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/utility/classhelpers.h"
  59
  60 #include "state_propagator_data_gpu_impl.h"
  61
  62 namespace gmx
  63 {
  64
  65 StatePropagatorDataGpu::Impl::Impl(const void            *pmeStream,
  66                                    const void            *localStream,
  67                                    const void            *nonLocalStream,
  68                                    const void            *deviceContext,
  69                                    GpuApiCallBehavior     transferKind,
  70                                    int                    paddingSize) :
  71     transferKind_(transferKind),
  72     paddingSize_(paddingSize)
  73 {
  74     static_assert(GMX_GPU != GMX_GPU_NONE, "This object should only be constructed on the GPU code-paths.");
  75     GMX_RELEASE_ASSERT(getenv("GMX_USE_GPU_BUFFER_OPS") == nullptr, "GPU buffer ops are not supported in this build.");
  76
  77     // TODO: Refactor when the StreamManager is introduced.
  78     if (GMX_GPU == GMX_GPU_OPENCL)
  79     {
  80         GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
  81         GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set in OpenCL builds.");
  82
  83         // The update stream is set to the PME stream in OpenCL, since PME stream is the only stream created in the PME context.
  84         pmeStream_      = *static_cast<const CommandStream*>(pmeStream);
  85         updateStream_   = *static_cast<const CommandStream*>(pmeStream);
  86         deviceContext_  = *static_cast<const DeviceContext*>(deviceContext);
  87         GMX_UNUSED_VALUE(localStream);
  88         GMX_UNUSED_VALUE(nonLocalStream);
  89     }
  90
  91     if (GMX_GPU == GMX_GPU_CUDA)
  92     {
  93         if (pmeStream != nullptr)
  94         {
  95             pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
  96         }
  97         if (localStream != nullptr)
  98         {
  99             localStream_ = *static_cast<const CommandStream*>(localStream);
 100         }
 101         if (nonLocalStream != nullptr)
 102         {
 103             nonLocalStream_ = *static_cast<const CommandStream*>(nonLocalStream);
 104         }
 105
 106         // TODO: The update stream should be created only when it is needed.
 107 #if (GMX_GPU == GMX_GPU_CUDA)
 108         cudaError_t stat;
 109         stat = cudaStreamCreate(&updateStream_);
 110         CU_RET_ERR(stat, "CUDA stream creation failed in StatePropagatorDataGpu");
 111 #endif
 112         GMX_UNUSED_VALUE(deviceContext);
 113     }
 114
 115     // Map the atom locality to the stream that will be used for coordinates,
 116     // velocities and forces transfers. Same streams are used for H2D and D2H copies.
 117     // Note, that nullptr stream is used here to indicate that the copy is not supported.
 118     xCopyStreams_[AtomLocality::Local]    = updateStream_;
 119     xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
 120     xCopyStreams_[AtomLocality::All]      = updateStream_;
 121
 122     vCopyStreams_[AtomLocality::Local]    = updateStream_;
 123     vCopyStreams_[AtomLocality::NonLocal] = nullptr;
 124     vCopyStreams_[AtomLocality::All]      = updateStream_;
 125
 126     fCopyStreams_[AtomLocality::Local]    = localStream_;
 127     fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
 128     fCopyStreams_[AtomLocality::All]      = nullptr;
 129 }
 130
 131 StatePropagatorDataGpu::Impl::~Impl()
 132 {
 133 }
 134
 135 void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
 136 {
 137     numAtomsLocal_ = numAtomsLocal;
 138     numAtomsAll_   = numAtomsAll;
 139
 140     int numAtomsPadded;
 141     if (paddingSize_ > 0)
 142     {
 143         numAtomsPadded = ((numAtomsAll_ + paddingSize_ - 1 ) / paddingSize_ )*paddingSize_;
 144     }
 145     else
 146     {
 147         numAtomsPadded = numAtomsAll_;
 148     }
 149
 150     reallocateDeviceBuffer(&d_x_, DIM*numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
 151
 152     const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
 153     if (paddingAllocationSize > 0)
 154     {
 155         // The PME stream is used here because the padding region of d_x_ is only in the PME task.
