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     // PME stream is used in OpenCL for H2D coordinate transfer
  81     updateStream_ = &deviceStreamManager.stream(
  82             GMX_GPU_OPENCL ? DeviceStreamType::Pme : DeviceStreamType::UpdateAndConstraints);
  83
  84     // Map the atom locality to the stream that will be used for coordinates,
  85     // velocities and forces transfers. Same streams are used for H2D and D2H copies.
  86     // Note, that nullptr stream is used here to indicate that the copy is not supported.
  87     xCopyStreams_[AtomLocality::Local]    = updateStream_;
  88     xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
  89     xCopyStreams_[AtomLocality::All]      = nullptr;
  90
  91     vCopyStreams_[AtomLocality::Local]    = updateStream_;
  92     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
  93     vCopyStreams_[AtomLocality::All]      = nullptr;
  94
  95     fCopyStreams_[AtomLocality::Local]    = localStream_;
  96     fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
  97     fCopyStreams_[AtomLocality::All]      = updateStream_;
  98
  99     copyInStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
 100     memsetStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
 101 }
 102
 103 StatePropagatorDataGpu::Impl::Impl(const DeviceStream*  pmeStream,
 104                                    const DeviceContext& deviceContext,
 105                                    GpuApiCallBehavior   transferKind,
 106                                    int                  allocationBlockSizeDivisor,
 107                                    gmx_wallcycle*       wcycle) :
 108     deviceContext_(deviceContext),
 109     transferKind_(transferKind),
 110     allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
 111     wcycle_(wcycle)
 112 {
 113     static_assert(
 114             GMX_GPU,
 115             "GPU state propagator data object should only be constructed on the GPU code-paths.");
 116
 117     GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
 118     pmeStream_      = pmeStream;
 119     localStream_    = pmeStream; // For clearing the force buffer
 120     nonLocalStream_ = nullptr;
 121     updateStream_   = nullptr;
 122
 123
 124     // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
 125     // This it temporary measure to make it safe to use this class in those cases.
 126     xCopyStreams_[AtomLocality::Local]    = pmeStream_;
 127     xCopyStreams_[AtomLocality::NonLocal] = nullptr;
 128     xCopyStreams_[AtomLocality::All]      = nullptr;
 129
 130     vCopyStreams_[AtomLocality::Local]    = nullptr;
 131     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
 132     vCopyStreams_[AtomLocality::All]      = nullptr;
 133
 134     fCopyStreams_[AtomLocality::Local]    = nullptr;
 135     fCopyStreams_[AtomLocality::NonLocal] = nullptr;
 136     fCopyStreams_[AtomLocality::All]      = nullptr;
 137 }
 138
 139 StatePropagatorDataGpu::Impl::~Impl() {}
 140
 141 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
 142 {
 143     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 144     wallcycle_sub_start_nocount(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 145
 146     numAtomsLocal_ = numAtomsLocal;
 147     numAtomsAll_   = numAtomsAll;
 148
 149     int numAtomsPadded;
 150     if (allocationBlockSizeDivisor_ > 0)
 151     {
 152         numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
 153                          * allocationBlockSizeDivisor_;
 154     }
 155     else
 156     {
 157         numAtomsPadded = numAtomsAll_;
 158     }
 159
 160     reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
 161
 162     const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
 163     if (paddingAllocationSize > 0)
 164     {
 165         // The PME stream is used here because the padding region of d_x_ is only in the PME task.
 166         clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
 167     }
 168
 169     reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
 170     const int d_fOldCapacity = d_fCapacity_;
 171     reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
 172
 173     // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
 174     // the force accumulation stage before syncing with the local stream. Only done in CUDA,
 175     // since the force buffer ops are not implemented in OpenCL.
