2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Definitions of interfaces for GPU state data propagator object.
39 * \author Artem Zhmurov <zhmurov@gmail.com>
41 * \ingroup module_mdtypes
47 #if GMX_GPU != GMX_GPU_NONE
49 # include "gromacs/gpu_utils/device_stream_manager.h"
50 # include "gromacs/gpu_utils/devicebuffer.h"
51 # include "gromacs/gpu_utils/gputraits.h"
52 # include "gromacs/math/vectypes.h"
53 # include "gromacs/mdtypes/state_propagator_data_gpu.h"
54 # include "gromacs/timing/wallcycle.h"
55 # include "gromacs/utility/classhelpers.h"
57 # include "state_propagator_data_gpu_impl.h"
63 StatePropagatorDataGpu::Impl::Impl(const DeviceStreamManager& deviceStreamManager,
64 GpuApiCallBehavior transferKind,
65 int allocationBlockSizeDivisor,
66 gmx_wallcycle* wcycle) :
67 deviceContext_(deviceStreamManager.context()),
68 transferKind_(transferKind),
69 allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
73 GMX_GPU != GMX_GPU_NONE,
74 "GPU state propagator data object should only be constructed on the GPU code-paths.");
76 // We need to keep local copies for re-initialization.
77 pmeStream_ = &deviceStreamManager.stream(DeviceStreamType::Pme);
78 localStream_ = &deviceStreamManager.stream(DeviceStreamType::NonBondedLocal);
79 nonLocalStream_ = &deviceStreamManager.stream(DeviceStreamType::NonBondedNonLocal);
80 // PME stream is used in OpenCL for H2D coordinate transfer
81 if (GMX_GPU == GMX_GPU_OPENCL)
83 updateStream_ = &deviceStreamManager.stream(DeviceStreamType::Pme);
87 updateStream_ = &deviceStreamManager.stream(DeviceStreamType::UpdateAndConstraints);
90 // Map the atom locality to the stream that will be used for coordinates,
91 // velocities and forces transfers. Same streams are used for H2D and D2H copies.
92 // Note, that nullptr stream is used here to indicate that the copy is not supported.
93 xCopyStreams_[AtomLocality::Local] = updateStream_;
94 xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
95 xCopyStreams_[AtomLocality::All] = updateStream_;
97 vCopyStreams_[AtomLocality::Local] = updateStream_;
98 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
99 vCopyStreams_[AtomLocality::All] = updateStream_;
101 fCopyStreams_[AtomLocality::Local] = localStream_;
102 fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
103 fCopyStreams_[AtomLocality::All] = updateStream_;
106 StatePropagatorDataGpu::Impl::Impl(const DeviceStream* pmeStream,
107 const DeviceContext& deviceContext,
108 GpuApiCallBehavior transferKind,
109 int allocationBlockSizeDivisor,
110 gmx_wallcycle* wcycle) :
111 deviceContext_(deviceContext),
112 transferKind_(transferKind),
113 allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
117 GMX_GPU != GMX_GPU_NONE,
118 "GPU state propagator data object should only be constructed on the GPU code-paths.");
120 GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
121 pmeStream_ = pmeStream;
122 localStream_ = pmeStream; // For clearing the force buffer
123 nonLocalStream_ = nullptr;
124 updateStream_ = nullptr;
127 // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
128 // This it temporary measure to make it safe to use this class in those cases.
129 xCopyStreams_[AtomLocality::Local] = pmeStream_;
130 xCopyStreams_[AtomLocality::NonLocal] = nullptr;
131 xCopyStreams_[AtomLocality::All] = pmeStream_;
133 vCopyStreams_[AtomLocality::Local] = nullptr;
134 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
135 vCopyStreams_[AtomLocality::All] = nullptr;
137 fCopyStreams_[AtomLocality::Local] = nullptr;
138 fCopyStreams_[AtomLocality::NonLocal] = nullptr;
139 fCopyStreams_[AtomLocality::All] = nullptr;
142 StatePropagatorDataGpu::Impl::~Impl() {}
144 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
146 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
147 wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
149 numAtomsLocal_ = numAtomsLocal;
150 numAtomsAll_ = numAtomsAll;
153 if (allocationBlockSizeDivisor_ > 0)
155 numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
156 * allocationBlockSizeDivisor_;
160 numAtomsPadded = numAtomsAll_;
163 reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
165 const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
166 if (paddingAllocationSize > 0)
168 // The PME stream is used here because the padding region of d_x_ is only in the PME task.
