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/gpu_utils/gpueventsynchronizer.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),
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 updateStream_ = &deviceStreamManager.stream(
82 GMX_GPU_OPENCL ? DeviceStreamType::Pme : DeviceStreamType::UpdateAndConstraints);
84 // Map the atom locality to the stream that will be used for coordinates,
85 // velocities and forces transfers. Same streams are used for H2D and D2H copies.
86 // Note, that nullptr stream is used here to indicate that the copy is not supported.
87 xCopyStreams_[AtomLocality::Local] = updateStream_;
88 xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
89 xCopyStreams_[AtomLocality::All] = nullptr;
91 vCopyStreams_[AtomLocality::Local] = updateStream_;
92 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
93 vCopyStreams_[AtomLocality::All] = nullptr;
95 fCopyStreams_[AtomLocality::Local] = localStream_;
96 fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
97 fCopyStreams_[AtomLocality::All] = updateStream_;
99 copyInStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
100 memsetStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
103 StatePropagatorDataGpu::Impl::Impl(const DeviceStream* pmeStream,
104 const DeviceContext& deviceContext,
105 GpuApiCallBehavior transferKind,
106 int allocationBlockSizeDivisor,
107 gmx_wallcycle* wcycle) :
108 deviceContext_(deviceContext),
109 transferKind_(transferKind),
110 allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
115 "GPU state propagator data object should only be constructed on the GPU code-paths.");
117 GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
118 pmeStream_ = pmeStream;
119 localStream_ = pmeStream; // For clearing the force buffer
120 nonLocalStream_ = nullptr;
121 updateStream_ = nullptr;
124 // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
125 // This it temporary measure to make it safe to use this class in those cases.
126 xCopyStreams_[AtomLocality::Local] = pmeStream_;
127 xCopyStreams_[AtomLocality::NonLocal] = nullptr;
128 xCopyStreams_[AtomLocality::All] = nullptr;
130 vCopyStreams_[AtomLocality::Local] = nullptr;
131 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
132 vCopyStreams_[AtomLocality::All] = nullptr;
134 fCopyStreams_[AtomLocality::Local] = nullptr;
135 fCopyStreams_[AtomLocality::NonLocal] = nullptr;
136 fCopyStreams_[AtomLocality::All] = nullptr;
139 StatePropagatorDataGpu::Impl::~Impl() {}
141 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
143 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
144 wallcycle_sub_start_nocount(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
146 numAtomsLocal_ = numAtomsLocal;
147 numAtomsAll_ = numAtomsAll;
150 if (allocationBlockSizeDivisor_ > 0)
152 numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
153 * allocationBlockSizeDivisor_;
157 numAtomsPadded = numAtomsAll_;
160 reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
162 const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
163 if (paddingAllocationSize > 0)
165 // The PME stream is used here because the padding region of d_x_ is only in the PME task.
166 clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
169 reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
170 const int d_fOldCapacity = d_fCapacity_;
171 reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
173 // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
174 // the force accumulation stage before syncing with the local stream. Only done in CUDA and
175 // SYCL, since the force buffer ops are not implemented in OpenCL.