 176     if (GMX_GPU_CUDA && 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 elemtnt 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     // markEvent is skipped in OpenCL as:
 334     //   - it's not needed, copy is done in the same stream as the only consumer task (PME)
 335     //   - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
 336     // TODO: remove this by adding an event-mark free flavor of this function
 337     if (GMX_GPU_CUDA)
 338     {
 339         xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 340     }
 341
 342     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 343     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 344 }
 345
 346 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(
 347         AtomLocality              atomLocality,
 348         const SimulationWorkload& simulationWork,
 349         const StepWorkload&       stepWork,
 350         GpuEventSynchronizer*     gpuCoordinateHaloLaunched)
 351 {
 352     // The provider of the coordinates may be different for local atoms. If the update is offloaded
 353     // and this is not a neighbor search step, then the consumer needs to wait for the update
 354     // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
 355     // the copy event. Non-local coordinates are provided by the GPU halo exchange (if active), otherwise by H2D copy.
 356     //
 357     // In OpenCL no events are used as coordinate sync is not necessary
 358     if (GMX_GPU_OPENCL)
 359     {
 360         return nullptr;
 361     }
 362     if (atomLocality == AtomLocality::NonLocal && stepWork.useGpuXHalo)
 363     {
 364         GMX_ASSERT(gpuCoordinateHaloLaunched != nullptr,
 365                    "GPU halo exchange is active but its completion event is null.");
 366         return gpuCoordinateHaloLaunched;
 367     }
 368     if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
 369     {
 370         GMX_ASSERT(xUpdatedOnDeviceEvent_ != nullptr, "The event synchronizer can not be nullptr.");
 371         return xUpdatedOnDeviceEvent_;
 372     }
 373     else
 374     {
 375         return &xReadyOnDevice_[atomLocality];
 376     }
 377 }
 378
 379 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 380 {
 381     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 382     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 383     xReadyOnDevice_[atomLocality].waitForEvent();
 384     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 385 }
 386
 387 void StatePropagatorDataGpu::Impl::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
 388 {
 389     GMX_ASSERT(xUpdatedOnDeviceEvent != nullptr, "The event synchronizer can not be nullptr.");
 390     xUpdatedOnDeviceEvent_ = xUpdatedOnDeviceEvent;
 391 }
 392
 393 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
 394                                                           AtomLocality             atomLocality,
 395                                                           GpuEventSynchronizer*    dependency)
 396 {
 397     GMX_ASSERT(atomLocality < AtomLocality::All,
 398                formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
 399                             "coordinate transfers, passed value is \"%s\"",
 400                             enumValueToString(atomLocality))
 401                        .c_str());
 402     const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
 403     GMX_ASSERT(deviceStream != nullptr,
 404                "No stream is valid for copying positions with given atom locality.");
 405
 406     if (dependency != nullptr)
 407     {
 408         dependency->enqueueWaitEvent(*deviceStream);
 409     }
 410
 411     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 412     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 413
 414     copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
 415     // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
 416     xReadyOnHost_[atomLocality].markEvent(*deviceStream);
 417
 418     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 419     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 420 }
 421
 422 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 423 {
 424     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 425     xReadyOnHost_[atomLocality].waitForEvent();
 426     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 427 }
 428
 429
 430 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
 431 {
 432     return d_v_;
 433 }
 434
 435 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 436                                                        AtomLocality atomLocality)
 437 {
 438     GMX_ASSERT(atomLocality == AtomLocality::Local,
 439                formatString("Wrong atom locality. Only Local is allowed for "
 440                             "velocity transfers, passed value is \"%s\"",
 441                             enumValueToString(atomLocality))
 442                        .c_str());
 443     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 444     GMX_ASSERT(deviceStream != nullptr,
 445                "No stream is valid for copying velocities with given atom locality.");
 446
 447     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 448     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 449
 450     copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
 451     /* Not marking the event, because it is not used anywhere.
 452      * Since we only use velocities on the device for update, and we launch the copy in
 453      * the "update" stream, that should be safe.
