MXNet master branch builds on Windows but fails to import in Python

[Tenchumaru]

I can successfully build MXNet for Windows -- it takes several hours -- but importing mxnet in Python fails.

>>> import mxnet
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "c:\github\incubator-mxnet\python\mxnet\__init__.py", line 23, in <module>
    from .context import Context, current_context, cpu, gpu, cpu_pinned
  File "c:\github\incubator-mxnet\python\mxnet\context.py", line 20, in <module>
    from .base import _LIB
  File "c:\github\incubator-mxnet\python\mxnet\base.py", line 315, in <module>
    ctypes.byref(ctypes.c_bool())))
OSError: exception: access violation writing 0x0000000000000000

I cloned the master branch of apache/incubator-mxnet from GitHub. The SHA is 5d247f1. I'm on Windows 10 version 21H1 build 19043.1288 as ver reports. I'm compiling with Visual Studio 2019, version 16.11.5. I'm using Python version 3.7.8, the version included with Visual Studio 2019. My CUDA Toolkit is version 11.4. My GPU is a RTX 2070 Super. This is in an Alienware laptop with an Intel i7 and 32 GiB of RAM. Here are my build commands executed in x64 Native Tools Command Prompt for VS 2019. I derived the cmake command line from this discussion.

MD build
CD build
SET OpenCV_DIR=C:\local\opencv\build
SET OpenBLAS_HOME=C:\local\OpenBLAS-0.3.18-x64
cmake -G "Visual Studio 16 2019" -A x64 -T cuda=11.4,host=x64 -DUSE_CUDA=1 -DUSE_CUDNN=1 -DUSE_OPENCV=1 -DUSE_OPENMP=1 -DUSE_BLAS=open -DUSE_LAPACK=1 -DUSE_DIST_KVSTORE=0 -DCUDA_ARCH_LIST=Turing -DCMAKE_CUDA_ARCHITECTURES=75 -DCUDNN_INCLUDE=C:\local\cudnn-11.4-windows-x64-v8.2.4.15\cuda\include -DCUDNN_LIBRARY=C:\local\cudnn-11.4-windows-x64-v8.2.4.15\cuda\lib\x64\cudnn.lib C:\GitHub\incubator-mxnet
msbuild -m -p:Configuration=Release;Platform=x64 mxnet.sln
CD ..
python -m pip install -e ./python

The cmake output has this note: You might need to add C:\local\opencv\build\x64\vc15\bin to your PATH to be able to run your applications. Before running Python, I set my path with PATH %PATH%;C:\local\opencv\build\x64\vc15\bin;C:\local\OpenBLAS-0.3.18-x64\bin.

The failure above is in MXSetFlushDenorms, a function in src\c_api\c_api.cc. Here is the compilation command for that file.

C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Tools\MSVC\14.29.30133\bin\HostX64\x64\CL.exe /c /I"C:\GitHub\incubator-mxnet\include" /I"C:\GitHub\incubator-mxnet\src" /I"C:\GitHub\incubator-mxnet\3rdparty\tvm\nnvm\include" /I"C:\GitHub\incubator-mxnet\3rdparty\tvm\include" /I"C:\GitHub\incubator-mxnet\3rdparty\dmlc-core\include" /I"C:\GitHub\incubator-mxnet\3rdparty\dlpack\include" /I"C:\GitHub\incubator-mxnet\3rdparty\mshadow" /I"C:\GitHub\incubator-mxnet\3rdparty\miniz" /I"C:\GitHub\incubator-mxnet\build\3rdparty\dmlc-core\include" /I"C:\local\OpenBLAS-0.3.18-x64\include" /IC:\local\opencv\build\include /I"C:\local\cudnn-11.4-windows-x64-v8.2.4.15\cuda\include" /I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\include" /I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\include" /nologo /W3 /WX- /diagnostics:column /MP /O2/Ob2 /D WIN32 /D _WINDOWS /D NDEBUG /D WIN32_LEAN_AND_MEAN /D DMLC_MODERN_THREAD_LOCAL=0 /D DMLC_LOG_STACK_TRACE_SIZE=0 /D DMLC_USE_CXX11 /D DMLC_STRICT_CXX11 /D DMLC_USE_CXX14 /D MSHADOW_IN_CXX11 /D _SCL_SECURE_NO_WARNINGS /D _CRT_SECURE_NO_WARNINGS /D MXNET_EXPORTS /D NNVM_EXPORTS /D NOMINMAX /D MSHADOW_USE_CBLAS=1 /D MSHADOW_USE_MKL=0 /D MXNET_USE_BLAS_OPEN=1 /D MXNET_USE_LIBJPEG_TURBO=0 /D MXNET_USE_OPENCV=1 /D MXNET_USE_OPENMP=1 /DMXNET_USE_LAPACK=1 /D USE_CUDNN /D MXNET_USE_CUDA=1 /D MXNET_USE_SIGNAL_HANDLER=1 /D MSHADOW_USE_CUDA=1 /D MSHADOW_FORCE_STREAM /D MSHADOW_USE_SSE=0 /D MSHADOW_USE_CUDNN /D MSHADOW_USE_F16C=0 /D MSHADOW_INT64_TENSOR_SIZE=1 /D __USE_XOPEN2K8 /D DMLC_CORE_USE_CMAKE /D DMLC_USE_CXX11=1 /D "CMAKE_INTDIR=\"Release\"" /D mxnet_75_EXPORTS /D _WINDLL /D _MBCS /Gm- /EHsc /MD /GS /Gy /fp:precise /Zc:wchar_t /Zc:forScope /Zc:inline /GR /openmp /std:c++17 /Fo"mxnet_75.dir\Release\\" /Fd"mxnet_75.dir\Release\vc142.pdb" /external:W3 /Gd /TP /errorReport:queue  /bigobj -openmp:experimental "C:\GitHub\incubator-mxnet\src\c_api\c_api.cc"

Since SUPPORT_FTZ_DMZ isn't defined, the function is essentially this.

int MXSetFlushDenorms(bool value, bool* prev_state) {
  API_BEGIN();
  *prev_state = false;
  API_END();
}

Might this be an ABI violation, such as __stdcall versus __cdecl? Am I missing some cmake flags?