Magnus Lundborg [Mon, 21 Sep 2020 10:02:18 +0000 (12:02 +0200)]
AWH: Remove random term from FEP tests
Test if it is the random numbers that are making tests fail on
some builds.
Fixes: #3660
Change-Id: I3e189d7770f0089ac5d74128e096461c5b55f44f
M. Eric Irrgang [Thu, 8 Oct 2020 17:47:33 +0000 (17:47 +0000)]
Clarify the GMX_BUILD_MDRUN_ONLY full testing instructions.
For full testing of an MDRUN_ONLY installation, the documentation
suggests installing over a non MDRUN_ONLY installation. This change
attempts to clarify the use case and normalize the important aspects
of the CMake command lines.
Fixes #3723
Andrey Alekseenko [Thu, 8 Oct 2020 11:14:07 +0000 (13:14 +0200)]
NBlibTest/IntegratorWorks: Initialize x's and v's
The coordinates and velocities of the test atom were left uninitialized.
This, on my machine, occasionally caused NbLibIntegratorTests to
timeout, getting stuck indefinitely inside `put_atoms_in_box` function.
Andrey Alekseenko [Thu, 8 Oct 2020 10:20:57 +0000 (10:20 +0000)]
Add explicit types to NB kernel type enums
So these enums could be used in template parameters
Andrey Alekseenko [Wed, 7 Oct 2020 12:54:22 +0000 (14:54 +0200)]
Set CUDA hardware constants for CC 7.5, 8.6
The devices from CC 7.5 and 8.6 have lower limits on # of threads/blocks
per SM compare to all other CC 5.x+ architectures.
Source: CUDA Occupancy Calculator,
https://docs.nvidia.com/cuda/cuda-occupancy-calculator/CUDA_Occupancy_Calculator.xls,
accessed 2020-10-07.
Andrey Alekseenko [Thu, 8 Oct 2020 09:24:09 +0000 (09:24 +0000)]
Device management: Add SYCL GpuEventSynchronizer
Andrey Alekseenko [Thu, 8 Oct 2020 08:51:36 +0000 (08:51 +0000)]
Activate SYCL build on CI
Artem Zhmurov [Wed, 7 Oct 2020 20:47:38 +0000 (23:47 +0300)]
Add Andrey Alekseenko to the list of authors
Andrey Alekseenko [Thu, 1 Oct 2020 14:54:38 +0000 (16:54 +0200)]
Add GpuRegionTimer for SYCL
Erik Lindahl [Wed, 7 Oct 2020 13:49:36 +0000 (15:49 +0200)]
New release, new quotes!
Mark Abraham [Wed, 7 Oct 2020 13:51:06 +0000 (13:51 +0000)]
Update to oneAPI beta09
M. Eric Irrgang [Wed, 7 Oct 2020 13:28:58 +0000 (13:28 +0000)]
Improve gmxPythonDiscovery.cmake
* Do a better job of urging CMake 3.15+ to use the requested Python
installation when a user specifies Python3_ROOT_DIR.
* Adopt another CMake 3.15 feature: Python3_FIND_VIRTUALENV is much more
readable than the older alternatives.
* Allow FindPython inputs to be set before entering the module (such as
with `-D` command line arguments) by checking before setting.
* Clarify which parts of this CMake module apply to the main FindPython
based detection and which parts exist to provide hints to other
embedded / downstream detection.
Refs #3720
Andrey Alekseenko [Tue, 6 Oct 2020 16:36:35 +0000 (18:36 +0200)]
Make LeapFrogGpu less tied to CUDA
- Extract common parts of LeapFrogGpu from .cuh into .h
- Make it use DeviceBuffers instead of raw pointers. For CUDA build,
they are the same, so no changes to other parts of the code necessary.
Paul Bauer [Tue, 6 Oct 2020 14:32:27 +0000 (16:32 +0200)]
Fix download location for regressiontests
Was still pointing to gerrit instead of the ftp server.
