src/gromacs/mdtypes/state_propagator_data_gpu_impl_gpu.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   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.
   8  *
   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.
  13  *
  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.
  18  *
  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.
  23  *
  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.
  31  *
  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.
  34  */
  35 /*! \internal \file
  36  *
  37  * \brief Definitions of interfaces for GPU state data propagator object.
  38  *
  39  * \author Artem Zhmurov <zhmurov@gmail.com>
  40  *
  41  * \ingroup module_mdtypes
  42  */
  43 #include "gmxpre.h"
  44
  45 #include "config.h"
  46
  47 #if GMX_GPU
  48
  49 #    include "gromacs/gpu_utils/device_stream_manager.h"
  50 #    include "gromacs/gpu_utils/devicebuffer.h"
  51 #    include "gromacs/math/vectypes.h"
  52 #    include "gromacs/mdtypes/state_propagator_data_gpu.h"
  53 #    include "gromacs/timing/wallcycle.h"
  54 #    include "gromacs/utility/classhelpers.h"
  55
  56 #    include "state_propagator_data_gpu_impl.h"
  57
  58
  59 namespace gmx
  60 {
  61
  62 StatePropagatorDataGpu::Impl::Impl(const DeviceStreamManager& deviceStreamManager,
  63                                    GpuApiCallBehavior         transferKind,
  64                                    int                        allocationBlockSizeDivisor,
  65                                    gmx_wallcycle*             wcycle) :
  66     deviceContext_(deviceStreamManager.context()),
  67     transferKind_(transferKind),
  68     allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
  69     wcycle_(wcycle)
  70 {
  71     static_assert(
  72             GMX_GPU,
  73             "GPU state propagator data object should only be constructed on the GPU code-paths.");
  74
  75     // We need to keep local copies for re-initialization.
  76     pmeStream_      = &deviceStreamManager.stream(DeviceStreamType::Pme);
  77     localStream_    = &deviceStreamManager.stream(DeviceStreamType::NonBondedLocal);
  78     nonLocalStream_ = &deviceStreamManager.stream(DeviceStreamType::NonBondedNonLocal);
  79     // PME stream is used in OpenCL for H2D coordinate transfer
  80     updateStream_ = &deviceStreamManager.stream(
  81             GMX_GPU_OPENCL ? DeviceStreamType::Pme : DeviceStreamType::UpdateAndConstraints);
  82
  83     // Map the atom locality to the stream that will be used for coordinates,
  84     // velocities and forces transfers. Same streams are used for H2D and D2H copies.
  85     // Note, that nullptr stream is used here to indicate that the copy is not supported.
  86     xCopyStreams_[AtomLocality::Local]    = updateStream_;
  87     xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
  88     xCopyStreams_[AtomLocality::All]      = updateStream_;
  89
  90     vCopyStreams_[AtomLocality::Local]    = updateStream_;
  91     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
  92     vCopyStreams_[AtomLocality::All]      = updateStream_;
  93
  94     fCopyStreams_[AtomLocality::Local]    = localStream_;
  95     fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
  96     fCopyStreams_[AtomLocality::All]      = updateStream_;
  97 }
  98
  99 StatePropagatorDataGpu::Impl::Impl(const DeviceStream*  pmeStream,
 100                                    const DeviceContext& deviceContext,
 101                                    GpuApiCallBehavior   transferKind,
 102                                    int                  allocationBlockSizeDivisor,
 103                                    gmx_wallcycle*       wcycle) :
 104     deviceContext_(deviceContext),
 105     transferKind_(transferKind),
 106     allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
 107     wcycle_(wcycle)
 108 {
 109     static_assert(
 110             GMX_GPU,
 111             "GPU state propagator data object should only be constructed on the GPU code-paths.");
 112
 113     GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
 114     pmeStream_      = pmeStream;
 115     localStream_    = pmeStream; // For clearing the force buffer
 116     nonLocalStream_ = nullptr;
 117     updateStream_   = nullptr;
 118
 119
 120     // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
 121     // This it temporary measure to make it safe to use this class in those cases.
