omp module logic troubled by newer cmake about autocmake

[loriab]

I'm going with

    set(_omp_enabled_for_enabled_lang OFF)

    if(DEFINED CMAKE_C_COMPILER_ID)
        if(OPENMP_FOUND OR OpenMP_C_FOUND)
            set(_omp_enabled_for_enabled_lang ON)
            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
        endif()
    endif()

    if(DEFINED CMAKE_CXX_COMPILER_ID)
        if(OPENMP_FOUND OR OpenMP_CXX_FOUND)
            set(_omp_enabled_for_enabled_lang ON)
            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
        endif()
    endif()

    if(DEFINED CMAKE_Fortran_COMPILER_ID)
        if(OPENMP_FOUND OR OpenMP_Fortran_FOUND)
            set(_omp_enabled_for_enabled_lang ON)
            set(CMAKE_Fortran_FLAGS "${CMAKE_Fortran_FLAGS} ${OpenMP_Fortran_FLAGS}")
        endif()
    endif()

    if(${_omp_enabled_for_enabled_lang})
        add_definitions(-DHAVE_OPENMP)
    endif()

[bast]

Thanks Lori! I will try to reproduce this.

[bast]

I believe there is a problem but at this moment cannot reproduce it with CMake 3.9.6 and GNU 8.1.0. I probably use a different compiler. Using which compilers do you observe this?

[loriab]

I'm on cmake 3.9.4 and gcc 7.2.0 (from conda). conda gfortran isn't providing omp_lib header or module, so that's probably why kitware findopenmp is failing to detect for Fortran, which starts the cascade.

[bast]

OK thanks. It would be nice if we could query OpenMP_<lang>_FOUND but these were only introduced in 3.9.

[bast]

We could for each language check whether either OpenMP_<lang>_FOUND or OpenMP_FOUND. Also the add_definitions(-DHAVE_OPENMP) seems kind of ugly from the modern CMake perspective. What is your opinion @robertodr?

[loriab]

Yes, I was just going to ask if there'd be forseeable trouble with if(${OPENMP_FOUND} OR ${OpenMP_C_FOUND}) per-language? I don't know what cmake wants in the way of quotes these days.

[bast]

I expect no trouble. Not sure what to do with the add_definitions but independently of it the current problem could be solved with:

if(DEFINED CMAKE_C_COMPILER_ID)
  if(OPENMP_FOUND OR OPENMP_C_FOUND)
    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
  endif()
endif()

# etc. for the other languages

On the other hand the "proper" way for 3.9 and above is actually this:
https://github.com/bast/cmake-recipes/blob/master/chapter-03/recipe-05/cxx-example/CMakeLists.txt

[robertodr]

The add_definitions should stay as-is, I think. @bast did copying the modern FindOpenMP.cmake module work for OpenMP detection? I think that would actually be the most reasonable choice...

[bast]

You mean shipping the latest FindOpenMP.cmake with autocmake and checking that it works also with 2.8? Just to make sure I understand the question.

[loriab]

I'll experiment with those lines, thanks.

I'm a huge fan of the imported target in general. But I don't dare do it here b/c of the problem I was discussing with @robertodr about needing to suppress lgomp linking for GCC+MKL (this is also the problem that led to the nasty table here. So basically I want the flags out of FindOpenMP but not the link libs.

[robertodr]

You can still get the flags out from the new module (though that might be on its way to deprecation)

[loriab]

By as-is on the definitions, you mean set when C found, @robertodr? Or potentially added three times if all langs enabled?

[robertodr]

Gee, what a mess! No we definitely don't want that, so I guess either they spill over to the client:

if(OpenMP_FOUND)
  target_compile_definitions(<target-name>
     PUBLIC
        HAVE_OPENMP
     )
endif()
# or a generator expression
target_compile_definitions(<target-name>
  PUBLIC
    "$<$<BOOL:${OPENMP_FOUND}>:HAVE_OPENMP>"
  )

or it spills over to the compiler flags:

set(CXX_FLAGS ${OpenMP_CXX_FLAGS} -DHAVE_OPENMP)

quite ugly in both cases, but I'd rather prefer the former: explicit is better than implicit (of sorts) and you get to choose the name of the preprocessor definition!

@bast, yes that's what I meant. I recall it didn't work out for FindMPI.cmake. (This thread is moving very fast, I'm losing half of the comments 😸 )

[robertodr]

Definitely more compact.

[loriab]

Well, 7 lines where there used to be 1. But I don't see another dumb way around it (classifying your sol'ns as "smart" cmake uses).

