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.");
75 GMX_RELEASE_ASSERT(getenv("GMX_USE_GPU_BUFFER_OPS") == nullptr, "GPU buffer ops are not supported in this build.");
77 // TODO: Refactor when the StreamManager is introduced.
78 if (GMX_GPU == GMX_GPU_OPENCL)
80 GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
81 GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set in OpenCL builds.");
83 // The update stream is set to the PME stream in OpenCL, since PME stream is the only stream created in the PME context.
84 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
85 updateStream_ = *static_cast<const CommandStream*>(pmeStream);
86 deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
87 GMX_UNUSED_VALUE(localStream);
88 GMX_UNUSED_VALUE(nonLocalStream);
91 if (GMX_GPU == GMX_GPU_CUDA)
93 if (pmeStream != nullptr)
95 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
97 if (localStream != nullptr)
99 localStream_ = *static_cast<const CommandStream*>(localStream);
101 if (nonLocalStream != nullptr)
103 nonLocalStream_ = *static_cast<const CommandStream*>(nonLocalStream);
106 // TODO: The update stream should be created only when it is needed.
107 #if (GMX_GPU == GMX_GPU_CUDA)
109 stat = cudaStreamCreate(&updateStream_);
110 CU_RET_ERR(stat, "CUDA stream creation failed in StatePropagatorDataGpu");
112 GMX_UNUSED_VALUE(deviceContext);
115 // Map the atom locality to the stream that will be used for coordinates,
116 // velocities and forces transfers. Same streams are used for H2D and D2H copies.
117 // Note, that nullptr stream is used here to indicate that the copy is not supported.
118 xCopyStreams_[AtomLocality::Local] = updateStream_;
119 xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
120 xCopyStreams_[AtomLocality::All] = updateStream_;
122 vCopyStreams_[AtomLocality::Local] = updateStream_;
123 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
124 vCopyStreams_[AtomLocality::All] = updateStream_;
126 fCopyStreams_[AtomLocality::Local] = localStream_;
127 fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
128 fCopyStreams_[AtomLocality::All] = nullptr;
131 StatePropagatorDataGpu::Impl::~Impl()
135 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
137 numAtomsLocal_ = numAtomsLocal;
138 numAtomsAll_ = numAtomsAll;
141 if (paddingSize_ > 0)
143 numAtomsPadded = ((numAtomsAll_ + paddingSize_ - 1 ) / paddingSize_ )*paddingSize_;
147 numAtomsPadded = numAtomsAll_;
150 reallocateDeviceBuffer(&d_x_, DIM*numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
152 const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
153 if (paddingAllocationSize > 0)
155 // The PME stream is used here because the padding region of d_x_ is only in the PME task.
156 clearDeviceBufferAsync(&d_x_, DIM*numAtomsAll_, DIM*paddingAllocationSize, pmeStream_);
159 reallocateDeviceBuffer(&d_v_, DIM*numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
160 reallocateDeviceBuffer(&d_f_, DIM*numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
164 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
166 int atomsStartAt = 0;
167 int numAtomsToCopy = 0;
168 switch (atomLocality)
170 case AtomLocality::All:
172 numAtomsToCopy = numAtomsAll_;
174 case AtomLocality::Local:
176 numAtomsToCopy = numAtomsLocal_;
178 case AtomLocality::NonLocal:
179 atomsStartAt = numAtomsLocal_;
180 numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
183 GMX_RELEASE_ASSERT(false, "Wrong range of atoms requested in GPU state data manager. Should be All, Local or NonLocal.");
185 GMX_ASSERT(atomsStartAt >= 0, "The first elemtnt to copy has negative index. Probably, the GPU propagator state was not initialized.");
186 GMX_ASSERT(numAtomsToCopy >= 0, "Number of atoms to copy is negative. Probably, the GPU propagator state was not initialized.");
187 return std::make_tuple(atomsStartAt, numAtomsToCopy);
190 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<float> d_data,
191 const gmx::ArrayRef<const gmx::RVec> h_data,
193 AtomLocality atomLocality,
194 CommandStream commandStream)
197 GMX_UNUSED_VALUE(dataSize);
199 GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
201 int atomsStartAt, numAtomsToCopy;
202 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
204 int elementsStartAt = atomsStartAt*DIM;
205 int numElementsToCopy = numAtomsToCopy*DIM;
207 if (numAtomsToCopy != 0)
209 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize, "The device allocation is smaller than requested copy range.");
210 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(), "The host buffer is smaller than the requested copy range.");
212 copyToDeviceBuffer(&d_data, reinterpret_cast<const float *>(&h_data.data()[atomsStartAt]),
213 elementsStartAt, numElementsToCopy,
