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
49 # include "gromacs/gpu_utils/device_stream_manager.h"
50 # include "gromacs/gpu_utils/devicebuffer.h"
51 # include "gromacs/math/vectypes.h"
52 # include "gromacs/mdtypes/state_propagator_data_gpu.h"
53 # include "gromacs/timing/wallcycle.h"
54 # include "gromacs/utility/classhelpers.h"
56 # include "state_propagator_data_gpu_impl.h"
62 StatePropagatorDataGpu::Impl::Impl(const DeviceStreamManager& deviceStreamManager,
63 GpuApiCallBehavior transferKind,
64 int allocationBlockSizeDivisor,
65 gmx_wallcycle* wcycle) :
66 deviceContext_(deviceStreamManager.context()),
67 transferKind_(transferKind),
68 allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
73 "GPU state propagator data object should only be constructed on the GPU code-paths.");
75 // We need to keep local copies for re-initialization.
76 pmeStream_ = &deviceStreamManager.stream(DeviceStreamType::Pme);
77 localStream_ = &deviceStreamManager.stream(DeviceStreamType::NonBondedLocal);
78 nonLocalStream_ = &deviceStreamManager.stream(DeviceStreamType::NonBondedNonLocal);
79 // PME stream is used in OpenCL for H2D coordinate transfer
80 updateStream_ = &deviceStreamManager.stream(
81 GMX_GPU_OPENCL ? DeviceStreamType::Pme : DeviceStreamType::UpdateAndConstraints);
83 // Map the atom locality to the stream that will be used for coordinates,
84 // velocities and forces transfers. Same streams are used for H2D and D2H copies.
85 // Note, that nullptr stream is used here to indicate that the copy is not supported.
86 xCopyStreams_[AtomLocality::Local] = updateStream_;
87 xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
88 xCopyStreams_[AtomLocality::All] = updateStream_;
90 vCopyStreams_[AtomLocality::Local] = updateStream_;
91 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
92 vCopyStreams_[AtomLocality::All] = updateStream_;
94 fCopyStreams_[AtomLocality::Local] = localStream_;
95 fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
96 fCopyStreams_[AtomLocality::All] = updateStream_;
99 StatePropagatorDataGpu::Impl::Impl(const DeviceStream* pmeStream,
100 const DeviceContext& deviceContext,
101 GpuApiCallBehavior transferKind,
102 int allocationBlockSizeDivisor,
103 gmx_wallcycle* wcycle) :
104 deviceContext_(deviceContext),
105 transferKind_(transferKind),
106 allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
111 "GPU state propagator data object should only be constructed on the GPU code-paths.");
113 GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
114 pmeStream_ = pmeStream;
115 localStream_ = pmeStream; // For clearing the force buffer
116 nonLocalStream_ = nullptr;
117 updateStream_ = nullptr;
120 // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
121 // This it temporary measure to make it safe to use this class in those cases.
122 xCopyStreams_[AtomLocality::Local] = pmeStream_;
123 xCopyStreams_[AtomLocality::NonLocal] = nullptr;
124 xCopyStreams_[AtomLocality::All] = pmeStream_;
126 vCopyStreams_[AtomLocality::Local] = nullptr;
127 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
128 vCopyStreams_[AtomLocality::All] = nullptr;
130 fCopyStreams_[AtomLocality::Local] = nullptr;
131 fCopyStreams_[AtomLocality::NonLocal] = nullptr;
132 fCopyStreams_[AtomLocality::All] = nullptr;
135 StatePropagatorDataGpu::Impl::~Impl() {}
137 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
139 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
140 wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
142 numAtomsLocal_ = numAtomsLocal;
143 numAtomsAll_ = numAtomsAll;
146 if (allocationBlockSizeDivisor_ > 0)
148 numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
149 * allocationBlockSizeDivisor_;
153 numAtomsPadded = numAtomsAll_;
156 reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
158 const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
159 if (paddingAllocationSize > 0)
161 // The PME stream is used here because the padding region of d_x_ is only in the PME task.
