directories.
    • The path to the CUDA bin directory must be added to the PATH environment variable so that nvcc is found.
    • The path to the cuDNN installation (path to folder that contains must be provided via the cuDNN_PATH environment variable, or --cudnn_home parameter.
  • Install TensorRT
    • The TensorRT execution provider for ONNX Runtime is built on TensorRT 7.x and is tested with TensorRT
    • The path to TensorRT installation must be provided via the --tensorrt_home parameter.

Build Instructions

.\build.bat --cudnn_home <path to cuDNN home> --cuda_home <path to CUDA home> --use_tensorrt --tensorrt_home <path to TensorRT home>
./ --cudnn_home <path to cuDNN e.g. /usr/lib/x86_64-linux-gnu/> --cuda_home <path to folder for CUDA e.g. /usr/local/cuda> --use_tensorrt --tensorrt_home <path to TensorRT home>

Dockerfile instructions are available here

NVIDIA Jetson TX1/TX2/Nano/Xavier

These instructions are for JetPack SDK 4.4.

  1. Clone the ONNX Runtime repo on the Jetson host

     git clone --recursive
  2. Specify the CUDA compiler, or add its location to the PATH.

    Cmake can’t automatically find the correct nvcc if it’s not in the PATH.

     export CUDACXX="/usr/local/cuda/bin/nvcc"


     export PATH="/usr/local/cuda/bin:${PATH}"
  3. Install the ONNX Runtime build dependencies on the Jetpack 4.4 host:

     sudo apt install -y --no-install-recommends \
       build-essential software-properties-common cmake libopenblas-dev \
       libpython3.6-dev python3-pip python3-dev
  4. Build the ONNX Runtime Python wheel:

     ./ --update --config Release --build --build_wheel \
     --use_cuda --cuda_home /usr/local/cuda --cudnn_home /usr/lib/aarch64-linux-gnu

    Note: You may add –use_tensorrt and –tensorrt_home options if you wish to use NVIDIA TensorRT (support is experimental), as well as any other options supported by script.


See more information on DNNL and MKL-ML here.

Build Instructions


DNNL: ./ --use_dnnl


Deprecation Notice

Deprecation Begins June 1, 2020
Removal Date December 1, 2020

Starting with the OpenVINO™ toolkit 2020.2 release, all of the features previously available through nGraph have been merged into the OpenVINO™ toolkit. As a result, all the features previously available through ONNX RT Execution Provider for nGraph have been merged with ONNX RT Execution Provider for OpenVINO™ toolkit.

Therefore, ONNX RT Execution Provider for nGraph will be deprecated starting June 1, 2020 and will be completely removed on December 1, 2020. Users are recommended to migrate to the ONNX RT Execution Provider for OpenVINO™ toolkit as the unified solution for all AI inferencing on Intel® hardware.

See more information on the nGraph Execution Provider here.

Build Instructions


.\build.bat --use_ngraph
./ --use_ngraph


See more information on the OpenVINO Execution Provider here.


  1. Install the Intel® Distribution of OpenVINOTM Toolkit Release 2020.3 for the appropriate OS and target hardware :

    Follow documentation for detailed instructions.

Although 2020.3 LTS is the recommended OpenVINO version, OpenVINO 2020.2 is also additionally supported.

  1. Configure the target hardware with specific follow on instructions:
    • To configure Intel® Processor Graphics(GPU) please follow these instructions: Windows, Linux
    • To configure Intel® MovidiusTM USB, please follow this getting started guide: Linux
    • To configure Intel® Vision Accelerator Design based on 8 MovidiusTM MyriadX VPUs, please follow this configuration guide: Windows, [Linux]( Follow steps 3 and 4 to complete the configuration.
    • To configure Intel® Vision Accelerator Design with an Intel® Arria® 10 FPGA, please follow this configuration guide: Linux
  2. Initialize the OpenVINO environment by running the setupvars script as shown below:
    • For Linux run:
       $ source <openvino_install_directory>/bin/
    • For Windows run:
       C:\ <openvino_install_directory>\bin\setupvars.bat
  3. Extra configuration step for Intel® Vision Accelerator Design based on 8 MovidiusTM MyriadX VPUs:
    • After setting the environment using setupvars script, follow these steps to change the default scheduler of VAD-M to Bypass:
      • Edit the hddl_service.config file from $HDDL_INSTALL_DIR/config/hddl_service.config and change the field “bypass_device_number” to 8.
      • Restart the hddl daemon for the changes to take effect.
      • Note that if OpenVINO was installed with root permissions, this file has to be changed with the same permissions.

