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/utility/classhelpers.h"
60 #include "state_propagator_data_gpu_impl.h"
65 StatePropagatorDataGpu::Impl::Impl(const void *pmeStream,
66 const void *localStream,
67 const void *nonLocalStream,
68 const void *deviceContext,
69 GpuApiCallBehavior transferKind,
71 transferKind_(transferKind),
72 paddingSize_(paddingSize)
74 static_assert(GMX_GPU != GMX_GPU_NONE, "This object should only be constructed on the GPU code-paths.");
76 // TODO: Refactor when the StreamManager is introduced.
77 if (GMX_GPU == GMX_GPU_OPENCL)
79 GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
80 GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set in OpenCL builds.");
82 // The update stream is set to the PME stream in OpenCL, since PME stream is the only stream created in the PME context.
83 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
84 updateStream_ = *static_cast<const CommandStream*>(pmeStream);
85 deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
86 GMX_UNUSED_VALUE(localStream);
87 GMX_UNUSED_VALUE(nonLocalStream);
90 if (GMX_GPU == GMX_GPU_CUDA)
92 if (pmeStream != nullptr)
94 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
96 if (localStream != nullptr)
98 localStream_ = *static_cast<const CommandStream*>(localStream);
100 if (nonLocalStream != nullptr)
102 nonLocalStream_ = *static_cast<const CommandStream*>(nonLocalStream);
105 // TODO: The update stream should be created only when it is needed.
106 #if (GMX_GPU == GMX_GPU_CUDA)
108 stat = cudaStreamCreate(&updateStream_);
109 CU_RET_ERR(stat, "CUDA stream creation failed in StatePropagatorDataGpu");
111 GMX_UNUSED_VALUE(deviceContext);
114 // Map the atom locality to the stream that will be used for coordinates,
115 // velocities and forces transfers. Same streams are used for H2D and D2H copies.
116 // Note, that nullptr stream is used here to indicate that the copy is not supported.
117 xCopyStreams_[AtomLocality::Local] = updateStream_;
118 xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
119 xCopyStreams_[AtomLocality::All] = updateStream_;
121 vCopyStreams_[AtomLocality::Local] = updateStream_;
122 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
123 vCopyStreams_[AtomLocality::All] = updateStream_;
125 fCopyStreams_[AtomLocality::Local] = localStream_;
126 fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
127 fCopyStreams_[AtomLocality::All] = nullptr;
130 StatePropagatorDataGpu::Impl::Impl(const void *pmeStream,
131 const void *deviceContext,
132 GpuApiCallBehavior transferKind,
134 transferKind_(transferKind),
135 paddingSize_(paddingSize)
137 static_assert(GMX_GPU != GMX_GPU_NONE, "This object should only be constructed on the GPU code-paths.");
139 if (GMX_GPU == GMX_GPU_OPENCL)
141 GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
142 deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
145 GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set.");
146 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
148 localStream_ = nullptr;
149 nonLocalStream_ = nullptr;
150 updateStream_ = nullptr;
153 // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
154 // This it temporary measure to make it safe to use this class in those cases.
155 xCopyStreams_[AtomLocality::Local] = pmeStream_;
156 xCopyStreams_[AtomLocality::NonLocal] = nullptr;
157 xCopyStreams_[AtomLocality::All] = pmeStream_;
159 vCopyStreams_[AtomLocality::Local] = nullptr;
160 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
161 vCopyStreams_[AtomLocality::All] = nullptr;
163 fCopyStreams_[AtomLocality::Local] = nullptr;
164 fCopyStreams_[AtomLocality::NonLocal] = nullptr;
165 fCopyStreams_[AtomLocality::All] = nullptr;
168 StatePropagatorDataGpu::Impl::~Impl()
172 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
174 numAtomsLocal_ = numAtomsLocal;
175 numAtomsAll_ = numAtomsAll;
178 if (paddingSize_ > 0)
180 numAtomsPadded = ((numAtomsAll_ + paddingSize_ - 1 ) / paddingSize_ )*paddingSize_;
184 numAtomsPadded = numAtomsAll_;
187 reallocateDeviceBuffer(&d_x_, DIM*numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
189 const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
190 if (paddingAllocationSize > 0)
192 // The PME stream is used here because the padding region of d_x_ is only in the PME task.
