Issues when building with on windows with Visual studio and CUDA, latest AMReX release 2025 #4867
Description
When trying to build on Windows (after generating with CMake) within Visual Studio a number of compile errors are obtained:
- "For this host platform/dialect, an extended lambda cannot be defined inside the 'if' or 'else' block of a constexpr if statement amrex_3d D:\mfem_testing\amrex_clean\amrex-25.12\Src\Base\AMReX_FBI.H 1261"
I was able to fix this issue by surrounding the "if constexpr" in line 1259 of AMReX_FBI.H by sth like
#ifdef _WIN32
if (...)
#else
if constexpr(...)
#endif
- "no instance of overloaded function "atomicAdd" matches the argument list amrex_3d D:\mfem_testing\amrex_clean\amrex-25.12\Src\Base\AMReX_GpuAtomic.H 143"
The problem here seems to be that the MSVSC compiler tries to compile also unused functions for the template "class BaseFab"
(AMRex_BaseFab.H) including the atomicAdd member functions. This becomes a problem when the following instantiations are created:
i) class TagBox final : public BaseFab in AMReX_TagBox.H
ii) BaseFab<GpuArray<Real,2>> in main.cpp of Tests/CommType
since for these template arguments the CUDA compiler can obviously find no specialization of atomicAdd (as mitigated in AMRex_GpuAtomic.H:143).
This issue can be solved by applying SFINAE to all the atomicAdd variants of the "class BaseFab" via sth like:
template
class BaseFab {
....
template <RunOn run_on AMREX_DEFAULT_RUNON, typename = std::enable_if_t<std::is_arithmetic_v>>
BaseFab& atomicAdd (const BaseFab& x) noexcept;
...
};
and
template
template <RunOn run_on, typename>
BaseFab&
BaseFab::atomicAdd (const BaseFab& x) noexcept {....}
Build configuration:
Windows version: 11 23H2 (10.0.22631 Build 22631.6060)
CUDA Toolkit: 12.8
MSVSC compiler version is 19.40.33818
Visual Studio version is 17.10.10
CMake (GUI): 3.31.10