M. Eric Irrgang [Thu, 12 Dec 2019 12:58:40 +0000 (15:58 +0300)]
Clean up Sphinx interpreted text in the default role.
* Fix typos or syntax for several text literals that Sphinx tries to
process as interpreted text (in the default role).
* Use Google docstring conventions to guide adjustment of gmxapi
docstrings.
Change-Id: If6eeede8dcc782c04fc4363005c38c746a8664ea
Paul Bauer [Fri, 20 Dec 2019 14:00:02 +0000 (15:00 +0100)]
Add files for Gitlab CI
Change-Id: I2508b6ba4a551a0768b1d5dfd742d97ebec37c7c
Alan Gray [Thu, 12 Dec 2019 11:06:36 +0000 (03:06 -0800)]
Add user guide doc for GPU buffer ops and comms env variables
Change-Id: Ib6701adbf6f4909a804394606a362fa209716497
Paul Bauer [Fri, 20 Dec 2019 13:28:10 +0000 (14:28 +0100)]
Fix wrong header in release notes
Fixes #3261
Change-Id: Icc31b7a273670e99c21c9bb60343e1ce5d532653
Paul Bauer [Tue, 17 Dec 2019 14:21:38 +0000 (15:21 +0100)]
GROMACS 2020 release candidate
Adjusted regresstiontest hash
Change-Id: I37cec8154843b8f8a2a9136051b50bce7eba5888
Szilárd Páll [Thu, 19 Dec 2019 19:44:35 +0000 (20:44 +0100)]
Enable buffer ops when GMX_GPU_DD_COMMS is set
Buffer ops is always a hard dependency for DD with GPU direct comm, so
enable it instead of requiring an extra env var to be set on mdrun
invocation.
Change-Id: Iea354f8ce636a115e0ed866228746e91f34df228
Paul Bauer [Thu, 19 Dec 2019 16:42:15 +0000 (17:42 +0100)]
Fix regressiontests failing with ctest
Setting the -ntomp variable and not -nt at the same time causes errors
when running the regressiontests through ctest.
This was introduced in I78925951dcb59a11af01bc64f0dea373fe172cbb while
updating things for Gitlab CI.
Fixes #3257
Change-Id: I19afd7817b5c637d6653fda8533f23720b45cc41
Alan Gray [Thu, 19 Dec 2019 13:32:53 +0000 (05:32 -0800)]
Release note for GPU direct communications
Change-Id: I4487f2ed6bc789e6cfb3e68d802ccedda0c51962
Alan Gray [Thu, 19 Dec 2019 09:45:58 +0000 (01:45 -0800)]
Fix PME tuning issue when PME-PP communications are active
If PME-PP direct GPU communication is active, the CPU-GPU
asynchronicity in the codepath causes the timing load balance ratio
used to trigger PME tuning to be unreliable. This change
unconditionally starts PME tuning when the PME-PP communication
feature is active (unless PME tuning has been de-activated with
the -notunepme option).
Partially addresses #3208
Change-Id: Iea173d19062dbd11d57ca9ceb4e52b9f20d4ff15
Szilárd Páll [Fri, 29 Nov 2019 00:10:10 +0000 (01:10 +0100)]
PP-PME load balancing improvements
Add a minimum number of nstlist tuning intervals and minimum time delay
at the beginning of the run before the load balancing starts. This allow
hardware clocks to ramp up and avoids having early measurements
overestimate rendering subsequent ones with different grid setups only
faster due to hardware warmup.
Also use global variables to adjust the number of measurements to be
skipped after switching configs.
Refs #3208
Fixes #2203
Change-Id: If835d2482e127caa51d50f45f25c19144d35efaa
M. Eric Irrgang [Thu, 19 Dec 2019 11:09:18 +0000 (14:09 +0300)]
Add issue tracking for deprecations.
Provide links in release notes to issue tracking for announced
deprecations.
Fixes #3251
Change-Id: I246651f8cb868eb1deecac86c2e7d6f0ec81b3f1
Berk Hess [Wed, 18 Dec 2019 18:23:43 +0000 (19:23 +0100)]
Fix mdrun appending invalidating energy files
When appending at mdrun restarts, energy files would become
unreadable at the appending point.
Fixes #3250
Change-Id: I7683e1a954fb8865acb5e77aae028ae717ae0ab3
Alan Gray [Wed, 18 Dec 2019 15:41:45 +0000 (07:41 -0800)]
Fix conditional on assertion for combining GPU Update and GPU coordinate send
Fixes a problem introduced in
9a3c2ce312c1e04634d636655388a8aaad53c8d1
which unintentionally blocks any use of GPU update with PME-PP
Communication because !stepWork.doNeighborSearch is missing from the
condition on the assertion.
Related to #3159 and #3160 which, when implemented, will remove the
restriction on coordinate send originating from the GPU (on non search steps).
