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
 100 StatePropagatorDataGpu::Impl::Impl(const DeviceStream*  pmeStream,
 101                                    const DeviceContext& deviceContext,
 102                                    GpuApiCallBehavior   transferKind,
 103                                    int                  allocationBlockSizeDivisor,
 104                                    gmx_wallcycle*       wcycle) :
 105     deviceContext_(deviceContext),
 106     transferKind_(transferKind),
 107     allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
 108     wcycle_(wcycle)
 109 {
 110     static_assert(
 111             GMX_GPU,
 112             "GPU state propagator data object should only be constructed on the GPU code-paths.");
 113
 114     GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
 115     pmeStream_      = pmeStream;
 116     localStream_    = pmeStream; // For clearing the force buffer
 117     nonLocalStream_ = nullptr;
 118     updateStream_   = nullptr;
 119
 120
 121     // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
 122     // This it temporary measure to make it safe to use this class in those cases.
 123     xCopyStreams_[AtomLocality::Local]    = pmeStream_;
 124     xCopyStreams_[AtomLocality::NonLocal] = nullptr;
 125     xCopyStreams_[AtomLocality::All]      = nullptr;
 126
 127     vCopyStreams_[AtomLocality::Local]    = nullptr;
 128     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
 129     vCopyStreams_[AtomLocality::All]      = nullptr;
 130
 131     fCopyStreams_[AtomLocality::Local]    = nullptr;
 132     fCopyStreams_[AtomLocality::NonLocal] = nullptr;
 133     fCopyStreams_[AtomLocality::All]      = nullptr;
 134 }
 135
 136 StatePropagatorDataGpu::Impl::~Impl() {}
 137
 138 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
 139 {
 140     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 141     wallcycle_sub_start_nocount(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 142
 143     numAtomsLocal_ = numAtomsLocal;
 144     numAtomsAll_   = numAtomsAll;
 145
 146     int numAtomsPadded;
 147     if (allocationBlockSizeDivisor_ > 0)
 148     {
 149         numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
 150                          * allocationBlockSizeDivisor_;
 151     }
 152     else
 153     {
 154         numAtomsPadded = numAtomsAll_;
 155     }
 156
 157     reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
 158
 159     const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
 160     if (paddingAllocationSize > 0)
 161     {
 162         // The PME stream is used here because the padding region of d_x_ is only in the PME task.
 163         clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
 164     }
 165
 166     reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
 167     const int d_fOldCapacity = d_fCapacity_;
 168     reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
 169
 170     // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
 171     // the force accumulation stage before syncing with the local stream. Only done in CUDA,
 172     // since the force buffer ops are not implemented in OpenCL.
 173     if (GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
 174     {
 175         clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
 176     }
 177
 178     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 179     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 180 }
 181
 182 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
 183 {
 184     int atomsStartAt   = 0;
 185     int numAtomsToCopy = 0;
 186     switch (atomLocality)
 187     {
 188         case AtomLocality::All:
 189             atomsStartAt   = 0;
 190             numAtomsToCopy = numAtomsAll_;
 191             break;
 192         case AtomLocality::Local:
 193             atomsStartAt   = 0;
 194             numAtomsToCopy = numAtomsLocal_;
 195             break;
 196         case AtomLocality::NonLocal:
 197             atomsStartAt   = numAtomsLocal_;
 198             numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
 199             break;
 200         default:
 201             GMX_RELEASE_ASSERT(false,
 202                                "Wrong range of atoms requested in GPU state data manager. Should "
 203                                "be All, Local or NonLocal.");
 204     }
 205     GMX_ASSERT(atomsStartAt >= 0,
 206                "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
 207                "was not initialized.");
 208     GMX_ASSERT(numAtomsToCopy >= 0,
 209                "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
 210                "initialized.");
 211     return std::make_tuple(atomsStartAt, numAtomsToCopy);
 212 }
 213
 214 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<RVec>                   d_data,
 215                                                 const gmx::ArrayRef<const gmx::RVec> h_data,
 216                                                 int                                  dataSize,
 217                                                 AtomLocality                         atomLocality,
 218                                                 const DeviceStream&                  deviceStream)
 219 {
 220     GMX_UNUSED_VALUE(dataSize);
 221
 222     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 223