Build Instructions

.\build.bat --config RelWithDebInfo --use_openvino <hardware_option>

Note: The default Windows CMake Generator is Visual Studio 2017, but you can also use the newer Visual Studio 2019 by passing --cmake_generator "Visual Studio 16 2019" to .\build.bat

./ --config RelWithDebInfo --use_openvino <hardware_option>

--use_openvino: Builds the OpenVINO Execution Provider in ONNX Runtime.

  • <hardware_option>: Specifies the default hardware target for building OpenVINO Execution Provider. This can be overriden dynamically at runtime with another option (refer to for more details on dynamic device selection). Below are the options for different Intel target devices.
Hardware Option Target Device  
CPU_FP32 Intel® CPUs  
GPU_FP32 Intel® Integrated Graphics  
GPU_FP16 Intel® Integrated Graphics with FP16 quantization of models  
 MYRIAD_FP16  Intel® MovidiusTM USB sticks  
 VAD-M_FP16  Intel® Vision Accelerator Design based on 8 MovidiusTM MyriadX VPUs  
 VAD-F_FP32  Intel® Vision Accelerator Design with an Intel® Arria® 10 FPGA  

For more information on OpenVINO Execution Provider's ONNX Layer support, Topology support, and Intel hardware enabled, please refer to the document in $onnxruntime_root/docs/execution_providers


See more information on the Nuphar Execution Provider here.


  • The Nuphar execution provider for ONNX Runtime is built and tested with LLVM 9.0.0. Because of TVM’s requirement when building with LLVM, you need to build LLVM from source. To build the debug flavor of ONNX Runtime, you need the debug build of LLVM.
    • Windows (Visual Studio 2017):
       REM download llvm source code 9.0.0 and unzip to \llvm\source\path, then install to \llvm\install\path
       cd \llvm\source\path
       mkdir build
       cd build
       cmake .. -G "Visual Studio 15 2017 Win64" -DLLVM_TARGETS_TO_BUILD=X86 -DLLVM_ENABLE_DIA_SDK=OFF
       msbuild llvm.sln /maxcpucount /p:Configuration=Release /p:Platform=x64
       cmake -DCMAKE_INSTALL_PREFIX=\llvm\install\path -DBUILD_TYPE=Release -P cmake_install.cmake

Note that following LLVM cmake patch is necessary to make the build work on Windows, Linux does not need to apply the patch. The patch is to fix the linking warning LNK4199 caused by this LLVM commit

diff --git "a/lib\\Support\\CMakeLists.txt" "b/lib\\Support\\CMakeLists.txt"
index 7dfa97c..6d99e71 100644
--- "a/lib\\Support\\CMakeLists.txt"
+++ "b/lib\\Support\\CMakeLists.txt"
@@ -38,12 +38,6 @@ elseif( CMAKE_HOST_UNIX )
 endif( MSVC OR MINGW )

-# Delay load shell32.dll if possible to speed up process startup.
-set (delayload_flags)
-if (MSVC)
-  set (delayload_flags delayimp -delayload:shell32.dll -delayload:ole32.dll)
 # Link Z3 if the user wants to build it.
@@ -187,7 +181,7 @@ add_llvm_library(LLVMSupport
-  LINK_LIBS ${system_libs} ${delayload_flags} ${Z3_LINK_FILES}
+  LINK_LIBS ${system_libs} ${Z3_LINK_FILES}

 set_property(TARGET LLVMSupport PROPERTY LLVM_SYSTEM_LIBS "${system_libs}")
  • Linux Download llvm source code 9.0.0 and unzip to /llvm/source/path, then install to /llvm/install/path
       cd /llvm/source/path
       mkdir build
       cd build
       make -j$(nproc)
       cmake -DCMAKE_INSTALL_PREFIX=/llvm/install/path -DBUILD_TYPE=Release -P cmake_install.cmake

Build Instructions

.\build.bat --use_tvm --use_llvm --llvm_path=\llvm\install\path\lib\cmake\llvm --use_mklml --use_nuphar --build_shared_lib --build_csharp --enable_pybind --config=Release
  • These instructions build the release flavor. The Debug build of LLVM would be needed to build with the Debug flavor of ONNX Runtime.
./ --use_tvm --use_llvm --llvm_path=/llvm/install/path/lib/cmake/llvm --use_mklml --use_nuphar --build_shared_lib --build_csharp --enable_pybind --config=Release

Dockerfile instructions are available here.