176 if ((GMX_GPU_CUDA || GMX_GPU_SYCL) && d_fCapacity_ != d_fOldCapacity)
178 clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
181 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
182 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
185 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
187 int atomsStartAt = 0;
188 int numAtomsToCopy = 0;
189 switch (atomLocality)
191 case AtomLocality::All:
193 numAtomsToCopy = numAtomsAll_;
195 case AtomLocality::Local:
197 numAtomsToCopy = numAtomsLocal_;
199 case AtomLocality::NonLocal:
200 atomsStartAt = numAtomsLocal_;
201 numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
204 GMX_RELEASE_ASSERT(false,
205 "Wrong range of atoms requested in GPU state data manager. Should "
206 "be All, Local or NonLocal.");
208 GMX_ASSERT(atomsStartAt >= 0,
209 "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
210 "was not initialized.");
211 GMX_ASSERT(numAtomsToCopy >= 0,
212 "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
214 return std::make_tuple(atomsStartAt, numAtomsToCopy);
217 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<RVec> d_data,
218 const gmx::ArrayRef<const gmx::RVec> h_data,
220 AtomLocality atomLocality,
221 const DeviceStream& deviceStream)
223 GMX_UNUSED_VALUE(dataSize);
225 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
227 GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
229 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
231 int atomsStartAt, numAtomsToCopy;
232 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
234 if (numAtomsToCopy != 0)
236 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
237 "The device allocation is smaller than requested copy range.");
238 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
239 "The host buffer is smaller than the requested copy range.");
241 copyToDeviceBuffer(&d_data,
242 reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
251 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
252 DeviceBuffer<RVec> d_data,
254 AtomLocality atomLocality,
255 const DeviceStream& deviceStream)
257 GMX_UNUSED_VALUE(dataSize);
259 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
261 GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
263 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
265 int atomsStartAt, numAtomsToCopy;
266 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
268 if (numAtomsToCopy != 0)
270 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
271 "The device allocation is smaller than requested copy range.");
272 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
273 "The host buffer is smaller than the requested copy range.");
275 copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]),
285 void StatePropagatorDataGpu::Impl::clearOnDevice(DeviceBuffer<RVec> d_data,
287 AtomLocality atomLocality,
288 const DeviceStream& deviceStream) const
290 GMX_UNUSED_VALUE(dataSize);
292 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
294 GMX_ASSERT(dataSize >= 0, "Trying to clear to device buffer before it was allocated.");
296 GMX_ASSERT(deviceStream.isValid(), "No stream is valid for clearing with given atom locality.");
298 int atomsStartAt, numAtomsToClear;
299 std::tie(atomsStartAt, numAtomsToClear) = getAtomRangesFromAtomLocality(atomLocality);
301 if (numAtomsToClear != 0)
303 GMX_ASSERT(atomsStartAt + numAtomsToClear <= dataSize,
304 "The device allocation is smaller than requested clear range.");
306 clearDeviceBufferAsync(&d_data, atomsStartAt, numAtomsToClear, deviceStream);
310 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
315 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
316 AtomLocality atomLocality)
318 GMX_ASSERT(atomLocality < AtomLocality::All,
319 formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
320 "coordinate transfers, passed value is \"%s\"",
321 enumValueToString(atomLocality))
324 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
325 GMX_ASSERT(deviceStream != nullptr,
326 "No stream is valid for copying positions with given atom locality.");
328 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
329 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
331 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
333 // markEvent is skipped in OpenCL as:
334 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
335 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
336 // TODO: remove this by adding an event-mark free flavor of this function
337 if (GMX_GPU_CUDA || GMX_GPU_SYCL)
339 xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
342 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
343 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
346 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(
347 AtomLocality atomLocality,
348 const SimulationWorkload& simulationWork,
349 const StepWorkload& stepWork,
350 GpuEventSynchronizer* gpuCoordinateHaloLaunched)
352 // The provider of the coordinates may be different for local atoms. If the update is offloaded
353 // and this is not a neighbor search step, then the consumer needs to wait for the update
354 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
355 // the copy event. Non-local coordinates are provided by the GPU halo exchange (if active), otherwise by H2D copy.
357 // In OpenCL no events are used as coordinate sync is not necessary
362 if (atomLocality == AtomLocality::NonLocal && stepWork.useGpuXHalo)
364 GMX_ASSERT(gpuCoordinateHaloLaunched != nullptr,
365 "GPU halo exchange is active but its completion event is null.");
366 return gpuCoordinateHaloLaunched;
368 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
370 GMX_ASSERT(xUpdatedOnDeviceEvent_ != nullptr, "The event synchronizer can not be nullptr.");
371 return xUpdatedOnDeviceEvent_;
375 if (stepWork.doNeighborSearch && xUpdatedOnDeviceEvent_)
377 /* On search steps, we do not consume the result of the GPU update
378 * but rather that of a H2D transfer. So, we reset the event triggered after
379 * update to avoid leaving it unconsumed.
380 * Unfortunately, we don't always have the event marked either (e.g., on the
381 * first step) so we just reset it here.