Change-Id: Ie9b1102386b2c1bc927c09e4d8ead2bd2321320d
Szilárd Páll [Tue, 10 Dec 2019 11:24:32 +0000 (12:24 +0100)]
Allow x D2H to overlap with GPU force compute
With GPU update coordinates are transferred back to the CPU every step
if there are forces to compute on the CPU. Originally this was
implemented with a back-to-back transfer launch and wait at the
beginning of do_force().
This change moves the CPU wait for the completion of the coordinate
transfer closer to the consumer tasks in order to avoid blocking GPU
force tasks' launch and allowing compute and transfer to overlap.
Fixes #3221
Change-Id: Ia6641147bbec1186b54c1445d36dc31000eae9c4
M. Eric Irrgang [Fri, 13 Dec 2019 16:41:48 +0000 (19:41 +0300)]
Update reference to uncrustify and clang-format.
Correct some copy-paste errors. Consolidate some references to
formatting tools and continue to transition away from uncrustify.
uncrustify references and infrastructure remain, along with Jenkins
documentation. These can be removed after transitioning from Jenkins to
GitLab Runner, pending clarification on how to distinguish between
documentation for all supported versions versus the current version.
There are also various references and workarounds related to uncrustify
in the source code. These can be updated in separate changes against the
post-release-2020 master branch.
Refs #3248
Change-Id: I3331f73921812a918d32dd6a045eabc4da4a6a6d
Berk Hess [Fri, 6 Dec 2019 13:39:28 +0000 (14:39 +0100)]
Change grompp decoupling SD warning to note
The warning that grompp issues when decoupling a molecule
without the use of the SD integrator has been changed to a note,
since there are valid use cases for using normal MD when not
running in the completely decoupled state.
Note that checking for actual decoupled VdW states requires a lot of
code and would still not be 100% correct.
Fixes #2767
Change-Id: I55c84afcbcb7b19bdb8a7a8bc586a65c50f8d0c1
Paul Bauer [Tue, 17 Dec 2019 17:22:16 +0000 (18:22 +0100)]
Add ability to set OMP and tMPI thread for ctest
For Gitlab CI we need at some points to be able to set the threads used
for running the regressiontests through ctest.
(cherry picked from commit
4e6b4134e62f17d64520a46e0aad7cbd60f778f3)
Change-Id: I78925951dcb59a11af01bc64f0dea373fe172cbb
Paul Bauer [Fri, 13 Dec 2019 15:51:44 +0000 (16:51 +0100)]
Ignore build host for CI builds
When moving the CMake build directory between hosts a warning was
generated. This does not integrate well with containerized CI.
(cherry picked from commit
33079aa495e7e3b7da78bd58de9c6f30edff51d4)
Change-Id: If0a775ab86d25a7d5646e824243892320d4ddbe3
Berk Hess [Mon, 16 Dec 2019 09:50:06 +0000 (10:50 +0100)]
Add note for COM removal with position restraints
Center of mass motion removal in combination with position restraints
will often cause artifacts. We should (automatically) disable this,
but this will disrupt workflows, so for now grompp will issue a note.
Refs #3244
Change-Id: If983e918d26f0af2fab96d44d0f2b1fc3cedb901
Artem Zhmurov [Mon, 16 Dec 2019 21:17:47 +0000 (22:17 +0100)]
Clean-up buffer-ops CUDA kernels
1. Adjust naming according to conventions.
2. Reduce number of type casting.
3. Use native CUDA types.
Fixes #3237.
Change-Id: I79da0dd1edf6c302661ba0954e4191fde1ebf2e8
Paul Bauer [Tue, 10 Dec 2019 10:11:26 +0000 (11:11 +0100)]
Test parallel mpi only with openmp
Can time out in other conditions
Found while working on Gitlab CI
Refs #3232
Change-Id: I7cf62ed913d2bc8c88317a674b0594dbe54ba394
Szilárd Páll [Tue, 17 Dec 2019 12:43:00 +0000 (13:43 +0100)]
Comment the use of waitCoordinatesCopiedToDevice
Tthis is to note that we are only using a barrier a mechanism to sync
with GPU tasks rather than to preventing a race between the copy and
other CPU tasks.
Change-Id: I8d432f7ef3e87e56beffa7f58230e50bcd59a666
Paul Bauer [Fri, 13 Dec 2019 15:50:21 +0000 (16:50 +0100)]
Set correct dependency for doc build on gmxapi
When building the documentation with the python package, we need to know
about the symbols in the package when building docs for e.g. the pdf
manual.
Change-Id: Ia9a248da5746a5d90736cd3f3aa345fe8ffd0c79
(cherry picked from commit
0aa283c6b578b1b6bf6673169697865142014368)
Paul Bauer [Fri, 13 Dec 2019 15:50:01 +0000 (16:50 +0100)]
Fix compiler warning about extra ;
Change-Id: I9718ab9d9772dfc7251bc1eb175be59544b2c2cb
(cherry picked from commit
6ef70b67bb06f03da97315170a44f76679d92a23)
Szilárd Páll [Fri, 6 Dec 2019 18:57:24 +0000 (19:57 +0100)]
Add wallcycle counting to StatePropagatorDataGpu
Launch overheads are counted in the main GPU launch overhead counter and
a separate subcounter is used for the launch and a main counter for the
CPU blocking wait timing.
