avoiding SVE instructions for GROMACS <= 2024.1 on Neoverse V1

[boegel]

We're configuring GROMACS <= 2024.1 with -DGMX_SIMD=ARM_NEON_ASIMD when building for aarch64/neoverse_v1 (done in EasyBuild hooks via #499), based on feedback received from the GROMACS developers after reporting test failures, see https://gitlab.com/gromacs/gromacs/-/issues/5057

Apparently the test failures are genuine and indicate a bug in the SVE support in GROMACS, which we circumvent by disabling those SVE-specific optimizations.

[mabraham]

A fix for GROMACS is available for testing at https://gitlab.com/gromacs/gromacs/-/merge_requests/4299

[trz42]

We've seen some issues with GROMACS 2024.4 with CUDA for NVIDIA Grace/Hopper. See #999

• in PR 999 we applied the change we previously used for neoverse_v1, i.e., adding the config option -DGMX_SIMD=ARM_NEON_ASIMD, and then no tests failed

  • running GROMACS demo (eessi-demo/gpu/GROMACS)

    GROMACS with NEON+CUDA: performance on GH node

    {EESSI 2023.06} [me@gh-node GROMACS]$ ./run-gpu.sh
                      :-) GROMACS - gmx mdrun, 2024.4 (-:
    
    Executable:   /cvmfs/software.eessi.io/versions/2023.06/software/linux/aarch64/nvidia/grace/accel/nvidia/cc90/software/GROMACS/2024.4-foss-2023b-CUDA-12.4.0/bin/gmx
    Data prefix:  /cvmfs/software.eessi.io/versions/2023.06/software/linux/aarch64/nvidia/grace/accel/nvidia/cc90/software/GROMACS/2024.4-foss-2023b-CUDA-12.4.0
    Working dir:  .../eessi-demo/gpu/GROMACS
    Command line:
    gmx mdrun -s ion_channel.tpr -maxh 0.50 -resethway -noconfout -nsteps 10000 -g logfile
    
    Compiled SIMD is ARM_NEON_ASIMD, but ARM_SVE might be faster (see log).
    Reading file ion_channel.tpr, VERSION 5.1.4 (single precision)
    Note: file tpx version 103, software tpx version 134
    Overriding nsteps with value passed on the command line: 10000 steps, 25 ps
    Changing nstlist from 10 to 100, rlist from 1 to 1.169
    
    
    Update groups can not be used for this system because there are three or more consecutively coupled constraints
    
    1 GPU selected for this run.
    Mapping of GPU IDs to the 2 GPU tasks in the 1 rank on this node:
      PP:0,PME:0
    PP tasks will do (non-perturbed) short-ranged interactions on the GPU
    PP task will update and constrain coordinates on the CPU
    PME tasks will do all aspects on the GPU
    Using 1 MPI thread
    
    Non-default thread affinity set, disabling internal thread affinity
    
    Using 72 OpenMP threads
    
    starting mdrun 'Protein'
    10000 steps,     25.0 ps.
    
    step 5000: resetting all time and cycle counters
    
               Core t (s)   Wall t (s)        (%)
       Time:     4795.567       66.630     7197.3
                 (ns/day)    (hour/ns)
    Performance:       16.212        1.480
    
    GROMACS reminds you: "There was no preconception on what to do" (Daft Punk)
    
    
    real    2m21.982s
    user    1m47.884s
    sys     0m29.845s
    

Without CUDA the same GROMACS version, same toolchain (foss/2023b, GCC 13.2.0) resulted in no failing tests. See #993

• in PR 993 we did use SVE instructions (and no tests failed)

  • running GROMACS demo (eessi-demo/gpu/GROMACS)

    GROMACS with SVE: performance on GH node

    {EESSI 2023.06} [me@gh-node GROMACS]$ ./run-cpu.sh
              :-) GROMACS - gmx mdrun, 2024.4-EasyBuild_4.9.4 (-:
    
    Executable:   /cvmfs/software.eessi.io/versions/2023.06/software/linux/aarch64/nvidia/grace/software/GROMACS/2024.4-foss-2023b/bin/gmx
    Data prefix:  /cvmfs/software.eessi.io/versions/2023.06/software/linux/aarch64/nvidia/grace/software/GROMACS/2024.4-foss-2023b
    Working dir:  .../eessi-demo/gpu/GROMACS
    Command line:
    gmx mdrun -s ion_channel.tpr -maxh 0.50 -resethway -noconfout -nsteps 10000 -g logfile
    
    Reading file ion_channel.tpr, VERSION 5.1.4 (single precision)
    Note: file tpx version 103, software tpx version 134
    Overriding nsteps with value passed on the command line: 10000 steps, 25 ps
    Changing nstlist from 10 to 50, rlist from 1 to 1.093
    
    
    Update groups can not be used for this system because there are three or more consecutively coupled constraints
    
    Using 72 MPI threads
    
    Non-default thread affinity set, disabling internal thread affinity
    
    Using 1 OpenMP thread per tMPI thread
    
    starting mdrun 'Protein'
    10000 steps,     25.0 ps.
    ... still running ...
    
    -------------------------------------------------------
    Program:     gmx mdrun, version 2024.4-EasyBuild_4.9.4
    Source file: src/gromacs/mdlib/resethandler.cpp (line 161)
    MPI rank:    62 (out of 72)
    
    Fatal error:
    PME tuning was still active when attempting to reset mdrun counters at step
    41. Try resetting counters later in the run, e.g. with gmx mdrun -resetstep.
    
    For more information and tips for troubleshooting, please check the GROMACS
    website at https://manual.gromacs.org/current/user-guide/run-time-errors.html
    -------------------------------------------------------
    
    real    22m6.133s
    user    21m53.362s
    sys     0m0.010s