Paul Bauer [Thu, 7 Jan 2021 13:08:41 +0000 (13:08 +0000)]
Write copyright output files as utf8
The script could fail otherwise when encodings didn't match.
Kevin Boyd [Tue, 5 Jan 2021 13:28:26 +0000 (13:28 +0000)]
Fix broken random seed generation.
On Ryzen 3k CPUs with the RDRAND microcode bug, the
std::mt19937_64 alternate path yields the same seed
on multiple calls that fall within the same second. Changed
the "seed" of the random seed to rescramble each time the
makeRandomSeed() function is called.
Fixes #3859
Szilárd Páll [Wed, 30 Dec 2020 12:27:34 +0000 (13:27 +0100)]
Bump version post-release to rc2
Szilárd Páll [Wed, 30 Dec 2020 16:28:34 +0000 (16:28 +0000)]
GROMACS 2021-rc1 release candidate
First release candidate.
Mark Abraham [Wed, 23 Dec 2020 13:44:37 +0000 (13:44 +0000)]
Deprecate .pdo file reading
Analysing such files is likely not relevant a decade later, and such
needs can be met by using an older version of GROMACS that did still
support reading these formats.
Fixes #3818
Alan Gray [Fri, 18 Dec 2020 07:45:10 +0000 (23:45 -0800)]
Stage bonded kernel atomics through shared memory
Fixes performance bug introduced in
01b2f20bd5 by staging energy step
atomics through shared memory rather than have all threads write
atomically directly to global memory.
Fixes #3443
Mark Abraham [Thu, 17 Dec 2020 11:10:35 +0000 (11:10 +0000)]
Update CI image to OneAPI 2021.1.1, add ICC tests.
Also, backport a change from MR !796 that silences the false-positive ICC
compiler warning.
Mark Abraham [Tue, 15 Dec 2020 17:01:19 +0000 (17:01 +0000)]
Update deprecation statuses
Various functionality has been agreed as deprecated at #3818 and
elsewhere. Those are now noted in the release notes, documentation,
and CMake status messages. The changes are all minor so there is
little need to make mdrun report when such configurations are used.
The deprecation content for release 2020 (and 2021) series is updated
to include the relevant content from release 2019 series. Comments
have been added to help us to this better in future.
Fixes #3818
Simon Christ [Mon, 30 Nov 2020 10:37:00 +0000 (10:37 +0000)]
Correct signs of lj and coulomb forces
Pascal Merz [Fri, 11 Dec 2020 13:07:23 +0000 (13:07 +0000)]
Restrict force frame comparison for FEP tests
Closes #3741.
Pascal Merz [Fri, 11 Dec 2020 12:22:44 +0000 (12:22 +0000)]
Fix trx frame reading from modular simulator checkpoint files
trx frames from checkpoints written by modular simulator could be
slightly off depending on the algorithms used (always up-to-date with
domain decomposition only). The global state is still written to
modular checkpoints (as there is a significant amount of setup work at
runner time that needs to be refactored before modular simulator can
get spawned completely without a global state). It is, however, not
guaranteed to be absolutely up to date, as it's not used in the
simulation continuation.
The trx frame reading is now handled by modular simulator if the
checkpoint was written by modular simulator. A test was added to
ensure that the final frame of the simulation is equivalent to the
checkpointed configuration.
Note that this was not affecting simulation restart (exact
continuation is tested also for modular simulator). It would, however,
affect other tools accessing the configuration from the checkpoint
file. The error was only present in master and 2021 beta, so it was
never in a released version.
trx frame reading by modular simulator required some refactoring of
the StatePropagatorData and FreeEnergyPerturbationData
checkpointing. The main read and write function was moved to the data
object itself rather than being owned by its element. This is likely a
better design decision anyway: The element orchestrates the
checkpointing during the simulations, but the data object is handling
the actual data reading / writing. This allows to create a dummy data
object, restore it from checkpoint, and read out the quantities
relevant for the trx frame.
Closes #3838
Berk Hess [Thu, 10 Dec 2020 13:34:59 +0000 (13:34 +0000)]
A check for perturbed listed pairs beyond rlist
Excluded atoms pairs should be within the pairlist cut-off, as
otherwise long-range corrections are not applied. This is problematic
for free-energy calculations where intra-molecular interactions are
not coupled and replaced by pair interactions. Now a fatal error
is generated when perturbed pair interactions are beyond rlist.
Also added PME grid correction for LJ excluded pairs beyond rvdw,
which was missing from #3403.
Added missing release note for the fix of #3403.