See more information on the DirectML execution provider here.


.\build.bat --use_dml


The DirectML execution provider supports building for both x64 and x86 architectures. DirectML is only supported on Windows.

ARM Compute Library

See more information on the ACL Execution Provider here.


  • Supported backend: i.MX8QM Armv8 CPUs
  • Supported BSP: i.MX8QM BSP
    • Install i.MX8QM BSP: source fsl-imx-xwayland-glibc-x86_64-fsl-image-qt5-aarch64-toolchain-4*.sh
  • Set up the build environment
    source /opt/fsl-imx-xwayland/4.*/environment-setup-aarch64-poky-linux
    alias cmake="/usr/bin/cmake -DCMAKE_TOOLCHAIN_FILE=$OECORE_NATIVE_SYSROOT/usr/share/cmake/OEToolchainConfig.cmake"
  • See Build ARM below for information on building for ARM devices

Build Instructions

  1. Configure ONNX Runtime with ACL support:
    cmake ../onnxruntime-arm-upstream/cmake -DONNX_CUSTOM_PROTOC_EXECUTABLE=/usr/bin/protoc -Donnxruntime_RUN_ONNX_TESTS=OFF -Donnxruntime_GENERATE_TEST_REPORTS=ON -Donnxruntime_DEV_MODE=ON -DPYTHON_EXECUTABLE=/usr/bin/python3 -Donnxruntime_USE_CUDA=OFF -Donnxruntime_USE_NSYNC=OFF -Donnxruntime_CUDNN_HOME= -Donnxruntime_USE_JEMALLOC=OFF -Donnxruntime_ENABLE_PYTHON=OFF -Donnxruntime_BUILD_CSHARP=OFF -Donnxruntime_BUILD_SHARED_LIB=ON -Donnxruntime_USE_EIGEN_FOR_BLAS=ON -Donnxruntime_USE_OPENBLAS=OFF -Donnxruntime_USE_ACL=ON -Donnxruntime_USE_DNNL=OFF -Donnxruntime_USE_MKLML=OFF -Donnxruntime_USE_OPENMP=ON -Donnxruntime_USE_TVM=OFF -Donnxruntime_USE_LLVM=OFF -Donnxruntime_ENABLE_MICROSOFT_INTERNAL=OFF -Donnxruntime_USE_BRAINSLICE=OFF -Donnxruntime_USE_NUPHAR=OFF -Donnxruntime_USE_EIGEN_THREADPOOL=OFF -Donnxruntime_BUILD_UNIT_TESTS=ON -DCMAKE_BUILD_TYPE=RelWithDebInfo

    The -Donnxruntime_USE_ACL=ON option will use, by default, the 19.05 version of the Arm Compute Library. To set the right version you can use: -Donnxruntime_USE_ACL_1902=ON, -Donnxruntime_USE_ACL_1905=ON or -Donnxruntime_USE_ACL_1908=ON;

  2. Build ONNX Runtime library, test and performance application:
    make -j 6
  3. Deploy ONNX runtime on the i.MX 8QM board

Build Instructions (Jetson Nano)

  1. Build ACL Library (skip if already built)
    cd ~
    git clone
    cd ComputeLibrary
    sudo apt install scons
    sudo apt install g++-arm-linux-gnueabihf
    scons -j8 arch=arm64-v8a  Werror=1 debug=0 asserts=0 neon=1 opencl=1 examples=1 build=native
  2. Set environment variables to set include directory and shared object library path.
    export CPATH=~/ComputeLibrary/include/:~/ComputeLibrary/
    export LD_LIBRARY_PATH=~/ComputeLibrary/build/
  3. Build onnxruntime with –use_acl flag
    ./ --use_acl


See more information on the ArmNN Execution Provider here.


