2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019, 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 # if GMX_GPU == GMX_GPU_CUDA
50 # include "gromacs/gpu_utils/cudautils.cuh"
52 # include "gromacs/gpu_utils/devicebuffer.h"
53 # if GMX_GPU == GMX_GPU_OPENCL
54 # include "gromacs/gpu_utils/oclutils.h"
56 # include "gromacs/math/vectypes.h"
57 # include "gromacs/mdtypes/state_propagator_data_gpu.h"
58 # include "gromacs/timing/wallcycle.h"
59 # include "gromacs/utility/classhelpers.h"
61 # include "state_propagator_data_gpu_impl.h"
67 StatePropagatorDataGpu::Impl::Impl(const void* pmeStream,
68 const void* localStream,
69 const void* nonLocalStream,
70 const void* deviceContext,
71 GpuApiCallBehavior transferKind,
73 gmx_wallcycle* wcycle) :
74 transferKind_(transferKind),
75 paddingSize_(paddingSize),
78 static_assert(GMX_GPU != GMX_GPU_NONE,
79 "This object should only be constructed on the GPU code-paths.");
81 // TODO: Refactor when the StreamManager is introduced.
82 if (GMX_GPU == GMX_GPU_OPENCL)
84 GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
85 GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set in OpenCL builds.");
87 // The update stream is set to the PME stream in OpenCL, since PME stream is the only stream created in the PME context.
88 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
89 updateStream_ = *static_cast<const CommandStream*>(pmeStream);
90 deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
91 GMX_UNUSED_VALUE(localStream);
92 GMX_UNUSED_VALUE(nonLocalStream);
95 if (GMX_GPU == GMX_GPU_CUDA)
97 if (pmeStream != nullptr)
99 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
101 if (localStream != nullptr)
103 localStream_ = *static_cast<const CommandStream*>(localStream);
105 if (nonLocalStream != nullptr)
107 nonLocalStream_ = *static_cast<const CommandStream*>(nonLocalStream);
110 // TODO: The update stream should be created only when it is needed.
111 # if (GMX_GPU == GMX_GPU_CUDA)
113 stat = cudaStreamCreate(&updateStream_);
114 CU_RET_ERR(stat, "CUDA stream creation failed in StatePropagatorDataGpu");
116 GMX_UNUSED_VALUE(deviceContext);
119 // Map the atom locality to the stream that will be used for coordinates,
120 // velocities and forces transfers. Same streams are used for H2D and D2H copies.
121 // Note, that nullptr stream is used here to indicate that the copy is not supported.
122 xCopyStreams_[AtomLocality::Local] = updateStream_;
123 xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
124 xCopyStreams_[AtomLocality::All] = updateStream_;
126 vCopyStreams_[AtomLocality::Local] = updateStream_;
127 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
128 vCopyStreams_[AtomLocality::All] = updateStream_;
130 fCopyStreams_[AtomLocality::Local] = localStream_;
131 fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
132 fCopyStreams_[AtomLocality::All] = updateStream_;
135 StatePropagatorDataGpu::Impl::Impl(const void* pmeStream,
136 const void* deviceContext,
137 GpuApiCallBehavior transferKind,
139 gmx_wallcycle* wcycle) :
140 transferKind_(transferKind),
141 paddingSize_(paddingSize),
144 static_assert(GMX_GPU != GMX_GPU_NONE,
145 "This object should only be constructed on the GPU code-paths.");
147 if (GMX_GPU == GMX_GPU_OPENCL)
149 GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
150 deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
153 GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set.");
154 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
156 localStream_ = nullptr;
157 nonLocalStream_ = nullptr;
158 updateStream_ = nullptr;
161 // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
162 // This it temporary measure to make it safe to use this class in those cases.
163 xCopyStreams_[AtomLocality::Local] = pmeStream_;
164 xCopyStreams_[AtomLocality::NonLocal] = nullptr;
165 xCopyStreams_[AtomLocality::All] = pmeStream_;
167 vCopyStreams_[AtomLocality::Local] = nullptr;
168 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
169 vCopyStreams_[AtomLocality::All] = nullptr;
171 fCopyStreams_[AtomLocality::Local] = nullptr;
172 fCopyStreams_[AtomLocality::NonLocal] = nullptr;
173 fCopyStreams_[AtomLocality::All] = nullptr;
176 StatePropagatorDataGpu::Impl::~Impl() {}
178 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
180 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
181 wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
183 numAtomsLocal_ = numAtomsLocal;
184 numAtomsAll_ = numAtomsAll;
187 if (paddingSize_ > 0)
189 numAtomsPadded = ((numAtomsAll_ + paddingSize_ - 1) / paddingSize_) * paddingSize_;
193 numAtomsPadded = numAtomsAll_;
196 reallocateDeviceBuffer(&d_x_, DIM * numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
198 const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
199 if (paddingAllocationSize > 0)
201 // The PME stream is used here because the padding region of d_x_ is only in the PME task.