 224     GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
 225
 226     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
 227
 228     int atomsStartAt, numAtomsToCopy;
 229     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 230
 231     if (numAtomsToCopy != 0)
 232     {
 233         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
 234                    "The device allocation is smaller than requested copy range.");
 235         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 236                    "The host buffer is smaller than the requested copy range.");
 237
 238         copyToDeviceBuffer(&d_data,
 239                            reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
 240                            atomsStartAt,
 241                            numAtomsToCopy,
 242                            deviceStream,
 243                            transferKind_,
 244                            nullptr);
 245     }
 246 }
 247
 248 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
 249                                                   DeviceBuffer<RVec>       d_data,
 250                                                   int                      dataSize,
 251                                                   AtomLocality             atomLocality,
 252                                                   const DeviceStream&      deviceStream)
 253 {
 254     GMX_UNUSED_VALUE(dataSize);
 255
 256     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 257
 258     GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
 259
 260     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
 261
 262     int atomsStartAt, numAtomsToCopy;
 263     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 264
 265     if (numAtomsToCopy != 0)
 266     {
 267         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
 268                    "The device allocation is smaller than requested copy range.");
 269         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 270                    "The host buffer is smaller than the requested copy range.");
 271
 272         copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]),
 273                              &d_data,
 274                              atomsStartAt,
 275                              numAtomsToCopy,
 276                              deviceStream,
 277                              transferKind_,
 278                              nullptr);
 279     }
 280 }
 281
 282 void StatePropagatorDataGpu::Impl::clearOnDevice(DeviceBuffer<RVec>  d_data,
 283                                                  int                 dataSize,
 284                                                  AtomLocality        atomLocality,
 285                                                  const DeviceStream& deviceStream) const
 286 {
 287     GMX_UNUSED_VALUE(dataSize);
 288
 289     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 290
 291     GMX_ASSERT(dataSize >= 0, "Trying to clear to device buffer before it was allocated.");
 292
 293     GMX_ASSERT(deviceStream.isValid(), "No stream is valid for clearing with given atom locality.");
 294
 295     int atomsStartAt, numAtomsToClear;
 296     std::tie(atomsStartAt, numAtomsToClear) = getAtomRangesFromAtomLocality(atomLocality);
 297
 298     if (numAtomsToClear != 0)
 299     {
 300         GMX_ASSERT(atomsStartAt + numAtomsToClear <= dataSize,
 301                    "The device allocation is smaller than requested clear range.");
 302
 303         clearDeviceBufferAsync(&d_data, atomsStartAt, numAtomsToClear, deviceStream);
 304     }
 305 }
 306
 307 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
 308 {
 309     return d_x_;
 310 }
 311
 312 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 313                                                         AtomLocality atomLocality)
 314 {
 315     GMX_ASSERT(atomLocality < AtomLocality::All,
 316                formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
 317                             "coordinate transfers, passed value is \"%s\"",
 318                             enumValueToString(atomLocality))
 319                        .c_str());
 320
 321     const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
 322     GMX_ASSERT(deviceStream != nullptr,
 323                "No stream is valid for copying positions with given atom locality.");
 324
 325     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 326     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 327
 328     copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
 329
 330     // markEvent is skipped in OpenCL as:
 331     //   - it's not needed, copy is done in the same stream as the only consumer task (PME)
 332     //   - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
 333     // TODO: remove this by adding an event-mark free flavor of this function
 334     if (GMX_GPU_CUDA)
 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*
 344 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
 345                                                                const SimulationWorkload& simulationWork,
 346                                                                const StepWorkload&       stepWork)
 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 always provided by the H2D copy.
 352     //
 353     // TODO: This should be reconsidered to support the halo exchange.