 454      */
 455
 456     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 457     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 458 }
 459
 460 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
 461 {
 462     GMX_ASSERT(atomLocality == AtomLocality::Local,
 463                formatString("Wrong atom locality. Only Local is allowed for "
 464                             "velocity transfers, passed value is \"%s\"",
 465                             enumValueToString(atomLocality))
 466                        .c_str());
 467     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 468     GMX_ASSERT(deviceStream != nullptr,
 469                "No stream is valid for copying velocities with given atom locality.");
 470
 471     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 472     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 473
 474     copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
 475     vReadyOnHost_[atomLocality].markEvent(*deviceStream);
 476
 477     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 478     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 479 }
 480
 481 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 482 {
 483     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 484     vReadyOnHost_[atomLocality].waitForEvent();
 485     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 486 }
 487
 488
 489 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
 490 {
 491     return d_f_;
 492 }
 493
 494 // Copy CPU forces to GPU using stream internal to this module to allow overlap
 495 // with GPU force calculations.
 496 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
 497                                                    AtomLocality atomLocality)
 498 {
 499     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 500     DeviceStream* deviceStream = copyInStream_.get();
 501     GMX_ASSERT(deviceStream != nullptr,
 502                "No stream is valid for copying forces with given atom locality.");
 503
 504     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 505     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 506
 507     copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
 508     fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 509
 510     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 511     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 512 }
 513
 514 void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
 515 {
 516     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 517     DeviceStream* deviceStream = memsetStream_.get();
 518
 519     GMX_ASSERT(dependency != nullptr, "Dependency is not valid for clearing forces.");
 520     dependency->enqueueWaitEvent(*deviceStream);
 521
 522     GMX_ASSERT(deviceStream != nullptr,
 523                "No stream is valid for clearing forces with given atom locality.");
 524
 525     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 526     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 527
 528     clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
 529
 530     fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 531
 532     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 533     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 534 }
 535
 536 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
 537                                                                                      SimulationWorkload simulationWork)
 538 {
 539     if (stepWork.useGpuFBufferOps && !simulationWork.useCpuPmePpCommunication)
 540     {
 541         return &fReducedOnDevice_[AtomLocality::Local];
 542     }
 543     else
 544     {
 545         return &fReadyOnDevice_[AtomLocality::Local];
 546     }
 547 }
 548
 549 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice(AtomLocality atomLocality)
 550 {
 551     return &fReducedOnDevice_[atomLocality];
 552 }
 553
 554 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReadyOnDevice(AtomLocality atomLocality)
 555 {
 556     return &fReadyOnDevice_[atomLocality];
 557 }
 558
 559 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
 560 {
 561     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 562     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 563     GMX_ASSERT(deviceStream != nullptr,
 564                "No stream is valid for copying forces with given atom locality.");
 565
 566     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 567     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 568
 569     copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
 570     fReadyOnHost_[atomLocality].markEvent(*deviceStream);
 571
 572     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 573     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 574 }
 575
 576 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
 577 {
 578     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 579     fReadyOnHost_[atomLocality].waitForEvent();
 580     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 581 }
 582
 583 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
 584 {
 585     return updateStream_;
 586 }
 587
 588 int StatePropagatorDataGpu::Impl::numAtomsLocal() const
 589 {
 590     return numAtomsLocal_;
 591 }
 592
 593 int StatePropagatorDataGpu::Impl::numAtomsAll() const
 594 {
 595     return numAtomsAll_;
 596 }
 597
 598
 599 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
