HDF5 conflicts

[KrisThielemans]

we're getting conflicts between HDF5 versions. Easiest route might be to use static linking when building SIRF and ISMRMRD.

[KrisThielemans]

confirmed with @johannesmayer that HDF5 conflicts easily happen even on the VM. We had the system HDF5, the SuperBuild HDF5 and ITK comes with HDF5 as well. Matlab does as well. All different versions. By the way, the GE converter needs to GE Orchestra version of HDF5...

HDF5 library has a habit of causing segfaults when there's a version mismatch anywhere, so this needs to be avoided.

I lean now towards relying on a system HDF5 by default (i.e. to not make the situation worse), and introduce static linking. The latter should be doable by adding -DHDF5_USE_STATIC_LIBRARIES=ON in the External*cmake. I don't think I'd put it in SIRF's CMakeLists.txt.

Anyone any other opinions? @bathomas, @casperdcl ?

[paskino]

Can't we rely only on the one we build?

[KrisThielemans]

we're trying to, but failing :-; I did notice there were a few HDF5_LIBRARIES missing in the External* files. @johannesmayer, maybe you could generate a PR with those fixes to the SuperBuild. However, even then we had conflicts until we deleted all our own HDF5 files, and just relied on the system one. I now think that relying on the system one is best, unless we cannot (e.g. because it's too old).

[KrisThielemans]

check if ITK can be build with a specific HDF5.

change option USE_SYSTEM_HDF5 depending on find_package

[KrisThielemans]

check if ITK can be build with a specific HDF5.

yes: -DITK_USE_SYSTEM_HDF5=ON

[KrisThielemans]

The situation on specifying HDF5 location is actually complicated.

At present we set HDF5_ROOT, HDF5_INCLUDE_DIRS and most of the time HDF5_LIBRARIES. However, if USE_SYSTEM_HDF5=OFF, HDF5_LIBRARIES is actually empty.

According to the FindHDF5 doc, only HDF5_ROOT should be used and the others will be ignored. However, with CMake 3.7.2, setting the others actually prepends them to the final version seen after find_package(HDF5), therefore taking precedence (which I guess is good, although unnecessary). I think with more recent CMake, those variables might be ignored, but I haven't checked properly.

Because of all this, I suggest removing the explicit setting of HDF5_INCLUDE_DIRS and HDF5_LIBRARIES.

However, this seems to mean that with current master, we do set HDF5_ROOT everywhere ok, except for ITK. But ITK by default builds its own HDF5, which one would hope doesn't conflict with anything else (libraries are called differently, e.g. libitkhdf5.a, and include files should only be found when using itkhdf5/hdf5.h etc). I haven't been able yet to let anything in SIRF depend on an itkhdf5 library.

All in all, it's a mistery why @johannesmayer had HDF5 conflicts, and that's of course worrying (as it means I don't know how to fix it).

@johannesmayer, do you remember what depended on the itkhdf5 library?

[KrisThielemans]

I've created #143 for this. I'm not 100% sure if this is a good idea though. On the other hand, we might only see trouble once using ITK to read HDF5 files.

[KrisThielemans]

this seems largely obsolete. We still have some HDF5 problems, but there is likely no tmuch here that is relevant.