  • Supported backend: i.MX8QM Armv8 CPUs
  • Supported BSP: i.MX8QM BSP
    • Install i.MX8QM BSP: source fsl-imx-xwayland-glibc-x86_64-fsl-image-qt5-aarch64-toolchain-4*.sh
  • Set up the build environment
    source /opt/fsl-imx-xwayland/4.*/environment-setup-aarch64-poky-linux
    alias cmake="/usr/bin/cmake -DCMAKE_TOOLCHAIN_FILE=$OECORE_NATIVE_SYSROOT/usr/share/cmake/OEToolchainConfig.cmake"
  • See Build ARM below for information on building for ARM devices

Build Instructions

./ --use_armnn

The Relu operator is set by default to use the CPU execution provider for better performance. To use the ArmNN implementation build with –armnn_relu flag

./ --use_armnn --armnn_relu

The Batch Normalization operator is set by default to use the CPU execution provider. To use the ArmNN implementation build with –armnn_bn flag

./ --use_armnn --armnn_bn


See more information on the RKNPU Execution Provider here.


  • Supported platform: RK1808 Linux
  • See Build ARM below for information on building for ARM devices
  • Use gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu instead of gcc-linaro-6.3.1-2017.05-x86_64_arm-linux-gnueabihf, and modify CMAKE_CXX_COMPILER & CMAKE_C_COMPILER in tool.cmake:
    set(CMAKE_CXX_COMPILER aarch64-linux-gnu-g++)
    set(CMAKE_C_COMPILER aarch64-linux-gnu-gcc)

Build Instructions

  1. Download rknpu_ddk to any directory.

  2. Build ONNX Runtime library and test:
     ./ --arm --use_rknpu --parallel --build_shared_lib --build_dir build_arm --config MinSizeRel --cmake_extra_defines RKNPU_DDK_PATH=<Path To rknpu_ddk> CMAKE_TOOLCHAIN_FILE=<Path To tool.cmake> ONNX_CUSTOM_PROTOC_EXECUTABLE=<Path To protoc>
  3. Deploy ONNX runtime and on the RK1808 board:


See more information on the Xilinx Vitis-AI execution provider here.

For instructions to setup the hardware environment: Hardware setup


./ --use_vitisai


The Vitis-AI execution provider is only supported on Linux.



Build Instructions

.\build.bat --use_openmp
./ --use_openmp



  • OpenBLAS
    • Windows: See build instructions here
    • Linux: Install the libopenblas-dev package sudo apt-get install libopenblas-dev

Build Instructions

.\build.bat --use_openblas
./ --use_openblas


OnnxRuntime supports build options for enabling debugging of intermediate tensor shapes and data.

Build Instructions

Set onnxruntime_DEBUG_NODE_INPUTS_OUTPUT to one of the values below.


./ --cmake_extra_defines onnxruntime_DEBUG_NODE_INPUTS_OUTPUTS=VALUE


.\build.bat --cmake_extra_defines onnxruntime_DEBUG_NODE_INPUTS_OUTPUTS=VALUE


  • 0: Disables this functionality if previously enabled; alternatively, delete CMakeCache.txt instead of setting this to 0
  • 1: Dump tensor input/output shapes for all nodes to stdout
  • 2: Dump tensor input/output shapes and output data for all nodes to stdout



Build Instructions

  • add --x86 argument when launching .\build.bat
  • Must be built on a x86 OS
  • add –x86 argument to


There are a few options for building for ARM.

Cross compiling for ARM with simulation (Linux/Windows)


This method rely on qemu user mode emulation. It allows you to compile using a desktop or cloud VM through instruction level simulation. You’ll run the build on x86 CPU and translate every ARM instruction to x86. This is much faster than compiling natively on a low-end ARM device and avoids out-of-memory issues that may be encountered. The resulting ONNX Runtime Python wheel (.whl) file is then deployed to an ARM device where it can be invoked in Python 3 scripts.

Here is an example for Raspberrypi3 and Raspbian. Note: this does not work for Raspberrypi 1 or Zero, and if your operating system is different from what the dockerfile uses, it also may not work.

The build process can take hours.

Cross compiling on Linux

Difficult, fast

This option is very fast and allows the package to be built in minutes, but is challenging to setup. If you have a large code base (e.g. you are adding a new execution provider to onnxruntime), this may be the only feasible method.

1. Get the corresponding toolchain.

TLDR; Go to, get “64-bit Armv8 Cortex-A, little-endian” and “Linux Targeted”, not “Bare-Metal Targeted”. Extract it to your build machine and add the bin folder to your $PATH env. Then skip this part.

