[Trilinos-Users] [Kokkos-Developers] [EXTERNAL] [12.2.1] MueLu/Kokkos OpenMP configuration/macros issue

Hoemmen, Mark mhoemme at sandia.gov
Sun Oct 18 19:25:24 EDT 2015

As far as I can tell, MueLu only checks Tpetra_INST_OPENMP.  I don't see
that MueLu uses Kokkos::OpenMP explicitly.


On 10/18/15, 12:55 AM, "Christian Trott" <crtrott at sandia.gov> wrote:

>Oh ok. That almost surely means one of two things:
>(i) MueLu is using Kokkos::OpenMP explicitly instead of
>(ii) Something screwed up the configuration whereas KOKKOS_HAVE_OPENMP
>is defined without enabling openmp in the compiler.
>MueLU would only check the KOKKOS_HAVE_OPENMP flag, which is probably
>the right thing to do.
>What we should add in Kokkos is a more meaningfull error (and early
>exit) if KOKKOS_HAVE_OPENMP is defined but _OPENMP is not.
>On 10/18/2015 12:30 AM, Denis Davydov wrote:
>> It¹s not that much about the usage of OpenMP but rather about
>>inconsistencies inside packages and their configuration when OpenMP=ON,
>> which in turn leads to compiler errors when using trilinos.
>> Roughly, It seems that Kokkos was clever to disable OpenMP when user
>>did not provide those compiler flags("-fope nmp"), whereas MueLu still
>>thinks that OpenMP is enabled
>> and thus the compiler error.
>> Denis.
>>> On 18 Oct 2015, at 08:18, Christian Trott <crtrott at sandia.gov> wrote:
>>> If what you suggests fixes it then you tried to compile with the
>>>OpenMP backend, without enabling OpenMP in the compiler (i.e. what Si
>>>indirectly suggested).
>>> We do not allow you to use the OpenMP backend if OpenMP is not enabled
>>>in the compiler. By removing our guards you circumvent that. While this
>>> that you might actually get your code to compiler, it is unlikely that
>>>it will run correctly.
>>> Christian
>Kokkos-Developers mailing list
>Kokkos-Developers at software.sandia.gov

More information about the Trilinos-Users mailing list