202 clearDeviceBufferAsync(&d_x_, DIM * numAtomsAll_, DIM * paddingAllocationSize, pmeStream_);
205 reallocateDeviceBuffer(&d_v_, DIM * numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
206 const int d_fOldCapacity = d_fCapacity_;
207 reallocateDeviceBuffer(&d_f_, DIM * numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
208 // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
209 // the force accumulation stage before syncing with the local stream. Only done in CUDA,
210 // since the force buffer ops are not implemented in OpenCL.
211 if (GMX_GPU == GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
213 clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, localStream_);
216 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
217 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
220 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
222 int atomsStartAt = 0;
223 int numAtomsToCopy = 0;
224 switch (atomLocality)
226 case AtomLocality::All:
228 numAtomsToCopy = numAtomsAll_;
230 case AtomLocality::Local:
232 numAtomsToCopy = numAtomsLocal_;
234 case AtomLocality::NonLocal:
235 atomsStartAt = numAtomsLocal_;
236 numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
239 GMX_RELEASE_ASSERT(false,
240 "Wrong range of atoms requested in GPU state data manager. Should "
241 "be All, Local or NonLocal.");
243 GMX_ASSERT(atomsStartAt >= 0,
244 "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
245 "was not initialized.");
246 GMX_ASSERT(numAtomsToCopy >= 0,
247 "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
249 return std::make_tuple(atomsStartAt, numAtomsToCopy);
252 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<float> d_data,
253 const gmx::ArrayRef<const gmx::RVec> h_data,
255 AtomLocality atomLocality,
256 CommandStream commandStream)
258 GMX_UNUSED_VALUE(dataSize);
260 GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
262 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
263 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
265 int atomsStartAt, numAtomsToCopy;
266 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
268 int elementsStartAt = atomsStartAt * DIM;
269 int numElementsToCopy = numAtomsToCopy * DIM;
271 if (numAtomsToCopy != 0)
273 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
274 "The device allocation is smaller than requested copy range.");
275 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
276 "The host buffer is smaller than the requested copy range.");
278 copyToDeviceBuffer(&d_data, reinterpret_cast<const float*>(&h_data.data()[atomsStartAt]),
279 elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
282 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
283 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
286 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
287 DeviceBuffer<float> d_data,
289 AtomLocality atomLocality,
290 CommandStream commandStream)
292 GMX_UNUSED_VALUE(dataSize);
294 GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
296 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
297 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
299 int atomsStartAt, numAtomsToCopy;
300 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
302 int elementsStartAt = atomsStartAt * DIM;
303 int numElementsToCopy = numAtomsToCopy * DIM;
305 if (numAtomsToCopy != 0)
307 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
308 "The device allocation is smaller than requested copy range.");
309 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
310 "The host buffer is smaller than the requested copy range.");
312 copyFromDeviceBuffer(reinterpret_cast<float*>(&h_data.data()[atomsStartAt]), &d_data,
313 elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
316 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
317 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
320 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getCoordinates()
325 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
326 AtomLocality atomLocality)
328 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
329 CommandStream commandStream = xCopyStreams_[atomLocality];
330 GMX_ASSERT(commandStream != nullptr,
331 "No stream is valid for copying positions with given atom locality.");
333 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
334 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
336 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, commandStream);
338 // markEvent is skipped in OpenCL as:
339 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
340 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
341 // TODO: remove this by adding an event-mark free flavor of this function
342 if (GMX_GPU == GMX_GPU_CUDA)
344 xReadyOnDevice_[atomLocality].markEvent(commandStream);
347 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
348 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
351 GpuEventSynchronizer*
352 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
353 const SimulationWorkload& simulationWork,
354 const StepWorkload& stepWork)
356 // The provider of the coordinates may be different for local atoms. If the update is offloaded
357 // and this is not a neighbor search step, then the consumer needs to wait for the update
358 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
359 // the copy event. Non-local coordinates are always provided by the H2D copy.
361 // TODO: This should be reconsidered to support the halo exchange.