Fixes #3403
Fixes #3809
Andrey Alekseenko [Wed, 9 Dec 2020 11:24:13 +0000 (11:24 +0000)]
Fix defining of EXCLUSION_FORCES in CUDA/OpenCL kernels
For the case of cut-off electrostatics + Ewald LJ + no energy calculation,
the EXCLUSION_FORCES macro should be active.
However, LJ_EWALD macro is (conditionally) set only later, and so is
always undefined when it is being checked for EXCLUSION_FORCES.
Mark Abraham [Wed, 9 Dec 2020 10:14:01 +0000 (10:14 +0000)]
Fix compilation and tests with icc 19.1.2
Closes #3822
Pascal Merz [Wed, 9 Dec 2020 09:51:32 +0000 (09:51 +0000)]
Fix and activate tests for kv-tree checkpoint functions
The tests in src/gromacs/fileio/tests/checkpoint.cpp were introduced
in
64bee06aa, but the file was never added to the CMakeLists.txt,
meaning they were never executed (or even compiled).
Added the file to the CMake target and fixed a few compilation
errors and warning to activate the tests.
Pascal Merz [Wed, 9 Dec 2020 09:23:13 +0000 (09:23 +0000)]
Fix dump of checkpoint files
With the addition of the MdModules and the modular checkpointing,
`gmx dump -cp` would end with an error when either of the two was
written to checkpoint. Reason was that the function dumping
checkpoint files was not aware of the additional blocks present.
As the new functionality checkpoints at the end of the files, the
remainder of the information was printed anyway, but the call ended
with an error.
This adds the two new checkpointing blocks to the function used
by gmx dump, and prints the kv-tree they are based on. To achieve
this, the kv-tree printing functionality had to be expanded to
allow for arrays of objects.
Finally, this change also fixes a wrong comment in the modular
simulator checkpoint helper - checkpointing happens on top of the
integration step, so _before_ the last step.
Pascal Merz [Sat, 5 Dec 2020 01:07:13 +0000 (18:07 -0700)]
Clarify checkpoint mismatch error, and enable in release
As of GROMACS 2021, checkpoint files written with the modular and the
legacy codepath need to be continued with the same code path. The
respective check was not executed in release mode, which would lead
to less descriptive errors.
Also reworded the error to guide users how they can solve the problem.
Pascal Merz [Fri, 4 Dec 2020 05:29:12 +0000 (22:29 -0700)]
Relax simulator comparison tests
The simulator comparison tests have proven to be too strict when running
them on more diverse hardware / compiler / parallelization combinations.
This doubles the tolerances for the energies of tests containing
constraints. Note that at 160 double precision ULPs, these are still
very tight for end-to-end tests, so the danger of relaxing them a bit
is low.
Closes #3836
Pascal Merz [Wed, 2 Dec 2020 21:38:48 +0000 (14:38 -0700)]
Print energy conservation in modular simulator
Paul Bauer [Wed, 2 Dec 2020 14:39:57 +0000 (14:39 +0000)]
Third GROMACS 2021 beta release
All features should now be stable before the release.
Pascal Merz [Wed, 2 Dec 2020 08:52:40 +0000 (01:52 -0700)]
Fixes post submit broken by
05f02ab82
Pascal Merz [Wed, 2 Dec 2020 00:30:34 +0000 (17:30 -0700)]
Separate simulator comparison tests
The simulator comparison tests ensure that legacy and modular
simulator code paths don't diverge while both exist. They are
currently part of the mdrun tests. Separating them has two
advantages:
* As the feature set of modular simulator grew, the run time
of these tests became the largest contributor to the run time of
mdrun-test. Keeping it separate makes it easy to keep the
run time in check and to disable the tests if necessary.
* It is useful to run the simulator comparison tests on
multiple ranks (at least locally). Not all other tests
in mdrun-test support more than one rank. Having a separate
binary simplifies running it with different parallelization schemes.
Pascal Merz [Wed, 11 Nov 2020 23:38:46 +0000 (16:38 -0700)]
Fix error when using simulator test with pre-defined environment variable
The simulator test compares the results of functionality implemented in
both the legacy and the modular simulator by running two simulations using
the environment variables GMX_DISABLE_MODULAR_SIMULATOR and
GMX_USE_MODULAR_SIMULATOR to cover both code paths. If one of these
variables was already present in the test sytem, the tests would fail with
an error.
This change backs up the current state of both variables, disables them,
runs the tests, and resets them to their previous state (if any).
