[BLOCKING] Latest XGBoost with CUDA fails to compile on Windows

[hcho3]

• OS: Windows Server 2012 R2
• Compiler: Visual Studio 2017 (15.0.0+26228.76)
• CUDA: 9.0
• CMake version: 3.12.0

I checked out the latest code from the master branch and tried to build and got this error:

13>common.cu.obj : error LNK2038: mismatch detected for 'RuntimeLibrary': value 'MD_DynamicRelease' doesn't match value 'MT_StaticRelease' in cli_main.obj [C:\Users\Administrator\Desktop\xgboost\build\runxgboost.vcxproj]
13>hist_util.cu.obj : error LNK2038: mismatch detected for 'RuntimeLibrary': value 'MD_DynamicRelease' doesn't match value 'MT_StaticRelease' in cli_main.obj [C:\Users\Administrator\Desktop\xgboost\build\runxgboost.vcxproj]
13>host_device_vector.cu.obj : error LNK2038: mismatch detected for 'RuntimeLibrary': value 'MD_DynamicRelease' doesn't match value 'MT_StaticRelease' in cli_main.obj [C:\Users\Administrator\Desktop\xgboost\build\runxgboost.vcxproj]
...
...
13>regression_obj.cu.obj : error LNK2001: unresolved external symbol __imp_acoshf 13>gpu_predictor.obj : error LNK2001: unresolved external symbol __imp_acoshf
13>updater_gpu.obj : error LNK2001: unresolved external symbol __imp_acoshf
13>updater_gpu_hist.obj : error LNK2001: unresolved external symbol __imp_acoshf
...

Get the full build log from MSBuild

I suspect that some /MD flags are not being replaced with /MT.

To reproduce:

mkdir build
cd build
cmake .. -G"Visual Studio 15 2017 Win64" -DUSE_CUDA=ON -DCMAKE_VERBOSE_MAKEFILE=ON
"C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvarsall.bat" amd64
msbuild xgboost.sln /m /p:Configuration=Release -fl -flp:logfile=MyProjectOutput.log;verbosity=diagnostic

[hcho3]

This time, I'm really hoping to set up CI for Windows for good.

Good news is that I finally managed to get SSH connection working with Windows EC2 workers. (Public key auth was quite tricky to get right.) Once this issue is fixed, I will submit a PR to add Windows GPU tests.

[hcho3]

Indeed, hist_util.cu is getting wrong compilation flag (/MD instead of /MT):

"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin\nvcc.exe" -gencode=arch=compute_35,code="sm_35,compute_35" -gencode=arch=compute_50,code="sm_50,compute_50" -gencode=arch=compute_52,code="sm_52,compute_52" -gencode=arch=compute_60,code="sm_60,compute_60" -gencode=arch=compute_61,code="sm_61,compute_61" -gencode=arch=compute_70,code="sm_70,compute_70" -gencode=arch=compute_70,code="compute_70,compute_70" --use-local-env --cl-version 2017 -ccbin "C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.10.25017\bin\HostX86\x64" -x cu -IC:\Users\Administrator\Desktop\xgboost\cub -IC:\Users\Administrator\Desktop\xgboost\include -I"C:\Users\Administrator\Desktop\xgboost\dmlc-core\include" -IC:\Users\Administrator\Desktop\xgboost\rabit\include -I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\include" --keep-dir x64\Release -maxrregcount=0 --machine 64 --compile -cudart static --expt-extended-lambda --expt-relaxed-constexpr -lineinfo -Xcompiler="/EHsc -Ob2 -openmp" -D_WINDOWS -DNDEBUG -DXGBOOST_USE_CUDA=1 -DDMLC_LOG_CUSTOMIZE=1 -DXGBOOST_MM_PREFETCH_PRESENT=1 -D"CMAKE_INTDIR="Release"" -DWIN32 -D_WINDOWS -DNDEBUG -DXGBOOST_USE_CUDA=1 -DDMLC_LOG_CUSTOMIZE=1 -DXGBOOST_MM_PREFETCH_PRESENT=1 -D"CMAKE_INTDIR="Release"" -D_MBCS -Xcompiler "/EHsc /W3 /nologo /O2 /FS /Zi /MD /GR" -o objxgboost.dir\Release/common/hist_util.cu.obj "C:\Users\Administrator\Desktop\xgboost\src\common\hist_util.cu" (TaskId:151)

[hcho3]

Update: The build is fixed if I manually replace

<Runtime>MD</Runtime>

with

<Runtime>MT</Runtime>

in all of the generated *.vcxproj files.

Now I'd like to find a way to force CMake to produce *.vcxproj files with MT runtime property.

[trivialfis]

The cuda source code is split into host and device source during compilation. So there are 3 different sources, c++ from cc file, c++ from cu files for host, and c++ from cu file for device. My guess is the second one went wrong because its flags are controlled by CMake CUDA language instead of CXX language.

[trivialfis]

We just need to copy the configuration for CMake CXX to CUDA lang.

[hcho3]

My guess is the second one went wrong because its flags are controlled by CMake CUDA language instead of CXX language.

Interesting.

We just need to copy the configuration for CMake CXX to CUDA lang.

@trivialfis Do you have a suggestion? So far I came up with #4463 and I'm wondering if you have a more elegant fix.