214 commandStream, transferKind_, nullptr);
218 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
219 DeviceBuffer<float> d_data,
221 AtomLocality atomLocality,
222 CommandStream commandStream)
225 GMX_UNUSED_VALUE(dataSize);
227 GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
229 int atomsStartAt, numAtomsToCopy;
230 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
232 int elementsStartAt = atomsStartAt*DIM;
233 int numElementsToCopy = numAtomsToCopy*DIM;
235 if (numAtomsToCopy != 0)
237 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize, "The device allocation is smaller than requested copy range.");
238 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(), "The host buffer is smaller than the requested copy range.");
240 copyFromDeviceBuffer(reinterpret_cast<float*>(&h_data.data()[atomsStartAt]), &d_data,
241 elementsStartAt, numElementsToCopy,
242 commandStream, transferKind_, nullptr);
246 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getCoordinates()
251 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
252 AtomLocality atomLocality)
254 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
255 CommandStream commandStream = xCopyStreams_[atomLocality];
256 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying positions with given atom locality.");
258 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, commandStream);
260 // markEvent is skipped in OpenCL as:
261 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
262 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
263 // TODO: remove this by adding an event-mark free flavor of this function
264 if (GMX_GPU == GMX_GPU_CUDA)
266 xReadyOnDevice_[atomLocality].markEvent(commandStream);
267 // TODO: Remove When event-based synchronization is introduced
268 gpuStreamSynchronize(commandStream);
272 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
273 const SimulationWorkload &simulationWork,
274 const StepWorkload &stepWork)
276 // The provider of the coordinates may be different for local atoms. If the update is offloaded
277 // and this is not a neighbor search step, then the consumer needs to wait for the update
278 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
279 // the copy event. Non-local coordinates are always provided by the H2D copy.
281 // TODO: This should be reconsidered to support the halo exchange.
283 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
285 return &xUpdatedOnDevice_;
289 return &xReadyOnDevice_[atomLocality];
293 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
295 return &xUpdatedOnDevice_;
298 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
299 AtomLocality atomLocality)
301 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
302 CommandStream commandStream = xCopyStreams_[atomLocality];
303 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying positions with given atom locality.");
305 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, commandStream);
306 // TODO: Remove When event-based synchronization is introduced
307 gpuStreamSynchronize(commandStream);
308 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
309 xReadyOnHost_[atomLocality].markEvent(commandStream);
312 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
314 xReadyOnHost_[atomLocality].waitForEvent();
318 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getVelocities()
323 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
324 AtomLocality atomLocality)
326 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
327 CommandStream commandStream = vCopyStreams_[atomLocality];
328 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying velocities with given atom locality.");
330 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, commandStream);
331 // TODO: Remove When event-based synchronization is introduced
332 gpuStreamSynchronize(commandStream);
333 vReadyOnDevice_[atomLocality].markEvent(commandStream);
336 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
338 return &vReadyOnDevice_[atomLocality];
342 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v,
343 AtomLocality atomLocality)
345 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
346 CommandStream commandStream = vCopyStreams_[atomLocality];
347 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying velocities with given atom locality.");
349 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, commandStream);
350 // TODO: Remove When event-based synchronization is introduced
351 gpuStreamSynchronize(commandStream);
352 vReadyOnHost_[atomLocality].markEvent(commandStream);
355 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
357 vReadyOnHost_[atomLocality].waitForEvent();
361 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getForces()