Change-Id: Ia49f06dd1438f6484470ea93ee93f4b26c708fb5
Andrey Alekseenko [Wed, 7 Oct 2020 09:38:35 +0000 (09:38 +0000)]
mdlib/tests/leapfrog: Make GPU tests more platform-agnostic
In preparation for adding SYCL implementation.
Also, clean-up includes a bit.
Paul Bauer [Tue, 6 Oct 2020 17:23:49 +0000 (19:23 +0200)]
Fix SYCL compilation
Would fail because device_management_common.cpp needed to include
gmxsycl.h transiently and fail compilation.
Change-Id: Idd19db8c328b3a9bd07924f72968a640158665d6
Paul Bauer [Wed, 7 Oct 2020 06:55:19 +0000 (08:55 +0200)]
Merge branch 'origin/release-2020' into merge-release-2020-into-master
Resolved Conflicts:
admin/gitlab-ci/archive.gitlab-ci.yml
admin/gitlab-ci/global.gitlab-ci.yml
admin/gitlab-ci/gromacs.gitlab-ci.yml
admin/gitlab-ci/lint.gitlab-ci.yml
admin/gitlab-ci/python-gmxapi.gitlab-ci.yml
admin/gitlab-ci/rules.gitlab-ci.yml
admin/gitlab-ci/sample_restraint-regression.gitlab-ci.yml
cmake/gmxVersionInfo.cmake
src/gromacs/fileio/checkpoint.cpp
src/gromacs/hardware/printhardware.cpp
src/gromacs/mdlib/md_support.cpp
src/gromacs/mdlib/trajectory_writing.cpp
src/gromacs/mdrun/runner.cpp
src/gromacs/modularsimulator/domdechelper.cpp
src/gromacs/modularsimulator/domdechelper.h
src/gromacs/modularsimulator/freeenergyperturbationelement.cpp
src/gromacs/modularsimulator/freeenergyperturbationelement.h
src/gromacs/modularsimulator/modularsimulator.cpp
src/gromacs/tools/trjcat.cpp
src/gromacs/topology/topology.h
src/gromacs/utility/fatalerror.cpp
src/gromacs/utility/futil.cpp
src/gromacs/utility/init.cpp
src/programs/mdrun/tests/CMakeLists.txt
tests/CMakeLists.txt
Change-Id: Icd5d9c78ff2cfb0598c3cb55b057487ca098a1f0
Artem Zhmurov [Tue, 6 Oct 2020 19:02:21 +0000 (22:02 +0300)]
Remove unused canComputeOnGpu(...) function from device management
This function was used in tests, but not needed anymore
Paul Bauer [Tue, 6 Oct 2020 15:08:18 +0000 (15:08 +0000)]
Set proper deviceVendor with SYCL
Change-Id: I7971f681b8e4348f2488b4d353ee187cfa5dba0b
Artem Zhmurov [Tue, 6 Oct 2020 12:15:54 +0000 (14:15 +0200)]
Simplify GPU usage decision function signatures
Some of GPU device usage decisions are made based on the number
of availabel GPU devices, not on their indices. Hence it is
natural to pass one integer instead of the vector.
Paul Bauer [Fri, 2 Oct 2020 09:32:41 +0000 (11:32 +0200)]
Prepare for GROMACS 2020.5
Reset DOI strings. Added new release note stubs.
Change-Id: I40615b30bc42b8d7e79bcbe6154d002412519abf
Paul Bauer [Tue, 6 Oct 2020 11:23:18 +0000 (11:23 +0000)]
GROMACS version 2020.4
Updated regtest hash and set DOI strings.
Change-Id: I42e99173615898b963c86b0a1db3770c5f071051
Andrey Alekseenko [Mon, 5 Oct 2020 10:03:45 +0000 (12:03 +0200)]
Fix logic in init_ewald_coulomb_force_table
Previously, destroyParamLookupTable was called only if buffer was
nullptr (uninitialized). Which looks like a typo, and leads to nothing,
because all implementation of destroyParamLookupTable would refuse to
work on nullptr. Changed the check to free the buffer only if there is
anything to free.