 156         clearDeviceBufferAsync(&d_x_, DIM*numAtomsAll_, DIM*paddingAllocationSize, pmeStream_);
 157     }
 158
 159     reallocateDeviceBuffer(&d_v_, DIM*numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
 160     reallocateDeviceBuffer(&d_f_, DIM*numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
 161
 162 }
 163
 164 std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality  atomLocality)
 165 {
 166     int atomsStartAt   = 0;
 167     int numAtomsToCopy = 0;
 168     switch (atomLocality)
 169     {
 170         case AtomLocality::All:
 171             atomsStartAt    = 0;
 172             numAtomsToCopy  = numAtomsAll_;
 173             break;
 174         case AtomLocality::Local:
 175             atomsStartAt    = 0;
 176             numAtomsToCopy  = numAtomsLocal_;
 177             break;
 178         case AtomLocality::NonLocal:
 179             atomsStartAt    = numAtomsLocal_;
 180             numAtomsToCopy  = numAtomsAll_ - numAtomsLocal_;
 181             break;
 182         default:
 183             GMX_RELEASE_ASSERT(false, "Wrong range of atoms requested in GPU state data manager. Should be All, Local or NonLocal.");
 184     }
 185     GMX_ASSERT(atomsStartAt   >= 0, "The first elemtnt to copy has negative index. Probably, the GPU propagator state was not initialized.");
 186     GMX_ASSERT(numAtomsToCopy >= 0, "Number of atoms to copy is negative. Probably, the GPU propagator state was not initialized.");
 187     return std::make_tuple(atomsStartAt, numAtomsToCopy);
 188 }
 189
 190 void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<float>                   d_data,
 191                                                 const gmx::ArrayRef<const gmx::RVec>  h_data,
 192                                                 int                                   dataSize,
 193                                                 AtomLocality                          atomLocality,
 194                                                 CommandStream                         commandStream)
 195 {
 196
 197     GMX_UNUSED_VALUE(dataSize);
 198
 199     GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
 200
 201     int atomsStartAt, numAtomsToCopy;
 202     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 203
 204     int elementsStartAt   = atomsStartAt*DIM;
 205     int numElementsToCopy = numAtomsToCopy*DIM;
 206
 207     if (numAtomsToCopy != 0)
 208     {
 209         GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize, "The device allocation is smaller than requested copy range.");
 210         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(), "The host buffer is smaller than the requested copy range.");
 211
 212         copyToDeviceBuffer(&d_data, reinterpret_cast<const float *>(&h_data.data()[atomsStartAt]),
 213                            elementsStartAt, numElementsToCopy,
 214                            commandStream, transferKind_, nullptr);
 215     }
 216 }
 217
 218 void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec>  h_data,
 219                                                   DeviceBuffer<float>       d_data,
 220                                                   int                       dataSize,
 221                                                   AtomLocality              atomLocality,
 222                                                   CommandStream             commandStream)
 223 {
 224
 225     GMX_UNUSED_VALUE(dataSize);
 226
 227     GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
 228
 229     int atomsStartAt, numAtomsToCopy;
 230     std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
 231
 232     int elementsStartAt   = atomsStartAt*DIM;
 233     int numElementsToCopy = numAtomsToCopy*DIM;
 234
 235     if (numAtomsToCopy != 0)
 236     {
 237         GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize, "The device allocation is smaller than requested copy range.");
 238         GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(), "The host buffer is smaller than the requested copy range.");
 239
 240         copyFromDeviceBuffer(reinterpret_cast<float*>(&h_data.data()[atomsStartAt]), &d_data,
 241                              elementsStartAt, numElementsToCopy,
 242                              commandStream, transferKind_, nullptr);
 243     }
 244 }
 245
 246 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getCoordinates()
 247 {
 248     return d_x_;
 249 }
 250
 251 void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec>  h_x,
 252                                                         AtomLocality                          atomLocality)
 253 {
 254     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 255     CommandStream commandStream = xCopyStreams_[atomLocality];
 256     GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying positions with given atom locality.");
 257
 258     copyToDevice(d_x_, h_x, d_xSize_, atomLocality, commandStream);
 259
 260     // markEvent is skipped in OpenCL as:
 261     //   - it's not needed, copy is done in the same stream as the only consumer task (PME)