Pascal Merz [Wed, 11 Nov 2020 22:52:22 +0000 (15:52 -0700)]
Modular simulator: Disable invalid code paths
This explicitly disables two invalid code paths for modular simulator:
* The new MTS is not implemented for modular simulator. It is only
implemented for integrator = md, which by default uses the legacy
code path. This default can however be overwritten by setting
the environment variable GMX_USE_MODULAR_SIMULATOR=ON, which would
result in errors due to the MTS data structures not being properly
initialized.
* The non-MTTK Parrinello-Rahman barostat is available in modular
simulator, but not in the legacy simulator. The case that this
barostat is chosen in combination with other options not compatible
with modular simulator needs to be caught explicitly.
Paul Bauer [Mon, 30 Nov 2020 16:07:42 +0000 (16:07 +0000)]
Add tests for newer GCC versions
Added gcc-9 + OpenCL (in release builds) and gcc-10 + OpenCL
during merge pipelines to replace previous gcc-7 + OpenCL builds.
Shuffled around previous gcc-8 + CUDA builds to now have
gcc-7 + CUDA 10.2 and gcc-8 + CUDA 11.0 instead.
Test OpenCL with internal clFFT during release builds.
M. Eric Irrgang [Fri, 27 Nov 2020 14:12:07 +0000 (17:12 +0300)]
Explicitly use xunit2 schema for pytest output.
Resolves #3717
M. Eric Irrgang [Sun, 29 Nov 2020 19:54:39 +0000 (19:54 +0000)]
Improve accessibility of pytest documentation.
* Update gmxapi install.rst.
* Consolidate (and prune) some README.md content in
python_packaging subdirectory.
Also ref #3706
Fixes #3702
Paul Bauer [Fri, 27 Nov 2020 14:37:28 +0000 (15:37 +0100)]
Fix build on CentOS 7
The compiler toolset coming with devtoolset-7 on CentOS still suffers
from https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58265. The workaround
for this has been removed before because other gcc-7 compilers are not
affected by this, but it is obviously still needed for CentOS.
Fixes #3826
M. Eric Irrgang [Fri, 27 Nov 2020 14:24:01 +0000 (17:24 +0300)]
Import setuptools early to suppress distutils warning.
Also apply PEP-8 formatting corrections.
Fixes #3715
Paul Bauer [Wed, 25 Nov 2020 10:30:08 +0000 (11:30 +0100)]
Fix cyclecounters on x32
A missing check meant that the code wouldn't compile due to an asm block
that is not compatible with 32 bit datatype ABI and x86_64 registers.
Fix provided by Nicholas Breen.
Fixes #3816
Paul Bauer [Wed, 25 Nov 2020 08:30:08 +0000 (09:30 +0100)]
Fix error with links in docs
Linkchecker complained about issues that I didn't fix correctly during
the latest merge.
Gilles Gouaillardet [Wed, 25 Nov 2020 07:06:46 +0000 (16:06 +0900)]
Fix frexp() for ARM SVE SIMD
Mark Abraham [Tue, 24 Nov 2020 15:26:49 +0000 (16:26 +0100)]
Stop GoogleTest requiring ancient CMake minimum versions
In practice the version is always new enough, and avoiding the
redundant minimum specifications means GROMACS developers will not be
spammed with useless warnings.
Erik Lindahl [Tue, 24 Nov 2020 11:48:54 +0000 (12:48 +0100)]
Fix frexp() for ARM neon SIMD.
Implements the same fix for arguments +-0 as previously done
for x86 and power implementations. After this, the only
remaining architecture is ARM SVE.
Refs #3773.
Paul Bauer [Tue, 24 Nov 2020 13:38:39 +0000 (14:38 +0100)]
Merge branch 'origin/release-2020' into merge-2020-into-2021
Resolved Conflicts:
admin/containers/scripted_gmx_docker_builds.py
docs/conf.cmakein.py
docs/dev-manual/change-management.rst
docs/dev-manual/contribute.rst
docs/dev-manual/releng/index.rst
docs/dev-manual/tools.rst
docs/install-guide/index.rst
docs/reference-manual/preface.rst
docs/release-notes/index.rst
scripts/GMXRC.bash.cmakein
src/gromacs/gmxana/gmx_covar.cpp
src/gromacs/gmxana/gmx_do_dssp.cpp
src/gromacs/gmxana/gmx_h2order.cpp
src/gromacs/gmxana/nsfactor.cpp
src/gromacs/hardware/detecthardware.cpp
src/gromacs/mdlib/coupling.cpp
src/gromacs/mdrun/md.cpp
src/gromacs/mdtypes/state.h
src/gromacs/trajectoryanalysis/modules/pairdist.cpp
src/gromacs/utility/basedefinitions.h
Paul Bauer [Tue, 24 Nov 2020 10:22:47 +0000 (11:22 +0100)]
Fix gmx h2order -d option
The code to implement this was never added.