Artem Zhmurov [Tue, 6 Oct 2020 09:37:49 +0000 (11:37 +0200)]
Use new GPU infrastructure in MDLib tests
This make use of common device testing infrastructure in MDLib tests,
where both GPU and CPU implementations are tested. The GPU runners
will now be executed on all the detected devices, not only on the
default one. Also, this will allow to use the MDLib tests in OpenCL
and SYCL, where proper device context object is needed.
Closes #3317
Closes #2254
Related #2092
Paul Bauer [Mon, 5 Oct 2020 15:33:02 +0000 (15:33 +0000)]
Always set DeviceVendor
Until now, only the OpenCL code would set the device Vendor information.
To prepare for later code that always needs this information, it is not
set also for the CUDA and SYCL code paths with (hopefully) reasonable
default values.
Change-Id: Ifb3180dff2c5b13aedd8e6808c172ea322d5cea8
Paul Bauer [Mon, 5 Oct 2020 14:45:54 +0000 (14:45 +0000)]
Fix debug message location
Also more release note fixes.
Change-Id: Id107ec9102f49dda0ba6c83a591fa14ae9b72c29
Paul Bauer [Mon, 5 Oct 2020 07:48:03 +0000 (09:48 +0200)]
Add muparser and CMake handling
Adds handling both for linking external muparser library and building
from source if it is not found.
Refs #3515
Change-Id: I09f725ea634d21e3a7d217cf99d28adb71d54473
Berk Hess [Mon, 5 Oct 2020 13:26:41 +0000 (15:26 +0200)]
Remove incorrect cycle suppression
Paul Bauer [Mon, 5 Oct 2020 06:35:30 +0000 (08:35 +0200)]
Fix doxygen warnings
Failed in master again for a while.
Change-Id: I8574794c8e2f9be5681e431c109df6fe91abce45
M. Eric Irrgang [Fri, 2 Oct 2020 16:41:01 +0000 (19:41 +0300)]
Don't override user provided CMAKE_OSX_DEPLOYMENT_TARGET.
Remove `FORCE` from the CMake `set()` in case users want to specify
another target, such as 10.15.
Allows users to resolve potential linker warnings by setting
CMAKE_OSX_DEPLOYMENT_TARGET on the CMake command line.
Pascal Merz [Mon, 5 Oct 2020 06:08:10 +0000 (06:08 +0000)]
Fix ForceBuffers compiler warning
When compiling in release mode (using clang-8), warning
src/gromacs/mdtypes/forcebuffers.h:128:10: warning: private field 'useForceMtsCombined_' is not used [-Wunused-private-field]
bool useForceMtsCombined_;
would be thrown because `useForceMtsCombined_` is only used within
`GMX_ASSERT` commands. Added a [[maybe_unused]] to keep compiler happy.
Mark Abraham [Fri, 2 Oct 2020 15:44:59 +0000 (15:44 +0000)]
Cleaned up CUDA compilation
Some .cpp source files used CUDA run-time headers that need
special handling when e.g. clang issues warnings from antiquated
constructs in older NVIDIA runtime headers. This is now handled
centrally in a similar manner to that required for SYCL compilation.
Dependency of the higher-level taskassignment module on GPU
API-specific compilation was removed, instead deploying some new
helper functions.
Paul Bauer [Fri, 2 Oct 2020 14:40:50 +0000 (14:40 +0000)]
Work-around for RDRAND returning -1 on Ryzen
Due to a bug, RDRAND can always return -1 instead of a random number.
This introduces a work-around to the problem by switching to PRNG in
case the RDRAND returns -1 twice in a row.
M. Eric Irrgang [Wed, 30 Sep 2020 20:32:20 +0000 (23:32 +0300)]
Handle gmxapi tmpi parameters better.
Use gmxconfig.json to support some additional testing support.
Configure thread allocation for gmxapi CI jobs more carefully.
Fixes #3710
M. Eric Irrgang [Wed, 30 Sep 2020 20:29:24 +0000 (23:29 +0300)]
Let gmxapi installation record the GROMACS library MPI configuration.