 600                                                GpuApiCallBehavior         transferKind,
 601                                                int            allocationBlockSizeDivisor,
 602                                                gmx_wallcycle* wcycle) :
 603     impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
 604 {
 605 }
 606
 607 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream*  pmeStream,
 608                                                const DeviceContext& deviceContext,
 609                                                GpuApiCallBehavior   transferKind,
 610                                                int                  allocationBlockSizeDivisor,
 611                                                gmx_wallcycle*       wcycle) :
 612     impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
 613 {
 614 }
 615
 616 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
 617
 618 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
 619
 620 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
 621
 622
 623 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
 624 {
 625     return impl_->reinit(numAtomsLocal, numAtomsAll);
 626 }
 627
 628 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
 629 {
 630     return impl_->getAtomRangesFromAtomLocality(atomLocality);
 631 }
 632
 633
 634 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
 635 {
 636     return impl_->getCoordinates();
 637 }
 638
 639 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 640                                                   AtomLocality                         atomLocality)
 641 {
 642     return impl_->copyCoordinatesToGpu(h_x, atomLocality);
 643 }
 644
 645 GpuEventSynchronizer*
 646 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality              atomLocality,
 647                                                          const SimulationWorkload& simulationWork,
 648                                                          const StepWorkload&       stepWork,
 649                                                          GpuEventSynchronizer* gpuCoordinateHaloLaunched)
 650 {
 651     return impl_->getCoordinatesReadyOnDeviceEvent(
 652             atomLocality, simulationWork, stepWork, gpuCoordinateHaloLaunched);
 653 }
 654
 655 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 656 {
 657     return impl_->waitCoordinatesCopiedToDevice(atomLocality);
 658 }
 659
 660 void StatePropagatorDataGpu::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
 661 {
 662     impl_->setXUpdatedOnDeviceEvent(xUpdatedOnDeviceEvent);
 663 }
 664
 665 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec>   h_x,
 666                                                     AtomLocality          atomLocality,
 667                                                     GpuEventSynchronizer* dependency)
 668 {
 669     return impl_->copyCoordinatesFromGpu(h_x, atomLocality, dependency);
 670 }
 671
 672 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 673 {
 674     return impl_->waitCoordinatesReadyOnHost(atomLocality);
 675 }
 676
 677
 678 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
 679 {
 680     return impl_->getVelocities();
 681 }
 682
 683 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 684                                                  AtomLocality                         atomLocality)
 685 {
 686     return impl_->copyVelocitiesToGpu(h_v, atomLocality);
 687 }
 688
 689 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
 690 {
 691     return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
 692 }
 693
 694 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 695 {
 696     return impl_->waitVelocitiesReadyOnHost(atomLocality);
 697 }
 698
 699
 700 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
 701 {
 702     return impl_->getForces();
 703 }
 704
 705 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
 706 {
 707     return impl_->copyForcesToGpu(h_f, atomLocality);
 708 }
 709
 710 void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
 711 {
 712     return impl_->clearForcesOnGpu(atomLocality, dependency);
 713 }
 714
 715 GpuEventSynchronizer* StatePropagatorDataGpu::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
 716                                                                                SimulationWorkload simulationWork)
 717 {
 718     return impl_->getLocalForcesReadyOnDeviceEvent(stepWork, simulationWork);
 719 }
 720
 721 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice(AtomLocality atomLocality)
 722 {
 723     return impl_->fReducedOnDevice(atomLocality);
 724 }
 725
 726 GpuEventSynchronizer* StatePropagatorDataGpu::fReadyOnDevice(AtomLocality atomLocality)
 727 {
 728     return impl_->fReadyOnDevice(atomLocality);
 729 }
 730
 731 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
 732 {
 733     return impl_->copyForcesFromGpu(h_f, atomLocality);
 734 }
 735
 736 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
 737 {
 738     return impl_->waitForcesReadyOnHost(atomLocality);
 739 }
 740
 741
 742 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
 743 {
 744     return impl_->getUpdateStream();
 745 }
 746
 747 int StatePropagatorDataGpu::numAtomsLocal() const
 748 {
 749     return impl_->numAtomsLocal();
 750 }
 751
 752 int StatePropagatorDataGpu::numAtomsAll() const
 753 {
 754     return impl_->numAtomsAll();
 755 }
 756
 757 } // namespace gmx
 758
 759 #endif // GMX_GPU