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