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/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"
  56
  57 #    include "state_propagator_data_gpu_impl.h"
  58
  59
  60 namespace gmx
  61 {
  62
  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),
  70     wcycle_(wcycle)
  71 {
  72     static_assert(
  73             GMX_GPU,
  74             "GPU state propagator data object should only be constructed on the GPU code-paths.");
  75
  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     updateStream_   = &deviceStreamManager.stream(DeviceStreamType::UpdateAndConstraints);
  81
  82     // Map the atom locality to the stream that will be used for coordinates,
  83     // velocities and forces transfers. Same streams are used for H2D and D2H copies.
  84     // Note, that nullptr stream is used here to indicate that the copy is not supported.
  85     xCopyStreams_[AtomLocality::Local]    = updateStream_;
  86     xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
  87     xCopyStreams_[AtomLocality::All]      = nullptr;
  88
  89     vCopyStreams_[AtomLocality::Local]    = updateStream_;
  90     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
  91     vCopyStreams_[AtomLocality::All]      = nullptr;
  92
  93     fCopyStreams_[AtomLocality::Local]    = localStream_;
  94     fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
  95     fCopyStreams_[AtomLocality::All]      = updateStream_;
  96
  97     copyInStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
  98     memsetStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
  99 }
 100
 101 StatePropagatorDataGpu::Impl::Impl(const DeviceStream*  pmeStream,
 102                                    const DeviceContext& deviceContext,
 103                                    GpuApiCallBehavior   transferKind,
 104                                    int                  allocationBlockSizeDivisor,
 105                                    gmx_wallcycle*       wcycle) :
 106     deviceContext_(deviceContext),
 107     transferKind_(transferKind),
 108     allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
 109     wcycle_(wcycle)
 110 {
 111     static_assert(
 112             GMX_GPU,
 113             "GPU state propagator data object should only be constructed on the GPU code-paths.");
 114
 115     GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
 116     pmeStream_      = pmeStream;
 117     localStream_    = pmeStream; // For clearing the force buffer
 118     nonLocalStream_ = nullptr;
 119     updateStream_   = nullptr;
 120
 121     isPmeOnly_ = true;
 122
 123     // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
 124     // This it temporary measure to make it safe to use this class in those cases.
 125     xCopyStreams_[AtomLocality::Local]    = pmeStream_;
 126     xCopyStreams_[AtomLocality::NonLocal] = nullptr;
 127     xCopyStreams_[AtomLocality::All]      = nullptr;
 128
 129     vCopyStreams_[AtomLocality::Local]    = nullptr;
 130     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
 131     vCopyStreams_[AtomLocality::All]      = nullptr;
 132
 133     fCopyStreams_[AtomLocality::Local]    = nullptr;
 134     fCopyStreams_[AtomLocality::NonLocal] = nullptr;
 135     fCopyStreams_[AtomLocality::All]      = nullptr;
 136 }
 137
 138 StatePropagatorDataGpu::Impl::~Impl() {}
 139
 140 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
 141 {
 142     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 143     wallcycle_sub_start_nocount(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 144
 145     numAtomsLocal_ = numAtomsLocal;
 146     numAtomsAll_   = numAtomsAll;
 147
 148     int numAtomsPadded;
 149     if (allocationBlockSizeDivisor_ > 0)
 150     {
 151         numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
 152                          * allocationBlockSizeDivisor_;
 153     }
 154     else
 155     {
 156         numAtomsPadded = numAtomsAll_;
 157     }
 158
 159     reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
 160
 161     const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
 162     if (paddingAllocationSize > 0)
 163     {
 164         // The PME stream is used here because the padding region of d_x_ is only in the PME task.
 165         clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
 166     }
 167
 168     reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
 169     const int d_fOldCapacity = d_fCapacity_;
 170     reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
 171
 172     // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
 173     // the force accumulation stage before syncing with the local stream. Only done in CUDA and
 174     // SYCL, since the force buffer ops are not implemented in OpenCL.