You can use GCC or Clang. Both work, but instructions here are based on GCC.

In GCC terms:

  • “build” describes the type of system on which GCC is being configured and compiled
  • “host” describes the type of system on which GCC runs. “target” to describe the type of system for which GCC produce code When not cross compiling, usually “build” = “host” = “target”. When you do cross compile, usually “build” = “host” != “target”. For example, you may build GCC on x86_64, then run GCC on x86_64, then generate binaries that target aarch64. In this case,”build” = “host” = x86_64 Linux, target is aarch64 Linux.

You can either build GCC from source code by yourself, or get a prebuilt one from a vendor like Ubuntu, linaro. Choosing the same compiler version as your target operating system is best. If ths is not possible, choose the latest stable one and statically link to the GCC libs.

When you get the compiler, run aarch64-linux-gnu-gcc -v This should produce an output like below:

Using built-in specs.
Target: aarch64-linux-gnu
Configured with: ../gcc-9.2.1-20190827/configure --bindir=/usr/bin --build=x86_64-redhat-linux-gnu --datadir=/usr/share --disable-decimal-float --disable-dependency-tracking --disable-gold --disable-libgcj --disable-libgomp --disable-libmpx --disable-libquadmath --disable-libssp --disable-libunwind-exceptions --disable-shared --disable-silent-rules --disable-sjlj-exceptions --disable-threads --with-ld=/usr/bin/aarch64-linux-gnu-ld --enable-__cxa_atexit --enable-checking=release --enable-gnu-unique-object --enable-initfini-array --enable-languages=c,c++ --enable-linker-build-id --enable-lto --enable-nls --enable-obsolete --enable-plugin --enable-targets=all --exec-prefix=/usr --host=x86_64-redhat-linux-gnu --includedir=/usr/include --infodir=/usr/share/info --libexecdir=/usr/libexec --localstatedir=/var --mandir=/usr/share/man --prefix=/usr --program-prefix=aarch64-linux-gnu- --sbindir=/usr/sbin --sharedstatedir=/var/lib --sysconfdir=/etc --target=aarch64-linux-gnu --with-bugurl= --with-gcc-major-version-only --with-isl --with-newlib --with-plugin-ld=/usr/bin/aarch64-linux-gnu-ld --with-sysroot=/usr/aarch64-linux-gnu/sys-root --with-system-libunwind --with-system-zlib --without-headers --enable-gnu-indirect-function --with-linker-hash-style=gnu
Thread model: single
gcc version 9.2.1 20190827 (Red Hat Cross 9.2.1-3) (GCC)

Check the value of --build, --host, --target, and if it has special args like --with-arch=armv8-a, --with-arch=armv6, --with-tune=arm1176jz-s, --with-fpu=vfp, --with-float=hard.

You must also know what kind of flags your target hardware need, which can differ greatly. For example, if you just get the normal ARMv7 compiler and use it for Raspberry Pi V1 directly, it won’t work because Raspberry Pi only has ARMv6. Generally every hardware vendor will provide a toolchain; check how that one was built.

A target env is identifed by:

  • Arch: x86_32, x86_64, armv6,armv7,arvm7l,aarch64,…
  • OS: bare-metal or linux.
  • Libc: gnu libc/ulibc/musl/…
  • ABI: ARM has mutilple ABIs like eabi, eabihf…

You can get all these information from the previous output, please be sure they are all correct.

2. Get a pre-compiled protoc:

Get this from and unzip after downloading. The version must match the one onnxruntime is using. Currently we are using 3.11.2.

3. (Optional) Setup sysroot to enable python extension. Skip if not using Python.

Dump the root file system of the target operating system to your build machine. We’ll call that folder “sysroot” and use it for build onnxruntime python extension. Before doing that, you should install python3 dev package(which contains the C header files) and numpy python package on the target machine first.

Below are some examples.