Add a `MPI` property to the Gromacs::gmxapi CMake target.
Add "gmx_mpi_type" to gmxconfig.json
Refs #2961
Paul Bauer [Fri, 2 Oct 2020 07:03:02 +0000 (09:03 +0200)]
Remove anonymous namespace in ARM SVE header
Caused complaints from clang-tidy.
Change-Id: I8c855fa26849a283a456b1315614cdc60041bbf0
Joe Jordan [Fri, 2 Oct 2020 10:23:04 +0000 (10:23 +0000)]
Add tests for setStringEntry
Mark Abraham [Fri, 2 Oct 2020 09:28:19 +0000 (09:28 +0000)]
Update GoogleTest to v1.8.1
Now known to work with recent versions of MSVC
Also, require proper language support when using GoogleTest on all
platforms. Populating the INTERFACE_COMPILE_DEFINITIONS with
GTEST_LANG_CXX11 will force GoogleTest to require the compiler to
provide std components like tuple even when compiling GROMACS
source files that include GoogleTest headers.
Fixes #3682
Christian Blau [Fri, 2 Oct 2020 07:50:52 +0000 (07:50 +0000)]
Fix FEP lambda interpolation for reruns
FEP lambda interpolation fix is now also applied when re-running simulations.
Closes #3585
Alan Gray [Fri, 2 Oct 2020 07:02:31 +0000 (07:02 +0000)]
Use workload data structures for GPU halo exchange triggers
Move GPU halo exchange trigger booleans and related conditionals into
workload data structures, and remove unnecessary assertion on GPU
buffer ops being active (since it is now automatically activated when
GPU halo exchange is active).
Partly addresses #3370
Gilles Gouaillardet [Fri, 2 Oct 2020 01:48:17 +0000 (10:48 +0900)]
arm_sve: fix decr3Hsimd()
Fix a gross typo introduced in gromacs/gromacs@
aabf8460e42c1702f38163d4163b6e59f6712737
Szilárd Páll [Thu, 1 Oct 2020 12:02:25 +0000 (12:02 +0000)]
Update mdrun-performance.rst
Gilles Gouaillardet [Wed, 16 Sep 2020 09:23:28 +0000 (18:23 +0900)]
docs: add ARM+SVE support to the release notes and SIMD support documentation
Gilles Gouaillardet [Wed, 16 Sep 2020 05:32:23 +0000 (14:32 +0900)]
hardware: detect ARM SVE architecture at runtime
- issue a warning if the application was built with NEON simd support
- check the vector length at runtime is the same than at cmake time,
and aborts otherwise
Gilles Gouaillardet [Wed, 16 Sep 2020 05:31:55 +0000 (14:31 +0900)]
cmake: use simd support for ARM+SVE
- detect when cmake is invoked on an ARM+SVE platform
- ARM_SVE simd can be forced with -DGMX_SIMD=ARM_SVE
- the SVE vector length can be forced with -DGMX_SIMD_ARM_SVE_LENGTH=<bits>
(default is 512)
Gilles Gouaillardet [Wed, 16 Sep 2020 05:24:47 +0000 (14:24 +0900)]
simd: add support for aarch64+sve (ARM_SVE)
support SVE vectors whose length is fixed at cmake time
e.g. no Vector Length Agnostic (VLA) support.
This is currently a fit for GCC 10 compilers
(via the -msve-vector-bits=<len> option), and will be supported
by LLVM 12 (based) compilers.
ARM_SVE support for GROMACS is contributed by the Research Organization
for Science Information and Technology (RIST).
Gilles Gouaillardet [Thu, 1 Oct 2020 21:11:39 +0000 (21:11 +0000)]
simd: revamp the (templated) decr3Hsimd() subroutine
decr3Hsimd<stride>(m, a0, a1, a2) replaces three decrHsimd():
- decrHsimd(m, a0);
- decrHsimd(m + stride, a1);
- decrHsimd(m + 2*stride, a2);
providing a given architecture the opportunity to perform additional optimizations.