Fixes #3820
Erik Lindahl [Tue, 24 Nov 2020 08:23:37 +0000 (08:23 +0000)]
Make sure frexp() returns correct for argument 0.0
Previous functions in gromacs have never called frexp
with argument 0.0, but cbrt() needs this. This will not
have caused any bugs, but since it results in FP
exceptions in debug builds we want to have it fixed
so our unit tests work.
This change adds the correct behavior for all x86
SIMD instruction sets, and it has also been tested
on all of them in both single and double, with the
exception for MIC (which should be correct, but we
cannot test it).
The ARM SIMD flavors have not yet been fixed yet
since the ARM build fails due to an apparent nblib
error. Once that is fixed, I will add it in a separate
change and close the issue.
Refs #3773.
Artem Zhmurov [Tue, 24 Nov 2020 08:02:48 +0000 (08:02 +0000)]
Improve handling of CUDA API errors
Consolidate the construction of the printed error message in one
place and make sure that the CUDA API information on error is
printed when the error is detected.
Paul Bauer [Mon, 23 Nov 2020 20:51:28 +0000 (20:51 +0000)]
Fixes for Intel and container build fixes
Generalizes the TSAN build stage to a compiler build stage, and uses
it for containing the installation process for the free Intel tools
Fixed or worked around several bugs
Updated install guide section on tested platforms to describe
what we actually do
Refs #3459, #3620
Change-Id: I28e4b816923395edead9d425cec5d85581e22b5e
Magnus Lundborg [Mon, 23 Nov 2020 06:55:13 +0000 (07:55 +0100)]
Properly check if there are perturbed constraints or masses
Instead of just checking, when running updates on GPU, if
there are lambda states that involve constraints or masses
(e.g. if using fep-lambdas) check if any constraints or masses
are actually perturbed.
Artem Zhmurov [Tue, 17 Nov 2020 14:22:41 +0000 (17:22 +0300)]
Make corresponding device active before each TestDevice is created
When populating the list of GPU devices for tests, the DeviceContext and
DeviceStream objects are created for each TestDevice. In CUDA, the
active device has to be explicitly set by the API since the DeviceStream
is created for the active device. Otherwise, all the streams are created
for the first device. When these streams are later used for the devices
other than first, CUDA API returns an invalid argument error. The bug only
affects unit tests in CUDA builds for systems with more than one GPU.
Does not affect OpenCL/SYCL builds, because they use DeviceContext to
attach the device to the stream (queue).
Fixes #3781, #3782 and #3805.
M. Eric Irrgang [Thu, 29 Oct 2020 21:07:19 +0000 (22:07 +0100)]
Look for the various possible names of the gromacs package.
Refs #3777
Artem Zhmurov [Mon, 23 Nov 2020 13:52:57 +0000 (16:52 +0300)]
Remove coolquote that sounds offensive without context
Joe Jordan [Tue, 17 Nov 2020 13:53:46 +0000 (13:53 +0000)]
Remove some extraneous compiler flags from nblib
Closes #3798
Gilles Gouaillardet [Fri, 13 Nov 2020 09:14:37 +0000 (09:14 +0000)]
Detect ARM SVE Vector length if possible
arm_sve: unless the SVE vector length is explicitly set via the
GMX_SIMD_ARM_SVE_LENGTH macro, detect the default vector length
and aborts if it cannot be found.
Thanks Olly Perks for the pointer.
Paul Bauer [Thu, 12 Nov 2020 10:45:27 +0000 (11:45 +0100)]
Fix thread mpi linking for mdrun only build
The linking would fail because the library include directory was not
added properly.
Fixes #3807
Paul Bauer [Thu, 12 Nov 2020 08:56:08 +0000 (09:56 +0100)]
Restore correct dynamics linking behaviour
Recent changes to the way shared libraries are managed introduced a bug
in setting up static linking by default when building mdrun only.
Fixes #3806
Paul Bauer [Wed, 11 Nov 2020 14:36:20 +0000 (15:36 +0100)]
Fix gmx do_dssp CHARMM support
The tool would not work correctly due to issues with detecting some
oxygen molecules.
Fixes #3568
Szilárd Páll [Wed, 11 Nov 2020 14:28:14 +0000 (14:28 +0000)]
Fix gcc 9 "warning: ‘unused’ attribute ignored"
gcc 9 seems to not like the unused because it would never warn anyway
about members, but clang does require it.