If the target OS is raspbian-buster, please download the RAW image from their website then run:

$ fdisk -l 2020-02-13-raspbian-buster.img

Disk 2020-02-13-raspbian-buster.img: 3.54 GiB, 3787456512 bytes, 7397376 sectors Units: sectors of 1 * 512 = 512 bytes Sector size (logical/physical): 512 bytes / 512 bytes I/O size (minimum/optimal): 512 bytes / 512 bytes Disklabel type: dos Disk identifier: 0xea7d04d6

Device Boot Start End Sectors Size Id Type
2020-02-13-raspbian-buster.img1   8192 532479 524288 256M c W95 FAT32 (LBA)
2020-02-13-raspbian-buster.img2   532480 7397375 6864896 3.3G 83 Linux

You’ll find the the root partition starts at the 532480 sector, which is 532480 * 512=272629760 bytes from the beginning.

Then run:

$ mkdir /mnt/pi
$ mount -r -o loop,offset=272629760 2020-02-13-raspbian-buster.img /mnt/pi

You’ll see all raspbian files at /mnt/pi. However you can’t use it yet. Because some of the symlinks are broken, you must fix them first. In /mnt/pi, run

$ find . -type l -exec realpath  {} \; |grep 'No such file'

It will show which are broken. Then you can fix them by running:

$ mkdir /mnt/pi2
$ cd /mnt/pi2
$ sudo tar -C /mnt/pi -cf - . | sudo tar --transform 'flags=s;s,^/,/mnt/pi2/,' -xf -

Then /mnt/pi2 is the sysroot folder you’ll use in the next step.

If the target OS is Ubuntu, you can get an image from But that image is in qcow2 format. Please convert it before run fdisk and mount.

qemu-img convert -p -O raw ubuntu-18.04-server-cloudimg-arm64.img ubuntu.raw

The remaining part is similar to raspbian.

If the target OS is manylinux2014, you can get it by: Install qemu-user-static from apt or dnf. Then run the docker Ubuntu:

docker run -v /usr/bin/qemu-aarch64-static:/usr/bin/qemu-aarch64-static -it --rm /bin/bash

The “-v /usr/bin/qemu-aarch64-static:/usr/bin/qemu-aarch64-static” arg is not needed on Fedora.

Then, inside the docker, run

cd /opt/python
./cp35-cp35m/bin/python -m pip install numpy==1.16.6
./cp36-cp36m/bin/python -m pip install numpy==1.16.6
./cp37-cp37m/bin/python -m pip install numpy==1.16.6
./cp38-cp38/bin/python -m pip install numpy==1.16.6

These commands will take a few hours because numpy doesn’t have a prebuilt package yet. When completed, open a second window and run

docker ps

From the output:

CONTAINER ID        IMAGE                                COMMAND             CREATED             STATUS              PORTS               NAMES
5a796e98db05   "/bin/bash"         3 minutes ago       Up 3 minutes                            affectionate_cannon

You’ll see the docker instance id is: 5a796e98db05. Use the following command to export the root filesystem as the sysroot for future use.

docker export 5a796e98db05 -o manylinux2014_aarch64.tar
4. Generate CMake toolchain file

Save the following content as tool.cmake

    SET(CMAKE_C_COMPILER aarch64-linux-gnu-gcc)
    SET(CMAKE_CXX_COMPILER aarch64-linux-gnu-g++)

If you don’t have a sysroot, you can delete the last line.

5. Run CMake and make

Append -DONNX_CUSTOM_PROTOC_EXECUTABLE=/path/to/protoc -DCMAKE_TOOLCHAIN_FILE=path/to/tool.cmake to your cmake args, run cmake and make to build it. If you want to build Python package as well, you can use cmake args like:

-Donnxruntime_GCC_STATIC_CPP_RUNTIME=ON -DCMAKE_BUILD_TYPE=Release -Dprotobuf_WITH_ZLIB=OFF -DCMAKE_TOOLCHAIN_FILE=path/to/tool.cmake -Donnxruntime_ENABLE_PYTHON=ON -DPYTHON_EXECUTABLE=/mnt/pi/usr/bin/python3 -Donnxruntime_BUILD_SHARED_LIB=OFF -Donnxruntime_DEV_MODE=OFF -DONNX_CUSTOM_PROTOC_EXECUTABLE=/path/to/protoc "-DPYTHON_INCLUDE_DIR=/mnt/pi/usr/include;/mnt/pi/usr/include/python3.7m" -DNUMPY_INCLUDE_DIR=/mnt/pi/folder/to/numpy/headers

After running cmake, run

$ make
6. (Optional) Build Python package

Copy the file from the source folder to the build folder and run

python3 bdist_wheel -p linux_aarch64

If targeting manylinux, unfortunately their tools do not work in the cross-compiling scenario. Run it in a docker like:

docker run  -v /usr/bin/qemu-aarch64-static:/usr/bin/qemu-aarch64-static -v `pwd`:/tmp/a -w /tmp/a --rm /opt/python/cp37-cp37m/bin/python3 bdist_wheel

This is not needed if you only want to target a specfic Linux distribution (i.e. Ubuntu).

