CMake: add HDF5 support

[mic84]

No description provided.

[mic84]

@eschnett can please try this out?

[eschnett]

This fails, for reasons probably unrelated to your work, but due to other changes. I receive this error message:

COMPILING CarpetX/CarpetX/src/io_silo.cxx
/__w/1/s/Cactus/arrangements/CarpetX/CarpetX/src/io_silo.cxx: In function 'bool CarpetX::operator==(const CarpetX::mesh_props_t&, const CarpetX::mesh_props_t&)':
/__w/1/s/Cactus/arrangements/CarpetX/CarpetX/src/io_silo.cxx:44:25: error: ambiguous overload for 'operator==' (operand types are 'std::tuple<amrex::IntVect>' and 'std::tuple<amrex::IntVect>')
   44 |     return p.to_tuple() == q.to_tuple();
      |            ~~~~~~~~~~~~ ^~ ~~~~~~~~~~~~
      |                      |               |
      |                      tuple<[...]>    tuple<[...]>
In file included from /usr/include/c++/9/functional:54,
                 from /usr/include/c++/9/pstl/glue_algorithm_defs.h:13,
                 from /usr/include/c++/9/algorithm:71,
                 from /__w/1/s/Cactus/arrangements/CarpetX/CarpetX/src/loop.hxx:8,
                 from /__w/1/s/Cactus/arrangements/CarpetX/CarpetX/src/driver.hxx:4,
                 from /__w/1/s/Cactus/arrangements/CarpetX/CarpetX/src/io_silo.cxx:3:
/usr/include/c++/9/tuple:1419:5: note: candidate: 'constexpr bool std::operator==(const std::tuple<_Tps ...>&, const std::tuple<_Elements ...>&) [with _TElements = {amrex::IntVect}; _UElements = {amrex::IntVect}]'
 1419 |     operator==(const tuple<_TElements...>& __t,
      |     ^~~~~~~~
In file included from /usr/local/include/AMReX_Gpu.H:29,
                 from /usr/local/include/AMReX_Reduce.H:5,
                 from /usr/local/include/AMReX_BaseFab.H:32,
                 from /usr/local/include/AMReX_TagBox.H:10,
                 from /usr/local/include/AMReX_AmrMesh.H:12,
                 from /usr/local/include/AMReX_AmrCore.H:8,
                 from /__w/1/s/Cactus/arrangements/CarpetX/CarpetX/src/driver.hxx:12,
                 from /__w/1/s/Cactus/arrangements/CarpetX/CarpetX/src/io_silo.cxx:3:
/usr/local/include/AMReX_GpuAllocators.H:141:5: note: candidate: 'bool amrex::operator==(const Allocator<T>&, const Allocator<U>&) [with Allocator = std::tuple; T = amrex::IntVect; U = amrex::IntVect]'
  141 |     operator==(Allocator<T> const&, Allocator<U> const&) noexcept
      |     ^~~~~~~~

It seems that AMReX provides an operator== that is too generic, and which overlaps with other definitions (AMReX_GpuAllocators.H:141):

    template <template <typename> class Allocator, class T, class U>
    bool
    operator==(Allocator<T> const&, Allocator<U> const&) noexcept
    {
        return true;
    }

I think this operator accepts essentially any templated class, and hence leads to ambiguities in many cases.

[WeiqunZhang]

I have seen operator overloading ambiguity (on a different type in a different code) with gcc-9.3.0 and c++17 or c++2a. gcc-9 with c++14 worked. gcc <9 and gcc 10 also worked. I believe it was a compiler bug in that case. Could try a different compiler or different standard if you are using C++17.

[WeiqunZhang]

Sorry, I didn't read the error message carefully. I think you are right. It is our fault.

[eschnett]

For the record, I'm using GCC 10 with C++ 17.

[WeiqunZhang]

The operator is fixed in #1050, I believe.

[mic84]

@eschnett can you try and let us know if it works?

[eschnett]

Trying it... I see that AMReX requires at least HDF5 1.12. At the same time, our code is stuck at 1.10 for some compatibility reason. (Or maybe that's just Ubuntu 20.04?) I will check and/or update our code, but that might take some time.

Is this version requirement new? It wasn't there yesterday.

[atmyers]

Let me double-check the minimum version required, it may be that we can relax this.

[atmyers]

It seems that we could drop this to at least 1.10.4.

[eschnett]

1.10.4 is what we currently use.

[mic84]

I modified CMake to accept HDF5 >= 1.10.4. Please try again.

[eschnett]

Another build error:

[ 46%] Building CXX object Src/CMakeFiles/amrex.dir/Base/AMReX_GpuControl.cpp.o
/cactus/src/amrex/Src/Base/AMReX_PlotFileUtil.cpp: In function 'void amrex::WriteMultiLevelPlotfileHDF5(const string&, int, const amrex::Vector<const amrex::MultiFab*>&, const amrex::Vector<std::__cxx11::basic_string<char> >&, const amrex::Vector<amrex::Geometry>&, amrex::Real, const amrex::Vector<int>&, const amrex::Vector<amrex::IntVect>&, const string&, const string&, const string&, const amrex::Vector<std::__cxx11::basic_string<char> >&)':
/cactus/src/amrex/Src/Base/AMReX_PlotFileUtil.cpp:765:9: error: 'H5Pset_fapl_mpio' was not declared in this scope; did you mean 'H5Pset_fapl_stdio'?
  765 |         H5Pset_fapl_mpio(fapl, MPI_COMM_SELF, MPI_INFO_NULL);
      |         ^~~~~~~~~~~~~~~~
      |         H5Pset_fapl_stdio

Do you require an HDF5 library that is configured with MPI? I'm currently using a stock serial Ubuntu HDF5 library.

[atmyers]

We do currently require a parallel hdf5 library. @mic84 is there a way we can check and enforce that in the build system?

[mic84]

@atmyers yes. Just submitted a commit to do just that. @eschnett can you try again one more time with both a parallel and serial hdf5 build? The serial build should now give you an error at configuration time, while the parallel build should be linked fine.

[eschnett]

The next build error I encounter is in the Silo library (again – that's what constraining me to HDF5 1.10.x), which doesn't handle a parallel HDF5 library. When using parallel HDF5, then Silo requires HDF5 <=1.8.12. I'm going to have to make some choices here.

[mic84]

Unfortunately, I am afraid that it cannot be fixed on the AMReX side. If you have no objections, I will merge this PR since I think that now everything is fine as far as AMReX is concerned.

[eschnett]

Please do.

-erik

On Wed, Jun 24, 2020 at 18:13 mic84 ***@***.***> wrote: Unfortunately, I am afraid that it cannot be fixed on the AMReX side. If you have no objections, I will merge this PR. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1048 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABUECSTFHM4YXWV6I6G6VDRYJ3APANCNFSM4OGFCSTA> .

-- Erik Schnetter <schnetter@gmail.com> http://www.perimeterinstitute.ca/personal/eschnetter/

[ax3l]

From a very quick check, I think AMReX also implements the C API - also since the C++ bindings of HDF5 do not support parallel I/O (yet).
(This is not the same CMake component naming convention as in FindMPI here.)

This is what I do in openPMD-api (with a different target name):

find_package(HDF5 1.10.4 REQUIRED COMPONENTS C)

# ...
target_link_libraries(amrex PUBLIC ${HDF5_LIBRARIES})
target_include_directories(amrex SYSTEM PUBLIC ${HDF5_INCLUDE_DIRS})
target_compile_definitions(amrex PUBLIC ${HDF5_DEFINITIONS})

Note that the C++ bindings are also not built by default with HDF5.