Note that this chnge introduces mdtypes->timing->mdtypes cyclic dependency,
the warning on which is suppressed.
Fixes #3207
Change-Id: I3b69df9e4888800b43712a42b863958db80f5caa
M. Eric Irrgang [Fri, 13 Dec 2019 16:14:43 +0000 (19:14 +0300)]
Be more specific with clang-format suggestions.
Refs #3238
Change-Id: If71687cbab9a9aa1c10e3501849932a0ea0f3112
Pascal Merz [Tue, 17 Dec 2019 04:24:40 +0000 (21:24 -0700)]
Fix exact continuation modular simulator
With modular simulator, the center of mass motion was mistakenly
removed also when continuing from checkpoint, breaking exact
continuation.
Change-Id: I4db837a1b11f1b44cfae4970230a51a2c05d84aa
Christian Blau [Mon, 9 Dec 2019 18:42:20 +0000 (19:42 +0100)]
Fix densityfitting stalling and energies
Density guided simulations would stall when there were no atoms on a
node. Further, the energy contributions were added for every
participating node, resulting in wrong energies.
Both issues are fixed with this patch.
fixes: #3230
refs: #3235
Change-Id: Iab773993fb3f98bd154acaa1d3971e655182a2b2
Mark Abraham [Wed, 31 Oct 2018 13:44:55 +0000 (14:44 +0100)]
Simplify gmxapi RPATH handling
Fixes #2722
Change-Id: Ic697f596285acdb5c7728b7ebc995ea44c82f3ac
Artem Zhmurov [Thu, 12 Dec 2019 16:30:27 +0000 (17:30 +0100)]
Clean up outdated comments in md.cpp
Some of the comments were left unattended as the code evolved.
Change-Id: I12c2e313645fcbb1a112f3f0c9f4c72e7943d9d6
Pascal Merz [Wed, 11 Dec 2019 01:49:24 +0000 (18:49 -0700)]
Test for expected output files
Ie7f03f5b fixes a bug where modular simulator would not write
final coordinates to file. This test checks whether some basic
output files are present (edr, trr, log, gro, cpt).
Change-Id: I02299d342667104ccc19f0af6d557bb8aa916494
Pascal Merz [Wed, 4 Dec 2019 06:55:36 +0000 (23:55 -0700)]
Fix final configuration writing for modular simulator
Modular simulator was not writing final configuration (.gro) file.
This commit fixes this, and removes an outdated sentence in the
documentation of ITrajectoryWriterClient.
I02299d34 introduces a test to avoid similar errors in the future.
Note that the complexity required to introduce this feature implies that
trajectory writing in the modular simulator needs some work for
GROMACS 2021.
Change-Id: Ie7f03f5beea2d78411c7a0cdfa85f71deeb622fa
Szilárd Páll [Mon, 2 Dec 2019 18:29:28 +0000 (19:29 +0100)]
Allow overlapping CPU force H2D with compute
The reduction orchestration code already uses explicit sync event
in all cases and StateGpu implements the ability to schedule force
H2D in a separate stream for the "All" locality.
Hence, this change switches for non-DD runs the CPU force H2D to be done
in the update stream to allow overlap with force work in the local
stream.
Refs #3170 #3029
Change-Id: Iceb9aac395335c062109d552d3f0289688a9c75f
Christian Blau [Tue, 22 Oct 2019 08:55:59 +0000 (10:55 +0200)]
Densityfitting - enforce similarity increase with adaptive force scaling
Previous to this patch, the force constant for density guided
simulations with adaptive force constant scaling was scaled up the same
amount that it was scaled down, not resulting in the announced enforced
increase in similarity. Now, the force constant is increased more strongly
than decreased, resulting in the desired enforced increase in similarity.
Change-Id: I5b149c79d91ea9e5ceeb5b0e7dce854bd6f34711
Artem Zhmurov [Thu, 5 Dec 2019 13:49:25 +0000 (14:49 +0100)]
Allow using GPU update with DD and update groups
The GPU update is now can be enabled for the supported DD cases
with the GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. Also
added the checks on whether the SHAKE algorithm was requested,
since SHAKE is not supported by the GPU update.
Refs. #3226, #3163.
Change-Id: I57e3ad3b8a571ec244989e888afd5cfcbaf9b75e
Artem Zhmurov [Wed, 11 Dec 2019 15:19:15 +0000 (16:19 +0100)]
Enable most of the pulling flavors with GPU update
Only constraints pulling is not supported, since it is not implemented
in GPU version of LINCS. There is no reason to disallow other flavors
of pulling.
