Open mathomp4 opened 1 year ago
Oh nuts. It looks like h5fortran
assumes HDF5 is built with CMake. I think this now means I need to CMake-ize HDF5, netCDF, netCDF-Fortran... Oy. Well, not impossible, but it'll probably mean a new major version as this would be a sort of "API change" to how to use Baselibs (i.e., find libraries in a new place).
It looks like h5fortran assumes HDF5 is built with CMake.
Is this true? I have both a system-provided HDF5 built with GCC and I build my own HDF5 with Autotools and Intel compiler. In both cases I can build h5fortran (as a neural-fortran dependency) with CMake. h5fortran does have some CMake magic inside to build the HDF5 for you if it doesn't find it.
The mechanism to detect HDF5 is defined in h5fortran's CMakeLists. It can find my OS-provided GCC build of HDF5 without additional information from me. For the Intel build of HDF5 from source, I need to pass HDF5_ROOT
variable to CMake so that h5fortran can find it.
@milancurcic Oooh. really? Okay. I'll see if I can figure it out. That'll make life easy for me if so!
This is a tracking issue to adding neural-fortran to Baselibs.
The main blocker at the moment is that neural-fortran uses
FetchContent
when building with CMake and that is sort of anti-compute node I think. I have an issue on their repo (https://github.com/modern-fortran/neural-fortran/issues/128) where I lay out some thoughts I have on how we might get this to work (though with the help of @milancurcic for sure!).