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,
75 "This object should only be constructed on the GPU code-paths.");
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] = updateStream_;
131 StatePropagatorDataGpu::Impl::Impl(const void* pmeStream,
132 const void* deviceContext,
133 GpuApiCallBehavior transferKind,
135 transferKind_(transferKind),
136 paddingSize_(paddingSize)
138 static_assert(GMX_GPU != GMX_GPU_NONE,
139 "This object should only be constructed on the GPU code-paths.");
141 if (GMX_GPU == GMX_GPU_OPENCL)
143 GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
144 deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
147 GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set.");
148 pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
150 localStream_ = nullptr;
151 nonLocalStream_ = nullptr;
152 updateStream_ = nullptr;
155 // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
156 // This it temporary measure to make it safe to use this class in those cases.
157 xCopyStreams_[AtomLocality::Local] = pmeStream_;
158 xCopyStreams_[AtomLocality::NonLocal] = nullptr;
159 xCopyStreams_[AtomLocality::All] = pmeStream_;
161 vCopyStreams_[AtomLocality::Local] = nullptr;
162 vCopyStreams_[AtomLocality::NonLocal] = nullptr;
163 vCopyStreams_[AtomLocality::All] = nullptr;
165 fCopyStreams_[AtomLocality::Local] = nullptr;
166 fCopyStreams_[AtomLocality::NonLocal] = nullptr;
167 fCopyStreams_[AtomLocality::All] = nullptr;
170 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_);
200 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
202 int atomsStartAt = 0;
203 int numAtomsToCopy = 0;
204 switch (atomLocality)
206 case AtomLocality::All:
208 numAtomsToCopy = numAtomsAll_;
210 case AtomLocality::Local:
212 numAtomsToCopy = numAtomsLocal_;
214 case AtomLocality::NonLocal:
215 atomsStartAt = numAtomsLocal_;
216 numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
219 GMX_RELEASE_ASSERT(false,
220 "Wrong range of atoms requested in GPU state data manager. Should "
221 "be All, Local or NonLocal.");
223 GMX_ASSERT(atomsStartAt >= 0,
224 "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
225 "was not initialized.");
226 GMX_ASSERT(numAtomsToCopy >= 0,
227 "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
229 return std::make_tuple(atomsStartAt, numAtomsToCopy);
232 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<float> d_data,
233 const gmx::ArrayRef<const gmx::RVec> h_data,
235 AtomLocality atomLocality,
236 CommandStream commandStream)
239 GMX_UNUSED_VALUE(dataSize);
241 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
243 GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
245 GMX_ASSERT(commandStream != nullptr,
246 "No stream is valid for copying with given atom locality.");
248 int atomsStartAt, numAtomsToCopy;
249 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
251 int elementsStartAt = atomsStartAt * DIM;
252 int numElementsToCopy = numAtomsToCopy * DIM;
254 if (numAtomsToCopy != 0)
256 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
257 "The device allocation is smaller than requested copy range.");
258 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
259 "The host buffer is smaller than the requested copy range.");
261 copyToDeviceBuffer(&d_data, reinterpret_cast<const float*>(&h_data.data()[atomsStartAt]),
262 elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
266 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
267 DeviceBuffer<float> d_data,
269 AtomLocality atomLocality,
270 CommandStream commandStream)
273 GMX_UNUSED_VALUE(dataSize);
275 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
277 GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
279 GMX_ASSERT(commandStream != nullptr,
280 "No stream is valid for copying with given atom locality.");
282 int atomsStartAt, numAtomsToCopy;
283 std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
285 int elementsStartAt = atomsStartAt * DIM;
286 int numElementsToCopy = numAtomsToCopy * DIM;
288 if (numAtomsToCopy != 0)
290 GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
291 "The device allocation is smaller than requested copy range.");
292 GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
293 "The host buffer is smaller than the requested copy range.");
295 copyFromDeviceBuffer(reinterpret_cast<float*>(&h_data.data()[atomsStartAt]), &d_data,
296 elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
300 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getCoordinates()
305 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
306 AtomLocality atomLocality)
308 copyToDevice(d_x_, h_x, d_xSize_, atomLocality, xCopyStreams_[atomLocality]);
310 // markEvent is skipped in OpenCL as:
311 // - it's not needed, copy is done in the same stream as the only consumer task (PME)
312 // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
313 // TODO: remove this by adding an event-mark free flavor of this function
314 if (GMX_GPU == GMX_GPU_CUDA)
316 xReadyOnDevice_[atomLocality].markEvent(xCopyStreams_[atomLocality]);
320 GpuEventSynchronizer*
321 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
322 const SimulationWorkload& simulationWork,
323 const StepWorkload& stepWork)
325 // The provider of the coordinates may be different for local atoms. If the update is offloaded
326 // and this is not a neighbor search step, then the consumer needs to wait for the update
327 // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
328 // the copy event. Non-local coordinates are always provided by the H2D copy.