 175     if ((bool(GMX_GPU_CUDA) || bool(GMX_GPU_SYCL)) && d_fCapacity_ != d_fOldCapacity)
 176     {
 177         clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
 178     }
 179
 180     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 181     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 182 }
 183
 184 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
 185 {
 186     int atomsStartAt   = 0;
 187     int numAtomsToCopy = 0;
 188     switch (atomLocality)
 189     {
 190         case AtomLocality::All:
 191             atomsStartAt   = 0;
 192             numAtomsToCopy = numAtomsAll_;
 193             break;
 194         case AtomLocality::Local:
 195             atomsStartAt   = 0;
 196             numAtomsToCopy = numAtomsLocal_;
 197             break;
 198         case AtomLocality::NonLocal:
 199             atomsStartAt   = numAtomsLocal_;
 200             numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
 201             break;
 202         default:
 203             GMX_RELEASE_ASSERT(false,
 204                                "Wrong range of atoms requested in GPU state data manager. Should "
 205                                "be All, Local or NonLocal.");
 206     }
 207     GMX_ASSERT(atomsStartAt >= 0,
 208                "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
 209                "was not initialized.");
 210     GMX_ASSERT(numAtomsToCopy >= 0,
 211                "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
 212                "initialized.");
 213     return std::make_tuple(atomsStartAt, numAtomsToCopy);
 214 }
 215
 216 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<RVec>                   d_data,
 217                                                 const gmx::ArrayRef<const gmx::RVec> h_data,
 218                                                 int                                  dataSize,
 219                                                 AtomLocality                         atomLocality,
 220                                                 const DeviceStream&                  deviceStream)
 221 {
 222     GMX_UNUSED_VALUE(dataSize);
 223
 224     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 225
 226     GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
 227
 228     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
 229
 230     int atomsStartAt, numAtomsToCopy;
 231     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 232
 233     if (numAtomsToCopy != 0)
 234     {
 235         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
 236                    "The device allocation is smaller than requested copy range.");
 237         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 238                    "The host buffer is smaller than the requested copy range.");
 239
 240         copyToDeviceBuffer(&d_data,
 241                            reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
 242                            atomsStartAt,
 243                            numAtomsToCopy,
 244                            deviceStream,
 245                            transferKind_,
 246                            nullptr);
 247     }
 248 }
 249
 250 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
 251                                                   DeviceBuffer<RVec>       d_data,
 252                                                   int                      dataSize,
 253                                                   AtomLocality             atomLocality,
 254                                                   const DeviceStream&      deviceStream)
 255 {
 256     GMX_UNUSED_VALUE(dataSize);
 257
 258     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 259
 260     GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
 261
 262     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
 263
 264     int atomsStartAt, numAtomsToCopy;
 265     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 266
 267     if (numAtomsToCopy != 0)
 268     {
 269         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
 270                    "The device allocation is smaller than requested copy range.");
 271         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 272                    "The host buffer is smaller than the requested copy range.");
 273
 274         copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]),
 275                              &d_data,
 276                              atomsStartAt,
 277                              numAtomsToCopy,
 278                              deviceStream,
 279                              transferKind_,
 280                              nullptr);
 281     }
 282 }
 283
 284 void StatePropagatorDataGpu::Impl::clearOnDevice(DeviceBuffer<RVec>  d_data,
 285                                                  int                 dataSize,
 286                                                  AtomLocality        atomLocality,
 287                                                  const DeviceStream& deviceStream) const
 288 {
 289     GMX_UNUSED_VALUE(dataSize);
 290
 291     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 292
 293     GMX_ASSERT(dataSize >= 0, "Trying to clear to device buffer before it was allocated.");
 294
 295     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for clearing with given atom locality.");
 296
 297     int atomsStartAt, numAtomsToClear;
 298     std::tie(atomsStartAt, numAtomsToClear) = getAtomRangesFromAtomLocality(atomLocality);
 299
 300     if (numAtomsToClear != 0)
 301     {
 302         GMX_ASSERT(atomsStartAt + numAtomsToClear <= dataSize,
 303                    "The device allocation is smaller than requested clear range.");
 304
 305         clearDeviceBufferAsync(&d_data, atomsStartAt, numAtomsToClear, deviceStream);
 306     }
 307 }
 308
 309 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
 310 {
 311     return d_x_;
 312 }
 313
 314 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 315                                                         AtomLocality atomLocality)
 316 {
 317     GMX_ASSERT(atomLocality < AtomLocality::All,
 318                formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
 319                             "coordinate transfers, passed value is \"%s\"",
 320                             enumValueToString(atomLocality))
 321                        .c_str());
 322
 323     const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
 324     GMX_ASSERT(deviceStream != nullptr,
 325                "No stream is valid for copying positions with given atom locality.");
 326
 327     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 328     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 329
 330     copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
 331
 332     // marking is skipped on the PME-rank mode as everything is on the same stream
 333     if (!isPmeOnly_)
 334     {
 335         xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 336     }
 337
 338     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 339     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 340 }
 341
 342 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(
 343         AtomLocality              atomLocality,
 344         const SimulationWorkload& simulationWork,
 345         const StepWorkload&       stepWork,
 346         GpuEventSynchronizer*     gpuCoordinateHaloLaunched)
 347 {
 348     // The provider of the coordinates may be different for local atoms. If the update is offloaded
 349     // and this is not a neighbor search step, then the consumer needs to wait for the update
 350     // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
 351     // the copy event. Non-local coordinates are provided by the GPU halo exchange (if active), otherwise by H2D copy.