Native compiling on Linux ARM device

Easy, slower

Docker build runs on a Raspberry Pi 3B with Raspbian Stretch Lite OS (Desktop version will run out memory when linking the .so file) will take 8-9 hours in total.

sudo apt-get update
sudo apt-get install -y \
    sudo \
    build-essential \
    curl \
    libcurl4-openssl-dev \
    libssl-dev \
    wget \
    python3 \
    python3-pip \
    python3-dev \
    git \

pip3 install --upgrade pip
pip3 install --upgrade setuptools
pip3 install --upgrade wheel
pip3 install numpy

* Build the latest cmake
mkdir /code
cd /code
tar zxf cmake-3.13.5.tar.gz

cd /code/cmake-3.13.5
./configure --system-curl
sudo make install

* Prepare onnxruntime Repo
cd /code
git clone --recursive

* Start the basic build
cd /code/onnxruntime
./ --config MinSizeRel --update --build

* Build Shared Library
./ --config MinSizeRel --build_shared_lib

* Build Python Bindings and Wheel
./ --config MinSizeRel --enable_pybind --build_wheel

* Build Output
ls -l /code/onnxruntime/build/Linux/MinSizeRel/*.so
ls -l /code/onnxruntime/build/Linux/MinSizeRel/dist/*.whl

Cross compiling on Windows

Using Visual C++ compilers

  1. Download and install Visual C++ compilers and libraries for ARM(64). If you have Visual Studio installed, please use the Visual Studio Installer (look under the section Individual components after choosing to modify Visual Studio) to download and install the corresponding ARM(64) compilers and libraries.

  2. Use .\build.bat and specify --arm or --arm64 as the build option to start building. Preferably use Developer Command Prompt for VS or make sure all the installed cross-compilers are findable from the command prompt being used to build using the PATH environmant variable.



The SDK and NDK packages can be installed via Android Studio or the sdkmanager command line tool. Android Studio is more convenient but a larger installation. The command line tools are smaller and usage can be scripted, but are a little more complicated to setup. They also require a Java runtime environment to be available.

General Info:

  • API levels:
  • Android ABIs:
  • System Images:
Android Studio

Install Android Studio from

Install any additional SDK Platforms if necessary

  • File->Settings->Appearance & Behavior->System Settings->Android SDK to see what is currently installed
    • Note that the SDK path you need to use as –android_sdk_path when building ORT is also on this configuration page
    • Most likely you don’t require additional SDK Platform packages as the latest platform can target earlier API levels.

Install an NDK version

  • File->Settings->Appearance & Behavior->System Settings->Android SDK
    • ‘SDK Tools’ tab
      • Select ‘Show package details’ checkbox at the bottom to see specific versions. By default the latest will be installed which should be fine.
  • The NDK path will be the ‘ndk/{version}’ subdirectory of the SDK path shown
    • e.g. if 21.1.6352462 is installed it will be {SDK path}/ndk/21.1.6352462
sdkmanager from command line tools
  • If necessary install the Java Runtime Environment and set the JAVA_HOME environment variable to point to it
    • Windows note: You MUST install the 64-bit version ( otherwise sdkmanager will only list x86 packages and the latest NDK is x64 only.
  • For sdkmanager to work it needs a certain directory structure. First create the top level directory for the Android infrastructure.
    • in our example we’ll call that .../Android/
  • Download the command line tools from the ‘Command line tools only’ section towards the bottom of
  • Create a directory called ‘cmdline-tools’ under your top level directory
    • giving .../Android/cmdline-tools
  • extract the ‘tools’ directory from the command line tools zip file into this directory
    • giving .../Android/cmdline-tools/tools
    • Windows note: preferably extract using 7-zip. If using the built in Windows zip extract tool you will need to fix the directory structure by moving the jar files from tools\lib\_ up to tools\lib
      • See
  • you should now be able to run Android/cmdline-tools/bin/sdkmanager[.bat] successfully
    • if you see an error about it being unable to save settings and the sdkmanager help text, your directory structure is incorrect.
      • see the final steps in this answer to double check:
  • Run .../Android/cmdline-tools/bin/sdkmanager --list to see the packages available

