Installation (WIN10) error: value 'MT_StaticRelease' doesn't match value 'MD_DynamicRelease'

User discussions
1

GROMACS version: 2023.2
GROMACS modification: No
OS: Windows 10 64 bit
GPU: GTX 1050

CMake command (newlines inserted for ease of reading):

cd C:\Users\Nicolas\Documents\Gromacs\gromacs-2023.2\build

cmake .. -G "Visual Studio 17 2022" 
-A x64 
-DCMAKE_INSTALL_PREFIX=C:\Users\Nicolas\Documents\Gromacs\GMX 
-DGMX_BUILD_OWN_FFTW=OFF  
-DGMX_FFT_LIBRARY=fftw3 
-DFFTWF_LIBRARY="C:\Users\Nicolas\Documents\Gromacs\fftw-install\lib\fftw3f.lib" 
-DFFTWF_INCLUDE_DIR="C:\Users\Nicolas\Documents\Gromacs\fftw-install\include" 
-DREGRESSIONTEST_DOWNLOAD=ON 
-DGMX_GPU=CUDA 
-DGMX_CUDA_TARGET_SM=61 
-DGMX_BUILD_SHARED_EXE=OFF

cmake --build . --config Release --parallel

The error I receive is

gromacs.lib(libgromacs_generated_detecthardware.cpp.obj) : error LNK2038: mismatch detected for 'RuntimeLibrary': value
 'MT_StaticRelease' doesn't match value 'MD_DynamicRelease' in gmx.obj [C:\Users\Nicolas\Documents\Gromacs\gromacs-2023
.2\build\src\programs\gmx.vcxproj]

There are a total of 49 of the linker errors.

I’ve also tried building without the -DGMX_BUILD_SHARED_EXE=OFF flag but I get the same error.

I see this post here (Link error when building Gromacs on windows with CUDA - #2 by SeLarin) mentions changing a certain cmake variable but I do not know which cmake file I should be making the changes in.

Thanks!

2

Hi!

Here is a more detailed discussion of this problem:
Build fails on windows when cuda is enabled (#4582) ¡ Issues ¡ GROMACS / GROMACS ¡ GitLab. They suggest setting it at the beginning of the top-level CMakeLists.txt file. There might be a few more modifications required, as mentioned there.

If you make it work, please describe (here or in Gitlab) which steps exactly you had to take. We don’t have any Windows + NVIDIA machines in the GROMACS core team, so reports from users are crucial for figuring out the best fix :)

1 Like
3

Hi,

Thank you for your help!

Adding the following line to the CMakeLists.txt file resulted in a successful build:
set(CMAKE_MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")

During previous attempts at building I also received these two errors (not all at once):
nvcc fatal : Unknown option '-std:c++17'
nvcc fatal : Unsupported gpu architecture 'compute_35'

To solve these issues, I built with the flag -DGMX_CUDA_TARGET_SM=61 to target my specific GPU architecture.

I also see you commented here that

However, although I am building GROMACS 2023.2 and it successfully detected CUDA 12.2, I still received the compute_35 error when I did not use the -DGMX_CUDA_TARGET_SM flag.

4

Great!

Could you please check if, in a clean build directory, you can build without setting DGMX_CUDA_TARGET_SM after replacing this line in cmake/gmxManageNvccConfig.cmake with the following five lines:

if (NOT MSVC)
    list(APPEND GMX_CUDA_NVCC_FLAGS "${CMAKE_CXX17_STANDARD_COMPILE_OPTION}")
else()
    list(APPEND GMX_CUDA_NVCC_FLAGS "-std=c++17")  # See #4582
endif()

Nothing wrong with using GMX_CUDA_TARGET_SM (it also makes the compilation faster), but it would be useful to understand the problem.

5

I freshly extracted gromacs-2023.2.tar.gz into a new directory (Gromacs-Test) and performed the modification you mentioned on that cmake file.