A simple wrapper is added to each architecture where
GMX_SIMD_HAVE_HSIMD_UTIL_{FLOAT,DOUBLE} is 1.
Refs !567
Andrey Alekseenko [Thu, 1 Oct 2020 20:10:36 +0000 (20:10 +0000)]
Remove ensureReferenceCount from OpenCL's GpuEventSynchronizer
M. Eric Irrgang [Thu, 1 Oct 2020 15:43:15 +0000 (15:43 +0000)]
Use updated images for Python CI jobs.
Also switches tool chains to account for rearrangement in the
CI testing matrix.
Fixes #3714
Artem Zhmurov [Thu, 1 Oct 2020 13:55:00 +0000 (15:55 +0200)]
Add missing include into the tests comparison helper
CLang may complain about not knowing what std::numeric_limits are
if <limits> is not included.
Change-Id: I74118c3f0082e22f634ae11bce542a0de709260b
Paul Bauer [Thu, 1 Oct 2020 11:50:30 +0000 (13:50 +0200)]
Break up MPI tests
Split PME from other mpi tests.
Due to time outs in CI :(
Change-Id: If9c9219ab6667ae98a9475ee03012d9aa81b28d4
Paul Bauer [Thu, 1 Oct 2020 12:12:24 +0000 (14:12 +0200)]
Fix post merge acceptance pipeline for release-2020
The logic didn't want to pick up the pipelines for merged commits.
Change-Id: I876d7ce975611bcc5466e5e984ed40de72a456ca
Berk Hess [Thu, 1 Oct 2020 12:02:27 +0000 (12:02 +0000)]
Move computeSlowForces into stepWork
Also moved a flag into forcerec and fixed documentation.
Christian Blau [Thu, 1 Oct 2020 11:52:26 +0000 (11:52 +0000)]
Fix the FEP lambda interpolation
Setting any lambda array, initial lambda or state != 0, the
interpolated lambda values were wrongly interpolated. This
patch fixes this behaviour.
Andrey Alekseenko [Thu, 1 Oct 2020 10:04:11 +0000 (10:04 +0000)]
Device management: Add SYCL DeviceBuffer
Pascal Merz [Thu, 1 Oct 2020 07:32:15 +0000 (07:32 +0000)]
Update modular simulator docs
Joe Jordan [Wed, 30 Sep 2020 20:22:59 +0000 (20:22 +0000)]
Add nblib sample script
The sample script argon-forces-integration.cpp gives an example of nblib client code.
Pascal Merz [Wed, 30 Sep 2020 18:26:04 +0000 (18:26 +0000)]
Regression-test-style Google tests for FEP
This set of tests check FEP simulations versus previous results
within the Google tests framework.
The systems are hand-picked to cover as many code paths as possible.
The reference results were carefully checked for correctness.
M. Eric Irrgang [Wed, 30 Sep 2020 12:39:37 +0000 (15:39 +0300)]
Restrict gmxapi pytest tests to `-nt 2`
Fixes #3704
The current change is a quick fix to alleviate pressure on the project's
automation resources. Hard-coded resource constraints should be removed
in a follow-up change.
Also, the new facility to the mdrun ResourceManager needs to be
considered in future work.
Andrey Alekseenko [Wed, 30 Sep 2020 15:49:11 +0000 (17:49 +0200)]
Squash SYCL DeviceContext and DeviceStream
Berk Hess [Wed, 30 Sep 2020 15:13:58 +0000 (15:13 +0000)]
Move computeSlowForces into stepWork
Also moved a flag into forcerec and fixed documentation.
Paul Bauer [Wed, 30 Sep 2020 13:22:48 +0000 (13:22 +0000)]
Update CI on 2020 branch
Now uses similar settings as master.
Change-Id: I2df3ab916930108afaa63830b7fd92680d837584
M. Eric Irrgang [Tue, 29 Sep 2020 14:27:03 +0000 (17:27 +0300)]
Specify OMP_NUM_THREADS for Pyhon CI scripts.
