[Build] Failure with OneDNN on Intel MacOS

[matteosal]

Describe the issue

I get errors trying to build with OneDNN onnxruntime_USE_DNNL=ON on an Intel MacOS machine.

Urgency

No response

Target platform

MacOSX-x86-64, Clang 13.1.6

Build script

cmake --compile-no-warning-as-error \
 `# GENERAL FLAGS` \
 -DCMAKE_INSTALL_PREFIX=$output_dir \
 -DCMAKE_BUILD_TYPE=Release \
 -DCMAKE_SKIP_BUILD_RPATH=ON \
 -DPYTHON_EXECUTABLE=$PYTHON3EXE \
 `# ONNXRUNTIME SPECIFIC` \
 -Donnxruntime_BUILD_SHARED_LIB=ON \
 -Donnxruntime_USE_DNNL=ON \
 -Donnxruntime_USE_TELEMETRY=OFF \
 -Donnxruntime_BUILD_UNIT_TESTS=OFF \
 -DProtobuf_USE_STATIC_LIBS=ON \
 $onnxruntime_dir/cmake

Error / output

I see this error during the CMake configuration phase:

CMake Error at onnxruntime_providers.cmake:629 (add_dependencies):
  The dependency target "onnxruntime_providers_shared" of target
  "onnxruntime_providers_dnnl" does not exist.

Followed by this error at build time:

onnxruntime/onnxruntime/core/providers/dnnl/dnnl_execution_provider.cc:14:10: fatal error: 'omp.h' file not found

Visual Studio Version

No response

GCC / Compiler Version

No response

[wschin]

Looks like you need to install openmp. Try this:

brew install libomp

[matteosal]

@wschin that would mean ONNXRuntime relies on an external installation of OpenMP in order to build with OneDNN, which doesn't sound quite right. There is no mention of this on the build instructions with the OneDNN provider here, and CMakeLists.txt doesn't expose any flag to set the OpenMP location, contrary to several other dependencies (e.g. onnxruntime_CUDA_HOME).

This looks more like ONNXRuntime is supposed to automatically take care of it but it's failing to do so for some reason, which is also supported by the fact that a CMake configuration error happens well before the compile error (as reported in the top comment).

[matteosal]

Any other take here? @RyanUnderhill maybe?