162 clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
165 reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
166 const int d_fOldCapacity = d_fCapacity_;
167 reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
169 // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
170 // the force accumulation stage before syncing with the local stream. Only done in CUDA,
171 // since the force buffer ops are not implemented in OpenCL.
172 if (GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
174 clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
177 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
178 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
181 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
183 int atomsStartAt = 0;
184 int numAtomsToCopy = 0;
185 switch (atomLocality)
187 case AtomLocality::All:
189 numAtomsToCopy = numAtomsAll_;
191 case AtomLocality::Local:
193 numAtomsToCopy = numAtomsLocal_;
195 case AtomLocality::NonLocal:
196 atomsStartAt = numAtomsLocal_;
197 numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
200 GMX_RELEASE_ASSERT(false,
201 "Wrong range of atoms requested in GPU state data manager. Should "
202 "be All, Local or NonLocal.");
204 GMX_ASSERT(atomsStartAt >= 0,
205 "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
206 "was not initialized.");
207 GMX_ASSERT(numAtomsToCopy >= 0,
208 "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
210 return std::make_tuple(atomsStartAt, numAtomsToCopy);
213 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<RVec> d_data,
214 const gmx::ArrayRef<const gmx::RVec> h_data,
216 AtomLocality atomLocality,
217 const DeviceStream& deviceStream)
219 GMX_UNUSED_VALUE(dataSize);
221 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
223 GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
225 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
226 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
227 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
229 int atomsStartAt, numAtomsToCopy;
230 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
232 if (numAtomsToCopy != 0)
234 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
235 "The device allocation is smaller than requested copy range.");
236 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
237 "The host buffer is smaller than the requested copy range.");
239 copyToDeviceBuffer(&d_data, reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
240 atomsStartAt, numAtomsToCopy, deviceStream, transferKind_, nullptr);
243 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
244 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
247 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
248 DeviceBuffer<RVec> d_data,
250 AtomLocality atomLocality,
251 const DeviceStream& deviceStream)
253 GMX_UNUSED_VALUE(dataSize);
255 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
257 GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
259 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
260 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
261 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
263 int atomsStartAt, numAtomsToCopy;
264 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
266 if (numAtomsToCopy != 0)
268 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
269 "The device allocation is smaller than requested copy range.");
270 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
271 "The host buffer is smaller than the requested copy range.");
273 copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]), &d_data,
274 atomsStartAt, numAtomsToCopy, deviceStream, transferKind_, nullptr);
277 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
278 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
281 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
286 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
287 AtomLocality atomLocality)
289 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
290 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
291 GMX_ASSERT(deviceStream != nullptr,
292 "No stream is valid for copying positions with given atom locality.");
294 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
295 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
297 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
299 // markEvent is skipped in OpenCL as:
300 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
301 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
302 // TODO: remove this by adding an event-mark free flavor of this function
305 xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
308 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
309 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
312 GpuEventSynchronizer*
313 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
314 const SimulationWorkload& simulationWork,
315 const StepWorkload& stepWork)
317 // The provider of the coordinates may be different for local atoms. If the update is offloaded
318 // and this is not a neighbor search step, then the consumer needs to wait for the update
319 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
320 // the copy event. Non-local coordinates are always provided by the H2D copy.
322 // TODO: This should be reconsidered to support the halo exchange.