363 // In OpenCL no events are used as coordinate sync is not necessary
364 if (GMX_GPU == GMX_GPU_OPENCL)
368 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
370 return &xUpdatedOnDevice_;
374 return &xReadyOnDevice_[atomLocality];
378 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
380 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
381 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
382 xReadyOnDevice_[atomLocality].waitForEvent();
383 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
386 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
388 return &xUpdatedOnDevice_;
391 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
393 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
394 CommandStream commandStream = xCopyStreams_[atomLocality];
395 GMX_ASSERT(commandStream != nullptr,
396 "No stream is valid for copying positions with given atom locality.");
398 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
399 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
401 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, commandStream);
402 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
403 xReadyOnHost_[atomLocality].markEvent(commandStream);
405 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
406 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
409 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
411 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
412 xReadyOnHost_[atomLocality].waitForEvent();
413 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
417 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getVelocities()
422 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
423 AtomLocality atomLocality)
425 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
426 CommandStream commandStream = vCopyStreams_[atomLocality];
427 GMX_ASSERT(commandStream != nullptr,
428 "No stream is valid for copying velocities with given atom locality.");
430 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
431 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
433 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, commandStream);
434 vReadyOnDevice_[atomLocality].markEvent(commandStream);
436 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
437 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
440 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
442 return &vReadyOnDevice_[atomLocality];
446 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
448 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
449 CommandStream commandStream = vCopyStreams_[atomLocality];
450 GMX_ASSERT(commandStream != nullptr,
451 "No stream is valid for copying velocities with given atom locality.");
453 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
454 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
456 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, commandStream);
457 vReadyOnHost_[atomLocality].markEvent(commandStream);
459 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
460 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
463 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
465 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
466 vReadyOnHost_[atomLocality].waitForEvent();
467 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
471 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getForces()
476 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
477 AtomLocality atomLocality)
479 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
480 CommandStream commandStream = fCopyStreams_[atomLocality];
481 GMX_ASSERT(commandStream != nullptr,
482 "No stream is valid for copying forces with given atom locality.");
484 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
485 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
487 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, commandStream);
488 fReadyOnDevice_[atomLocality].markEvent(commandStream);
490 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
491 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
494 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
495 bool useGpuFBufferOps)
497 if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
499 return &fReducedOnDevice_;
503 return &fReadyOnDevice_[atomLocality];
507 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
509 return &fReducedOnDevice_;
512 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
514 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
515 CommandStream commandStream = fCopyStreams_[atomLocality];
516 GMX_ASSERT(commandStream != nullptr,
517 "No stream is valid for copying forces with given atom locality.");
519 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
520 wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
522 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, commandStream);
523 fReadyOnHost_[atomLocality].markEvent(commandStream);
525 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
526 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
529 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
531 wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
532 fReadyOnHost_[atomLocality].waitForEvent();
533 wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
536 void* StatePropagatorDataGpu::Impl::getUpdateStream()
538 return &updateStream_;
541 int StatePropagatorDataGpu::Impl::numAtomsLocal()
543 return numAtomsLocal_;
546 int StatePropagatorDataGpu::Impl::numAtomsAll()
552 StatePropagatorDataGpu::StatePropagatorDataGpu(const void* pmeStream,
553 const void* localStream,
554 const void* nonLocalStream,
555 const void* deviceContext,
556 GpuApiCallBehavior transferKind,
558 gmx_wallcycle* wcycle) :
559 impl_(new Impl(pmeStream, localStream, nonLocalStream, deviceContext, transferKind, paddingSize, wcycle))
563 StatePropagatorDataGpu::StatePropagatorDataGpu(const void* pmeStream,
564 const void* deviceContext,
565 GpuApiCallBehavior transferKind,
567 gmx_wallcycle* wcycle) :
568 impl_(new Impl(pmeStream, deviceContext, transferKind, paddingSize, wcycle))
572 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
574 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
576 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
579 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
581 return impl_->reinit(numAtomsLocal, numAtomsAll);
584 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
586 return impl_->getAtomRangesFromAtomLocality(atomLocality);
590 DeviceBuffer<float> StatePropagatorDataGpu::getCoordinates()
592 return impl_->getCoordinates();
595 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
596 AtomLocality atomLocality)
598 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
601 GpuEventSynchronizer*
602 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
603 const SimulationWorkload& simulationWork,
604 const StepWorkload& stepWork)
606 return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
609 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
611 return impl_->waitCoordinatesCopiedToDevice(atomLocality);
614 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
616 return impl_->xUpdatedOnDevice();
619 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
621 return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
624 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
626 return impl_->waitCoordinatesReadyOnHost(atomLocality);
630 DeviceBuffer<float> StatePropagatorDataGpu::getVelocities()
632 return impl_->getVelocities();
635 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
636 AtomLocality atomLocality)
638 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
641 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
643 return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
646 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
648 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
651 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
653 return impl_->waitVelocitiesReadyOnHost(atomLocality);
657 DeviceBuffer<float> StatePropagatorDataGpu::getForces()
659 return impl_->getForces();
662 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
664 return impl_->copyForcesToGpu(h_f, atomLocality);
667 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
668 bool useGpuFBufferOps)
670 return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
673 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
675 return impl_->fReducedOnDevice();
678 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
680 return impl_->copyForcesFromGpu(h_f, atomLocality);
683 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
685 return impl_->waitForcesReadyOnHost(atomLocality);
689 void* StatePropagatorDataGpu::getUpdateStream()
691 return impl_->getUpdateStream();
694 int StatePropagatorDataGpu::numAtomsLocal()
696 return impl_->numAtomsLocal();
699 int StatePropagatorDataGpu::numAtomsAll()
701 return impl_->numAtomsAll();
706 #endif // GMX_GPU == GMX_GPU_NONE