330 // TODO: This should be reconsidered to support the halo exchange.
332 // In OpenCL no events are used as coordinate sync is not necessary
333 if (GMX_GPU == GMX_GPU_OPENCL)
337 if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
339 return &xUpdatedOnDevice_;
343 return &xReadyOnDevice_[atomLocality];
347 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
349 GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
350 xReadyOnDevice_[atomLocality].waitForEvent();
353 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
355 return &xUpdatedOnDevice_;
358 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
360 copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, xCopyStreams_[atomLocality]);
361 // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
362 xReadyOnHost_[atomLocality].markEvent(xCopyStreams_[atomLocality]);
365 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
367 xReadyOnHost_[atomLocality].waitForEvent();
371 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getVelocities()
376 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
377 AtomLocality atomLocality)
379 copyToDevice(d_v_, h_v, d_vSize_, atomLocality, vCopyStreams_[atomLocality]);
380 vReadyOnDevice_[atomLocality].markEvent(vCopyStreams_[atomLocality]);
383 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
385 return &vReadyOnDevice_[atomLocality];
389 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
391 copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, vCopyStreams_[atomLocality]);
392 vReadyOnHost_[atomLocality].markEvent(vCopyStreams_[atomLocality]);
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 copyToDevice(d_f_, h_f, d_fSize_, atomLocality, fCopyStreams_[atomLocality]);
410 fReadyOnDevice_[atomLocality].markEvent(fCopyStreams_[atomLocality]);
413 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
414 bool useGpuFBufferOps)
416 if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
418 return &fReducedOnDevice_;
422 return &fReadyOnDevice_[atomLocality];
426 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
428 return &fReducedOnDevice_;
431 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
433 copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, fCopyStreams_[atomLocality]);
434 fReadyOnHost_[atomLocality].markEvent(fCopyStreams_[atomLocality]);
437 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
439 fReadyOnHost_[atomLocality].waitForEvent();
442 void* StatePropagatorDataGpu::Impl::getUpdateStream()
444 return &updateStream_;
447 int StatePropagatorDataGpu::Impl::numAtomsLocal()
449 return numAtomsLocal_;
452 int StatePropagatorDataGpu::Impl::numAtomsAll()
458 StatePropagatorDataGpu::StatePropagatorDataGpu(const void* pmeStream,
459 const void* localStream,
460 const void* nonLocalStream,
461 const void* deviceContext,
462 GpuApiCallBehavior transferKind,
464 impl_(new Impl(pmeStream, localStream, nonLocalStream, deviceContext, transferKind, paddingSize))
468 StatePropagatorDataGpu::StatePropagatorDataGpu(const void* pmeStream,
469 const void* deviceContext,
470 GpuApiCallBehavior transferKind,
472 impl_(new Impl(pmeStream, deviceContext, transferKind, paddingSize))
476 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
478 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
480 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
483 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
485 return impl_->reinit(numAtomsLocal, numAtomsAll);
488 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
490 return impl_->getAtomRangesFromAtomLocality(atomLocality);
494 DeviceBuffer<float> StatePropagatorDataGpu::getCoordinates()
496 return impl_->getCoordinates();
499 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
500 AtomLocality atomLocality)
502 return impl_->copyCoordinatesToGpu(h_x, atomLocality);
505 GpuEventSynchronizer*
506 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
507 const SimulationWorkload& simulationWork,
508 const StepWorkload& stepWork)
510 return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
513 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
515 return impl_->waitCoordinatesCopiedToDevice(atomLocality);
518 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
520 return impl_->xUpdatedOnDevice();
523 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
525 return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
528 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
530 return impl_->waitCoordinatesReadyOnHost(atomLocality);
534 DeviceBuffer<float> StatePropagatorDataGpu::getVelocities()
536 return impl_->getVelocities();
539 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
540 AtomLocality atomLocality)
542 return impl_->copyVelocitiesToGpu(h_v, atomLocality);
545 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
547 return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
550 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
552 return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
555 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
557 return impl_->waitVelocitiesReadyOnHost(atomLocality);
561 DeviceBuffer<float> StatePropagatorDataGpu::getForces()
563 return impl_->getForces();
566 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
568 return impl_->copyForcesToGpu(h_f, atomLocality);
571 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
572 bool useGpuFBufferOps)
574 return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
577 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
579 return impl_->fReducedOnDevice();
582 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
584 return impl_->copyForcesFromGpu(h_f, atomLocality);
587 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
589 return impl_->waitForcesReadyOnHost(atomLocality);
593 void* StatePropagatorDataGpu::getUpdateStream()
595 return impl_->getUpdateStream();
598 int StatePropagatorDataGpu::numAtomsLocal()
600 return impl_->numAtomsLocal();
603 int StatePropagatorDataGpu::numAtomsAll()
605 return impl_->numAtomsAll();
610 #endif // GMX_GPU == GMX_GPU_NONE