324 // In OpenCL no events are used as coordinate sync is not necessary
329 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
331 return &xUpdatedOnDevice_;
335 return &xReadyOnDevice_[atomLocality];
339 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
341 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
342 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
343 xReadyOnDevice_[atomLocality].waitForEvent();
344 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
347 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
349 return &xUpdatedOnDevice_;
352 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
354 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
355 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
356 GMX_ASSERT(deviceStream != nullptr,
357 "No stream is valid for copying positions with given atom locality.");
359 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
360 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
362 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
363 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
364 xReadyOnHost_[atomLocality].markEvent(*deviceStream);
366 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
367 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
370 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
372 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
373 xReadyOnHost_[atomLocality].waitForEvent();
374 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
378 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
383 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
384 AtomLocality atomLocality)
386 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
387 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
388 GMX_ASSERT(deviceStream != nullptr,
389 "No stream is valid for copying velocities with given atom locality.");
391 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
392 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
394 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
395 vReadyOnDevice_[atomLocality].markEvent(*deviceStream);
397 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
398 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
401 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
403 return &vReadyOnDevice_[atomLocality];
407 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
409 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
410 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
411 GMX_ASSERT(deviceStream != nullptr,
412 "No stream is valid for copying velocities with given atom locality.");
414 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
415 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
417 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
418 vReadyOnHost_[atomLocality].markEvent(*deviceStream);
420 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
421 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
424 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
426 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
427 vReadyOnHost_[atomLocality].waitForEvent();
428 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
432 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
437 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
438 AtomLocality atomLocality)
440 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
441 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
442 GMX_ASSERT(deviceStream != nullptr,
443 "No stream is valid for copying forces with given atom locality.");
445 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
446 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
448 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
449 fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
451 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
452 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
455 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
456 bool useGpuFBufferOps)
458 if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
460 return &fReducedOnDevice_;
464 return &fReadyOnDevice_[atomLocality];
468 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
470 return &fReducedOnDevice_;
473 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
475 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
476 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
477 GMX_ASSERT(deviceStream != nullptr,
478 "No stream is valid for copying forces with given atom locality.");
480 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
481 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
483 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
484 fReadyOnHost_[atomLocality].markEvent(*deviceStream);
486 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
487 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
490 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
492 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
493 fReadyOnHost_[atomLocality].waitForEvent();
494 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
497 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
499 return updateStream_;
502 int StatePropagatorDataGpu::Impl::numAtomsLocal()
504 return numAtomsLocal_;
507 int StatePropagatorDataGpu::Impl::numAtomsAll()
513 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
514 GpuApiCallBehavior transferKind,
515 int allocationBlockSizeDivisor,
516 gmx_wallcycle* wcycle) :
517 impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
521 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream* pmeStream,
522 const DeviceContext& deviceContext,
523 GpuApiCallBehavior transferKind,
524 int allocationBlockSizeDivisor,
525 gmx_wallcycle* wcycle) :
526 impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
530 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
532 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
534 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
537 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
539 return impl_->reinit(numAtomsLocal, numAtomsAll);
542 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
544 return impl_->getAtomRangesFromAtomLocality(atomLocality);
548 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
550 return impl_->getCoordinates();
553 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
554 AtomLocality atomLocality)
556 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
559 GpuEventSynchronizer*
560 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
561 const SimulationWorkload& simulationWork,
562 const StepWorkload& stepWork)
564 return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
567 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
569 return impl_->waitCoordinatesCopiedToDevice(atomLocality);
572 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
574 return impl_->xUpdatedOnDevice();
577 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
579 return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
582 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
584 return impl_->waitCoordinatesReadyOnHost(atomLocality);
588 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
590 return impl_->getVelocities();
593 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
594 AtomLocality atomLocality)
596 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
599 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
601 return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
604 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
606 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
609 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
611 return impl_->waitVelocitiesReadyOnHost(atomLocality);
615 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
617 return impl_->getForces();
620 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
622 return impl_->copyForcesToGpu(h_f, atomLocality);
625 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
626 bool useGpuFBufferOps)
628 return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
631 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
633 return impl_->fReducedOnDevice();
636 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
638 return impl_->copyForcesFromGpu(h_f, atomLocality);
641 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
643 return impl_->waitForcesReadyOnHost(atomLocality);
647 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
649 return impl_->getUpdateStream();
652 int StatePropagatorDataGpu::numAtomsLocal()
654 return impl_->numAtomsLocal();
657 int StatePropagatorDataGpu::numAtomsAll()
659 return impl_->numAtomsAll();