 354     //
 355     // In OpenCL no events are used as coordinate sync is not necessary
 356     if (GMX_GPU_OPENCL)
 357     {
 358         return nullptr;
 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         return &xReadyOnDevice_[atomLocality];
 368     }
 369 }
 370
 371 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 372 {
 373     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 374     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 375     xReadyOnDevice_[atomLocality].waitForEvent();
 376     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 377 }
 378
 379 void StatePropagatorDataGpu::Impl::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
 380 {
 381     GMX_ASSERT(xUpdatedOnDeviceEvent != nullptr, "The event synchronizer can not be nullptr.");
 382     xUpdatedOnDeviceEvent_ = xUpdatedOnDeviceEvent;
 383 }
 384
 385 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
 386 {
 387     GMX_ASSERT(atomLocality < AtomLocality::All,
 388                formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
 389                             "coordinate transfers, passed value is \"%s\"",
 390                             enumValueToString(atomLocality))
 391                        .c_str());
 392     const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
 393     GMX_ASSERT(deviceStream != nullptr,
 394                "No stream is valid for copying positions with given atom locality.");
 395
 396     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 397     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 398
 399     copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
 400     // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
 401     xReadyOnHost_[atomLocality].markEvent(*deviceStream);
 402
 403     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 404     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 405 }
 406
 407 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 408 {
 409     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 410     xReadyOnHost_[atomLocality].waitForEvent();
 411     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 412 }
 413
 414
 415 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
 416 {
 417     return d_v_;
 418 }
 419
 420 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 421                                                        AtomLocality atomLocality)
 422 {
 423     GMX_ASSERT(atomLocality == AtomLocality::Local,
 424                formatString("Wrong atom locality. Only Local is allowed for "
 425                             "velocity transfers, passed value is \"%s\"",
 426                             enumValueToString(atomLocality))
 427                        .c_str());
 428     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 429     GMX_ASSERT(deviceStream != nullptr,
 430                "No stream is valid for copying velocities with given atom locality.");
 431
 432     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 433     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 434
 435     copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
 436     /* Not marking the event, because it is not used anywhere.
 437      * Since we only use velocities on the device for update, and we launch the copy in
 438      * the "update" stream, that should be safe.
 439      */
 440
 441     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 442     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 443 }
 444
 445 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
 446 {
 447     GMX_ASSERT(atomLocality == AtomLocality::Local,
 448                formatString("Wrong atom locality. Only Local is allowed for "
 449                             "velocity transfers, passed value is \"%s\"",
 450                             enumValueToString(atomLocality))
 451                        .c_str());
 452     const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
 453     GMX_ASSERT(deviceStream != nullptr,
 454                "No stream is valid for copying velocities with given atom locality.");
 455
 456     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 457     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 458
 459     copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
 460     vReadyOnHost_[atomLocality].markEvent(*deviceStream);
 461
 462     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 463     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 464 }
 465
 466 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 467 {
 468     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 469     vReadyOnHost_[atomLocality].waitForEvent();
 470     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 471 }
 472
 473
 474 DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
 475 {
 476     return d_f_;
 477 }
 478
 479 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
 480                                                    AtomLocality atomLocality)
 481 {
 482     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 483     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 484     GMX_ASSERT(deviceStream != nullptr,
 485                "No stream is valid for copying forces with given atom locality.");
 486
 487     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 488     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 489
 490     copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
 491     fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
 492
 493     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 494     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 495 }
 496
 497 void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality)
 498 {
 499     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 500     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 501     GMX_ASSERT(deviceStream != nullptr,
 502                "No stream is valid for clearing forces with given atom locality.");
 503
 504     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 505     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 506
 507     clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
 508
 509     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 510     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 511 }
 512
 513 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
 514                                                                                 bool useGpuFBufferOps)
 515 {