 122     xCopyStreams_[AtomLocality::Local]    = pmeStream_;
 123     xCopyStreams_[AtomLocality::NonLocal] = nullptr;
 124     xCopyStreams_[AtomLocality::All]      = pmeStream_;
 125
 126     vCopyStreams_[AtomLocality::Local]    = nullptr;
 127     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
 128     vCopyStreams_[AtomLocality::All]      = nullptr;
 129
 130     fCopyStreams_[AtomLocality::Local]    = nullptr;
 131     fCopyStreams_[AtomLocality::NonLocal] = nullptr;
 132     fCopyStreams_[AtomLocality::All]      = nullptr;
 133 }
 134
 135 StatePropagatorDataGpu::Impl::~Impl() {}
 136
 137 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
 138 {
 139     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 140     wallcycle_sub_start_nocount(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 141
 142     numAtomsLocal_ = numAtomsLocal;
 143     numAtomsAll_   = numAtomsAll;
 144
 145     int numAtomsPadded;
 146     if (allocationBlockSizeDivisor_ > 0)
 147     {
 148         numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
 149                          * allocationBlockSizeDivisor_;
 150     }
 151     else
 152     {
 153         numAtomsPadded = numAtomsAll_;
 154     }
 155
 156     reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
 157
 158     const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
 159     if (paddingAllocationSize > 0)
 160     {
 161         // The PME stream is used here because the padding region of d_x_ is only in the PME task.
 162         clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
 163     }
 164
 165     reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
 166     const int d_fOldCapacity = d_fCapacity_;
 167     reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
 168
 169     // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
 170     // the force accumulation stage before syncing with the local stream. Only done in CUDA,
 171     // since the force buffer ops are not implemented in OpenCL.
 172     if (GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
 173     {
 174         clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
 175     }
 176
 177     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 178     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 179 }
 180
 181 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
 182 {
 183     int atomsStartAt   = 0;
 184     int numAtomsToCopy = 0;
 185     switch (atomLocality)
 186     {
 187         case AtomLocality::All:
 188             atomsStartAt   = 0;
 189             numAtomsToCopy = numAtomsAll_;
 190             break;
 191         case AtomLocality::Local:
 192             atomsStartAt   = 0;
 193             numAtomsToCopy = numAtomsLocal_;
 194             break;
 195         case AtomLocality::NonLocal:
 196             atomsStartAt   = numAtomsLocal_;
 197             numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
 198             break;
 199         default:
 200             GMX_RELEASE_ASSERT(false,
 201                                "Wrong range of atoms requested in GPU state data manager. Should "
 202                                "be All, Local or NonLocal.");
 203     }
 204     GMX_ASSERT(atomsStartAt >= 0,
 205                "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
 206                "was not initialized.");
 207     GMX_ASSERT(numAtomsToCopy >= 0,
 208                "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
 209                "initialized.");
 210     return std::make_tuple(atomsStartAt, numAtomsToCopy);
 211 }
 212
 213 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<RVec>                   d_data,
 214                                                 const gmx::ArrayRef<const gmx::RVec> h_data,
 215                                                 int                                  dataSize,
 216                                                 AtomLocality                         atomLocality,
 217                                                 const DeviceStream&                  deviceStream)
 218 {
 219     GMX_UNUSED_VALUE(dataSize);
 220
 221     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 222
 223     GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
 224
 225     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
 226
 227     int atomsStartAt, numAtomsToCopy;
 228     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 229
 230     if (numAtomsToCopy != 0)
 231     {
 232         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
 233                    "The device allocation is smaller than requested copy range.");
 234         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 235                    "The host buffer is smaller than the requested copy range.");
 236
 237         copyToDeviceBuffer(&d_data,
 238                            reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
 239                            atomsStartAt,
 240                            numAtomsToCopy,
 241                            deviceStream,
 242                            transferKind_,
 243                            nullptr);
 244     }
 245 }
 246
 247 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
 248                                                   DeviceBuffer<RVec>       d_data,
 249                                                   int                      dataSize,
 250                                                   AtomLocality             atomLocality,
 251                                                   const DeviceStream&      deviceStream)
 252 {
 253     GMX_UNUSED_VALUE(dataSize);
 254
 255     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 256
 257     GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
 258
 259     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
 260
 261     int atomsStartAt, numAtomsToCopy;
 262     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 263
 264     if (numAtomsToCopy != 0)
 265     {
 266         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
 267                    "The device allocation is smaller than requested copy range.");
 268         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 269                    "The host buffer is smaller than the requested copy range.");
 270
 271         copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]),
 272                              &d_data,
 273                              atomsStartAt,
 274                              numAtomsToCopy,
 275                              deviceStream,
 276                              transferKind_,
 277                              nullptr);
 278     }
 279 }
 280
 281 void StatePropagatorDataGpu::Impl::clearOnDevice(DeviceBuffer<RVec>  d_data,
 282                                                  int                 dataSize,
 283                                                  AtomLocality        atomLocality,
 284                                                  const DeviceStream& deviceStream) const
 285 {
 286     GMX_UNUSED_VALUE(dataSize);
 287
 288     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 289
 290     GMX_ASSERT(dataSize >= 0, "Trying to clear to device buffer before it was allocated.");
 291
 292     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for clearing with given atom locality.");
 293