* Explicitly propagate OMP_NUM_THREADS through mpiexec.
* Apply self-consistent CI memory parameters.
* Only use 1 OMP thread for gmxapi Python tests. We want to use two tMPI
ranks, and no extra threads. tMPI threads are set separately.
Refs #3704
Mark Abraham [Wed, 30 Sep 2020 10:29:20 +0000 (12:29 +0200)]
Fix warnings issued by clang with nvcc
Host-side code sometimes uses antiquated code that clang warns
about. Previous attempts to suppress some of these were broken by the
changes to GMX_GPU (and also used the wrong cache variable). New
warnings have also emerged.
Also fixed a related list of warning suppressions to be a proper
CMake list, rather than a hacky list.
Andrey Alekseenko [Wed, 30 Sep 2020 09:41:15 +0000 (11:41 +0200)]
Fix Doxygen formula generation
Paul Bauer [Wed, 30 Sep 2020 09:41:54 +0000 (11:41 +0200)]
Revert "Add new GROMACS logo and use it in the manual."
This reverts commit
4120bf7734d9a235c7db720a2c882bb6c11470b6.
Legal reasons
Pascal Merz [Wed, 30 Sep 2020 07:25:45 +0000 (07:25 +0000)]
Fix bug resetting mdatoms masses to lambda=0 state
When DD is used, the mdatoms masses were reset to their lambda=0 state
after every DD step. This is fixed by making the DomDecHelper aware of
the FEP element.
Note that the use of MDAtoms should likely be revisited to allow for a
more elegant solution, but this would likely require changes which are
less local than the proposed solution here. Further development is
therefore deferred to #3700.
Pascal Merz [Tue, 29 Sep 2020 21:39:04 +0000 (15:39 -0600)]
Fix #3655
This fixes #3655 by addressing some readability issues in
velocityscalingtemperaturecoupling.cpp, and adding some trivial
const qualifiers in the simulator tests.
Closes #3655
Pascal Merz [Wed, 30 Sep 2020 03:26:19 +0000 (21:26 -0600)]
Fix #3694
This fixes #3694 by moving the initialization of the `isInputCompatible`
boolean to its first use in modularsimulator.cpp.
Closes #3694
Paul Bauer [Tue, 29 Sep 2020 11:40:31 +0000 (13:40 +0200)]
Add new GROMACS logo and use it in the manual.
Also added as a highlight in the release notes.
Change-Id: Icac8e5c0134302d4c251cc9e6f5428bca766717d
Artem Zhmurov [Wed, 30 Sep 2020 05:55:14 +0000 (05:55 +0000)]
Restore destructor for GpuTest and gmx_hw_info_t
Needed to able to forward-declare DeviceInformation in headers
Pascal Merz [Tue, 29 Sep 2020 19:47:35 +0000 (19:47 +0000)]
Fix bug resetting mdatoms masses to lambda=0 state
Following the reorganization in !384, the TopologyHolder reset the masses
in mdatoms after they were set by the FreeEnergyPerturbationElement. This
lead to having the masses equal to state A independently of the actual
lambda state.
This change fixes this bug by moving the mass setting into the setup
phase ensuring that it is set by the FEP element independently of the
build order of the elements.
When DD is used, the mdatoms masses were reset to their lambda=0 state
after every DD step. This is fixed by making the DomDecHelper aware of
the FEP element.
Note that the use of MDAtoms should likely be revisited to allow for a
more elegant solution, but this would likely require changes which are
less local than the proposed solution here. Further development is
therefore deferred to #3700.
Paul Bauer [Tue, 29 Sep 2020 17:06:22 +0000 (17:06 +0000)]
Fix doxygen warnings
Add doxygen comments into the recent code
Artem Zhmurov [Tue, 29 Sep 2020 14:24:21 +0000 (14:24 +0000)]
Don't use fah_fsync
There's no need in a custom fsync functionality for F@H core.
