GROMACS 2023.1 release notes#
This version was released on April 21st, 2023. These release notes document the changes that have taken place in GROMACS since the previous 2023 version, to fix known issues. It also incorporates all fixes made in version 2022.5 and earlier, which you can find described in the Release notes.
Fixes where mdrun could behave incorrectly#
Parallelization of TPI an normal modes working again#
When running TPI or normal mode analysis with multiple MPI ranks mdrun would exit with an assertion failure.
The AWH metric could be incorrect for free-energy lambda dimensions#
When different lambda components had different values for the same lambda point index, the AWH metric used dH/dlambda as input that used the derivative with respect to all lambda components. Note that this only affected the metric, not the sampling nor the free energy values.
Fix checkpointing of expanded ensemble simulations with domain decomposition#
Expanded-ensemble simulations can now restart from a checkpoint when running multiple PP ranks.
Fix PME pipelining support in SYCL#
When PME pipelining was used, long-range PME electrostatics were producing incorrect results in SYCL.
Only runs with >=3 GPUs and with direct GPU communication enabled
(GMX_ENABLE_DIRECT_GPU_COMM env. var.) are affected.
Fix checkpointing of AWH friction metric for dimensions > 1.#
The friction metric checkpoint i/o was wrong for dimensions > 1. This did not affect the AWH PMF or sampling, but would result in nonsense if the AWH friction tensors were used to calculate the diffusion in dimensions > 1.
Fixes for gmx tools#
Fix crash in gmx solvate when using solvent box in PDB format#
Now a PDB file can be passed to the -cs option in gmx solvate.
In previous releases (since at least 2016) this lead to a segfault.
Fix creating index file from another index file#
gmx make_ndx can again accept an index file alone as input, without associated structure file.
Allow selection of energy term by full name in gmx energy#
It is now possible to select energy terms by full name. This is specifically helpful for terms starting with a number, such as “1/Viscosity” or “2CosZ*Vel-X”, which could previously only be selected reliably by number.
Fix early crash in gmx anaeig#
An internal change in GROMACS 2023 caused improper handling of optional program arguments leading to a crash in the program. This might have affected some other analysis tools.
Fixes that affect portability#
Fixed GMX_USE_TNG=off build#
GROMACS can again be built without TNG support.
Fixed abnormal termination during gmx startup#
GROMACS made a call to std::filesystem::equivalent in a less than
perfectly robust manner. This caused gmx to stop execution
when linked against the (atypical) libc++ standard library when
the build directory no longer existed.
Fixed CPU FFT with MKL 2023.0#
Previously, GROMACS would fail during the initialization of CPU FFT when it was compiled with oneMKL 2023.0. This is fixed now.
Miscellaneous#
Workaround for strange compiler behavior to improve SYCL bonded kernel performance#
For some SYCL targets (most notably, hipSYCL for AMD GPUs with ROCm 5.x), a very inefficient code was generated for bonded kernels. Now, bonded force calculation on GPU is expected to be up to 3 times faster.
Restored OpenMP acceleration of pulling routines#
During internal code reorganization, OpenMP acceleration was accidentally disabled for pulling force calculation in GROMACS 2023. This is now fixed.
Added support for new cuFFTMp interface#
The interface to the cuFFTMp library has changed with its latest release in the the NVIDIA HPC SDK version 23.3, which is required for NVIDIA Hopper GPU support. We have now added default support to the new interface, while retaining support for the legacy interface.
Document workaround when MPI detection fails#
MPI is an optional dependency of gmxapi even when building GROMACS without support for an MPI library. CMake’s mechanism to find MPI can choke on broken MPI installations in ways that could be confusing. Now a work-around is documented for the convenience of a user who was not intending to use MPI.