Building the OpenMOLCAS interface fails due to missing GlobalArrays headers/libraries detection

[behnle]

I tried to build the latest QCMaquis as standalone program with the option BUILD_OPENMOLCAS_INTERFACE. This fails as the cmake script does not test for the availability of the GlobalArrays toolkit and also does not pass the installation directory to the respective include/linker path:

 /home-link/user/builddir_own_modules/qcmaquis_standalone/qcmaquis/dmrg/lib/interfaces/openmolcas>/qcmaquis_interface.f90:1071:25:

 1071 |     Call GA_Brdcst(MT_DBL, [res], storage_size(res)/8, 0)
      |                         1
 Error: Symbol ‘mt_dbl’ at (1) has no IMPLICIT type

QCMaquis was configured as

cmake . -Bobjdir -DBUILD_OPENMOLCAS_INTERFACE="ON" -DTESTS="ON" -DLAPACK_64_BIT="ON" -DBLAS_LAPACK_SELECTOR="openblas" -DGA="ON" -DOPENBLASROOT=/opt/bwhpc/common/group/numlib/openblas/0.3.17-int64-gnu-11.1 -DBOOSTROOT=/opt/bwhpc/common/group/lib/boost/1.78-gnu-11.1

which successfully finishes.

Toolchain:

• GNU GCC 11.1.0
• cmake 3.20.2
• OpenBLAS 0.3.17 compiled with full 64bit interface
• GSL 2.7
• HDF5 1.12.0
• Boost 1.78 (i also had to fiddle with CMakeLists.txt to make cmake recognize my boost installation)
• GlobalArrays 5.8 built against OpenMPI 4.1.2

Full cmake and make outputs are attached. Output was slightly anonymized.

Am i doing something wrong? Is this option only meant to be used in conjunction with an OpenMOLCAS build?

cmake.log make.log

[stknecht]

Dear (@behnle) Stefan, three remarks:

• your -DBOOSTROOT is wrong if I am not mistaken. The variable to pass should be -DBOOST_ROOT. Could that be the problem with the Boost recognition of cmake?
• more importantly: the OpenMolcas interface does NOT support GA outside of OpenMolcas, hence the linking issue. We do not use it at all within QCMaquis and if you ask me, OpenMolcas should not do it either (but that's my personal opinion)...
• what do you mean by "and also does not pass the installation directory to the respective include/linker"? Does that refer to the GA installation.

If you want to run QCMaquis calculations within OpenMolcas, I would recommend to compile a sequential OpenMolcas version. QCMaquis does not yet support MPI threads, that's something my colleague Alberto and I are currently finishing up.

[behnle]

Dear (@stknecht) Stefan,

• more importantly: the OpenMolcas interface does NOT support GA outside of OpenMolcas,

OK, this fully explains the problem. I thought that it might be possible to compile QCMaquis separately and then pass the installation directory to OpenMolcas like a "plugin".

• your -DBOOSTROOT is wrong if I am not mistaken. The variable to pass should be -DBOOST_ROOT. Could that be the problem with the Boost recognition of cmake?

According to https://cmake.org/cmake/help/latest/module/FindBoost.html both variants are possible. The issue with not recognizing boost turned out to be a combination of building boost with --build-type=complete and setting set (Boost_NO_BOOST_CMAKE ON).

• what do you mean by "and also does not pass the installation directory to the respective include/linker"? Does that refer to the GA installation.

I expected cmake to add the GA lib to -L and include to -I. But if i am misusing the -DBUILD_OPENMOLCAS_INTERFACE flag, the fault is on my side. Maybe it could be documented a bit more prominent that this flag is NOT allowed in "standalone" builds. Thanks anyway.

[stknecht]

@behnle : thanks for your feedback. I did not know about the -DBOOSTROOT option. I will make a note to add a warning that the GA option is only meant to be used in connection with OpenMolcas. Just to be clear: you can use -DBUILD_OPENMOLCAS_INTERFACE=ON in a standalone build of QCMaquis (I am using it myself in a different project), but not -DGA=ON.