Change-Id: Iceca52bba5e040bd241a156f9cf58ab3556acdac
Mark Abraham [Wed, 11 Dec 2019 14:46:51 +0000 (15:46 +0100)]
Document DOMAINDECOMP(cr) correctly
It is likely there are numerous cases where this is used mistakenly
when havePPDomainDecomposition expresses the real intent. If so,
runs with 1 PP and 1 PME rank may have buggy behaviour.
Refs #2395
Change-Id: I07be73a6c690887b3043140a2a78ae6fe6bb17f1
Artem Zhmurov [Wed, 11 Dec 2019 10:53:31 +0000 (11:53 +0100)]
Clarify what setting zero to some output mdp options does
Change-Id: I36dc059e982c3358cbfa55596e71093ce3208446
M. Eric Irrgang [Wed, 11 Dec 2019 13:50:33 +0000 (16:50 +0300)]
gmxapi landing page update.
In response to user feedback, replace some content that was confusing
instead of informative.
Change-Id: Ida8c12ba2caf4ef224b9276570bc719c9000e2bd
M. Eric Irrgang [Wed, 11 Dec 2019 13:45:36 +0000 (16:45 +0300)]
gmxapi documentation readability improvements.
In response to user feedback, provide a "quick install" overview section.
Rearrange and update some content.
Change-Id: I2a1bbf1a1efbda769e330b85925ed01dbcdae29e
M. Eric Irrgang [Wed, 11 Dec 2019 13:39:56 +0000 (16:39 +0300)]
gmxapi documentation formatting improvements.
Improve some formatting mark-up. Use normative Sphinx roles for
commands, Python modules, environment variables, and file paths.
Change-Id: Iff1d4b09da13c421da809f882e73c8404eb70f38
Artem Zhmurov [Mon, 9 Dec 2019 18:20:03 +0000 (19:20 +0100)]
Set GPU update to false in case of a rerun
Rerun does not require the numerical integration of equation of
motion, hence the useGpuUpdate should be false.
Fixes #3231.
Change-Id: I9027641e9dc32fb464ebbb6148c11c530c0eb2cd
M. Eric Irrgang [Fri, 6 Dec 2019 12:29:57 +0000 (15:29 +0300)]
Finalize gmxapi 0.1
Set 0.1 release metadata for Python package and C++ library.
Note that 0.1 is API identical to 0.0.8, but we are
taking the next step to the targeted semantic versioning,
and patch releases should no longer affect the APIs.
However, we are not yet at a point where we can have
clear stability guarantees between minor versions.
Change-Id: I4f4a6ba84353730e2962b730cd1e656bae4e377a
Mark Abraham [Fri, 6 Dec 2019 09:45:21 +0000 (10:45 +0100)]
Avoid 386 O3 codegen bug
Refs #3205
Change-Id: I2aa5018bde089b11497892dd68a7f9715e69e20e
M. Eric Irrgang [Thu, 5 Dec 2019 10:48:18 +0000 (13:48 +0300)]
Add notes on MPI with gmxapi Python package.
Change-Id: I7ecd33fcea482b8f300873fc7fdcea2c3958215e
Artem Zhmurov [Fri, 6 Dec 2019 09:35:17 +0000 (10:35 +0100)]
Pass array ref instead of casting to basic array in PME redistribute
Change-Id: I98b3c211b2669266f8bc6b37580b8f48087e3645
Mark Abraham [Wed, 4 Dec 2019 16:07:59 +0000 (17:07 +0100)]
Provide more feedback when listed forces tests fail
Refs #3205
Change-Id: Ic012aeaf8b35af88b78e109acabb6048d593cab3
M. Eric Irrgang [Thu, 5 Dec 2019 10:46:50 +0000 (13:46 +0300)]
gmxapi Python package usage prose.
* Provide a walk through of basic gmxapi Python usage.
* Update one of the helper functions in test_subgraph.py to
make sure that a documented example is tested.
Change-Id: I2ad3368fe8a39307bcd7e676dc36627e30ab7ac1
M. Eric Irrgang [Thu, 5 Dec 2019 14:47:07 +0000 (17:47 +0300)]
Improve pytest fixtures for directory management.
* Add a pytest command line option to control whether to remove
temporary directories always, never, or conditionally upon test success.
* Encapsulate context managers and test fixtures for tempdir removal,
temporary changes of working directory, and access to the new pytest
command line argument.
* Use a context manager to use the temporary testing directories
for working directory, as well. Works around potential issues
with GROMACS creating temporary files in different directories
(or filesystems) from where they are relevant.
This change makes identical changes to 3 different conftest.py files.
Consolidation into the gmxapi.testsupport module is deferred to a
separate change.