Signed-off-by: Dmitry Moskalchuk <dm@crystax.net>
ejjordan [Tue, 29 Sep 2020 09:36:43 +0000 (11:36 +0200)]
Fail gracefully when nblib cannot be built
Artem Zhmurov [Tue, 29 Sep 2020 14:08:18 +0000 (14:08 +0000)]
Remove two-stage initialization in DeviceStream
A default constructor was needed for DeviceStream as an intermediate
measure to make the refactoring easier. It now can be eliminated.
Joe Jordan [Tue, 29 Sep 2020 13:48:33 +0000 (13:48 +0000)]
Add nblib backend: Part 2 of 2
Paul Bauer [Tue, 29 Sep 2020 13:00:15 +0000 (13:00 +0000)]
Add post submit rules to CI
Use those for resource intensive CUDA tests
Change-Id: I9c23139656fd91a833c67dd20d76a1429e91c254
Alan Gray [Tue, 29 Sep 2020 11:27:18 +0000 (11:27 +0000)]
Redevelopment of GPU Force Reduction/Buffer Ops
Introduces a new purpose-build class for GPU force reduction, which
replaces the previous force buffer ops mechanism.
Refs #3370
M. Eric Irrgang [Mon, 28 Sep 2020 18:11:28 +0000 (21:11 +0300)]
Clarifications for ResourceAssignment interface.
Add and update some documentation. Improve error messages. Improve
file naming.
Refs #3644
M. Eric Irrgang [Sat, 26 Sep 2020 15:05:25 +0000 (18:05 +0300)]
Support ResourceAssignment interface for tMPI library builds.
Fixes #3644
M. Eric Irrgang [Mon, 28 Sep 2020 12:55:50 +0000 (15:55 +0300)]
Allow MPI-enabled clients to provide the library communicator.
Provide library helpers and template headers to allow clients to
createContext() with explitly assigned resources.
Relates to #3644
M. Eric Irrgang [Mon, 28 Sep 2020 12:35:43 +0000 (15:35 +0300)]
Use MpiContextManager to provide MPI communicator.
* Give MpiContextManager a MPI_Comm and make it movable so that it can
be provided while creating the ContextImpl.
* In gmxapi::ContextImpl::launch(), use an MpiContextManager instance as
the source for the SimulationContext communicator.
Refs #3644
M. Eric Irrgang [Mon, 28 Sep 2020 12:24:03 +0000 (15:24 +0300)]
Relocate MpiContextManager.
* Change the scope of the MpiContextManager instance from the Session
to the ContextImpl.
* Tighten up some details to help maintain clarity over subsequent
changes.
Refs #3644
M. Eric Irrgang [Sat, 26 Sep 2020 09:54:32 +0000 (12:54 +0300)]
Interface updates allowing client to provide MPI communicator.
gmxapi::MpiContextManager is acquired earlier in API initialization
and has a larger role in setting up the SimulationContext as a member
of gmxapi::ContextImpl. The MpiContextManager can be acquired with
appropriate default values for the various MPI-related GROMACS build
variants, or it can be initialized with user-provided resources.
gmxapi::Context creation method has been updated to accept optional
non-default resources from the client.
A template header helps MPI-enabled client code to provide resources
to createContext(). The client requires minimal awareness of whether
GROMACS was built for thread-MPI or an installed MPI library.
Thread-MPI details are isolated from any translation unit that might
include the client MPI tool chain.
Refs #3644
Alan Gray [Mon, 28 Sep 2020 20:53:07 +0000 (20:53 +0000)]
Remove single-dimension limitation from GPU halo exchange
Allows GPU halo exchange to be active when the number of dimensions is
greater than 1, thus simplifying the codepath logic.
Partly addresses #3370
Joe Jordan [Mon, 28 Sep 2020 17:21:12 +0000 (17:21 +0000)]
Add nblib backend: Part 1 of 2
Paul Bauer [Mon, 28 Sep 2020 11:43:37 +0000 (13:43 +0200)]
Change download for regressiontest away from gerrit
The 2020 branch still used the gerrit download links for getting the
regressiontests for users, leading to some reported instability.