 294     int atomsStartAt, numAtomsToClear;
 295     std::tie(atomsStartAt, numAtomsToClear) = getAtomRangesFromAtomLocality(atomLocality);
 296
 297     if (numAtomsToClear != 0)
 298     {
 299         GMX_ASSERT(atomsStartAt + numAtomsToClear <= dataSize,
 300                    "The device allocation is smaller than requested clear range.");
 301
 302         clearDeviceBufferAsync(&d_data, atomsStartAt, numAtomsToClear, deviceStream);
 303     }
 304 }
 305
 306 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
 307 {
 308     return d_x_;
 309 }
 310
 311 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 312                                                         AtomLocality atomLocality)
 313 {
 314     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 315     const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
 316     GMX_ASSERT(deviceStream != nullptr,
 317                "No stream is valid for copying positions with given atom locality.");
 318
 319     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 320     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 321
 322     copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
 323
 324     // markEvent is skipped in OpenCL as:
 325     //   - it's not needed, copy is done in the same stream as the only consumer task (PME)
 326     //   - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
 327     // TODO: remove this by adding an event-mark free flavor of this function
 328     if (GMX_GPU_CUDA)
 329     {
 330         xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 331     }
 332
 333     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 334     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 335 }
 336
 337 GpuEventSynchronizer*
 338 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
 339                                                                const SimulationWorkload& simulationWork,
 340                                                                const StepWorkload&       stepWork)
 341 {
 342     // The provider of the coordinates may be different for local atoms. If the update is offloaded
 343     // and this is not a neighbor search step, then the consumer needs to wait for the update
 344     // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
 345     // the copy event. Non-local coordinates are always provided by the H2D copy.
 346     //
 347     // TODO: This should be reconsidered to support the halo exchange.
 348     //
 349     // In OpenCL no events are used as coordinate sync is not necessary
 350     if (GMX_GPU_OPENCL)
 351     {
 352         return nullptr;
 353     }
 354     if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
 355     {
 356         return &xUpdatedOnDevice_;
 357     }
 358     else
 359     {
 360         return &xReadyOnDevice_[atomLocality];
 361     }
 362 }
 363
 364 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 365 {
 366     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 367     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 368     xReadyOnDevice_[atomLocality].waitForEvent();
 369     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 370 }
 371
 372 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
 373 {
 374     return &xUpdatedOnDevice_;
 375 }
 376
 377 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
 378 {
 379     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 380     const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
 381     GMX_ASSERT(deviceStream != nullptr,
 382                "No stream is valid for copying positions with given atom locality.");
 383
 384     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 385     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 386
 387     copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
 388     // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
 389     xReadyOnHost_[atomLocality].markEvent(*deviceStream);
 390
 391     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 392     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 393 }
 394
 395 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 396 {
 397     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 398     xReadyOnHost_[atomLocality].waitForEvent();
 399     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 400 }
 401
 402
 403 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
 404 {
 405     return d_v_;
 406 }
 407
 408 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 409                                                        AtomLocality atomLocality)
 410 {
 411     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 412     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 413     GMX_ASSERT(deviceStream != nullptr,
 414                "No stream is valid for copying velocities with given atom locality.");
 415
 416     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 417     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 418
 419     copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
 420     vReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 421
 422     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 423     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 424 }
 425
 426 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
 427 {
 428     return &vReadyOnDevice_[atomLocality];
 429 }
 430
 431
 432 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
 433 {
 434     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 435     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 436     GMX_ASSERT(deviceStream != nullptr,
 437                "No stream is valid for copying velocities with given atom locality.");
 438
 439     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 440     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 441
 442     copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
 443     vReadyOnHost_[atomLocality].markEvent(*deviceStream);
 444
 445     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 446     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 447 }
 448
 449 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 450 {
 451     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 452     vReadyOnHost_[atomLocality].waitForEvent();
 453     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 454 }
 455
 456
 457 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
 458 {
 459     return d_f_;
 460 }
 461
 462 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
 463                                                    AtomLocality atomLocality)
 464 {
 465     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 466     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 467     GMX_ASSERT(deviceStream != nullptr,
 468                "No stream is valid for copying forces with given atom locality.");
 469
 470     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 471     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 472
 473     copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
 474     fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 475
 476     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 477     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 478 }
 479
 480 void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality)
 481 {
