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, 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 #    if GMX_GPU == GMX_GPU_OPENCL
  54 #        include "gromacs/gpu_utils/oclutils.h"
  55 #    endif
  56 #    include "gromacs/math/vectypes.h"
  57 #    include "gromacs/mdtypes/state_propagator_data_gpu.h"
  58 #    include "gromacs/timing/wallcycle.h"
  59 #    include "gromacs/utility/classhelpers.h"
  60
  61 #    include "state_propagator_data_gpu_impl.h"
  62
  63
  64 namespace gmx
  65 {
  66
  67 StatePropagatorDataGpu::Impl::Impl(const void*        pmeStream,
  68                                    const void*        localStream,
  69                                    const void*        nonLocalStream,
  70                                    const void*        deviceContext,
  71                                    GpuApiCallBehavior transferKind,
  72                                    int                paddingSize,
  73                                    gmx_wallcycle*     wcycle) :
  74     transferKind_(transferKind),
  75     paddingSize_(paddingSize),
  76     wcycle_(wcycle)
  77 {
  78     static_assert(GMX_GPU != GMX_GPU_NONE,
  79                   "This object should only be constructed on the GPU code-paths.");
  80
  81     // TODO: Refactor when the StreamManager is introduced.
  82     if (GMX_GPU == GMX_GPU_OPENCL)
  83     {
  84         GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
  85         GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set in OpenCL builds.");
  86
  87         // The update stream is set to the PME stream in OpenCL, since PME stream is the only stream created in the PME context.
  88         pmeStream_     = *static_cast<const CommandStream*>(pmeStream);
  89         updateStream_  = *static_cast<const CommandStream*>(pmeStream);
  90         deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
  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_ = *static_cast<const CommandStream*>(pmeStream);
 100         }
 101         if (localStream != nullptr)
 102         {
 103             localStream_ = *static_cast<const CommandStream*>(localStream);
 104         }
 105         if (nonLocalStream != nullptr)
 106         {
 107             nonLocalStream_ = *static_cast<const CommandStream*>(nonLocalStream);
 108         }
 109
 110         // TODO: The update stream should be created only when it is needed.
 111 #    if (GMX_GPU == GMX_GPU_CUDA)
 112         cudaError_t stat;
 113         stat = cudaStreamCreate(&updateStream_);
 114         CU_RET_ERR(stat, "CUDA stream creation failed in StatePropagatorDataGpu");
 115 #    endif
 116         GMX_UNUSED_VALUE(deviceContext);
 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 void*        pmeStream,
 136                                    const void*        deviceContext,
 137                                    GpuApiCallBehavior transferKind,
 138                                    int                paddingSize,
 139                                    gmx_wallcycle*     wcycle) :
 140     transferKind_(transferKind),
 141     paddingSize_(paddingSize),
 142     wcycle_(wcycle)
 143 {
 144     static_assert(GMX_GPU != GMX_GPU_NONE,
 145                   "This object should only be constructed on the GPU code-paths.");
 146
 147     if (GMX_GPU == GMX_GPU_OPENCL)
 148     {
 149         GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
 150         deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
 151     }
 152
 153     GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set.");
 154     pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
 155
 156     localStream_    = nullptr;
 157     nonLocalStream_ = nullptr;
 158     updateStream_   = nullptr;
 159
 160
 161     // Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
 162     // This it temporary measure to make it safe to use this class in those cases.
 163     xCopyStreams_[AtomLocality::Local]    = pmeStream_;
 164     xCopyStreams_[AtomLocality::NonLocal] = nullptr;
 165     xCopyStreams_[AtomLocality::All]      = pmeStream_;
 166
 167     vCopyStreams_[AtomLocality::Local]    = nullptr;
 168     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
 169     vCopyStreams_[AtomLocality::All]      = nullptr;
 170
 171     fCopyStreams_[AtomLocality::Local]    = nullptr;
 172     fCopyStreams_[AtomLocality::NonLocal] = nullptr;
 173     fCopyStreams_[AtomLocality::All]      = nullptr;
 174 }
 175
 176 StatePropagatorDataGpu::Impl::~Impl() {}
 177
 178 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
 179 {
 180     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 181     wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 182
 183     numAtomsLocal_ = numAtomsLocal;
 184     numAtomsAll_   = numAtomsAll;
 185
 186     int numAtomsPadded;
 187     if (paddingSize_ > 0)
 188     {
 189         numAtomsPadded = ((numAtomsAll_ + paddingSize_ - 1) / paddingSize_) * paddingSize_;
 190     }
 191     else
 192     {
 193         numAtomsPadded = numAtomsAll_;
 194     }
 195
 196     reallocateDeviceBuffer(&d_x_, DIM * numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
 197
 198     const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
 199     if (paddingAllocationSize > 0)
 200     {
 201         // The PME stream is used here because the padding region of d_x_ is only in the PME task.