 352
 353     if (atomLocality == AtomLocality::NonLocal && stepWork.useGpuXHalo)
 354     {
 355         GMX_ASSERT(gpuCoordinateHaloLaunched != nullptr,
 356                    "GPU halo exchange is active but its completion event is null.");
 357         return gpuCoordinateHaloLaunched;
 358     }
 359     if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
 360     {
 361         GMX_ASSERT(xUpdatedOnDeviceEvent_ != nullptr, "The event synchronizer can not be nullptr.");
 362         return xUpdatedOnDeviceEvent_;
 363     }
 364     else
 365     {
 366         if (stepWork.doNeighborSearch && xUpdatedOnDeviceEvent_)
 367         {
 368             /* On search steps, we do not consume the result of the GPU update
 369              * but rather that of a H2D transfer. So, we reset the event triggered after
 370              * update to avoid leaving it unconsumed.
 371              * Unfortunately, we don't always have the event marked either (e.g., on the
 372              * first step) so we just reset it here.
 373              * See Issue #3988. */
 374             xUpdatedOnDeviceEvent_->reset();
 375         }
 376         return &xReadyOnDevice_[atomLocality];
 377     }
 378 }
 379
 380 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 381 {
 382     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 383     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 384     xReadyOnDevice_[atomLocality].waitForEvent();
 385     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 386 }
 387
 388 void StatePropagatorDataGpu::Impl::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
 389 {
 390     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 391     xReadyOnDevice_[atomLocality].consume();
 392 }
 393
 394 void StatePropagatorDataGpu::Impl::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
 395 {
 396     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 397     xReadyOnDevice_[atomLocality].reset();
 398 }
 399
 400 void StatePropagatorDataGpu::Impl::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
 401 {
 402     GMX_ASSERT(xUpdatedOnDeviceEvent != nullptr, "The event synchronizer can not be nullptr.");
 403     xUpdatedOnDeviceEvent_ = xUpdatedOnDeviceEvent;
 404 }
 405
 406 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
 407                                                           AtomLocality             atomLocality,
 408                                                           GpuEventSynchronizer*    dependency)
 409 {
 410     GMX_ASSERT(atomLocality < AtomLocality::All,
 411                formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
 412                             "coordinate transfers, passed value is \"%s\"",
 413                             enumValueToString(atomLocality))
 414                        .c_str());
 415     const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
 416     GMX_ASSERT(deviceStream != nullptr,
 417                "No stream is valid for copying positions with given atom locality.");
 418
 419     if (dependency != nullptr)
 420     {
 421         dependency->enqueueWaitEvent(*deviceStream);
 422     }
 423
 424     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 425     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 426
 427     copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
 428     // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
 429     xReadyOnHost_[atomLocality].markEvent(*deviceStream);
 430
 431     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 432     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 433 }
 434
 435 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 436 {
 437     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 438     xReadyOnHost_[atomLocality].waitForEvent();
 439     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 440 }
 441
 442
 443 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
 444 {
 445     return d_v_;
 446 }
 447
 448 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 449                                                        AtomLocality atomLocality)
 450 {
 451     GMX_ASSERT(atomLocality == AtomLocality::Local,
 452                formatString("Wrong atom locality. Only Local is allowed for "
 453                             "velocity transfers, passed value is \"%s\"",
 454                             enumValueToString(atomLocality))
 455                        .c_str());
 456     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 457     GMX_ASSERT(deviceStream != nullptr,
 458                "No stream is valid for copying velocities with given atom locality.");
 459
 460     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 461     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 462
 463     copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
 464     /* Not marking the event, because it is not used anywhere.