169 clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
172 reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
173 const int d_fOldCapacity = d_fCapacity_;
174 reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
175 // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
176 // the force accumulation stage before syncing with the local stream. Only done in CUDA,
177 // since the force buffer ops are not implemented in OpenCL.
178 if (GMX_GPU == GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
180 clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
183 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
184 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
187 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
189 int atomsStartAt = 0;
190 int numAtomsToCopy = 0;
191 switch (atomLocality)
193 case AtomLocality::All:
195 numAtomsToCopy = numAtomsAll_;
197 case AtomLocality::Local:
199 numAtomsToCopy = numAtomsLocal_;
201 case AtomLocality::NonLocal:
202 atomsStartAt = numAtomsLocal_;
203 numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
206 GMX_RELEASE_ASSERT(false,
207 "Wrong range of atoms requested in GPU state data manager. Should "
208 "be All, Local or NonLocal.");
210 GMX_ASSERT(atomsStartAt >= 0,
211 "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
212 "was not initialized.");
213 GMX_ASSERT(numAtomsToCopy >= 0,
214 "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
216 return std::make_tuple(atomsStartAt, numAtomsToCopy);
219 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<RVec> d_data,
220 const gmx::ArrayRef<const gmx::RVec> h_data,
222 AtomLocality atomLocality,
223 const DeviceStream& deviceStream)
225 GMX_UNUSED_VALUE(dataSize);
227 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
229 GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
231 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
232 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
233 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
235 int atomsStartAt, numAtomsToCopy;
236 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
238 if (numAtomsToCopy != 0)
240 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
241 "The device allocation is smaller than requested copy range.");
242 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
243 "The host buffer is smaller than the requested copy range.");
245 copyToDeviceBuffer(&d_data, reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
246 atomsStartAt, numAtomsToCopy, deviceStream, transferKind_, nullptr);
249 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
250 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
253 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
254 DeviceBuffer<RVec> d_data,
256 AtomLocality atomLocality,
257 const DeviceStream& deviceStream)
259 GMX_UNUSED_VALUE(dataSize);
261 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
263 GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
265 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
266 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
267 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
269 int atomsStartAt, numAtomsToCopy;
270 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
272 if (numAtomsToCopy != 0)
274 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
275 "The device allocation is smaller than requested copy range.");
276 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
277 "The host buffer is smaller than the requested copy range.");
279 copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]), &d_data,
280 atomsStartAt, numAtomsToCopy, deviceStream, transferKind_, nullptr);
283 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
284 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
287 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
292 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
293 AtomLocality atomLocality)
295 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
296 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
297 GMX_ASSERT(deviceStream != nullptr,
298 "No stream is valid for copying positions with given atom locality.");
300 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
301 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
303 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
305 // markEvent is skipped in OpenCL as:
306 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
307 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
308 // TODO: remove this by adding an event-mark free flavor of this function
309 if (GMX_GPU == GMX_GPU_CUDA)
311 xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
314 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
315 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
318 GpuEventSynchronizer*
319 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
320 const SimulationWorkload& simulationWork,
321 const StepWorkload& stepWork)
323 // The provider of the coordinates may be different for local atoms. If the update is offloaded
324 // and this is not a neighbor search step, then the consumer needs to wait for the update
325 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
326 // the copy event. Non-local coordinates are always provided by the H2D copy.
328 // TODO: This should be reconsidered to support the halo exchange.
