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