 465      * Since we only use velocities on the device for update, and we launch the copy in
 466      * the "update" stream, that should be safe.
 467      */
 468
 469     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 470     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 471 }
 472
 473 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
 474 {
 475     GMX_ASSERT(atomLocality == AtomLocality::Local,
 476                formatString("Wrong atom locality. Only Local is allowed for "
 477                             "velocity transfers, passed value is \"%s\"",
 478                             enumValueToString(atomLocality))
 479                        .c_str());
 480     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 481     GMX_ASSERT(deviceStream != nullptr,
 482                "No stream is valid for copying velocities with given atom locality.");
 483
 484     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 485     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 486
 487     copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
 488     vReadyOnHost_[atomLocality].markEvent(*deviceStream);
 489
 490     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 491     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 492 }
 493
 494 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 495 {
 496     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 497     vReadyOnHost_[atomLocality].waitForEvent();
 498     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 499 }
 500
 501
 502 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
 503 {
 504     return d_f_;
 505 }
 506
 507 // Copy CPU forces to GPU using stream internal to this module to allow overlap
 508 // with GPU force calculations.
 509 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
 510                                                    AtomLocality atomLocality)
 511 {
 512     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 513     DeviceStream* deviceStream = copyInStream_.get();
 514     GMX_ASSERT(deviceStream != nullptr,
 515                "No stream is valid for copying forces with given atom locality.");
 516
 517     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 518     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 519
 520     copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
 521     fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 522
 523     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 524     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 525 }
 526
 527 void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
 528 {
 529     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 530     DeviceStream* deviceStream = memsetStream_.get();
 531
 532     GMX_ASSERT(dependency != nullptr, "Dependency is not valid for clearing forces.");
 533     dependency->enqueueWaitEvent(*deviceStream);
 534
 535     GMX_ASSERT(deviceStream != nullptr,
 536                "No stream is valid for clearing forces with given atom locality.");
 537
 538     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 539     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 540
 541     clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
 542
 543     fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 544
 545     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 546     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 547 }
 548
 549 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
 550                                                                                      SimulationWorkload simulationWork)
 551 {
 552     if (stepWork.useGpuFBufferOps && !simulationWork.useCpuPmePpCommunication)
 553     {
 554         return &fReducedOnDevice_[AtomLocality::Local];
 555     }
 556     else
 557     {
 558         return &fReadyOnDevice_[AtomLocality::Local];
 559     }
 560 }
 561
 562 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice(AtomLocality atomLocality)
 563 {
 564     return &fReducedOnDevice_[atomLocality];
 565 }
 566
 567 void StatePropagatorDataGpu::Impl::consumeForcesReducedOnDeviceEvent(AtomLocality atomLocality)
 568 {
 569     fReducedOnDevice_[atomLocality].consume();
 570 }
 571
 572 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReadyOnDevice(AtomLocality atomLocality)
 573 {
 574     return &fReadyOnDevice_[atomLocality];
 575 }
 576
 577 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
 578 {
 579     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 580     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 581     GMX_ASSERT(deviceStream != nullptr,
 582                "No stream is valid for copying forces with given atom locality.");
 583
 584     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 585     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 586
 587     copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
 588     fReadyOnHost_[atomLocality].markEvent(*deviceStream);
 589
 590     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 591     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 592 }
 593
 594 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
 595 {
 596     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 597     fReadyOnHost_[atomLocality].waitForEvent();
 598     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 599 }
 600
 601 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
 602 {
 603     return updateStream_;
 604 }
 605
 606 int StatePropagatorDataGpu::Impl::numAtomsLocal() const
 607 {
 608     return numAtomsLocal_;
 609 }
 610
 611 int StatePropagatorDataGpu::Impl::numAtomsAll() const
 612 {
 613     return numAtomsAll_;
 614 }
 615
 616
 617 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
 618                                                GpuApiCallBehavior         transferKind,