  • Install the SDK Platform
    • Generally installing the latest is fine. You pick an API level when compiling the code and the latest platform will support many recent API levels
      • e.g. sdkmanager --install "platforms;android-29"
    • This will install into the ‘platforms’ directory of our top level directory
      • so the ‘Android’ directory in our example
    • The SDK path to use as –android_sdk_path when building is this top level directory
  • Install the NDK
    • Find the available NDK versions by running sdkmanager --list
    • Install
      • you can install a specific version or the latest (called ‘ndk-bundle’)
      • e.g. sdkmanager --install "ndk;21.1.6352462"
        • NDK path in our example with this install would be .../Android/ndk/21.1.6352462
      • NOTE: If you install the ndk-bundle package the path will be .../Android/ndk-bundle as there’s no version number

Android Build Instructions

Cross compiling on Windows

The Ninja generator needs to be used to build on Windows as the Visual Studio generator doesn’t support Android.

./build.bat --android --android_sdk_path <android sdk path> --android_ndk_path <android ndk path> --android_abi <android abi, e.g., arm64-v8a (default) or armeabi-v7a> --android_api <android api level, e.g., 27 (default)> --cmake_generator Ninja

e.g. using the paths from our example

./build.bat --android --android_sdk_path .../Android --android_ndk_path .../Android/ndk/21.1.6352462 --android_abi arm64-v8a --android_api 27 --cmake_generator Ninja
Cross compiling on Linux and macOS
./ --android --android_sdk_path <android sdk path> --android_ndk_path <android ndk path> --android_abi <android abi, e.g., arm64-v8a (default) or armeabi-v7a> --android_api <android api level, e.g., 27 (default)>
Build Android Archive (AAR)

Android Archive (AAR) files, which can be imported directly in Android Studio, will be generated in your_build_dir/java/build/outputs/aar, by using the above building commands with --build_java

Android NNAPI Execution Provider

If you want to use NNAPI Execution Provider on Android, see NNAPI Execution Provider.

Build Instructions

Android NNAPI Execution Provider can be built using building commands in Android Build instructions with --use_nnapi


See more information on the MIGraphX Execution Provider here.


  • Install ROCM
    • The MIGraphX execution provider for ONNX Runtime is built and tested with ROCM3.3
  • Install MIGraphX
    • The path to MIGraphX installation must be provided via the --migraphx_home parameter.

Build Instructions

./ --config <Release|Debug|RelWithDebInfo> --use_migraphx --migraphx_home <path to MIGraphX home>

Dockerfile instructions are available here



The default NVIDIA GPU build requires CUDA runtime libraries installed on the system:

  • CUDA 10.1
  • cuDNN 7.6.2
  • NCCL v2.4.8 (download v2.4.8 from the Legacy downloads page)
  • OpenMPI
    tar zxf openmpi-4.0.0.tar.gz
    cd openmpi-4.0.0
    ./configure --enable-orterun-prefix-by-default
    make -j $(nproc) all
    sudo make install
    sudo ldconfig

    Build instructions

  1. Checkout this code repo with git clone

  2. Set the environment variables: adjust the path for location your build machine
     export CUDA_HOME=<location for CUDA libs> # e.g. /usr/local/cuda
     export CUDNN_HOME=<location for cuDNN libs> # e.g. /usr/local/cuda
     export CUDACXX=<location for NVCC> #e.g. /usr/local/cuda/bin/nvcc
     export PATH=<location for openmpi/bin/>:$PATH
     export LD_LIBRARY_PATH=<location for openmpi/lib/>:$LD_LIBRARY_PATH
     export MPI_CXX_INCLUDE_PATH=<location for openmpi/include/>
     source <location of the mpivars script> # e.g. /data/intel/impi/2018.3.222/intel64/bin/
  3. Create the ONNX Runtime wheel

    • Change to the ONNX Runtime repo base folder: cd onnxruntime
    • Run ./ --enable_training --use_cuda --config=RelWithDebInfo --build_wheel

    This produces the .whl file in ./build/Linux/RelWithDebInfo/dist for ONNX Runtime Training.