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