[robertodr]

Don't know it mine was "smart". Yours is compact wrt spilling variables over to other CMake files.

[bast]

For more compactness we can add a foreach looping over the languages. I think we should rename this module to omp-flags since it's really only about flags. I will test backwards compatibility of recent find module and sleep on this issue and then tomorrow comment more and/or send a PR based on our findings.

[bast]

Using latest FindOpenMP.cmake with older versions of CMake does not work. Already 3.5 chokes on it.

[loriab]

Don't consider Psi4 as a an impediment to bumping min supported cmake version.

For one, all the repos have their own copies of the autocmake file. For second, someday when need is pressing and the blas/omp detection landscape is cleaner, I'll probably bump the p4 ecosystem to 3.9 or whereever all the Intel/OMP/MPI goodies are. conda is distributing 3.11 now, so access not a problem.

[bast]

We (the editorial we :-) are not super sure yet which minimum we want to support (#243) but until we settle that, probably as far back as 2.8 and mostly because of more conservative codes than Psi4. I will send a PR and let you both review it and we go from there.

[loriab]

FYI, cmake 3.11 (but not 3.9) gets around my initial problem by letting you specify a subset of the enabled langs to search for find_package(OpenMP COMPONENTS C CXX) and thus still merit the overall OpenMP_FOUND.

[loriab]

3.10 is the exact introduction date of the components mentioned above.

fyi, libraries returned in the OpenMP::OpenMP_CXX target for gcc are gomp;pthreads and for icpc are iomp5;pthreads, along with corresponding lang-specific COMPILE_FLAGS

 Properties for TARGET OpenMP::OpenMP_CXX:
   OpenMP::OpenMP_CXX.INTERFACE_COMPILE_OPTIONS = "$<$<COMPILE_LANGUAGE:CXX>:-fopenmp>"
   OpenMP::OpenMP_CXX.INTERFACE_LINK_LIBRARIES = "/home/psilocaluser/toolchainconda/envs/p4dev36/x86_64-conda_cos6-linux-gnu/sysroot/lib/libgomp.so;/home/psilocaluser/toolchainconda/envs/p4dev36/x86_64-conda_cos6-linux-gnu/sysroot/usr/lib/libpthread.so"

and all that can be copied into a new target via

    include(CMakePrintHelpers)
    if(TARGET OpenMP::OpenMP_CXX)
        add_library(omp_cxx INTERFACE)
        get_property(_ill TARGET OpenMP::OpenMP_CXX PROPERTY INTERFACE_LINK_LIBRARIES)
        get_property(_ico TARGET OpenMP::OpenMP_CXX PROPERTY INTERFACE_COMPILE_OPTIONS)
        set_property(TARGET omp_cxx PROPERTY INTERFACE_COMPILE_OPTIONS "${_ico}")
        set_property(TARGET omp_cxx PROPERTY INTERFACE_LINK_LIBRARIES "${_ill}")
        cmake_print_properties(TARGETS OpenMP::OpenMP_CXX omp_cxx
                               PROPERTIES INTERFACE_COMPILE_DEFINITIONS INTERFACE_COMPILE_OPTIONS INTERFACE_COMPILE_FEATURES INTERFACE_INCLUDE_DIRS INTERFACE_LINK_LIBRARIES)

        message("FL: ${OpenMP_CXX_FLAGS}")
        message("LN: ${OpenMP_CXX_LIB_NAMES}")
        message("LI: ${OpenMP_CXX_LIBRARY}")
        message("LS: ${OpenMP_CXX_LIBRARIES}")
        # FL: -fopenmp
        # LN: gomp;pthread
        # LI: 
        # LS: /home/psilocaluser/toolchainconda/envs/p4dev36/x86_64-conda_cos6-linux-gnu/sysroot/lib/libgomp.so;/home/psilocaluser/toolchainconda/envs/p4dev36/x86_64-conda_cos6-linux-gnu/sysroot/usr/lib/libpthread.so
    endif()

[bast]

Thank you for the information! I will revise #248.

[loriab]

I pinged you on psi4/psi4#1031 and would be glad of any comments on that new OpenMP-finding scheme. Most everything is centered in https://github.com/loriab/psi4/tree/openblas/external/common/lapack . The FindTargetOpenMP.cmake file within is most relevant to this autocmake issue. It finds an OpenMP target with modern cmake or constructs one to compensate in old cmake. I know autocmake doesn't deal much with targets, but the usual variables can just read off the target. Also the psi one is CXX only, so extensions would need be be made for other langs.

[bast]

Thanks! I will look at it. Also I will finally restart work on this issue hopefully today.