Serial unit test failure

[ibaned]

@prwolfe wrote the following as trilinos/Trilinos#2793 :

Kokkos unit test failing with MPICH

When building with MPICH3.2 I see a failure in KokkosCore_UnitTest_Serial.exe (the serial.cxx test) due to comparison of two doubles. It looks like they are off by less than 10-9 and I'm not sure what if any tolerance is acceptable in this case.

@trilinos/kokkos

Expectations

This test should pass with all mpi variants

Current Behavior

This test passes in openmpi builds

Steps to Reproduce

To recreate this you can use the PR template we are working on. I did attempt to recreate this In a kooks stand-alone build without sucess.

source ${TrilinosSource}/cmake/std/sems/PullRequestIntel17.0.1TestingEnv.sh
cmake -C ${TrilinosSource}/cmake/std/PullRequestLinuxIntelTestingSettings.cmake -DTrilinos_ENABLE_Kokkos=ON ${TrilinosSource}
cd pacakges/kokkos/core/unit_tests
make

[prwolfe]

Dan, Did I miss how to recreate this in a direct kokkos build? I didn’t want to submit to the kokkos issue list without being able to reproduce it there. Paul From: Dan Ibanez <notifications@github.com> Reply-To: kokkos/kokkos <reply@reply.github.com> Date: Tuesday, May 22, 2018 at 8:03 AM To: kokkos/kokkos <kokkos@noreply.github.com> Cc: "Wolfenbarger, Paul R" <prwolfe@sandia.gov>, Mention <mention@noreply.github.com> Subject: [EXTERNAL] [kokkos/kokkos] Serial unit test failure (#1632) @prwolfe<https://github.com/prwolfe> wrote the following as trilinos/Trilinos#2793<trilinos/Trilinos#2793> : Kokkos unit test failing with MPICH When building with MPICH3.2 I see a failure in KokkosCore_UnitTest_Serial.exe (the serial.cxx test) due to comparison of two doubles. It looks like they are off by less than 10-9 and I'm not sure what if any tolerance is acceptable in this case. Expectations This test should pass with all mpi variants Current Behavior This test passes in openmpi builds Steps to Reproduce To recreate this you can use the PR template we are working on. I did attempt to recreate this In a kooks stand-alone build without sucess. source ${TrilinosSource}/cmake/std/sems/PullRequestIntel17.0.1TestingEnv.sh cmake -C ${TrilinosSource}/cmake/std/PullRequestLinuxIntelTestingSettings.cmake -DTrilinos_ENABLE_Kokkos=ON ${TrilinosSource} cd pacakges/kokkos/core/unit_tests make — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#1632>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AHTpCRKe3pIdaXkw5slCamXEhBCtfnl5ks5t1BqTgaJpZM4UItc6>.

[ibaned]

@prwolfe I think I fixed this will pull request #1729. It doesn't have so much to do with MPICH as opposed to how the compiler optimizes to different ways of computing the same number. This fix should show up in the next Kokkos release.

[prwolfe]

Thanks @ibaned ! I will watch for it.