Build Failed on Windows with CUDA >= 12

Summary

The compilation process encounters errors on Windows with CUDA version above 12

Exact steps to reproduce

Within a Powershell with CUDA > 12

PS D:\depots\gromacs_gitlab> cmake -S . -B .\build\ -D GMX_GPU=CUDA
PS D:\depots\gromacs_gitlab> cmake --build .\build --parallel 4
Building CUDA object src\gromacs\CMakeFiles\libgromacs.dir\nbnxm\cuda\nbnxm_cuda.cu.obj
FAILED: src/gromacs/CMakeFiles/libgromacs.dir/nbnxm/cuda/nbnxm_cuda.cu.obj
C:\PROGRA~1\NVIDIA~2\CUDA\v12.6\bin\nvcc.exe -forward-unknown-to-host-compiler -ccbin=C:\PROGRA~2\MICROS~3\2019\COMMUN~1\VC\Tools\MSVC\1429~1.301\bin\Hostx64\x64\cl.exe -DGMX_DOUBLE=0 -DHAVE_CONFIG_H -DMUPARSER_STATIC -DNOMINMAX -DTMPI_EXPORTS -DTMPI_USE_VISIBILITY -DUSE_STD_INTTYPES_H -ID:\depots\gromacs_gitlab\src\include -ID:\depots\gromacs_gitlab\build\src\include -ID:\depots\gromacs_gitlab\src -ID:\depots\gromacs_gitlab\src\external\tng_io\include -ID:\depots\gromacs_gitlab\build\tng\include -ID:\depots\gromacs_gitlab\api\legacy\include -ID:\depots\gromacs_gitlab\build\api\legacy\include -ID:\depots\gromacs_gitlab\src\gromacs\math\include -ID:\depots\gromacs_gitlab\src\gromacs\timing\include -ID:\depots\gromacs_gitlab\src\gromacs\utility\include -ID:\depots\gromacs_gitlab\src\gromacs\pbcutil\include -ID:\depots\gromacs_gitlab\src\gromacs\pulling\include -ID:\depots\gromacs_gitlab\src\gromacs\topology\include -ID:\depots\gromacs_gitlab\src\gromacs\linearalgebra\include -ID:\depots\gromacs_gitlab\src\gromacs\simd\include -ID:\depots\gromacs_gitlab\src\gromacs\taskassignment\include -isystem D:\depots\gromacs_gitlab\src\external\thread_mpi\include -isystem D:\depots\gromacs_gitlab\src\external -isystem D:\depots\gromacs_gitlab\src\gromacs\..\external\rpc_xdr -isystem "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.6\include" -isystem D:\depots\gromacs_gitlab\src\external\muparser\include -isystem D:\depots\gromacs_gitlab\src\external\lmfit -isystem D:\depots\gromacs_gitlab\src\external\colvars -D_WINDOWS -Xcompiler=" /GR /EHsc" -Xcompiler=" -Zi -Ob0 -Od /RTC1" -std=c++17 "--generate-code=arch=compute_35,code=[compute_35,sm_35]" "--generate-code=arch=compute_37,code=[compute_37,sm_37]" "--generate-code=arch=compute_50,code=[compute_50,sm_50]" "--generate-code=arch=compute_52,code=[compute_52,sm_52]" "--generate-code=arch=compute_60,code=[compute_60,sm_60]" "--generate-code=arch=compute_61,code=[compute_61,sm_61]" "--generate-code=arch=compute_70,code=[compute_70,sm_70]" "--generate-code=arch=compute_75,code=[compute_75,sm_75]" "--generate-code=arch=compute_80,code=[compute_80,sm_80]" "--generate-code=arch=compute_86,code=[compute_86,sm_86]" "--generate-code=arch=compute_89,code=[compute_89,sm_89]" "--generate-code=arch=compute_90,code=[compute_90,sm_90]" "--generate-code=arch=compute_53,code=[compute_53,sm_53]" "--generate-code=arch=compute_80,code=[compute_80,sm_80]" -Xcompiler=-MDd -Wno-deprecated-gpu-targets -use_fast_math -Xptxas=-warn-double-usage -Xptxas=-Werror -Xcompiler=-openmp -Xcompiler=\"/DWIN32\" -Xcompiler=\"/D_WINDOWS\" -Xcompiler=\"/GR\" -Xcompiler=\"/EHsc\" -diag-suppress=177 -MD -MT src\gromacs\CMakeFiles\libgromacs.dir\nbnxm\cuda\nbnxm_cuda.cu.obj -MF src\gromacs\CMakeFiles\libgromacs.dir\nbnxm\cuda\nbnxm_cuda.cu.obj.d -x cu -c D:\depots\gromacs_gitlab\src\gromacs\nbnxm\cuda\nbnxm_cuda.cu -o src\gromacs\CMakeFiles\libgromacs.dir\nbnxm\cuda\nbnxm_cuda.cu.obj -Xcompiler=-Fdsrc\gromacs\CMakeFiles\libgromacs.dir\libgromacs.pdb,-FS
nvcc fatal   : Unsupported gpu architecture 'compute_35'

For developers: Why is this important?

Currently, none Windows users with CUDA >12 can build GROMACS with default options unless the GMX_CUDA_TARGET_SM option is provided.

This happends because CUDA 12 has dropped support for compute 3.X and how it's managed by GROMACS in gmxManageNvccConfig.cmake.

For Windows, the nvcc flags test is disabled so by default, GROMACS try to compile for all compute architectures listed:

-- Found CUDAToolkit: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.6/include (found suitable version "12.6.68", minimum required is "11.0")
-- The CUDA compiler identification is NVIDIA 12.6.68
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Check for working CUDA compiler: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.6/bin/nvcc.exe - skipped
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done

Possible fixes

In my opinion, the easiest way to fix it is to remove the OR NOT WIN32 in the various if statements in gmxManageNvccConfig.cmake:

-- Found CUDAToolkit: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.6/include (found suitable version "12.6.68", minimum required is "11.0")
-- Check for working NVCC/C++ compiler combination with nvcc 'C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.6/bin/nvcc.exe'
-- Check for working NVCC/C++ compiler combination - works
-- Checking if nvcc accepts flags --generate-code=arch=compute_35,code=sm_35
-- Checking if nvcc accepts flags --generate-code=arch=compute_35,code=sm_35 - No
-- Checking if nvcc accepts flags --generate-code=arch=compute_37,code=sm_37
-- Checking if nvcc accepts flags --generate-code=arch=compute_37,code=sm_37 - No
-- Checking if nvcc accepts flags --generate-code=arch=compute_50,code=sm_50
-- Checking if nvcc accepts flags --generate-code=arch=compute_50,code=sm_50 - Success
-- Checking if nvcc accepts flags --generate-code=arch=compute_52,code=sm_52
-- Checking if nvcc accepts flags --generate-code=arch=compute_52,code=sm_52 - Success
-- Checking if nvcc accepts flags --generate-code=arch=compute_60,code=sm_60
-- Checking if nvcc accepts flags --generate-code=arch=compute_60,code=sm_60 - Success
...
-- The CUDA compiler identification is NVIDIA 12.6.68
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Check for working CUDA compiler: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.6/bin/nvcc.exe - skipped
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done