Refs #3227
Change-Id: I703d1e3806f7b4d8f2ae3b80c39ec64acd9cf3b4
Christian Blau [Wed, 4 Dec 2019 15:52:01 +0000 (16:52 +0100)]
Relax cross-correlation test precision
Relase the floating point precision in cross-correlation evaluations on
large data arrays further. i386 fails here with debian experimental build.
Refs #3205
Change-Id: I86c9ad1804da5be98671ac14a5bc37ea5636aad0
Mark Abraham [Mon, 5 Nov 2018 18:07:20 +0000 (19:07 +0100)]
Added OpenCL build to release matrix
Added missing information for one build.
Change-Id: I30da8ce4e19e104182ee082cc18a83e318069b3c
Mark Abraham [Wed, 4 Dec 2019 15:25:35 +0000 (16:25 +0100)]
Removed unnecessary link_directories command for gtest
This warns at link time on MacOS Catalina with macports
clang-9 and cmake 3.15.5
Change-Id: If493f72133c0388e6119ea7a30aa1dc5c5830c7a
Christian Blau [Wed, 4 Dec 2019 12:08:10 +0000 (13:08 +0100)]
Remove cutoff-scheme group from default mdout.mdp output
The .mdp default output file had a reference to the group scheme in it,
which might prove confusing to users that assume that this is still
available.
Change-Id: I434f15e46dfc0a8ebffe8d826855dc8e67168fba
Paul Bauer [Tue, 26 Nov 2019 09:37:15 +0000 (10:37 +0100)]
Third beta release of GROMACS 2020
Updated regressiontest hash
Change-Id: I96be9b7b9f3f94f6eddf1fd4603cc829fd3e1848
Paul Bauer [Tue, 3 Dec 2019 10:21:47 +0000 (11:21 +0100)]
Revert "Enable GPU update with DD when GPU comm features are enabled"
This reverts commit
b88f3eadf90261555508f6a4df64ce184141a66a.
Reason for revert: Mark found the code segfaulting under some conditions with domain decomposition and update enabled, so this is not ready to be used yet.
Refs #3226
Change-Id: I45cf9564a87595d445057da071a8fca5c60f9d9a
Paul Bauer [Tue, 3 Dec 2019 11:25:54 +0000 (12:25 +0100)]
Fix nightly build by correctly defining several
variables as bool instead of int.
Change-Id: Ia22fe3acdc55131ea40c80b104293273cafb1017
Alan Gray [Tue, 5 Nov 2019 12:52:21 +0000 (04:52 -0800)]
Enable GPU update with DD when GPU comm features are enabled
Changes assertions to allow GPU update to run with domain
decomposition, only when GPU PME-PP communication and GPU halo
exchange features are enabled through their respective environment
variables.
Added test coverage to the gpuupdate matrix
Change-Id: Ie05dac64eada8f704c3a33d0eb6e5bc74dfcc61e
Paul Bauer [Fri, 29 Nov 2019 15:00:52 +0000 (16:00 +0100)]
Set update to run on GPU by default if possible
Change-Id: I1a2052968c8d8dd8020744bf15f974bc0d92a6ef
Jonathan Vincent [Wed, 6 Nov 2019 14:55:10 +0000 (06:55 -0800)]
Heuristics for switching between CUDA spread/gather kernels
The various CUDA spread/gather kernels perform better in different circumstances,
so heuristics are used to control which one is selected.
Refs #3189
Change-Id: I32c0726021a48dc8721e337f8f41e9c9d334e05c
Berk Hess [Mon, 2 Dec 2019 13:15:08 +0000 (14:15 +0100)]
Add check for coupled constraints for CUDA update
The CUDA LINCS code has a limit on the number of coupled constraints.
This is now checked during in the GPU update decision function.
Change-Id: I5eee96b82f3f5196b64a5e815eae78d4ed367a80
Berk Hess [Mon, 2 Dec 2019 10:43:16 +0000 (11:43 +0100)]
Extract LINCS CUDA helper functions
Change-Id: I00fb311053aa9c13036b401107399409f9a01599
Christian Blau [Mon, 2 Dec 2019 13:52:27 +0000 (14:52 +0100)]
Relax test tolerance for cross correlation measure for densities
Some compilers do not perform the optimisations that are necessary to
reach default real tolerance when calculating the cross correlation of
larger three dimensional densities (100x100x100 voxels)
This patch loosens the requirement on the precision of cross correlation
evaluations.
refs #3205
Change-Id: Iad1bb1b17bb6935b2ce2537bcbb6c3902172a6bc
Paul Bauer [Fri, 29 Nov 2019 15:10:05 +0000 (16:10 +0100)]
Use DOI for citing in reference manual section
Change-Id: I84ee722345669afbca972697ebcc76d874af1b46
Mark Abraham [Thu, 28 Nov 2019 16:49:26 +0000 (17:49 +0100)]
Further improve taskassignment exception handling
It was still posssible that the exit() from one thread-MPI rank would
cut off another rank while the latter's error output before it
finished being written to the terminal stream.