330 // In OpenCL no events are used as coordinate sync is not necessary
331 if (GMX_GPU == GMX_GPU_OPENCL)
335 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
337 return &xUpdatedOnDevice_;
341 return &xReadyOnDevice_[atomLocality];
345 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
347 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
348 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
349 xReadyOnDevice_[atomLocality].waitForEvent();
350 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
353 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
355 return &xUpdatedOnDevice_;
358 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
360 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
361 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
362 GMX_ASSERT(deviceStream != nullptr,
363 "No stream is valid for copying positions with given atom locality.");
365 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
366 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
368 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
369 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
370 xReadyOnHost_[atomLocality].markEvent(*deviceStream);
372 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
373 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
376 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
378 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
379 xReadyOnHost_[atomLocality].waitForEvent();
380 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
384 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
389 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
390 AtomLocality atomLocality)
392 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
393 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
394 GMX_ASSERT(deviceStream != nullptr,
395 "No stream is valid for copying velocities with given atom locality.");
397 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
398 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
400 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
401 vReadyOnDevice_[atomLocality].markEvent(*deviceStream);
403 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
404 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
407 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
409 return &vReadyOnDevice_[atomLocality];
413 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
415 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
416 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
417 GMX_ASSERT(deviceStream != nullptr,
418 "No stream is valid for copying velocities with given atom locality.");
420 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
421 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
423 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
424 vReadyOnHost_[atomLocality].markEvent(*deviceStream);
426 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
427 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
430 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
432 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
433 vReadyOnHost_[atomLocality].waitForEvent();
434 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
438 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
443 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
444 AtomLocality atomLocality)
446 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
447 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
448 GMX_ASSERT(deviceStream != nullptr,
449 "No stream is valid for copying forces with given atom locality.");
451 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
452 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
454 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
455 fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
457 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
458 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
461 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
462 bool useGpuFBufferOps)
464 if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
466 return &fReducedOnDevice_;
470 return &fReadyOnDevice_[atomLocality];
474 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
476 return &fReducedOnDevice_;
479 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
481 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
482 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
483 GMX_ASSERT(deviceStream != nullptr,
484 "No stream is valid for copying forces with given atom locality.");
486 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
487 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
489 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
490 fReadyOnHost_[atomLocality].markEvent(*deviceStream);
492 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
493 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
496 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
498 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
499 fReadyOnHost_[atomLocality].waitForEvent();
500 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
503 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
505 return updateStream_;
508 int StatePropagatorDataGpu::Impl::numAtomsLocal()
510 return numAtomsLocal_;
513 int StatePropagatorDataGpu::Impl::numAtomsAll()
519 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
520 GpuApiCallBehavior transferKind,
521 int allocationBlockSizeDivisor,
522 gmx_wallcycle* wcycle) :
523 impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
527 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream* pmeStream,
528 const DeviceContext& deviceContext,
529 GpuApiCallBehavior transferKind,
530 int allocationBlockSizeDivisor,
531 gmx_wallcycle* wcycle) :
532 impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
536 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
538 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
540 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
543 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
545 return impl_->reinit(numAtomsLocal, numAtomsAll);
548 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
550 return impl_->getAtomRangesFromAtomLocality(atomLocality);
554 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
556 return impl_->getCoordinates();
559 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
560 AtomLocality atomLocality)
562 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
565 GpuEventSynchronizer*
566 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
567 const SimulationWorkload& simulationWork,
568 const StepWorkload& stepWork)
570 return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
573 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
575 return impl_->waitCoordinatesCopiedToDevice(atomLocality);
578 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
580 return impl_->xUpdatedOnDevice();
583 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
585 return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
588 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
590 return impl_->waitCoordinatesReadyOnHost(atomLocality);
594 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
596 return impl_->getVelocities();
599 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
600 AtomLocality atomLocality)
602 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
605 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
607 return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
610 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
612 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
615 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
617 return impl_->waitVelocitiesReadyOnHost(atomLocality);
621 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
623 return impl_->getForces();
626 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
628 return impl_->copyForcesToGpu(h_f, atomLocality);
631 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
632 bool useGpuFBufferOps)
634 return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
637 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
639 return impl_->fReducedOnDevice();
642 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
644 return impl_->copyForcesFromGpu(h_f, atomLocality);
647 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
649 return impl_->waitForcesReadyOnHost(atomLocality);
653 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
655 return impl_->getUpdateStream();
658 int StatePropagatorDataGpu::numAtomsLocal()
660 return impl_->numAtomsLocal();
663 int StatePropagatorDataGpu::numAtomsAll()
665 return impl_->numAtomsAll();
670 #endif // GMX_GPU == GMX_GPU_NONE