2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020,2021, 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.");
227 int atomsStartAt, numAtomsToCopy;
228 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
230 if (numAtomsToCopy != 0)
232 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
233 "The device allocation is smaller than requested copy range.");
234 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
235 "The host buffer is smaller than the requested copy range.");
237 copyToDeviceBuffer(&d_data,
238 reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
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.");
261 int atomsStartAt, numAtomsToCopy;
262 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
264 if (numAtomsToCopy != 0)
266 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
267 "The device allocation is smaller than requested copy range.");
268 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
269 "The host buffer is smaller than the requested copy range.");
271 copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]),
281 void StatePropagatorDataGpu::Impl::clearOnDevice(DeviceBuffer<RVec> d_data,
283 AtomLocality atomLocality,
284 const DeviceStream& deviceStream)
286 GMX_UNUSED_VALUE(dataSize);
288 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
290 GMX_ASSERT(dataSize >= 0, "Trying to clear to device buffer before it was allocated.");
292 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for clearing with given atom locality.");
294 int atomsStartAt, numAtomsToClear;
295 std::tie(atomsStartAt, numAtomsToClear) = getAtomRangesFromAtomLocality(atomLocality);
297 if (numAtomsToClear != 0)
299 GMX_ASSERT(atomsStartAt + numAtomsToClear <= dataSize,
300 "The device allocation is smaller than requested clear range.");
302 clearDeviceBufferAsync(&d_data, atomsStartAt, numAtomsToClear, deviceStream);
306 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
311 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
312 AtomLocality atomLocality)
314 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
315 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
316 GMX_ASSERT(deviceStream != nullptr,
317 "No stream is valid for copying positions with given atom locality.");
319 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
320 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
322 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
324 // markEvent is skipped in OpenCL as:
325 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
326 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
327 // TODO: remove this by adding an event-mark free flavor of this function
330 xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
333 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
334 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
337 GpuEventSynchronizer*
338 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
339 const SimulationWorkload& simulationWork,
340 const StepWorkload& stepWork)
342 // The provider of the coordinates may be different for local atoms. If the update is offloaded
343 // and this is not a neighbor search step, then the consumer needs to wait for the update
344 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
345 // the copy event. Non-local coordinates are always provided by the H2D copy.
347 // TODO: This should be reconsidered to support the halo exchange.
349 // In OpenCL no events are used as coordinate sync is not necessary
354 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
356 return &xUpdatedOnDevice_;
360 return &xReadyOnDevice_[atomLocality];
364 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
366 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
367 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
368 xReadyOnDevice_[atomLocality].waitForEvent();
369 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
372 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
374 return &xUpdatedOnDevice_;
377 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
379 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
380 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
381 GMX_ASSERT(deviceStream != nullptr,
382 "No stream is valid for copying positions with given atom locality.");
384 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
385 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
387 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
388 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
389 xReadyOnHost_[atomLocality].markEvent(*deviceStream);
391 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
392 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
395 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
397 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
398 xReadyOnHost_[atomLocality].waitForEvent();
399 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
403 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
408 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
409 AtomLocality atomLocality)
411 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
412 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
413 GMX_ASSERT(deviceStream != nullptr,
414 "No stream is valid for copying velocities with given atom locality.");
416 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
417 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
419 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
420 vReadyOnDevice_[atomLocality].markEvent(*deviceStream);
422 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
423 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
426 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
428 return &vReadyOnDevice_[atomLocality];
432 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
434 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
435 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
436 GMX_ASSERT(deviceStream != nullptr,
437 "No stream is valid for copying velocities with given atom locality.");
439 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
440 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
442 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
443 vReadyOnHost_[atomLocality].markEvent(*deviceStream);
445 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
446 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
449 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
451 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
452 vReadyOnHost_[atomLocality].waitForEvent();
453 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
457 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
462 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
463 AtomLocality atomLocality)
465 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
466 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
467 GMX_ASSERT(deviceStream != nullptr,
468 "No stream is valid for copying forces with given atom locality.");
470 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
471 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
473 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
474 fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
476 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
477 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
480 void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality)
482 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
483 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
484 GMX_ASSERT(deviceStream != nullptr,
485 "No stream is valid for clearing forces with given atom locality.");
487 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
488 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
490 clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
492 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
493 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
496 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
497 bool useGpuFBufferOps)
499 if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
501 return &fReducedOnDevice_;
505 return &fReadyOnDevice_[atomLocality];
509 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
511 return &fReducedOnDevice_;
514 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
516 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
517 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
518 GMX_ASSERT(deviceStream != nullptr,
519 "No stream is valid for copying forces with given atom locality.");
521 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
522 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
524 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
525 fReadyOnHost_[atomLocality].markEvent(*deviceStream);
527 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
528 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
531 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
533 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
534 fReadyOnHost_[atomLocality].waitForEvent();
535 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
538 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
540 return updateStream_;
543 int StatePropagatorDataGpu::Impl::numAtomsLocal()
545 return numAtomsLocal_;
548 int StatePropagatorDataGpu::Impl::numAtomsAll()
554 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
555 GpuApiCallBehavior transferKind,
556 int allocationBlockSizeDivisor,
557 gmx_wallcycle* wcycle) :
558 impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
562 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream* pmeStream,
563 const DeviceContext& deviceContext,
564 GpuApiCallBehavior transferKind,
565 int allocationBlockSizeDivisor,
566 gmx_wallcycle* wcycle) :
567 impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
571 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
573 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
575 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
578 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
580 return impl_->reinit(numAtomsLocal, numAtomsAll);
583 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
585 return impl_->getAtomRangesFromAtomLocality(atomLocality);
589 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
591 return impl_->getCoordinates();
594 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
595 AtomLocality atomLocality)
597 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
600 GpuEventSynchronizer*
601 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
602 const SimulationWorkload& simulationWork,
603 const StepWorkload& stepWork)
605 return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
608 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
610 return impl_->waitCoordinatesCopiedToDevice(atomLocality);
613 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
615 return impl_->xUpdatedOnDevice();
618 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
620 return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
623 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
625 return impl_->waitCoordinatesReadyOnHost(atomLocality);
629 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
631 return impl_->getVelocities();
634 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
635 AtomLocality atomLocality)
637 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
640 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
642 return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
645 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
647 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
650 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
652 return impl_->waitVelocitiesReadyOnHost(atomLocality);
656 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
658 return impl_->getForces();
661 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
663 return impl_->copyForcesToGpu(h_f, atomLocality);
666 void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality)
668 return impl_->clearForcesOnGpu(atomLocality);
671 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
672 bool useGpuFBufferOps)
674 return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
677 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
679 return impl_->fReducedOnDevice();
682 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
684 return impl_->copyForcesFromGpu(h_f, atomLocality);
687 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
689 return impl_->waitForcesReadyOnHost(atomLocality);
693 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
695 return impl_->getUpdateStream();
698 int StatePropagatorDataGpu::numAtomsLocal()
700 return impl_->numAtomsLocal();
703 int StatePropagatorDataGpu::numAtomsAll()
705 return impl_->numAtomsAll();