This solution is still deficient in some aspects, but it will
always work on single-node testing (ie. including Jenkins) and
often work usefully in multi-node testing.
Change-Id: Ib41e012b9f56591c80046b44891dd77af5a2f938
Berk Hess [Fri, 29 Nov 2019 16:22:15 +0000 (17:22 +0100)]
Merge "Merge branch release-2019 into release-2020" into release-2020
Paul Bauer [Fri, 29 Nov 2019 13:21:44 +0000 (14:21 +0100)]
Fix unaligned access with AVX512
Found and debugged by Gilles Gouaillardet
Change-Id: I9203593c94eb745020274b126601d30e49d7aac2
Paul Bauer [Fri, 29 Nov 2019 10:32:17 +0000 (11:32 +0100)]
Fix build error without thread or regular mpi
Some more unused variable warnings that I haven't fixed yet.
Fixes #3218
Change-Id: Id47718bb0af97d3a494afa133af8b54ba71b9d5d
Artem Zhmurov [Thu, 28 Nov 2019 18:59:36 +0000 (19:59 +0100)]
Print OpenCL error info in case of failed assertion in device buffer management
Having error code availible is helpful for debugging.
Change-Id: I71c808c9171eae3f87b72253f3e81f0e2e2c22a2
Mark Abraham [Fri, 29 Nov 2019 12:52:13 +0000 (13:52 +0100)]
Merge branch release-2019 into release-2020
Change-Id: I138d910b0d5cb8555460f87345c4e312e528a3ce
Artem Zhmurov [Thu, 28 Nov 2019 15:09:48 +0000 (16:09 +0100)]
Make Leap-Frog CUDA kernel selection safer
This is a small improvment on a recent bug fix, which adds:
1. An implicit boolean on whether the kernel with temperature
coupling velocity scaling should be selected.
2. Adds an assertion on the correctness of an input.
Original bug fix:
b362a880f75d93e9252affdad12ed5448fbac517
Change-Id: Idd981337f2f53f4d3a803ce50162b2f9c753b860
Artem Zhmurov [Thu, 28 Nov 2019 09:59:31 +0000 (10:59 +0100)]
Rename some variables in GPU Update
The temperature and pressure coupling booleans in GPU version of
update were renamed to better represent their meaning. Minor
refactoring in the UpdateConstraintsCUDA CPU-stub class.
This change is just a refactoring, mostly renaming.
Change-Id: I039722a5ef827a20264b6d4ac6c6436227094e09
Paul Bauer [Wed, 27 Nov 2019 15:35:16 +0000 (16:35 +0100)]
Check for using correct hwloc headers and runtime
Also add assertion in the code to prevent errors from linking
against the wrong library while running.
Fixes #3200
Change-Id: Ib2f2861702e111f67c38b0c9d65ccbe4c81a0ccd
Artem Zhmurov [Thu, 28 Nov 2019 16:41:43 +0000 (17:41 +0100)]
Clear device force buffer if it was reallocated
The force buffer on the device is assumed to be clear at the start of
the step. This was not ensured if the buffer was realocated after e.g.
domain decomposition.
Change-Id: I4677f71edc9479d972a6d9471b8cbe2c377f8827
Berk Hess [Thu, 28 Nov 2019 15:21:53 +0000 (16:21 +0100)]
Update documenation for NPT GPU update
Change-Id: Iaa0ee0555621955c6eb7276d1b9f097420c4e70a
Artem Zhmurov [Tue, 12 Nov 2019 10:54:35 +0000 (11:54 +0100)]
Fix pressure coupling with GPU update
This fixes the pressure coupling Parrinello-Rahman and
Berendsen schemes when Update is offloaded to the GPU.
The coordinate scaling is performed on GPU. The PBC
data is also copied to device when periodic box is
updated by the pressure coupling.
Fixes #3182.
Change-Id: I57226c9f0fa249cb018d38b653a6ecb3ffd9c7c2
Paul Bauer [Thu, 28 Nov 2019 12:28:57 +0000 (13:28 +0100)]
Fix postsubmit
Change-Id: I47496ad6e82f9b67060979f916d5d5510fdba182
Berk Hess [Thu, 28 Nov 2019 10:58:48 +0000 (11:58 +0100)]
Remove assertion failure in AWH without initial stage
When not using the initial stage in AWH, we should not check
for coverage and histogram anomalies at step 0.
Fixes #3217
Change-Id: Iee19c4740c02ce56c0b43ad89efb6e37f4ad1a0d
Berk Hess [Thu, 28 Nov 2019 09:23:45 +0000 (10:23 +0100)]
Fix bug with GPU pressure scaling at non T-scaling steps
Only affected Parrinello-Rahman coupling, which is not enabled yet.
Change-Id: Ie081ffe2a925de92361c43b1f5de8e1fe324a486
Christian Blau [Wed, 27 Nov 2019 11:10:03 +0000 (12:10 +0100)]
Fix check-pointing for density guided simulations
Previously, checkpointing for density-guided simulations did not take
into account that the checkpointing occurs in the middle of the md-loop.