Paul Bauer [Wed, 11 Nov 2020 13:56:29 +0000 (13:56 +0000)]
Fix segmentation fault in gmx pairdist
When using the pairdist tool with a coordinate in space as selection,
the tool would overwrite the selection group of the selection by default
to implement reference grouping. This would change the selection type of
the coordinate from INDEX_UNKNOWN to INDEX_ALL by default (or to any
other user supplied reference grouping). The change of the selection
type would then later cause a segmentation fault, as the underlying
machinery would try to access the non-existing atoms belonging to the
selection.
This changes the mechanism to no longer overwrite the selection group
type if the current one is INDEX_UNKNOWN.
Introduces test case with command line selection that would crash
without the fix.
Fixes #3762
Paul Bauer [Wed, 11 Nov 2020 10:30:50 +0000 (10:30 +0000)]
Second GROMACS 2021 beta release
Feature freeze for 2021 branch.
Berk Hess [Wed, 11 Nov 2020 07:42:53 +0000 (08:42 +0100)]
Update gmx msd documentation
Removed comment on weighted fitting for determining the diffusion
coefficients. This weighting had been removed at some point, but
the documentation had not been updated.
Andrey Alekseenko [Tue, 10 Nov 2020 15:21:45 +0000 (16:21 +0100)]
Fix "GMX_CUDA|OCL_NB_EWALD_TWINCUT is disabling twin cut-off"
Paul Bauer [Tue, 10 Nov 2020 16:45:03 +0000 (17:45 +0100)]
Disable failing NB-LIB listed forces tests
Tests are not stable enough to be run in all configurations and need
some more digging into for a proper solution.
Refs #3795
kanduri [Tue, 10 Nov 2020 16:54:53 +0000 (16:54 +0000)]
NBLIB: Fix build warning from uninitialized array
Szilárd Páll [Tue, 10 Nov 2020 12:33:23 +0000 (12:33 +0000)]
GPU codepath cycle counting fixes
* eliminated double-counting in bonded module initialization
* moved the bonded kernel launch counting into the gpuBoned module
* avoid counting ~0 CPU update work on the GPU path by conditionally
starting the ewcUPDATE counter
* also added some minimal improvements to the wallcycle debugging
tooling which should facilitate finding invalid nesting and missing
stop calls.
Refs #3764
Mark Abraham [Mon, 9 Nov 2020 15:16:01 +0000 (16:16 +0100)]
Add move operations to GpuHaloExchange
This avoids needing to put the objects in a unique_ptr to put them in
a container. Also made a collection of them use std::array for
clarity.
Berk Hess [Mon, 9 Nov 2020 20:37:31 +0000 (21:37 +0100)]
Fix dHdl and foreign energy clearing at checkpointing
The code copying dH/dl and foreign lambda energies to the checkpoint
data structures copied the wrong way around leading to zero values
being written to energy file for steps in between the last energy
frame and the checkpoint.
Fixes #3763
Pascal Merz [Tue, 10 Nov 2020 03:01:39 +0000 (20:01 -0700)]
Fix MTTK constant energy bugs
Fixes #3796
This also fixes an unclear comment relating to the barostat
Nose-Hoover chain variables in t_state.
Andrey Alekseenko [Mon, 9 Nov 2020 17:57:21 +0000 (17:57 +0000)]
Add SYCL implementation of LeapFrogGpu
An implementation of LeapFrogGpu integrator using SYCL. Also, some additions to SYCL's DeviceBuffer.
Code pretty much follows CUDA version.
For float3, I had to use gmx::RVec instead of SYCL's built-in cl::sycl::float3, since latter is 16 bytes.
Joe Jordan [Mon, 9 Nov 2020 14:55:16 +0000 (14:55 +0000)]
Add backend for nblib listed forces API
This adds the back-end for the nblib API defined in
listed_forces/bondtypes.h and listed_forces/definitions.h.
Additionally, header and source are added for the private
implementations of these files. These are implementation details,
and are not visible to or interactable with by the user, so they
are not part of the public API. This backend is a ground up refactoring
of gromacs, copying the force and energy kernels but rewriting the disipatch.
There are tests that compare forces and energies to gromacs.
Each function in this commit has at least one test for correctness,
though in most cases there are many more since the functionality is
cumulative. All of this functionality has undergone extensive design
review and code review already, so it should not be necessary to go
through the whole change with a fine-tooth comb. Suggestions on how
implementations could be improved are welcome. However, they will
likely need to be addressed as follow-up issues unless they are
trivial to fix.