382 * See Issue #3988. */
383 xUpdatedOnDeviceEvent_->reset();
385 return &xReadyOnDevice_[atomLocality];
389 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
391 wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
392 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
393 xReadyOnDevice_[atomLocality].waitForEvent();
394 wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
397 void StatePropagatorDataGpu::Impl::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
399 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
400 xReadyOnDevice_[atomLocality].consume();
403 void StatePropagatorDataGpu::Impl::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
405 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
406 xReadyOnDevice_[atomLocality].reset();
409 void StatePropagatorDataGpu::Impl::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
411 GMX_ASSERT(xUpdatedOnDeviceEvent != nullptr, "The event synchronizer can not be nullptr.");
412 xUpdatedOnDeviceEvent_ = xUpdatedOnDeviceEvent;
415 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
416 AtomLocality atomLocality,
417 GpuEventSynchronizer* dependency)
419 GMX_ASSERT(atomLocality < AtomLocality::All,
420 formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
421 "coordinate transfers, passed value is \"%s\"",
422 enumValueToString(atomLocality))
424 const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
425 GMX_ASSERT(deviceStream != nullptr,
426 "No stream is valid for copying positions with given atom locality.");
428 if (dependency != nullptr)
430 dependency->enqueueWaitEvent(*deviceStream);
433 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
434 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
436 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
437 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
438 xReadyOnHost_[atomLocality].markEvent(*deviceStream);
440 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
441 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
444 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
446 wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
447 xReadyOnHost_[atomLocality].waitForEvent();
448 wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
452 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
457 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
458 AtomLocality atomLocality)
460 GMX_ASSERT(atomLocality == AtomLocality::Local,
461 formatString("Wrong atom locality. Only Local is allowed for "
462 "velocity transfers, passed value is \"%s\"",
463 enumValueToString(atomLocality))
465 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
466 GMX_ASSERT(deviceStream != nullptr,
467 "No stream is valid for copying velocities with given atom locality.");
469 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
470 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
472 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
473 /* Not marking the event, because it is not used anywhere.
474 * Since we only use velocities on the device for update, and we launch the copy in
475 * the "update" stream, that should be safe.
478 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
479 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
482 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
484 GMX_ASSERT(atomLocality == AtomLocality::Local,
485 formatString("Wrong atom locality. Only Local is allowed for "
486 "velocity transfers, passed value is \"%s\"",
487 enumValueToString(atomLocality))
489 const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
490 GMX_ASSERT(deviceStream != nullptr,
491 "No stream is valid for copying velocities with given atom locality.");
493 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
494 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
496 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
497 vReadyOnHost_[atomLocality].markEvent(*deviceStream);
499 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
500 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
503 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
505 wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
506 vReadyOnHost_[atomLocality].waitForEvent();
507 wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
511 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
516 // Copy CPU forces to GPU using stream internal to this module to allow overlap
517 // with GPU force calculations.
518 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
519 AtomLocality atomLocality)
521 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
522 DeviceStream* deviceStream = copyInStream_.get();
523 GMX_ASSERT(deviceStream != nullptr,
524 "No stream is valid for copying forces with given atom locality.");
526 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
527 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
529 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
530 fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
532 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
533 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
536 void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
538 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
539 DeviceStream* deviceStream = memsetStream_.get();
541 GMX_ASSERT(dependency != nullptr, "Dependency is not valid for clearing forces.");
542 dependency->enqueueWaitEvent(*deviceStream);
544 GMX_ASSERT(deviceStream != nullptr,
545 "No stream is valid for clearing forces with given atom locality.");
547 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
548 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
550 clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
552 fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
554 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
555 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
558 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
559 SimulationWorkload simulationWork)