Now, the state is stored before the force-calculations, so that
simulations are restarted correctly.
refs #2282
fixes #3215
Change-Id: Ide4ebf94f2801acbe476a5cd348af66d0a6f28ac
Artem Zhmurov [Wed, 27 Nov 2019 22:23:33 +0000 (23:23 +0100)]
Combine two booleans in md loop
Two booleans that were supposed to have identical meaning have being
created in the md loop. This patch conbines them into one. Note that
the scope of the boolean that is responsible for the D2H coordinates
copy have expanded, which represents correct behavior.
Change-Id: I421b18ee81f4cf9cce19cdd60023cb097d85a5d7
Mark Abraham [Wed, 27 Nov 2019 08:56:37 +0000 (09:56 +0100)]
Disable GPU PME-PP communication with mdrun -pmefft cpu
The MdrunMpiTest with two ranks defaults to using a separate PME rank,
and that attempts to run all flavors of -pmefft. However only -pmefft
gpu is targeted for any of the new GPU development, so the development
feature flags need to account for that.
SimulationWorkload is also improved to be aware of this fact.
Change-Id: I57c8e8ce81fdb7aaf65de89feebb369d3f2f10eb
Artem Zhmurov [Tue, 12 Nov 2019 09:55:02 +0000 (10:55 +0100)]
Add method for coordinates scaling into GPU update-constrain class
The coordinates scaling is needed on the pressure coupling step in
case the box size has changed.
Change-Id: I0bf7c7a2f77c1bd42dd0713f689096eaca409f88
M. Eric Irrgang [Wed, 6 Nov 2019 15:07:56 +0000 (18:07 +0300)]
Add gmxapi_extension tests to 'tests' target.
Improve GMX_PYTHON_PACKAGE convenience targets. When GMX_PYTHON_PACKAGE
is ON, the subprojects in the python_packaging subdirectory are built.
When testing is enabled, the sample_restraint C++ unit tests were
transitively included in the 'check' target, but no mechanism was in
place to ensure they were built before the CTests and google tests
were run. This change adds the test binaries as dependencies of the
'tests' target so that the test binaries are built through normal
machinery when GROMACS is built with testing enabled and
GMX_PYTHON_PACKAGE=ON.
Change-Id: I751655dc6b6d4d58b5bbb9f5f522e26382b4f0a2
Alan Gray [Tue, 22 Oct 2019 14:23:11 +0000 (07:23 -0700)]
Event-based Dependency for GPU Force Halo Exchange
Introduces new event recorded when exchanged forces are ready on GPU,
and passes this into force buffer ops using dependencyList. Removes previous
mechanism of forcing local stream to wait on non-local stream.
Addresses part of #3093
Refs #3194
Change-Id: I768898839e5c6a653894d5eb80354f0e423e06ed
Mark Abraham [Mon, 25 Nov 2019 12:10:50 +0000 (13:10 +0100)]
Improve handling of task-assignment exceptions
Resolves several current TODOs to improve the reporting of error
messages derived from exceptions. This improves the robustness of the
regressiontests logic that depends on parsing the error message to
understand whether and how to retry running mdrun.
Change-Id: I8c9eac95a2ef4861e90350e77fb9beafb08e9d2e
Mark Abraham [Tue, 26 Nov 2019 09:54:14 +0000 (10:54 +0100)]
Relax tolerance for bd tests
The implementation is somehow not very reproducible.
Refs #3205
Change-Id: Ibf9de00d6e67b8e2237cc8726f300b2bf8ad2327
Alan Gray [Tue, 26 Nov 2019 10:12:02 +0000 (02:12 -0800)]
Bug fix for nonlocal D2H coordinate transfer synchronization
Fixes bug introduced in I2e2ba1b6436f087d1f2fef4ff876445814a724e7,
which replaced the NB-module D2H nonlocal coordinate copy launch with
a new State Propagator module equivalent, but did not change the
corresponding event wait call, such that the dependency was not
satisfied. This change replaces the NB event wait with the correct
State Propagator event wait.
Change-Id: I7c812974c145d315fa6516b2bbea39164439728e
Paul Bauer [Tue, 26 Nov 2019 10:29:04 +0000 (11:29 +0100)]
Fix unused variable warning in ewald code
Fixes #3193
Change-Id: Id3f8422ec95e6864553a6e9f03a144fbe4ed8674
Berk Hess [Mon, 25 Nov 2019 14:43:05 +0000 (15:43 +0100)]
Fix issues with AWH with periodic pulling
Removed fatal error with AWH with periodic pull-geometry 'direction'
when the distance was within 2% of half the box size.
Changed an assertion failure when the AWH interval was larger than
the box size to a fatal error.
Clarified the documentation for pull geometry 'direction-periodic'.