 202         clearDeviceBufferAsync(&d_x_, DIM * numAtomsAll_, DIM * paddingAllocationSize, pmeStream_);
 203     }
 204
 205     reallocateDeviceBuffer(&d_v_, DIM * numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
 206     const int d_fOldCapacity = d_fCapacity_;
 207     reallocateDeviceBuffer(&d_f_, DIM * numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
 208     // Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
 209     // the force accumulation stage before syncing with the local stream. Only done in CUDA,
 210     // since the force buffer ops are not implemented in OpenCL.
 211     if (GMX_GPU == GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
 212     {
 213         clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, localStream_);
 214     }
 215
 216     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 217     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 218 }
 219
 220 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
 221 {
 222     int atomsStartAt   = 0;
 223     int numAtomsToCopy = 0;
 224     switch (atomLocality)
 225     {
 226         case AtomLocality::All:
 227             atomsStartAt   = 0;
 228             numAtomsToCopy = numAtomsAll_;
 229             break;
 230         case AtomLocality::Local:
 231             atomsStartAt   = 0;
 232             numAtomsToCopy = numAtomsLocal_;
 233             break;
 234         case AtomLocality::NonLocal:
 235             atomsStartAt   = numAtomsLocal_;
 236             numAtomsToCopy = numAtomsAll_ - numAtomsLocal_;
 237             break;
 238         default:
 239             GMX_RELEASE_ASSERT(false,
 240                                "Wrong range of atoms requested in GPU state data manager. Should "
 241                                "be All, Local or NonLocal.");
 242     }
 243     GMX_ASSERT(atomsStartAt >= 0,
 244                "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
 245                "was not initialized.");
 246     GMX_ASSERT(numAtomsToCopy >= 0,
 247                "Number of atoms to copy is negative. Probably, the GPU propagator state was not "
 248                "initialized.");
 249     return std::make_tuple(atomsStartAt, numAtomsToCopy);
 250 }
 251
 252 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<float>                  d_data,
 253                                                 const gmx::ArrayRef<const gmx::RVec> h_data,
 254                                                 int                                  dataSize,
 255                                                 AtomLocality                         atomLocality,
 256                                                 CommandStream                        commandStream)
 257 {
 258     GMX_UNUSED_VALUE(dataSize);
 259
 260     GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
 261
 262     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 263     wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 264
 265     int atomsStartAt, numAtomsToCopy;
 266     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 267
 268     int elementsStartAt   = atomsStartAt * DIM;
 269     int numElementsToCopy = numAtomsToCopy * DIM;
 270
 271     if (numAtomsToCopy != 0)
 272     {
 273         GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
 274                    "The device allocation is smaller than requested copy range.");
 275         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 276                    "The host buffer is smaller than the requested copy range.");
 277
 278         copyToDeviceBuffer(&d_data, reinterpret_cast<const float*>(&h_data.data()[atomsStartAt]),
 279                            elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
 280     }
 281
 282     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 283     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 284 }
 285
 286 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
 287                                                   DeviceBuffer<float>      d_data,
 288                                                   int                      dataSize,
 289                                                   AtomLocality             atomLocality,
 290                                                   CommandStream            commandStream)
 291 {
 292     GMX_UNUSED_VALUE(dataSize);
 293
 294     GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
 295
 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     int elementsStartAt   = atomsStartAt * DIM;
 303     int numElementsToCopy = numAtomsToCopy * DIM;
 304
 305     if (numAtomsToCopy != 0)
 306     {
 307         GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
 308                    "The device allocation is smaller than requested copy range.");
 309         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
 310                    "The host buffer is smaller than the requested copy range.");
 311
 312         copyFromDeviceBuffer(reinterpret_cast<float*>(&h_data.data()[atomsStartAt]), &d_data,
 313                              elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
 314     }
 315
 316     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 317     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 318 }
 319
 320 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getCoordinates()
 321 {
 322     return d_x_;
 323 }
 324
 325 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 326                                                         AtomLocality atomLocality)
 327 {
 328     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 329     CommandStream commandStream = xCopyStreams_[atomLocality];
 330     GMX_ASSERT(commandStream != nullptr,
 331                "No stream is valid for copying positions with given atom locality.");
 332
 333     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 334     wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 335
 336     copyToDevice(d_x_, h_x, d_xSize_, atomLocality, commandStream);
 337
 338     // markEvent is skipped in OpenCL as:
 339     //   - it's not needed, copy is done in the same stream as the only consumer task (PME)
 340     //   - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
 341     // TODO: remove this by adding an event-mark free flavor of this function
 342     if (GMX_GPU == GMX_GPU_CUDA)
 343     {
 344         xReadyOnDevice_[atomLocality].markEvent(commandStream);
 345     }
 346
 347     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 348     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 349 }
 350
 351 GpuEventSynchronizer*
 352 StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
 353                                                                const SimulationWorkload& simulationWork,
 354                                                                const StepWorkload&       stepWork)
 355 {
 356     // The provider of the coordinates may be different for local atoms. If the update is offloaded
 357     // and this is not a neighbor search step, then the consumer needs to wait for the update
 358     // to complete. Otherwise, the coordinates are copied from the host and we need to wait for
 359     // the copy event. Non-local coordinates are always provided by the H2D copy.