561 if (stepWork.useGpuFBufferOps && !simulationWork.useCpuPmePpCommunication)
563 return &fReducedOnDevice_[AtomLocality::Local];
567 return &fReadyOnDevice_[AtomLocality::Local];
571 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice(AtomLocality atomLocality)
573 return &fReducedOnDevice_[atomLocality];
576 void StatePropagatorDataGpu::Impl::consumeForcesReducedOnDeviceEvent(AtomLocality atomLocality)
578 fReducedOnDevice_[atomLocality].consume();
581 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReadyOnDevice(AtomLocality atomLocality)
583 return &fReadyOnDevice_[atomLocality];
586 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
588 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
589 const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
590 GMX_ASSERT(deviceStream != nullptr,
591 "No stream is valid for copying forces with given atom locality.");
593 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
594 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
596 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
597 fReadyOnHost_[atomLocality].markEvent(*deviceStream);
599 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
600 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
603 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
605 wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
606 fReadyOnHost_[atomLocality].waitForEvent();
607 wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
610 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
612 return updateStream_;
615 int StatePropagatorDataGpu::Impl::numAtomsLocal() const
617 return numAtomsLocal_;
620 int StatePropagatorDataGpu::Impl::numAtomsAll() const
626 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
627 GpuApiCallBehavior transferKind,
628 int allocationBlockSizeDivisor,
629 gmx_wallcycle* wcycle) :
630 impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
634 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream* pmeStream,
635 const DeviceContext& deviceContext,
636 GpuApiCallBehavior transferKind,
637 int allocationBlockSizeDivisor,
638 gmx_wallcycle* wcycle) :
639 impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
643 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
645 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
647 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
650 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
652 return impl_->reinit(numAtomsLocal, numAtomsAll);
655 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
657 return impl_->getAtomRangesFromAtomLocality(atomLocality);
661 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
663 return impl_->getCoordinates();
666 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
667 AtomLocality atomLocality)
669 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
672 GpuEventSynchronizer*
673 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
674 const SimulationWorkload& simulationWork,
675 const StepWorkload& stepWork,
676 GpuEventSynchronizer* gpuCoordinateHaloLaunched)
678 return impl_->getCoordinatesReadyOnDeviceEvent(
679 atomLocality, simulationWork, stepWork, gpuCoordinateHaloLaunched);
682 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
684 return impl_->waitCoordinatesCopiedToDevice(atomLocality);
687 void StatePropagatorDataGpu::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
689 return impl_->consumeCoordinatesCopiedToDeviceEvent(atomLocality);
692 void StatePropagatorDataGpu::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
694 return impl_->resetCoordinatesCopiedToDeviceEvent(atomLocality);
697 void StatePropagatorDataGpu::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
699 impl_->setXUpdatedOnDeviceEvent(xUpdatedOnDeviceEvent);
702 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x,
703 AtomLocality atomLocality,
704 GpuEventSynchronizer* dependency)
706 return impl_->copyCoordinatesFromGpu(h_x, atomLocality, dependency);
709 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
711 return impl_->waitCoordinatesReadyOnHost(atomLocality);
715 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
717 return impl_->getVelocities();
720 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
721 AtomLocality atomLocality)
723 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
726 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
728 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
731 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
733 return impl_->waitVelocitiesReadyOnHost(atomLocality);
737 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
739 return impl_->getForces();
742 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
744 return impl_->copyForcesToGpu(h_f, atomLocality);
747 void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
749 return impl_->clearForcesOnGpu(atomLocality, dependency);
752 GpuEventSynchronizer* StatePropagatorDataGpu::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
753 SimulationWorkload simulationWork)
755 return impl_->getLocalForcesReadyOnDeviceEvent(stepWork, simulationWork);
758 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice(AtomLocality atomLocality)
760 return impl_->fReducedOnDevice(atomLocality);
763 void StatePropagatorDataGpu::consumeForcesReducedOnDeviceEvent(AtomLocality atomLocality)
765 impl_->consumeForcesReducedOnDeviceEvent(atomLocality);
768 GpuEventSynchronizer* StatePropagatorDataGpu::fReadyOnDevice(AtomLocality atomLocality)
770 return impl_->fReadyOnDevice(atomLocality);
773 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
775 return impl_->copyForcesFromGpu(h_f, atomLocality);
778 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
780 return impl_->waitForcesReadyOnHost(atomLocality);
784 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
786 return impl_->getUpdateStream();
789 int StatePropagatorDataGpu::numAtomsLocal() const
791 return impl_->numAtomsLocal();
794 int StatePropagatorDataGpu::numAtomsAll() const
796 return impl_->numAtomsAll();