Paul Bauer [Mon, 9 Nov 2020 12:38:18 +0000 (12:38 +0000)]
Use images from GitLab docker repository
Docker is reducing the maximum number of pulls from their storage, and
we are already now hitting the limit with heavy CI activity.
Now uses images from GitLab Docker Repository instead.
Fixes #3788
Andrey Alekseenko [Mon, 9 Nov 2020 11:30:10 +0000 (11:30 +0000)]
Fix intemittent MdrunIOTests segfault
In Modular Simulator, we wish to have two copies of
ComputeGlobalsElement<ComputeGlobalsAlgorithm::VelocityVerlet>. To avoid
making two separate objects, a `static thread_local` singleton was
introduced. However, it caused issues when the ModularSimulator was
built repeatedly in the same process, due to stale pointer persisting in
a static variable.
This commit adds a key-value storage to
ModularSimulatorAlgorithmBuilderHelper, and makes ComputeGlobalsElement
use it instead of static variables. This ensures proper lifetime of
cached pointers and can be used by other Elements for similar goals.
Closes #3791
Artem Zhmurov [Mon, 9 Nov 2020 06:20:17 +0000 (09:20 +0300)]
Remove source file that is no longer in use
All CUDA data transfer calls now use the same wrappers, the
NBNXM-specific wrapers can be removed.
Szilárd Páll [Mon, 9 Nov 2020 10:33:38 +0000 (10:33 +0000)]
Fix cycle counting in StatePropagatorDataGpu
Double-counting resulted in broken/truncated performance acounting
table.
Fixes #3764
Berk Hess [Mon, 9 Nov 2020 10:10:29 +0000 (10:10 +0000)]
Move implementation to cpp file
Fixes issue with missing GMX_ASSERT as well.
Alan Gray [Mon, 9 Nov 2020 09:48:34 +0000 (09:48 +0000)]
Use dynamically allocated unique pointers for GPU halo tests
The tests were intermittently failing in the enqueuing of a remote
GPU's event to a local GPU's stream because the remote object
(allocated on the stack in the innermost loop) was already destroyed (due
to the asynchronicity of the operation). This change allows the objects
to persist, and is the same way we already do it in the main code.
Fixes intermittent failures appearing in !773 and eleswhere.
Andrey Alekseenko [Fri, 6 Nov 2020 12:19:56 +0000 (13:19 +0100)]
Suggest environment variable to limit visible SYCL devices
As with OpenCL, there is no portable way to do it, so we suggest the
currently working Intel option.
Artem Zhmurov [Fri, 6 Nov 2020 07:26:26 +0000 (10:26 +0300)]
Fix doxygen comments formattig in update.h
Paul Bauer [Thu, 5 Nov 2020 16:18:24 +0000 (16:18 +0000)]
Allow pdb2gmx to handle cyclic molecules
This has been a long outstanding issue with the code not being able to
do this.
Cyclic molecules are detected by checking if chain ends are within bonding
distance from each other. This does not detect cyclic molecules over PBC
boundaries.
Initial patch proposed by Boris Timofeev.
Resolves #3571
Change-Id: Ic25e46a12fef5fb46748dc0497f469b5834d4886
Paul Bauer [Thu, 5 Nov 2020 11:35:58 +0000 (12:35 +0100)]
Allow CUDA build tests to succeed without GPU
Added early return to haloexchange GPU pathway test to not try to run
the tests without a GPU (i.e. device list is empty).
Refs #3789
Mark Abraham [Thu, 5 Nov 2020 09:10:34 +0000 (10:10 +0100)]
Stop showing wrong README on GitHub front page
Paul Bauer [Thu, 5 Nov 2020 10:27:15 +0000 (10:27 +0000)]
Limit parallelism in Github CI
Remove the -j option from CTest, avoiding issues with trying to run the
test binaries in parallel.
Set environment variable so that the build knows it is under CI, and
limits the number of thread-MPI and if needed OpenMP threads.
Activate regressiontest testing.
Berk Hess [Thu, 5 Nov 2020 09:57:25 +0000 (09:57 +0000)]
Disable MTS with SD integrator
The SD integrator requires different damping coefficients for
the different MTS forces and we currently don't have the different
MTS forces available separately in the integrator.
Fixes #3775
Magnus Lundborg [Tue, 3 Nov 2020 20:35:14 +0000 (21:35 +0100)]
Set pbcatom_input in order to pass later checks
pbcatom_input was not set. This meant that pbcatom checks in
set_pull_init() failed for large pull groups.