 360     //
 361     // TODO: This should be reconsidered to support the halo exchange.
 362     //
 363     // In OpenCL no events are used as coordinate sync is not necessary
 364     if (GMX_GPU == GMX_GPU_OPENCL)
 365     {
 366         return nullptr;
 367     }
 368     if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
 369     {
 370         return &xUpdatedOnDevice_;
 371     }
 372     else
 373     {
 374         return &xReadyOnDevice_[atomLocality];
 375     }
 376 }
 377
 378 void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 379 {
 380     wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 381     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 382     xReadyOnDevice_[atomLocality].waitForEvent();
 383     wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 384 }
 385
 386 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
 387 {
 388     return &xUpdatedOnDevice_;
 389 }
 390
 391 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
 392 {
 393     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 394     CommandStream commandStream = xCopyStreams_[atomLocality];
 395     GMX_ASSERT(commandStream != nullptr,
 396                "No stream is valid for copying positions with given atom locality.");
 397
 398     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 399     wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 400
 401     copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, commandStream);
 402     // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
 403     xReadyOnHost_[atomLocality].markEvent(commandStream);
 404
 405     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 406     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 407 }
 408
 409 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 410 {
 411     wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 412     xReadyOnHost_[atomLocality].waitForEvent();
 413     wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 414 }
 415
 416
 417 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getVelocities()
 418 {
 419     return d_v_;
 420 }
 421
 422 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 423                                                        AtomLocality atomLocality)
 424 {
 425     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 426     CommandStream commandStream = vCopyStreams_[atomLocality];
 427     GMX_ASSERT(commandStream != nullptr,
 428                "No stream is valid for copying velocities with given atom locality.");
 429
 430     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 431     wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 432
 433     copyToDevice(d_v_, h_v, d_vSize_, atomLocality, commandStream);
 434     vReadyOnDevice_[atomLocality].markEvent(commandStream);
 435
 436     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 437     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 438 }
 439
 440 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
 441 {
 442     return &vReadyOnDevice_[atomLocality];
 443 }
 444
 445
 446 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
 447 {
 448     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 449     CommandStream commandStream = vCopyStreams_[atomLocality];
 450     GMX_ASSERT(commandStream != nullptr,
 451                "No stream is valid for copying velocities with given atom locality.");
 452
 453     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 454     wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 455
 456     copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, commandStream);
 457     vReadyOnHost_[atomLocality].markEvent(commandStream);
 458
 459     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 460     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 461 }
 462
 463 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 464 {
 465     wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 466     vReadyOnHost_[atomLocality].waitForEvent();
 467     wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 468 }
 469
 470
 471 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getForces()
 472 {
 473     return d_f_;
 474 }
 475
 476 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
 477                                                    AtomLocality atomLocality)
 478 {
 479     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 480     CommandStream commandStream = fCopyStreams_[atomLocality];
 481     GMX_ASSERT(commandStream != nullptr,
 482                "No stream is valid for copying forces with given atom locality.");
 483
 484     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 485     wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 486
 487     copyToDevice(d_f_, h_f, d_fSize_, atomLocality, commandStream);
 488     fReadyOnDevice_[atomLocality].markEvent(commandStream);
 489
 490     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 491     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 492 }
 493
 494 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
 495                                                                                 bool useGpuFBufferOps)
 496 {
 497     if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
 498     {
 499         return &fReducedOnDevice_;
 500     }
 501     else
 502     {
 503         return &fReadyOnDevice_[atomLocality];
 504     }
 505 }
 506
 507 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
 508 {
 509     return &fReducedOnDevice_;
 510 }
 511
 512 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
 513 {
 514     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 515     CommandStream commandStream = fCopyStreams_[atomLocality];
 516     GMX_ASSERT(commandStream != nullptr,
 517                "No stream is valid for copying forces with given atom locality.");
 518
 519     wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
 520     wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 521
 522     copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, commandStream);