 619                                                int            allocationBlockSizeDivisor,
 620                                                gmx_wallcycle* wcycle) :
 621     impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
 622 {
 623 }
 624
 625 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream*  pmeStream,
 626                                                const DeviceContext& deviceContext,
 627                                                GpuApiCallBehavior   transferKind,
 628                                                int                  allocationBlockSizeDivisor,
 629                                                gmx_wallcycle*       wcycle) :
 630     impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
 631 {
 632 }
 633
 634 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
 635
 636 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
 637
 638 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
 639
 640
 641 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
 642 {
 643     return impl_->reinit(numAtomsLocal, numAtomsAll);
 644 }
 645
 646 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
 647 {
 648     return impl_->getAtomRangesFromAtomLocality(atomLocality);
 649 }
 650
 651
 652 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
 653 {
 654     return impl_->getCoordinates();
 655 }
 656
 657 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 658                                                   AtomLocality                         atomLocality)
 659 {
 660     return impl_->copyCoordinatesToGpu(h_x, atomLocality);
 661 }
 662
 663 GpuEventSynchronizer*
 664 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality              atomLocality,
 665                                                          const SimulationWorkload& simulationWork,
 666                                                          const StepWorkload&       stepWork,
 667                                                          GpuEventSynchronizer* gpuCoordinateHaloLaunched)
 668 {
 669     return impl_->getCoordinatesReadyOnDeviceEvent(
 670             atomLocality, simulationWork, stepWork, gpuCoordinateHaloLaunched);
 671 }
 672
 673 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 674 {
 675     return impl_->waitCoordinatesCopiedToDevice(atomLocality);
 676 }
 677
 678 void StatePropagatorDataGpu::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
 679 {
 680     return impl_->consumeCoordinatesCopiedToDeviceEvent(atomLocality);
 681 }
 682
 683 void StatePropagatorDataGpu::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
 684 {
 685     return impl_->resetCoordinatesCopiedToDeviceEvent(atomLocality);
 686 }
 687
 688 void StatePropagatorDataGpu::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
 689 {
 690     impl_->setXUpdatedOnDeviceEvent(xUpdatedOnDeviceEvent);
 691 }
 692
 693 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec>   h_x,
 694                                                     AtomLocality          atomLocality,
 695                                                     GpuEventSynchronizer* dependency)
 696 {
 697     return impl_->copyCoordinatesFromGpu(h_x, atomLocality, dependency);
 698 }
 699
 700 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 701 {
 702     return impl_->waitCoordinatesReadyOnHost(atomLocality);
 703 }
 704
 705
 706 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
 707 {
 708     return impl_->getVelocities();
 709 }
 710
 711 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 712                                                  AtomLocality                         atomLocality)
 713 {
 714     return impl_->copyVelocitiesToGpu(h_v, atomLocality);
 715 }
 716
 717 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
 718 {
 719     return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
 720 }
 721
 722 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 723 {
 724     return impl_->waitVelocitiesReadyOnHost(atomLocality);
 725 }
 726
 727
 728 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
 729 {
 730     return impl_->getForces();
 731 }
 732
 733 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
 734 {
 735     return impl_->copyForcesToGpu(h_f, atomLocality);
 736 }
 737
 738 void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
 739 {
 740     return impl_->clearForcesOnGpu(atomLocality, dependency);
 741 }
 742
 743 GpuEventSynchronizer* StatePropagatorDataGpu::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
 744                                                                                SimulationWorkload simulationWork)
 745 {
 746     return impl_->getLocalForcesReadyOnDeviceEvent(stepWork, simulationWork);
 747 }
 748
 749 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice(AtomLocality atomLocality)
 750 {
 751     return impl_->fReducedOnDevice(atomLocality);
 752 }
 753
 754 void StatePropagatorDataGpu::consumeForcesReducedOnDeviceEvent(AtomLocality atomLocality)
 755 {
 756     impl_->consumeForcesReducedOnDeviceEvent(atomLocality);
 757 }
 758
 759 GpuEventSynchronizer* StatePropagatorDataGpu::fReadyOnDevice(AtomLocality atomLocality)
 760 {
 761     return impl_->fReadyOnDevice(atomLocality);
 762 }
 763
 764 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
 765 {
 766     return impl_->copyForcesFromGpu(h_f, atomLocality);
 767 }
 768
 769 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
 770 {
 771     return impl_->waitForcesReadyOnHost(atomLocality);
 772 }
 773
 774
 775 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
 776 {
 777     return impl_->getUpdateStream();
 778 }
 779
 780 int StatePropagatorDataGpu::numAtomsLocal() const
 781 {
 782     return impl_->numAtomsLocal();
 783 }
 784
 785 int StatePropagatorDataGpu::numAtomsAll() const
 786 {
 787     return impl_->numAtomsAll();
 788 }
 789
 790 } // namespace gmx
 791
 792 #endif // GMX_GPU