 262     //   - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
 263     // TODO: remove this by adding an event-mark free flavor of this function
 264     if (GMX_GPU == GMX_GPU_CUDA)
 265     {
 266         xReadyOnDevice_[atomLocality].markEvent(commandStream);
 267         // TODO: Remove When event-based synchronization is introduced
 268         gpuStreamSynchronize(commandStream);
 269     }
 270 }
 271
 272 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality  atomLocality)
 273 {
 274     return &xReadyOnDevice_[atomLocality];
 275 }
 276
 277 void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec>  h_x,
 278                                                           AtomLocality              atomLocality)
 279 {
 280     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 281     CommandStream commandStream = xCopyStreams_[atomLocality];
 282     GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying positions with given atom locality.");
 283
 284     copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, commandStream);
 285     // TODO: Remove When event-based synchronization is introduced
 286     gpuStreamSynchronize(commandStream);
 287     // Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
 288     xReadyOnHost_[atomLocality].markEvent(commandStream);
 289 }
 290
 291 void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality  atomLocality)
 292 {
 293     xReadyOnHost_[atomLocality].waitForEvent();
 294 }
 295
 296
 297 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getVelocities()
 298 {
 299     return d_v_;
 300 }
 301
 302 void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec>  h_v,
 303                                                        AtomLocality                          atomLocality)
 304 {
 305     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 306     CommandStream commandStream = vCopyStreams_[atomLocality];
 307     GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying velocities with given atom locality.");
 308
 309     copyToDevice(d_v_, h_v, d_vSize_, atomLocality, commandStream);
 310     // TODO: Remove When event-based synchronization is introduced
 311     gpuStreamSynchronize(commandStream);
 312     vReadyOnDevice_[atomLocality].markEvent(commandStream);
 313 }
 314
 315 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality  atomLocality)
 316 {
 317     return &vReadyOnDevice_[atomLocality];
 318 }
 319
 320
 321 void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec>  h_v,
 322                                                          AtomLocality              atomLocality)
 323 {
 324     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 325     CommandStream commandStream = vCopyStreams_[atomLocality];
 326     GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying velocities with given atom locality.");
 327
 328     copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, commandStream);
 329     // TODO: Remove When event-based synchronization is introduced
 330     gpuStreamSynchronize(commandStream);
 331     vReadyOnHost_[atomLocality].markEvent(commandStream);
 332 }
 333
 334 void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality  atomLocality)
 335 {
 336     vReadyOnHost_[atomLocality].waitForEvent();
 337 }
 338
 339
 340 DeviceBuffer<float> StatePropagatorDataGpu::Impl::getForces()
 341 {
 342     return d_f_;
 343 }
 344
 345 void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec>  h_f,
 346                                                    AtomLocality                          atomLocality)
 347 {
 348     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 349     CommandStream commandStream = fCopyStreams_[atomLocality];
 350     GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying forces with given atom locality.");
 351
 352     copyToDevice(d_f_, h_f, d_fSize_, atomLocality, commandStream);
 353     // TODO: Remove When event-based synchronization is introduced
 354     gpuStreamSynchronize(commandStream);
 355     fReadyOnDevice_[atomLocality].markEvent(commandStream);
 356 }
 357
 358 GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality  atomLocality)
 359 {
 360     return &fReadyOnDevice_[atomLocality];
 361 }
 362
 363
 364 void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec>  h_f,
 365                                                      AtomLocality              atomLocality)
 366 {
 367     GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
 368     CommandStream commandStream = fCopyStreams_[atomLocality];
 369     GMX_ASSERT(commandStream != nullptr, "No stream is valid for copying forces with given atom locality.");
 370
 371     copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, commandStream);
 372     // TODO: Remove When event-based synchronization is introduced
 373     gpuStreamSynchronize(commandStream);
 374     fReadyOnHost_[atomLocality].markEvent(commandStream);
 375 }
 376
 377 void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality  atomLocality)
 378 {
 379     fReadyOnHost_[atomLocality].waitForEvent();
 380 }
 381
