GROMACS 2021.2 release notes
----------------------------

This version was released on May 5th, 2021. These release notes
document the changes that have taken place in GROMACS since the
previous 2021.1 version, to fix known issues. It also incorporates all
fixes made in version 2020.6 and earlier, which you can find described
in the :ref:`release-notes`.

.. Note to developers!
   Please use """"""" to underline the individual entries for fixed issues in the subfolders,
   otherwise the formatting on the webpage is messed up.
   Also, please use the syntax :issue:`number` to reference issues on GitLab, without the
   a space between the colon and number!

Fixes where mdrun could behave incorrectly
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Removed a potential race condition with GPU update
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

Fixed possible (but so far unobserved) race condition in coordinate copy when
using GPU update with dipole moment calculation.

:issue:`4024`

Avoided issues with global reduction observed with md-vv
""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

The new implementation for md-vv in the modular simulator could
produce floating-point exceptions computing values on non-master ranks
that were never used. This is now fixed by avoiding that
computation. The other integrators were unaffected because they
over-wrote the values computed.

:issue:`4031`

Prohibited SETTLE interactions for atoms with perturbed masses
""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

Older implementations produced varying degrees of wrong results because
this has never been implemented. Now both ``mdrun`` and ``grompp``
refuse to handle such a system, suggesting using normal constraints.

:issue:`3959`

Rerun now writes pull output correctly
""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

Refactoring omitted to preserve that pullf.xvg and pullx.xvg files
should be written during a rerun. All 2019 and 2020 versions were
affected, as well as 2021 and 2021.1. The pull output files are
now written as they used to be in 2018 and earlier.

:issue:`4043`

Fixes for ``gmx`` tools
^^^^^^^^^^^^^^^^^^^^^^^

Fix incorrect behaviour with single residue chains in pdb2gmx
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

The code for chcking for cyclic molecules could lead to single residue chains
incorrectly to be assigned as circular molecules.

:issue:`4029`

Fix grompp check for position restraints with absolute reference
""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

Fixed that grompp with position restraints would always issue a warning about
using an absolute reference, even when an absolute reference was not used.

:issue:`3996`

Fix error when using VMD plugin
"""""""""""""""""""""""""""""""

Tools would crash with a C++ library assertion because the plugin loading
code incorrectly tried to construct a string from nullptr.

:issue:`3055`

Fix file permissions with ``gmx solvate`` and ``gmx genion``
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

These used to write temporary files with Unix permissions 0600. Now
they respect the umask of the process (typically 0644).

:issue:`4040`

Fixes that affect portability
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Support for Intel oneAPI compiler 2021.2
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

Fixed compiler infinity math and MKL flags.

Fix Apple OpenCL build
""""""""""""""""""""""

:issue:`4008`

Fixed compilation issue with GCC 11
""""""""""""""""""""""""""""""""""""

:issue:`4039`

Miscellaneous
^^^^^^^^^^^^^

Fix bond type in GROMOS force fields
""""""""""""""""""""""""""""""""""""

The bond type for C and +N in [ACE] was incorrect.

:issue:`3995`


Allow PME on CPU in runs with domain decomposition and GPU update
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""

Relaxed a limitation which prevented running parallel runs with domain
decomposition and GPU update to use the CPU for PME (as long as combined
PP-PME ranks are used). This allows parallel runs to scale when the CPU
resources are sufficient for PME.

:issue:`4035`