cannot build shared library on Windows

[jcar87]

On Windows with msvc (Visual Studio 2022), with BUILD_SHARED_LIBS=ON, the build fails with:

  Auto build dll exports
LINK : fatal error LNK1189: library limit of 65535 objects exceeded [C:\workspace\cci_prod_PR-29156\conan-home\p\b\ginkg9e01d3cc3e3da\b\build\core\ginkgo_core.vcxproj]

I suspect this is a known consequence of setting CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS to True

But the question stands: is this a regression and is this expected to work, or should this just be a static build only?

[yhmtsai]

We still expect it works under some restriction due to the limitation of LNK1189.
We have tested the Windows build in Github Action.
Could you share the environment and the detailed.log in your build folder?

[spiralbit]

@jcar87 Here is a BAT script I wrote to get the cmake configuration and build to work on Windows. I had the same linker problem with MSVC of the number of objects, which I think is a hard limit and cannot be changed.

call "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat"

set CXX=
set CC=
set CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1
set CUDA_PATH_V13_0=
set CUDA_PATH_V13_1=

rd /s /q ginkgo-build
mkdir ginkgo-build
cd ginkgo-build

cmake -G "Ninja" ^
    -DCMAKE_CXX_FLAGS="/bigobj /EHsc /wd4244" ^
    -DCMAKE_C_FLAGS="/bigobj" ^
    -DBUILD_SHARED_LIBS=OFF ^
    -DGINKGO_BUILD_CUDA=ON ^
    -DCMAKE_BUILD_TYPE=Release ^
    -DCUDAToolkit_ROOT="C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.1" ^
    -DCMAKE_CUDA_COMPILER="C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.1/bin/nvcc.exe" ^
    -DGINKGO_CUDA_ARCHITECTURES=120 ^
    -DCMAKE_CUDA_FLAGS="-diag-suppress 221" ^
    -DCMAKE_CXX_COMPILER="C:/Program Files/Microsoft Visual Studio/2022/Community/VC/Tools/MSVC/14.44.35207/bin/Hostx64/x64/cl.exe" ^
    c:\users\dch88\source\repos\ginkgo

cd ..

The key line is BUILD_SHARED_LIBS=OFF. If you must have DLLs on Windows then there are two options: 1) Use a different compiler like GCC under MinGW or 2) Change the repo around somehow to result in smaller linkage files. Developing the script above was not entirely easy or quick - there were several iterations of getting cmake happy then compiling OK, then running the tests. With this configuration 15 tests still fail for me, so I'm currently building on Ubuntu under WSL2 to see if that results in 100% test pass.

I'm wondering if we should have a file in the documentation somewhere which list these combinations of options for particular hardware+OS+configuration preferences ? Not having a default compilation for Windows 11 + NVidia 13.1 (RTX 5070 card) which I could fall back on seemed like something which could be improved on. I notice on another thread someone is cooking up a mixture of options for a Cray build. All these could be captured. In the "Days of Make" we might have Makefile.Linux, Makefile.BSD, Makefile.WindowsGCC etc.

Could we have a set of platform specific notes under doc/install/* or something like that? I would be happy to have a go at producing something for Windows and Ubuntu under WSL2