USER-INTEL: Wrong energy for pe/atom when running LAMMPS with "-sf intel" flag

Dear fellow LAMMPS experts,

I recently compiled LAMMPS (stable release 12Dec2018) for our cluster with
the USER-INTEL package added to the installation (Intel 17.0.2 compiler
+ MKL, OpenMPI 3.1.2)

I ran a small test calculation (Inputfile and used potential are
attached to this mail) which
returned results identical to an older LAMMPS version installed on the
cluster.
However, enabling the USER-INTEL package, i.e. adding the `-sf intel`
flag to the run command,
pe/atom produces a very wrong energy and the forces after the relaxation
differ as well
(see the two attached stdout-logs for the "default" run without any
flags and the "intel" run with `-sf intel` set)

Is there something I am missing? I know from the docs that calculating
stress/atom is not possible
using the USER-INTEL package, does this also apply to pe/atom?

Thanks for your help!

Cheers,
Andreas

Al_fcc_cohesive_energy_alat.in (1.86 KB)

Al1.eam.fs (739 KB)

default_run_stdout.log (2.95 KB)

intel_run_stdout.log (2.99 KB)

Dear fellow LAMMPS experts,

I recently compiled LAMMPS (stable release 12Dec2018) for our cluster with
the USER-INTEL package added to the installation (Intel 17.0.2 compiler

  • MKL, OpenMPI 3.1.2)

I ran a small test calculation (Inputfile and used potential are
attached to this mail) which
returned results identical to an older LAMMPS version installed on the
cluster.
However, enabling the USER-INTEL package, i.e. adding the -sf intel
flag to the run command,
pe/atom produces a very wrong energy and the forces after the relaxation
differ as well
(see the two attached stdout-logs for the “default” run without any
flags and the “intel” run with -sf intel set)

Is there something I am missing? I know from the docs that calculating
stress/atom is not possible
using the USER-INTEL package, does this also apply to pe/atom?

yes

Thanks for the quick response!

Well that’s good to know! Maybe it would be good to have pe/atom raise an error if used with -sf intel (i think stress/atom does this) I’ll open an issue for this on github.

Thanks for clearing that up!