366 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
367 AtomLocality atomLocality)
369 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
370 CommandStream commandStream = fCopyStreams_[atomLocality];
371 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying forces with given atom locality.");
373 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, commandStream);
374 fReadyOnDevice_[atomLocality].markEvent(commandStream);
377 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
378 bool useGpuFBufferOps)
380 if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
382 return &fReducedOnDevice_;
386 return &fReadyOnDevice_[atomLocality];
390 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
392 return &fReducedOnDevice_;
395 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f,
396 AtomLocality atomLocality)
398 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
399 CommandStream commandStream = fCopyStreams_[atomLocality];
400 GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying forces with given atom locality.");
402 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, commandStream);
403 fReadyOnHost_[atomLocality].markEvent(commandStream);
406 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
408 fReadyOnHost_[atomLocality].waitForEvent();
411 void* StatePropagatorDataGpu::Impl::getUpdateStream()
413 return updateStream_;
416 int StatePropagatorDataGpu::Impl::numAtomsLocal()
418 return numAtomsLocal_;
421 int StatePropagatorDataGpu::Impl::numAtomsAll()
428 StatePropagatorDataGpu::StatePropagatorDataGpu(const void *pmeStream,
429 const void *localStream,
430 const void *nonLocalStream,
431 const void *deviceContext,
432 GpuApiCallBehavior transferKind,
434 : impl_(new Impl(pmeStream,
443 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu && /* other */) noexcept = default;
445 StatePropagatorDataGpu &StatePropagatorDataGpu::operator=(StatePropagatorDataGpu && /* other */) noexcept = default;
447 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
450 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
452 return impl_->reinit(numAtomsLocal, numAtomsAll);
455 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
457 return impl_->getAtomRangesFromAtomLocality(atomLocality);
461 DeviceBuffer<float> StatePropagatorDataGpu::getCoordinates()
463 return impl_->getCoordinates();
466 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
467 AtomLocality atomLocality)
469 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
472 GpuEventSynchronizer* StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
473 const SimulationWorkload &simulationWork,
474 const StepWorkload &stepWork)
476 return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
479 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
481 return impl_->xUpdatedOnDevice();
484 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x,
485 AtomLocality atomLocality)
487 return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
490 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
492 return impl_->waitCoordinatesReadyOnHost(atomLocality);
496 DeviceBuffer<float> StatePropagatorDataGpu::getVelocities()
498 return impl_->getVelocities();
501 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
502 AtomLocality atomLocality)
504 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
507 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
509 return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
512 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v,
513 AtomLocality atomLocality)
515 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
518 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
520 return impl_->waitVelocitiesReadyOnHost(atomLocality);
524 DeviceBuffer<float> StatePropagatorDataGpu::getForces()
526 return impl_->getForces();
529 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
530 AtomLocality atomLocality)
532 return impl_->copyForcesToGpu(h_f, atomLocality);
535 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
536 bool useGpuFBufferOps)
538 return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
541 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
543 return impl_->fReducedOnDevice();
546 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f,
547 AtomLocality atomLocality)
549 return impl_->copyForcesFromGpu(h_f, atomLocality);
552 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
554 return impl_->waitForcesReadyOnHost(atomLocality);
558 void* StatePropagatorDataGpu::getUpdateStream()
560 return impl_->getUpdateStream();
563 int StatePropagatorDataGpu::numAtomsLocal()
565 return impl_->numAtomsLocal();
568 int StatePropagatorDataGpu::numAtomsAll()
570 return impl_->numAtomsAll();
575 #endif // GMX_GPU == GMX_GPU_NONE