193 clearDeviceBufferAsync(&d_x_, DIM*numAtomsAll_, DIM*paddingAllocationSize, pmeStream_);
196 reallocateDeviceBuffer(&d_v_, DIM*numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
197 reallocateDeviceBuffer(&d_f_, DIM*numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
201 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
203 int atomsStartAt = 0;
204 int numAtomsToCopy = 0;
205 switch (atomLocality)
207 case AtomLocality::All:
209 numAtomsToCopy = numAtomsAll_;
211 case AtomLocality::Local:
213 numAtomsToCopy = numAtomsLocal_;
215 case AtomLocality::NonLocal:
216 atomsStartAt = numAtomsLocal_;
217 numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
220 GMX_RELEASE_ASSERT(false, "Wrong range of atoms requested in GPU state data manager. Should be All, Local or NonLocal.");
222 GMX_ASSERT(atomsStartAt >= 0, "The first elemtnt to copy has negative index. Probably, the GPU propagator state was not initialized.");
223 GMX_ASSERT(numAtomsToCopy >= 0, "Number of atoms to copy is negative. Probably, the GPU propagator state was not initialized.");
224 return std::make_tuple(atomsStartAt, numAtomsToCopy);
227 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<float> d_data,
228 const gmx::ArrayRef<const gmx::RVec> h_data,
230 AtomLocality atomLocality,
231 CommandStream commandStream)
234 GMX_UNUSED_VALUE(dataSize);
236 GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
238 int atomsStartAt, numAtomsToCopy;
239 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
241 int elementsStartAt = atomsStartAt*DIM;
242 int numElementsToCopy = numAtomsToCopy*DIM;
244 if (numAtomsToCopy != 0)
246 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize, "The device allocation is smaller than requested copy range.");
247 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(), "The host buffer is smaller than the requested copy range.");
249 copyToDeviceBuffer(&d_data, reinterpret_cast<const float *>(&h_data.data()[atomsStartAt]),
250 elementsStartAt, numElementsToCopy,
251 commandStream, transferKind_, nullptr);
255 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
256 DeviceBuffer<float> d_data,
258 AtomLocality atomLocality,
259 CommandStream commandStream)
262 GMX_UNUSED_VALUE(dataSize);
264 GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
266 int atomsStartAt, numAtomsToCopy;
267 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
269 int elementsStartAt = atomsStartAt*DIM;
270 int numElementsToCopy = numAtomsToCopy*DIM;
272 if (numAtomsToCopy != 0)
274 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize, "The device allocation is smaller than requested copy range.");
275 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(), "The host buffer is smaller than the requested copy range.");
277 copyFromDeviceBuffer(reinterpret_cast<float*>(&h_data.data()[atomsStartAt]), &d_data,
278 elementsStartAt, numElementsToCopy,
279 commandStream, transferKind_, nullptr);
283 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getCoordinates()
288 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
289 AtomLocality atomLocality)
291 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
292 CommandStream commandStream = xCopyStreams_[atomLocality];
293 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying positions with given atom locality.");
295 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, commandStream);
297 // markEvent is skipped in OpenCL as:
298 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
299 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
300 // TODO: remove this by adding an event-mark free flavor of this function
301 if (GMX_GPU == GMX_GPU_CUDA)
303 xReadyOnDevice_[atomLocality].markEvent(commandStream);
307 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
308 const SimulationWorkload &simulationWork,
309 const StepWorkload &stepWork)
311 // The provider of the coordinates may be different for local atoms. If the update is offloaded
312 // and this is not a neighbor search step, then the consumer needs to wait for the update
313 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
314 // the copy event. Non-local coordinates are always provided by the H2D copy.
316 // TODO: This should be reconsidered to support the halo exchange.
