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