Skip to content

Latest commit

 

History

History
166 lines (122 loc) · 7.08 KB

BUILDING.md

File metadata and controls

166 lines (122 loc) · 7.08 KB
Raw

Building RAD

There are two ways to build the dependencies for RAD:

  1. Use the built-in script, or
  2. Build the dependencies manually.

It is recommended that you build the dependencies yourself if you require specific configurations for OpenCV and ONNXRuntime or if you are going to be sharing these libraries with other projects. If you need to build a specific version of RAD, then you're encouraged to use the built-in script.

Using the Script

Starting with RAD 2.3.0, a Python script is provided to automatically build all dependencies. The script requires Python 3.12+ and can be used as follows:

python3 -m venv .venv
. .venv/bin/activate # or .venv/Scripts/activate for Windows
pip install -r requirements.txt

python3 ./tools/build_dependencies.py <install-root>

The script will automatically clone, build, and install the correct versions for all dependencies in the <install-root> directory. Once the script finishes, you can build RAD by appending -DCMAKE_PREFIX_PATH=<install-root> to the list of CMake options when configuring the build system.

Manually Building

Below is a quick guide on the configurations used to build the libraries RAD depends on. Note that while RAD does depend on Zeus it (and its dependencies) can be pulled directly through CMake if you don't wish to build everything yourself.

RAD only supports Debug and Release builds, so it is recommended that all SDKs that are built from source are built for both of these configurations.

Installing OneAPI TBB

RAD uses TBB internally for multi-threading operations, and it is also recommended that OpenCV be built with TBB to avoid clashes between different threading systems. As a result, TBB should be installed first. To do so, clone the code from here and checkout the version that RAD supports (as noted in the README). When building, it is recommended that you set TBB_TEST to false in order to speed up compilation. Once it is built, install it using the install target.

Installing OpenCV

OpenCV can be cloned from these two repositories:

Once the appropriate version of OpenCV Is checked out, it is time to build OpenCV. If the contrib module was cloned, please set OPENCV_EXTRA_MODULES_PATH to the corresponding directory. The following variables must also be set:

  1. WITH_TBB=ON: builds OpenCV with TBB as the threading back-end
  2. BUILD_TESTS=OFF: disable unit tests
  3. BUILD_opencv_world: build OpenCV as a single DLL (optional).

Once cmake runs, build and install OpenCV as usual.

Installing ONNXRuntime

ONNXRuntime is an optional dependecy for RAD that is only used for inferencing neural networks. It is important to note that CMake cannot link with the pre-built libraries as they don't contain the correct setup to allow CMake to properly include them.

Before continuing, please ensure you have all the required dependencies as listed here. Once you do, clone the code:

git clone https://github.com/Microsoft/onnxruntime.git --branch <version> --depth 1 --recurse-submodules
cd onnxruntime

The latest supported version of ONNXRuntime can be found in the project's README. Once the code is checked out, we can build the library depending on which backend is required. Note that we must build both Debug and Release configurations of ONNXRuntime (for Windows only).

Building for CPU

Windows

Open a terminal and navigate to the directory where the code is checked out. Then enter:

.\build.bat --config <config> --target install --build_shared_lib --parallel --skip_submodule_sync --skip_tests --cmake_extra_defines CMAKE_INSTALL_PREFIX=<install-path> onnxruntime_BUILD_UNIT_TESTS=OFF

Where <config> is the config you wish to build and <install-path> is the path where the library should be installed to.

Note: If you wish to install to the system directories, you must use an administrator prompt to do so.

Linux

Open a terminal and navigate tot he code where the code is checked out. Then enter:

./build.sh --config <config> --target install --build_shared_lib --parallel --skip_submodule_sync --skip_tests --cmake_extra_defines CMAKE_INSTALL_PREFIX=<install-path> onnxruntime_BUILD_UNIT_TESTS=OFF

Building for DirectML (Windows only)

The process for building with DML is identical to the CPU process, with the exception that for all build commands, you must add the --use_dml flag. Once both versions of ONNXRuntime have been built and installed, navigate to onnxruntime/build/Windows/Release/Release and copy DirectML.dll to the bin folder in the install location. This will ensure that RAD can automatically copy the DLL when building through the provided CopySharedLibs script.

Troubleshooting

In certain Windows platforms, building ONNXRuntime can fail due to compiler errors. The issue is tracked here. Below are some workarounds to the compiler errors that may arise:

  • Compiler errors appear regarding undefined symbols for STRSAFE_LPSTR and similar aliases. To solve this, navigate to onnxruntime/build/Windows/<Debug/Release>/_deps/microsoft-wil-src/include/wil and open result_macros.h. Immediately after the include block that contains #include <strsafe.h>, copy the following:

    // Deprecated, use the non STRSAFE_ prefixed types instead (e.g. LPSTR or PSTR) as they are the same as these.
typedef _Null_terminated_ char* STRSAFE_LPSTR;
typedef _Null_terminated_ const char* STRSAFE_LPCSTR;
typedef _Null_terminated_ wchar_t* STRSAFE_LPWSTR;
typedef _Null_terminated_ const wchar_t* STRSAFE_LPCWSTR;
typedef _Null_terminated_ const wchar_t UNALIGNED* STRSAFE_LPCUWSTR;

// Deprecated, use the base types instead.
// Strings where the string is NOT guaranteed to be null terminated (does not have _Null_terminated_).
typedef const char* STRSAFE_PCNZCH;
typedef const wchar_t* STRSAFE_PCNZWCH;
typedef const wchar_t UNALIGNED* STRSAFE_PCUNZWCH;

  • Linker errors appear regarding: IID_ID3D12Device, IID_ID3D12RootSignature, and IID_ID3D12PipelineState. Navigate to onnxruntime/onnxruntime/core/providers/dml/DMLExecutionProvider/src/Operators and open DmlDFT.h. Replace any instances of those types with __uuidof(<name-without-IID>). Also open DmlGridSample.h and do the same.

  • Linker errors appear regarding: DXCORE_ADAPTER_ATTRIBUTE_D3D12_GRAPHICS and DXCORE_ADAPTER_ATTRIBUTE_D3D12_CORE_COMPUTE. Navigate to onnxruntime/onnxruntime/core/providers/dml and open dml_provider_factory.cc. At the top of the file, after the final include, copy the following:

    #if !defined(INITGUID)
#define INITGUID
#include <guiddef.h>
#endif
DEFINE_GUID(DXCORE_ADAPTER_ATTRIBUTE_D3D12_GRAPHICS, 0x0c9ece4d, 0x2f6e, 0x4f01, 0x8c, 0x96, 0xe8, 0x9e, 0x33, 0x1b, 0x47, 0xb1);
DEFINE_GUID(DXCORE_ADAPTER_ATTRIBUTE_D3D12_CORE_COMPUTE, 0x248e2800, 0xa793, 0x4724, 0xab, 0xaa, 0x23, 0xa6, 0xde, 0x1b, 0xe0, 0x90);