Change-Id: I61f937b28ae1a7f6637514254f504a67315f0c37
Andrey Alekseenko [Mon, 28 Sep 2020 15:35:39 +0000 (15:35 +0000)]
Fix error in ensureReferenceCount
We were overwriting input value and comparing it to itself.
Mark Abraham [Mon, 28 Sep 2020 13:09:06 +0000 (15:09 +0200)]
Fix check for BitScanReverse64
Berk Hess [Mon, 28 Sep 2020 14:24:26 +0000 (14:24 +0000)]
Fix dispersion correction table generation
Dispersion correction table generation was incorrectly skipped
during PME tuning, which lead to an assertion failure with LJ-PME.
Fixes #3677
Joe Jordan [Mon, 28 Sep 2020 10:35:59 +0000 (10:35 +0000)]
Add initial nblib public API
Giovanni Bussi [Mon, 28 Sep 2020 07:46:02 +0000 (07:46 +0000)]
Fixed incorrect enum
Harmless, since both have value 1, but confusing
Artem Zhmurov [Mon, 28 Sep 2020 06:00:40 +0000 (06:00 +0000)]
add detection of CPX 5300 w/ 2FMAs
Signed-off-by: Hammond, Jeff R <jeff.r.hammond@intel.com>
Pascal Merz [Sun, 27 Sep 2020 11:23:20 +0000 (11:23 +0000)]
Compare trajectories to reference data
This introduces functionality to compare trajectories (x/v/f/box)
to reference data. It also allows to define the maximum number of
frames which should be checked. The latter allows to have a single
reference data, but decide based on the simulation setup (e.g.
parallelization scheme) how many frames are expected to match
before numerical divergence gets too large. To allow this, the
reference data functionality needs to give the option to ignore
unused reference frames.
Pascal Merz [Sun, 27 Sep 2020 06:55:11 +0000 (00:55 -0600)]
Roll back "Ensure the modular simulator runs only when it should"
The fact that GMX_USE_MODULAR_SIMULATOR does not lead to an error
when the run is not supported is a feature, not a bug. It allows
to run the entire test set using modular simulator wherever is is
implemented by using "export GMX_USE_MODULAR_SIMULATOR=ON; make check",
even if modular simulator is disabled by default. It is the only way
to make sure that modular simulator is a valid replacement for the
legacy code path.
Similarly, using "export GMX_DISABLE_MODULAR_SIMULATOR=OFF; make check"
allows to check the legacy code path, even if modular simulator is used
by default.
Andrey Alekseenko [Fri, 25 Sep 2020 20:41:08 +0000 (20:41 +0000)]
Fix #3668
Following the implementation in Google's cpu_features tool made by an
Intel engineer used for original implementation, we only check for the
presence of the second AVX-512 FMA unit if the CPU supports AVX-512F
(foundation), which should be present for all AVX-512 implementations.
That's distinct from the way cpu_features tool does AVX detection, but
should be no less reliable.
Pascal Merz [Fri, 25 Sep 2020 16:37:48 +0000 (16:37 +0000)]
Replace FramesToCompare enum by MaxNumFrames
This allows to explicitly chose a maximum number of frames to be
tested, rather than only chosing to compare the first frame or
all frames. It also implements to chose the number of frames for
tests comparing to reference data rather than only when comparing
simulation results.
Mark Abraham [Fri, 25 Sep 2020 15:21:43 +0000 (17:21 +0200)]
Ensure the modular simulator runs only when it should
Previous checks were ineffective if an unsupported inputrec
was accompanied by GMX_USE_MODULAR_SIMULATOR.
Change-Id: I74e4c177b6b4d7dc25bb3ca32a7dff5a6027276e
Paul Bauer [Fri, 25 Sep 2020 13:15:44 +0000 (15:15 +0200)]
Fix bad logic in regressiontest branch selection
Change-Id: I1eed8a5228362d34af9426a73c2a9d06eae7a714