This first cmake command succeeded without errors. I removed the DGMX_CUDA_TARGET_SM flag as requested.

cmake .. -G "Visual Studio 17 2022" -A x64 -DCMAKE_INSTALL_PREFIX=C:\Users\Nicolas\Documents\Gromacs-Test\GMX -DGMX_BUILD_OWN_FFTW=OFF  -DGMX_FFT_LIBRARY=fftw3 -DFFTWF_LIBRARY="C:\Users\Nicolas\Documents\Gromacs-Test\fftw-install\lib\fftw3f.lib" -DFFTWF_INCLUDE_DIR="C:\Users\Nicolas\Documents\Gromacs-Test\fftw-install\include" -DREGRESSIONTEST_DOWNLOAD=ON -DGMX_GPU=CUDA -DGMX_BUILD_SHARED_EXE=OFF

However, the second cmake command (unsure what the proper names for these commands are) failed.
cmake --build . --config Release --parallel

Below is the abbreviated output.

  Generating release version information
  Building Custom Rule C:/Users/Nicolas/Documents/Gromacs-Test/gromacs-2023.2/CMakeLists.txt
  Building NVCC (Device) object src/gromacs/CMakeFiles/libgromacs.dir/nbnxm/cuda/Release/libgromacs_generated_nbnxm_cud
  a.cu.obj
  nvcc fatal   : Unsupported gpu architecture 'compute_35'
  CMake Error at libgromacs_generated_nbnxm_cuda.cu.obj.Release.cmake:224 (message):
    Error generating
    C:/Users/Nicolas/Documents/Gromacs-Test/gromacs-2023.2/build/src/gromacs/CMakeFiles/libgromacs.dir/nbnxm/cuda/Relea
  se/libgromacs_generated_nbnxm_cuda.cu.obj

I have not touched the directory after this should you want to view some of the intermediate files.

6

Thank you for checking!

So, the problem is that we were skipping the compiler flag check on Windows during CMake run, and, with CUDA 12, it became a problem at compile time.

One more thing to try: could you please edit cmake/gmxManageNvccConfig.cmake to remove AND NOT WIN32 and OR WIN32 from lines 134 and 168, respectively? Again, in a clean directory.

Before:

    # If the check has already been run, do not re-run it
    if (NOT ${_flags_cache_variable_name} AND NOT WIN32)
        message(STATUS "Checking if nvcc accepts flags ${ARGN}")

.....

    # Append the flags to the output variable if they have been tested to work
    if (${_flags_cache_variable_name} OR WIN32)
        list(APPEND ${_output_variable_name_to_append_to} ${ARGN})

After:

    # If the check has already been run, do not re-run it
    if (NOT ${_flags_cache_variable_name})
        message(STATUS "Checking if nvcc accepts flags ${ARGN}")

.....

    # Append the flags to the output variable if they have been tested to work
    if (${_flags_cache_variable_name})
        list(APPEND ${_output_variable_name_to_append_to} ${ARGN})
7

Hi,

I removed those conditions from the if statements in that CMake file.

With just that edit, I obtain the following abbreviated output.

Generating release version information
  Building Custom Rule C:/Users/Nicolas/Documents/Gromacs-Test/gromacs-2023.2/CMakeLists.txt
  Building NVCC (Device) object src/gromacs/CMakeFiles/libgromacs.dir/nbnxm/cuda/Release/libgromacs_generated_nbnxm_cud
  a.cu.obj
  nvcc fatal   : Unknown option '-std:c++17'
  CMake Error at libgromacs_generated_nbnxm_cuda.cu.obj.Release.cmake:224 (message):
    Error generating
    C:/Users/Nicolas/Documents/Gromacs-Test/gromacs-2023.2/build/src/gromacs/CMakeFiles/libgromacs.dir/nbnxm/cuda/Relea
  se/libgromacs_generated_nbnxm_cuda.cu.obj

With your suggested edits to gmxManageNvccConfig.cmake from both of your posts, the error was the same as in my OP (which makes sense). I might try building again with the CMAKE_MSVC_RUNTIME_LIBRARY variable set properly. Should probably give my laptop a break first though.

8

Yes, sorry for being unclear. I meant that all the fixes should be combined:

  • APPEND GMX_CUDA_NVCC_FLAGS "-std=c++17" to solve the “Unknown option ‘-std:c++17’” error,
  • CMAKE_MSVC_RUNTIME_LIBRARY to solve the “MT_StaticRelease’ doesn’t match” error,
  • and the change around _flags_cache_variable_name to, hopefully, solve the issue of " Unsupported gpu architecture 'compute_35'" without the need to manually set GMX_CUDA_TARGET_SM.

those are all independent issues.