318 // In OpenCL no events are used as coordinate sync is not necessary
319 if (GMX_GPU == GMX_GPU_OPENCL)
323 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
325 return &xUpdatedOnDevice_;
329 return &xReadyOnDevice_[atomLocality];
333 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
335 return &xUpdatedOnDevice_;
338 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
339 AtomLocality atomLocality)
341 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
342 CommandStream commandStream = xCopyStreams_[atomLocality];
343 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying positions with given atom locality.");
345 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, commandStream);
346 // TODO: Remove When event-based synchronization is introduced
347 gpuStreamSynchronize(commandStream);
348 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
349 xReadyOnHost_[atomLocality].markEvent(commandStream);
352 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
354 xReadyOnHost_[atomLocality].waitForEvent();
358 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getVelocities()
363 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
364 AtomLocality atomLocality)
366 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
367 CommandStream commandStream = vCopyStreams_[atomLocality];
368 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying velocities with given atom locality.");
370 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, commandStream);
371 // TODO: Remove When event-based synchronization is introduced
372 gpuStreamSynchronize(commandStream);
373 vReadyOnDevice_[atomLocality].markEvent(commandStream);
376 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
378 return &vReadyOnDevice_[atomLocality];
382 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v,
383 AtomLocality atomLocality)
385 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
386 CommandStream commandStream = vCopyStreams_[atomLocality];
387 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying velocities with given atom locality.");
389 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, commandStream);
390 // TODO: Remove When event-based synchronization is introduced
391 gpuStreamSynchronize(commandStream);
392 vReadyOnHost_[atomLocality].markEvent(commandStream);
395 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
397 vReadyOnHost_[atomLocality].waitForEvent();
401 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getForces()
406 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
407 AtomLocality atomLocality)
409 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
410 CommandStream commandStream = fCopyStreams_[atomLocality];
411 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying forces with given atom locality.");
413 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, commandStream);
414 fReadyOnDevice_[atomLocality].markEvent(commandStream);
417 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
418 bool useGpuFBufferOps)
420 if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
422 return &fReducedOnDevice_;
426 return &fReadyOnDevice_[atomLocality];
430 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
432 return &fReducedOnDevice_;
435 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f,
436 AtomLocality atomLocality)
438 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
439 CommandStream commandStream = fCopyStreams_[atomLocality];
440 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying forces with given atom locality.");
442 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, commandStream);
443 fReadyOnHost_[atomLocality].markEvent(commandStream);
446 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
448 fReadyOnHost_[atomLocality].waitForEvent();
451 void* StatePropagatorDataGpu::Impl::getUpdateStream()
453 return updateStream_;
456 int StatePropagatorDataGpu::Impl::numAtomsLocal()
458 return numAtomsLocal_;
461 int StatePropagatorDataGpu::Impl::numAtomsAll()
468 StatePropagatorDataGpu::StatePropagatorDataGpu(const void *pmeStream,
469 const void *localStream,
470 const void *nonLocalStream,
471 const void *deviceContext,
472 GpuApiCallBehavior transferKind,
474 : impl_(new Impl(pmeStream,
483 StatePropagatorDataGpu::StatePropagatorDataGpu(const void *pmeStream,
484 const void *deviceContext,
485 GpuApiCallBehavior transferKind,
487 : impl_(new Impl(pmeStream,
494 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu && /* other */) noexcept = default;
496 StatePropagatorDataGpu &StatePropagatorDataGpu::operator=(StatePropagatorDataGpu && /* other */) noexcept = default;
498 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
501 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
503 return impl_->reinit(numAtomsLocal, numAtomsAll);
506 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
508 return impl_->getAtomRangesFromAtomLocality(atomLocality);
512 DeviceBuffer<float> StatePropagatorDataGpu::getCoordinates()
514 return impl_->getCoordinates();
517 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
518 AtomLocality atomLocality)
520 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
523 GpuEventSynchronizer* StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
524 const SimulationWorkload &simulationWork,
525 const StepWorkload &stepWork)
527 return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
530 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
532 return impl_->xUpdatedOnDevice();
535 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x,
536 AtomLocality atomLocality)
538 return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
541 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
543 return impl_->waitCoordinatesReadyOnHost(atomLocality);
547 DeviceBuffer<float> StatePropagatorDataGpu::getVelocities()
549 return impl_->getVelocities();
552 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
553 AtomLocality atomLocality)
555 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
558 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
560 return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
563 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v,
564 AtomLocality atomLocality)
566 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
569 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
571 return impl_->waitVelocitiesReadyOnHost(atomLocality);
575 DeviceBuffer<float> StatePropagatorDataGpu::getForces()
577 return impl_->getForces();
580 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
581 AtomLocality atomLocality)
583 return impl_->copyForcesToGpu(h_f, atomLocality);
586 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
587 bool useGpuFBufferOps)
589 return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
592 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
594 return impl_->fReducedOnDevice();
597 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f,
598 AtomLocality atomLocality)
600 return impl_->copyForcesFromGpu(h_f, atomLocality);
603 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
605 return impl_->waitForcesReadyOnHost(atomLocality);
609 void* StatePropagatorDataGpu::getUpdateStream()
611 return impl_->getUpdateStream();
614 int StatePropagatorDataGpu::numAtomsLocal()
616 return impl_->numAtomsLocal();
619 int StatePropagatorDataGpu::numAtomsAll()
621 return impl_->numAtomsAll();
626 #endif // GMX_GPU == GMX_GPU_NONE