Refs #2946
Change-Id: I447a9a490a5883139afc268fd367b8b249591fbc
Alan Gray [Sat, 2 Nov 2019 19:24:43 +0000 (12:24 -0700)]
Fix race condition in GPU X halo exchange with GPU update
When the GPU Update featured is enabled on multi-GPU with the GPU halo
exchange feature, it was possible that the GPU receiving the halo data
would still be working on the previous timestep and writing to the
same coordinate buffer in the GPU update, causing a race
condition. This change replaces the pre-existing synchronization
mechanism of a MPI Send/Recv of a boolean flag with the same but
exchanging the events that are recorded when the coordinates are ready
on the device. The remote event is then enqueued to the stream that
the GPU uses for pushing the data, satisfying the dependency of the
push on the remote update being completed on the GPU.
Change-Id: Ie16a93dcc3064d73da1e560d82b4e76870d03e52
Christian Blau [Mon, 25 Nov 2019 15:18:40 +0000 (16:18 +0100)]
Fix duplicate CONNECT records in pdb files
When adding connect records in pdb files, it was not checked if records
already existed and were thus duplicated in output. This patch fixes
this behavior.
Fix originally provided by Boris Timofeev.
refs #3206
Change-Id: I5850f4b6279b37df07282acc4416f433099d90c8
Artem Zhmurov [Tue, 12 Nov 2019 17:33:37 +0000 (18:33 +0100)]
Use same matrix for Parrinello-Rahman and Berendsen coordinates scaling
Currently, Parrinello-Rahman uses the matrix created in the md.cpp,
whereas it is created when needed internally for the Berendsen.
This commit renames the matrix and make use of it in both barostats.
Change-Id: I66e010a1ed7efbc93c8a2240ccb7b9e174fd7e45
(cherry picked from commit
abab855fdceee437c33010ebbe544afe1f88db87)
Christian Blau [Fri, 22 Nov 2019 16:51:23 +0000 (17:51 +0100)]
Read mrc density map format with all endianesses.
Swaps the endianess of reading mrc density maps if header is not sane.
refs: #2282
Change-Id: I417e402b2ab801c22d9963fabb0e60b186288651
Christian Blau [Thu, 21 Nov 2019 15:48:07 +0000 (16:48 +0100)]
Check mrc header for sanity
Add functionality to check mrc density map headers for sanity.
Required for endianess correction, because the endianess stamp is not
guaranteed to be right for most .mrc maps.
Change-Id: I8956bcbc4b9fd7b3b6346c3b4cdd0022c9e3332a
Artem Zhmurov [Thu, 21 Nov 2019 16:33:57 +0000 (17:33 +0100)]
Fix index mapping issue with LINCS on GPU
In some cases (e.g. when DD is used), coupled constraints blocks
can have constraint indices that are interleaved with other blocks.
This was not taken into account when setting up the LINCS GPU data
structures. This patch fixes that by ensuring that all coupled
constraints are added to the mapping array one after the other.
Add note in decideWhetherToUseGpuForUpdate() that we need to add
a check for this when update auto will choose gpu in some cases.
Also renamed spaceNeeded by numCoupledConstraints to avoid confusion.
Fixes #3211.
Change-Id: Ib402c5cc9fc20116496034cdc11be921586bbd15
Christian Blau [Tue, 29 Oct 2019 14:42:46 +0000 (15:42 +0100)]
InMemorySerializer with endianess swap.
Adds an InMemorySerializer that swaps endianess. Required for reading
data that is not handled with XDR serializers and was written with
different endianess, e.g., mrc files for density guided simulations.
refs #2282
Change-Id: Ice96fe8684bb0b920c22dc7d3c2b2087260eb9ad
Artem Zhmurov [Sat, 23 Nov 2019 17:34:10 +0000 (18:34 +0100)]
New cool quote
Change-Id: I8998f56f73a595df0e1b547d307eef65afe7944d
Mark Abraham [Thu, 21 Nov 2019 15:03:21 +0000 (16:03 +0100)]
New cool quote
Change-Id: I96e52f16aeaf2e41db3a68b889dcd0afae966828
Vedran Miletić [Fri, 26 Jul 2019 14:48:58 +0000 (16:48 +0200)]
Fix building on Solaris with GCC
GROMACS now compiles on Solaris x86, specifically illumos distribution
openindiana. GCC complained about a missing sqrt() variant which was
addressed with static_cast and linker complained about missing
libsocket which was addressed with SunOS-specific entry in
target_link_libraries.
Refs #3050
Change-Id: Ic14a9327fa353270ef0e787b4d2357e86e71da53
Artem Zhmurov [Wed, 30 Oct 2019 10:15:04 +0000 (11:15 +0100)]
Add D2H copies for the checkpointing
If the update is offloaded, the current coordinates and velocities
are stored on the device and have to be copied for the checkpointing.
Change-Id: I391bb83e1c70a17c1679f479234a157ce606a908