 382 void* StatePropagatorDataGpu::Impl::getUpdateStream()
 383 {
 384     return updateStream_;
 385 }
 386
 387 int StatePropagatorDataGpu::Impl::numAtomsLocal()
 388 {
 389     return numAtomsLocal_;
 390 }
 391
 392 int StatePropagatorDataGpu::Impl::numAtomsAll()
 393 {
 394     return numAtomsAll_;
 395 }
 396
 397
 398
 399 StatePropagatorDataGpu::StatePropagatorDataGpu(const void        *pmeStream,
 400                                                const void        *localStream,
 401                                                const void        *nonLocalStream,
 402                                                const void        *deviceContext,
 403                                                GpuApiCallBehavior transferKind,
 404                                                int                paddingSize)
 405     : impl_(new Impl(pmeStream,
 406                      localStream,
 407                      nonLocalStream,
 408                      deviceContext,
 409                      transferKind,
 410                      paddingSize))
 411 {
 412 }
 413
 414 StatePropagatorDataGpu::StatePropagatorDataGpu(StatePropagatorDataGpu && /* other */) noexcept = default;
 415
 416 StatePropagatorDataGpu &StatePropagatorDataGpu::operator=(StatePropagatorDataGpu && /* other */) noexcept = default;
 417
 418 StatePropagatorDataGpu::~StatePropagatorDataGpu() = default;
 419
 420
 421 void StatePropagatorDataGpu::reinit(int numAtomsLocal, int numAtomsAll)
 422 {
 423     return impl_->reinit(numAtomsLocal, numAtomsAll);
 424 }
 425
 426 std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality  atomLocality)
 427 {
 428     return impl_->getAtomRangesFromAtomLocality(atomLocality);
 429 }
 430
 431
 432 DeviceBuffer<float> StatePropagatorDataGpu::getCoordinates()
 433 {
 434     return impl_->getCoordinates();
 435 }
 436
 437 void StatePropagatorDataGpu::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec>  h_x,
 438                                                   AtomLocality                          atomLocality)
 439 {
 440     return impl_->copyCoordinatesToGpu(h_x, atomLocality);
 441 }
 442
 443 GpuEventSynchronizer* StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality  atomLocality)
 444 {
 445     return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality);
 446 }
 447
 448 void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec>  h_x,
 449                                                     AtomLocality         atomLocality)
 450 {
 451     return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
 452 }
 453
 454 void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality  atomLocality)
 455 {
 456     return impl_->waitCoordinatesReadyOnHost(atomLocality);
 457 }
 458
 459
 460 DeviceBuffer<float> StatePropagatorDataGpu::getVelocities()
 461 {
 462     return impl_->getVelocities();
 463 }
 464
 465 void StatePropagatorDataGpu::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec>  h_v,
 466                                                  AtomLocality                          atomLocality)
 467 {
 468     return impl_->copyVelocitiesToGpu(h_v, atomLocality);
 469 }
 470
 471 GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality  atomLocality)
 472 {
 473     return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
 474 }
 475
 476 void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec>  h_v,
 477                                                    AtomLocality         atomLocality)
 478 {
 479     return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
 480 }
 481
 482 void StatePropagatorDataGpu::waitVelocitiesReadyOnHost(AtomLocality  atomLocality)
 483 {
 484     return impl_->waitVelocitiesReadyOnHost(atomLocality);
 485 }
 486
 487
 488 DeviceBuffer<float> StatePropagatorDataGpu::getForces()
 489 {
 490     return impl_->getForces();
 491 }
 492
 493 void StatePropagatorDataGpu::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec>  h_f,
 494                                              AtomLocality                          atomLocality)
 495 {
 496     return impl_->copyForcesToGpu(h_f, atomLocality);
 497 }
 498
 499 GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality  atomLocality)
 500 {
 501     return impl_->getForcesReadyOnDeviceEvent(atomLocality);
 502 }
 503
 504 void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec>  h_f,
 505                                                AtomLocality         atomLocality)
 506 {
 507     return impl_->copyForcesFromGpu(h_f, atomLocality);
 508 }
 509
 510 void StatePropagatorDataGpu::waitForcesReadyOnHost(AtomLocality  atomLocality)
 511 {
 512     return impl_->waitForcesReadyOnHost(atomLocality);
 513 }
 514
 515
 516 void* StatePropagatorDataGpu::getUpdateStream()
 517 {
 518     return impl_->getUpdateStream();
 519 }
 520
 521 int StatePropagatorDataGpu::numAtomsLocal()
 522 {
 523     return impl_->numAtomsLocal();
 524 }
 525
 526 int StatePropagatorDataGpu::numAtomsAll()
 527 {
 528     return impl_->numAtomsAll();
 529 }
 530
 531 }      // namespace gmx
 532
 533 #endif // GMX_GPU == GMX_GPU_NONE