 523     fReadyOnHost_[atomLocality].markEvent(commandStream);
 524
 525     wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
 526     wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
 527 }
 528
 529 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
 530 {
 531     wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 532     fReadyOnHost_[atomLocality].waitForEvent();
 533     wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
 534 }
 535
 536 void* StatePropagatorDataGpu::Impl::getUpdateStream()
 537 {
 538     return &updateStream_;
 539 }
 540
 541 int StatePropagatorDataGpu::Impl::numAtomsLocal()
 542 {
 543     return numAtomsLocal_;
 544 }
 545
 546 int StatePropagatorDataGpu::Impl::numAtomsAll()
 547 {
 548     return numAtomsAll_;
 549 }
 550
 551
 552 StatePropagatorDataGpu::StatePropagatorDataGpu(const void*        pmeStream,
 553                                                const void*        localStream,
 554                                                const void*        nonLocalStream,
 555                                                const void*        deviceContext,
 556                                                GpuApiCallBehavior transferKind,
 557                                                int                paddingSize,
 558                                                gmx_wallcycle*     wcycle) :
 559     impl_(new Impl(pmeStream, localStream, nonLocalStream, deviceContext, transferKind, paddingSize, wcycle))
 560 {
 561 }
 562
 563 StatePropagatorDataGpu::StatePropagatorDataGpu(const void*        pmeStream,
 564                                                const void*        deviceContext,
 565                                                GpuApiCallBehavior transferKind,
 566                                                int                paddingSize,
 567                                                gmx_wallcycle*     wcycle) :
 568     impl_(new Impl(pmeStream, deviceContext, transferKind, paddingSize, wcycle))
 569 {
 570 }
 571
 572 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu&& /* other */) noexcept = default;
 573
 574 StatePropagatorDataGpu& StatePropagatorDataGpu::operator=(StatePropagatorDataGpu&& /* other */) noexcept = default;
 575
 576 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
 577
 578
 579 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
 580 {
 581     return impl_->reinit(numAtomsLocal, numAtomsAll);
 582 }
 583
 584 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
 585 {
 586     return impl_->getAtomRangesFromAtomLocality(atomLocality);
 587 }
 588
 589
 590 DeviceBuffer<float> StatePropagatorDataGpu::getCoordinates()
 591 {
 592     return impl_->getCoordinates();
 593 }
 594
 595 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
 596                                                   AtomLocality                         atomLocality)
 597 {
 598     return impl_->copyCoordinatesToGpu(h_x, atomLocality);
 599 }
 600
 601 GpuEventSynchronizer*
 602 StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality              atomLocality,
 603                                                          const SimulationWorkload& simulationWork,
 604                                                          const StepWorkload&       stepWork)
 605 {
 606     return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
 607 }
 608
 609 void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
 610 {
 611     return impl_->waitCoordinatesCopiedToDevice(atomLocality);
 612 }
 613
 614 GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
 615 {
 616     return impl_->xUpdatedOnDevice();
 617 }
 618
 619 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
 620 {
 621     return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
 622 }
 623
 624 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
 625 {
 626     return impl_->waitCoordinatesReadyOnHost(atomLocality);
 627 }
 628
 629
 630 DeviceBuffer<float> StatePropagatorDataGpu::getVelocities()
 631 {
 632     return impl_->getVelocities();
 633 }
 634
 635 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
 636                                                  AtomLocality                         atomLocality)
 637 {
 638     return impl_->copyVelocitiesToGpu(h_v, atomLocality);
 639 }
 640
 641 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
 642 {
 643     return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
 644 }
 645
 646 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
 647 {
 648     return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
 649 }
 650
 651 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
 652 {
 653     return impl_->waitVelocitiesReadyOnHost(atomLocality);
 654 }
 655
 656
 657 DeviceBuffer<float> StatePropagatorDataGpu::getForces()
 658 {
 659     return impl_->getForces();
 660 }
 661
 662 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f, AtomLocality atomLocality)
 663 {
 664     return impl_->copyForcesToGpu(h_f, atomLocality);
 665 }
 666
 667 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
 668                                                                           bool useGpuFBufferOps)
 669 {
 670     return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
 671 }
 672
 673 GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
 674 {
 675     return impl_->fReducedOnDevice();
 676 }
 677
 678 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)
 679 {
 680     return impl_->copyForcesFromGpu(h_f, atomLocality);
 681 }
 682
 683 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality atomLocality)
 684 {
 685     return impl_->waitForcesReadyOnHost(atomLocality);
 686 }
 687
 688
 689 void* StatePropagatorDataGpu::getUpdateStream()
 690 {
 691     return impl_->getUpdateStream();
 692 }
 693
 694 int StatePropagatorDataGpu::numAtomsLocal()
 695 {
 696     return impl_->numAtomsLocal();
 697 }
 698
 699 int StatePropagatorDataGpu::numAtomsAll()
 700 {
 701     return impl_->numAtomsAll();
 702 }
 703
 704 } // namespace gmx
 705
 706 #endif // GMX_GPU == GMX_GPU_NONE