 516     if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
 517     {
 518         return &fReducedOnDevice_;
 519     }
 520     else
 521     {
 522         return &fReadyOnDevice_[atomLocality];
 523     }
 524 }
 525
 526 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
 527 {
 528     return &fReducedOnDevice_;
 529 }
 530
 531 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
 532 {
 533     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 534     const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
 535     GMX_ASSERT(deviceStream != nullptr,
 536                "No stream is valid for copying forces with given atom locality.");
 537
 538     wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
 539     wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 540
 541     copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
 542     fReadyOnHost_[atomLocality].markEvent(*deviceStream);
 543
 544     wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
 545     wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
 546 }
 547
 548 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
 549 {
 550     wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 551     fReadyOnHost_[atomLocality].waitForEvent();
 552     wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
 553 }
 554
 555 const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
 556 {
 557     return updateStream_;
 558 }
 559
 560 int StatePropagatorDataGpu::Impl::numAtomsLocal() const
 561 {
 562     return numAtomsLocal_;
 563 }
 564
 565 int StatePropagatorDataGpu::Impl::numAtomsAll() const
 566 {
 567     return numAtomsAll_;
 568 }
 569
 570
 571 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
 572                                                GpuApiCallBehavior         transferKind,
 573                                                int            allocationBlockSizeDivisor,
 574                                                gmx_wallcycle* wcycle) :
 575     impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
 576 {
 577 }
 578
 579 StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream*  pmeStream,
 580                                                const DeviceContext& deviceContext,
 581                                                GpuApiCallBehavior   transferKind,
 582                                                int                  allocationBlockSizeDivisor,
 583                                                gmx_wallcycle*       wcycle) :
 584     impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
 585 {
 586 }
 587
 588 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
 589
 590 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
 591
 592 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
 593
 594
 595 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
 596 {
 597     return impl_->reinit(numAtomsLocal, numAtomsAll);
 598 }
 599
 600 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
 601 {
 602     return impl_->getAtomRangesFromAtomLocality(atomLocality);
 603 }
 604
 605
 606 DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
 607 {
 608     return impl_->getCoordinates();
 609 }
 610
 611 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 612                                                   AtomLocality                         atomLocality)
 613 {
 614     return impl_->copyCoordinatesToGpu(h_x, atomLocality);
 615 }
 616
 617 GpuEventSynchronizer*
 618 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality              atomLocality,
 619                                                          const SimulationWorkload& simulationWork,
 620                                                          const StepWorkload&       stepWork)
 621 {
 622     return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
 623 }
 624
 625 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 626 {
 627     return impl_->waitCoordinatesCopiedToDevice(atomLocality);
 628 }
 629
 630 void StatePropagatorDataGpu::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
 631 {
 632     impl_->setXUpdatedOnDeviceEvent(xUpdatedOnDeviceEvent);
 633 }
 634
 635 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
 636 {
 637     return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
 638 }
 639
 640 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 641 {
 642     return impl_->waitCoordinatesReadyOnHost(atomLocality);
 643 }
 644
 645
 646 DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
 647 {
 648     return impl_->getVelocities();
 649 }
 650
 651 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 652                                                  AtomLocality                         atomLocality)
 653 {
 654     return impl_->copyVelocitiesToGpu(h_v, atomLocality);
 655 }
 656
 657 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
 658 {
 659     return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
 660 }
 661
 662 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 663 {
 664     return impl_->waitVelocitiesReadyOnHost(atomLocality);
 665 }
 666
 667
 668 DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
 669 {
 670     return impl_->getForces();
 671 }
 672
 673 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
 674 {
 675     return impl_->copyForcesToGpu(h_f, atomLocality);
 676 }
 677
 678 void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality)
 679 {
 680     return impl_->clearForcesOnGpu(atomLocality);
 681 }
 682
 683 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
 684                                                                           bool useGpuFBufferOps)
 685 {
 686     return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
 687 }
 688
 689 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
 690 {
 691     return impl_->fReducedOnDevice();
 692 }
 693
 694 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
 695 {
 696     return impl_->copyForcesFromGpu(h_f, atomLocality);
 697 }
 698
 699 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
 700 {
 701     return impl_->waitForcesReadyOnHost(atomLocality);
 702 }
 703
 704
 705 const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
 706 {
 707     return impl_->getUpdateStream();
 708 }
 709
 710 int StatePropagatorDataGpu::numAtomsLocal() const
 711 {
 712     return impl_->numAtomsLocal();
 713 }
 714
 715 int StatePropagatorDataGpu::numAtomsAll() const
 716 {
 717     return impl_->numAtomsAll();
 718 }
 719
 720 } // namespace gmx
 721
 722 #endif // GMX_GPU