Fixes: #3786
Berk Hess [Wed, 4 Nov 2020 19:31:12 +0000 (20:31 +0100)]
Copy position restraint parameters with MTS
Fixes #3784
Mark Abraham [Thu, 5 Nov 2020 08:00:48 +0000 (08:00 +0000)]
Improve issues with CUDA interacting with std flags
GCC 7 needs a work-around just like older CUDA does.
Removed unnecessary use of C++17 features in compat/pointers.h as we
do in other parts of the code while CUDA compilation must be within
C++14.
Refs #3608
Fixes #3783, #3780
Alan Gray [Wed, 4 Nov 2020 18:56:03 +0000 (18:56 +0000)]
Allow GPU update without GPU DD
Mark Abraham [Wed, 4 Nov 2020 15:40:32 +0000 (16:40 +0100)]
Fix use of config.h in halo exchange tests
Mark Abraham [Wed, 4 Nov 2020 14:16:50 +0000 (14:16 +0000)]
Remove logging from hardware detection
This should be deferred to separate responsibilities, in particular
so that hardware detection can run before logging functionality
is available. Subsequent logging of the results restores the value
to the user from understanding what mdrun could see about their
hardware.
Refs #3650, #3768
Joe Jordan [Wed, 4 Nov 2020 13:28:54 +0000 (13:28 +0000)]
NBLIB listed forces API update
- added a listed forces calculator for the user-facing API
- expanded documentation
- reduced boilerplate
- addressed review comments
Alan Gray [Tue, 3 Nov 2020 14:01:51 +0000 (14:01 +0000)]
Add more halo exchange unit test variations
Addition of several 1D and 2D variations with different pulse configs.
Follow up will extend to 3D.
Szilárd Páll [Tue, 3 Nov 2020 12:47:16 +0000 (12:47 +0000)]
Add missing wallcycle_stop
On the GPU update path the Update CPU counter was started and not
correctly stopped.
Refs #3764
Alan Gray [Tue, 3 Nov 2020 12:12:57 +0000 (12:12 +0000)]
Extend halo exchange unit test to GPU codepath
Test will additionally re-perform halo exchange on GPU codepath (for
CUDA thread-MPI builds).
Gilles Gouaillardet [Tue, 3 Nov 2020 11:23:27 +0000 (11:23 +0000)]
timing: try to directly retrieve gmx_cycles_calibrate() from the hardware
The timer period can be directly retrieved from the hardware on ARMv8
and recent x86_64 platforms.
Try this first, and fall back on the current guessestimate on older x86_64 platforms
Magnus Lundborg [Mon, 2 Nov 2020 15:53:15 +0000 (16:53 +0100)]
Remove TODO regarding AWH cover threshold along lambda axis.
The 1.0 factor should be OK for a FEP lambda axis. It should be OK
in theory and tests show it works.
Change-Id: Iafe39ab0663ec227e96e40f1b2073818610e93ae
Closes: #3634
Artem Zhmurov [Tue, 3 Nov 2020 07:43:32 +0000 (07:43 +0000)]
Make the message why GPU bondeds are not available more clear
Closes #3728
Paul Bauer [Fri, 30 Oct 2020 18:15:17 +0000 (19:15 +0100)]
Add Clang CUDA build to CI
Runs tests during post-merge acceptance, but compiled for all merge
requests.
ejjordan [Mon, 2 Nov 2020 11:05:11 +0000 (12:05 +0100)]
Update to nblib non-bonded backend
* added more documentation
* more consistent naming
* cleaner, more consistent interfaces
* add more checks
* expand test coverage
* fix bugs
Andrey Alekseenko [Thu, 29 Oct 2020 18:21:16 +0000 (19:21 +0100)]
Rework handling of PATH-like variables in GMXRC.bash/zsh
Old version was not robust to various kinds of trailing colons, as
experienced in, e.g., #3219.
To make it all safer, a single function, that removes the
old path and adds the new one to the beginning, is introduced.
Tested with:
- GNU bash 4.4.20
- GNU bash 5.0.17
- zsh 5.8
- dash 0.5.8
```
$ replace_in_path ":" "/new" "/old"
/new:
$ replace_in_path "" "/new" "/old"
/new
$ replace_in_path "/old" "/new" "/old"
/new
$ replace_in_path "/old:" "/new" "/old"
/new:
$ replace_in_path "/old:/other" "/new" "/old"
/new:/other
$ replace_in_path ":/other" "/new" "/old"
/new::/other
$ replace_in_path "/other:" "/new" "/old"
/new:/other:
$ replace_in_path "/other:/old" "/new" "/old"
/new:/other
```
Full test:
```
$ bash -c 'MANPATH=''; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man
$ bash -c 'MANPATH=':'; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:
$ bash -c 'MANPATH='/abc'; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc
$ bash -c 'MANPATH='/abc:'; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc:
$ bash -c 'MANPATH=':'; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:
$ bash -c 'MANPATH=''; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man
$ bash -c 'MANPATH='/abc'; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc
$ bash -c 'MANPATH='/abc:'; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc:
```
Closes #2628, #3219
M. Eric Irrgang [Thu, 29 Oct 2020 20:48:06 +0000 (21:48 +0100)]
Remove an unnecessary transitive dependency of imported CMake target.