 482     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 483     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 484     GMX_ASSERT(deviceStream != nullptr,
 485                "No stream is valid for clearing forces with given atom locality.");
 486
 487     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 488     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 489
 490     clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
 491
 492     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 493     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 494 }
 495
 496 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
 497                                                                                 bool useGpuFBufferOps)
 498 {
 499     if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
 500     {
 501         return &fReducedOnDevice_;
 502     }
 503     else
 504     {
 505         return &fReadyOnDevice_[atomLocality];
 506     }
 507 }
 508
 509 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
 510 {
 511     return &fReducedOnDevice_;
 512 }
 513
 514 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
 515 {
 516     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 517     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 518     GMX_ASSERT(deviceStream != nullptr,
 519                "No stream is valid for copying forces with given atom locality.");
 520
 521     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 522     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 523
 524     copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
 525     fReadyOnHost_[atomLocality].markEvent(*deviceStream);
 526
 527     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 528     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 529 }
 530
 531 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
 532 {
 533     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 534     fReadyOnHost_[atomLocality].waitForEvent();
 535     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 536 }
 537
 538 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
 539 {
 540     return updateStream_;
 541 }
 542
 543 int StatePropagatorDataGpu::Impl::numAtomsLocal() const
 544 {
 545     return numAtomsLocal_;
 546 }
 547
 548 int StatePropagatorDataGpu::Impl::numAtomsAll() const
 549 {
 550     return numAtomsAll_;
 551 }
 552
 553
 554 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
 555                                                GpuApiCallBehavior         transferKind,
 556                                                int            allocationBlockSizeDivisor,
 557                                                gmx_wallcycle* wcycle) :
 558     impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
 559 {
 560 }
 561
 562 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream*  pmeStream,
 563                                                const DeviceContext& deviceContext,
 564                                                GpuApiCallBehavior   transferKind,
 565                                                int                  allocationBlockSizeDivisor,
 566                                                gmx_wallcycle*       wcycle) :
 567     impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
 568 {
 569 }
 570
 571 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
 572
 573 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
 574
 575 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
 576
 577
 578 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
 579 {
 580     return impl_->reinit(numAtomsLocal, numAtomsAll);
 581 }
 582
 583 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
 584 {
 585     return impl_->getAtomRangesFromAtomLocality(atomLocality);
 586 }
 587
 588
 589 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
 590 {
 591     return impl_->getCoordinates();
 592 }
 593
 594 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 595                                                   AtomLocality                         atomLocality)
 596 {
 597     return impl_->copyCoordinatesToGpu(h_x, atomLocality);
 598 }
 599
 600 GpuEventSynchronizer*
 601 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality              atomLocality,
 602                                                          const SimulationWorkload& simulationWork,
 603                                                          const StepWorkload&       stepWork)
 604 {
 605     return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
 606 }
 607
 608 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 609 {
 610     return impl_->waitCoordinatesCopiedToDevice(atomLocality);
 611 }
 612
 613 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
 614 {
 615     return impl_->xUpdatedOnDevice();
 616 }
 617
 618 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
 619 {
 620     return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
 621 }
 622
 623 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 624 {
 625     return impl_->waitCoordinatesReadyOnHost(atomLocality);
 626 }
 627
 628
 629 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
 630 {
 631     return impl_->getVelocities();
 632 }
 633
 634 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 635                                                  AtomLocality                         atomLocality)
 636 {
 637     return impl_->copyVelocitiesToGpu(h_v, atomLocality);
 638 }
 639
 640 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
 641 {
 642     return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
 643 }
 644
 645 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
 646 {
 647     return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
 648 }
 649
 650 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 651 {
 652     return impl_->waitVelocitiesReadyOnHost(atomLocality);
 653 }
 654
 655
 656 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
 657 {
 658     return impl_->getForces();
 659 }
 660
 661 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
 662 {
 663     return impl_->copyForcesToGpu(h_f, atomLocality);
 664 }
 665
 666 void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality)
 667 {
 668     return impl_->clearForcesOnGpu(atomLocality);
 669 }
 670
 671 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
 672                                                                           bool useGpuFBufferOps)
 673 {
 674     return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
 675 }
 676
 677 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
 678 {
 679     return impl_->fReducedOnDevice();
 680 }
 681
 682 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
 683 {
 684     return impl_->copyForcesFromGpu(h_f, atomLocality);
 685 }
 686
 687 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
 688 {
 689     return impl_->waitForcesReadyOnHost(atomLocality);
 690 }
 691
 692
 693 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
 694 {
 695     return impl_->getUpdateStream();
 696 }
 697
 698 int StatePropagatorDataGpu::numAtomsLocal() const
 699 {
 700     return impl_->numAtomsLocal();
 701 }
 702
 703 int StatePropagatorDataGpu::numAtomsAll() const
 704 {
 705     return impl_->numAtomsAll();
 706 }
 707
 708 } // namespace gmx
 709
 710 #endif // GMX_GPU