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