Refs #3777
Paul Bauer [Fri, 30 Oct 2020 17:26:33 +0000 (17:26 +0000)]
Fix unused variable warnings
Paul Bauer [Fri, 30 Oct 2020 12:58:59 +0000 (13:58 +0100)]
Add missing fix for Clang CUDA
Was missing from the previous list of fixes (don't know how that
happened).
Paul Bauer [Fri, 30 Oct 2020 12:38:29 +0000 (12:38 +0000)]
Fix build issues with Clang CUDA
Fixed missing changes in CMake property name
Apply properties correctly for test library.
Fix wrongly attributed gpu_utils test file.
Joe Jordan [Fri, 30 Oct 2020 11:36:34 +0000 (11:36 +0000)]
Introduce nblib versioning
Andrey Alekseenko [Thu, 29 Oct 2020 18:21:16 +0000 (19:21 +0100)]
Rework handling of PATH-like variables in GMXRC.bash/zsh
Old version was not robust to various kinds of trailing colons, as
exerienced in, e.g., #3219.
To make it all safer, a single function, that removes the
old path and adds the new one to the beginning, is introduced.
Tested with:
- GNU bash 4.4.20
- GNU bash 5.0.17
- zsh 5.8
- dash 0.5.8
```
$ replace_in_path ":" "/new" "/old"
/new:
$ replace_in_path "" "/new" "/old"
/new
$ replace_in_path "/old" "/new" "/old"
/new
$ replace_in_path "/old:" "/new" "/old"
/new:
$ replace_in_path "/old:/other" "/new" "/old"
/new:/other
$ replace_in_path ":/other" "/new" "/old"
/new::/other
$ replace_in_path "/other:" "/new" "/old"
/new:/other:
$ replace_in_path "/other:/old" "/new" "/old"
/new:/other
```
Full test:
```
$ bash -c 'MANPATH=''; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man
$ bash -c 'MANPATH=':'; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:
$ bash -c 'MANPATH='/abc'; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc
$ bash -c 'MANPATH='/abc:'; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc:
$ bash -c 'MANPATH=':'; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:
$ bash -c 'MANPATH=''; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man
$ bash -c 'MANPATH='/abc'; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc
$ bash -c 'MANPATH='/abc:'; . scripts/GMXRC.bash; . scripts/GMXRC.bash; . scripts/GMXRC.bash; echo $MANPATH;'
/usr/local/gromacs/share/man:/abc:
```
Closes #2628, #3219
Szilárd Páll [Tue, 27 Oct 2020 10:49:58 +0000 (11:49 +0100)]
Clang format
Szilárd Páll [Tue, 27 Oct 2020 10:49:24 +0000 (11:49 +0100)]
Add fvec doxygen
Szilárd Páll [Thu, 6 Feb 2020 23:40:12 +0000 (00:40 +0100)]
Fix AMD OpenCL float3 array optimization bug
Because float3 by OpenCL spec is 16-byte, when used as an array type
the allocation needs to optimized to avoid unnecessary register use.
The nbnxm kernels use a float3 i-force accumulator array in registers.
Starting with ROCm 2.3 the AMD OpenCL compiler regressed and lost
its ability to effectively optimize code that uses float3 register
arrays. The large amount of extra registers used limits the kernel
occupancy and significantly impacts performance.
Only the AMD platform is affected, other vendors' compilers are able to
do the necessary transformations to avoid the extra register use.
This change converts the float3 array to a float[3] saving 8*4 bytes
register space. This improves nonbonded kernel performance
on an AMD Vega GPU by 25% and 40% for the most common flavor of the
Ewald and RF force-only kernels, respectively.
Note that eliminating the rest of the non-array use of float3 has no
significant impact.
Paul Bauer [Thu, 29 Oct 2020 16:47:33 +0000 (17:47 +0100)]
Fix CUDA code issues
Remove unused template.
Fix variable initialization order and remove obsolete variable.
Fix missing newline.