Benchmarking is an exact science as the results may change depending on the machine used to compute the figures. There is not necessarily an exact correlation between the processing time and the algorithm cost. The results may also depend on the options used to compile a library (CPU, GPU, MKL, …). Next sections gives some details on how it was done.
./configure --enable-optimizations --with-ensurepip=install --enable-shared --prefix=/opt/bin
This is due to a feature which requests to be able to interpret
Python inside a package itself and more specifically: Embedding the Python interpreter.
Then the environment variable
LD_LIBRARY_PATH must be set to
the location of the shard libraries,
/opt/bin in the previous example.
The following issue might appear:
UserWarning: Cannot load onnxruntime.capi. Error: 'libnnvm_compiler.so: cannot open shared object file: No such file or directory'
To build onnxruntime:
git clone https://github.com/Microsoft/onnxruntime.git --recursive export LD_LIBRARY_PATH=/usr/local/Python-3.6.8 python3.6 ./onnxruntime/tools/ci_build/build.py --build_dir ./onnxruntime/build/debian36 --config Release --enable_pybind --build_wheel --use_mkldnn --use_openmp --build_shared_lib export LD_LIBRARY_PATH=/usr/local/Python-3.7.2 python3.7 ./onnxruntime/tools/ci_build/build.py --build_dir ./onnxruntime/build/debian37 --config Release --enable_pybind --build_wheel --use_mkldnn --use_openmp --build_shared_lib
If the wheel then, it is possible to just copy the files into the python distribution:
cp -r ./onnxruntime/build/debian36/Release/onnxruntime /usr/local/lib/python3.6/site-packages/ cp -r ./onnxruntime/build/debian37/Release/onnxruntime /usr/local/lib/python3.7/site-packages/