src/gromacs/domdec/gpuhaloexchange_impl.cu

   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 Implements GPU halo exchange using CUDA.
  38  *
  39  *
  40  * \author Alan Gray <alang@nvidia.com>
  41  *
  42  * \ingroup module_domdec
  43  */
  44 #include "gmxpre.h"
  45
  46 #include "gpuhaloexchange_impl.cuh"
  47
  48 #include "config.h"
  49
  50 #include <assert.h>
  51 #include <stdio.h>
  52
  53 #include "gromacs/domdec/domdec.h"
  54 #include "gromacs/domdec/domdec_struct.h"
  55 #include "gromacs/domdec/gpuhaloexchange.h"
  56 #include "gromacs/gpu_utils/cudautils.cuh"
  57 #include "gromacs/gpu_utils/devicebuffer.h"
  58 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
  59 #include "gromacs/gpu_utils/vectype_ops.cuh"
  60 #include "gromacs/pbcutil/ishift.h"
  61
  62 #include "domdec_internal.h"
  63
  64 namespace gmx
  65 {
  66
  67 //! Number of CUDA threads in a block
  68 //TODO Optimize this through experimentation
  69 constexpr static int c_threadsPerBlock = 256;
  70
  71 template <bool usePBC>
  72 __global__ void packSendBufKernel(float3 * __restrict__       dataPacked,
  73                                   const float3 * __restrict__ data,
  74                                   const int * __restrict__    map,
  75                                   const int                   mapSize,
  76                                   const float3                coordinateShift)
  77 {
  78     int           threadIndex          = blockIdx.x*blockDim.x+threadIdx.x;
  79     float3       *gm_dataDest          = &dataPacked[threadIndex];
  80     const float3 *gm_dataSrc           = &data[map[threadIndex]];
  81
  82     if (threadIndex < mapSize)
  83     {
  84         if (usePBC)
  85         {
  86             *gm_dataDest = *gm_dataSrc + coordinateShift;
  87         }
  88         else
  89         {
  90             *gm_dataDest = *gm_dataSrc;
  91         }
  92
  93     }
  94
  95     return;
  96 }
  97
  98 /*! \brief unpack non-local force data buffer on the GPU using pre-populated "map" containing index information
  99  * \param[out] data        full array of force values
 100  * \param[in]  dataPacked  packed array of force values to be transferred
 101  * \param[in]  map         array of indices defining mapping from full to packed array
 102  * \param[in]  mapSize     number of elements in map array
 103  */
 104 template <bool accumulate>
 105 __global__ void unpackRecvBufKernel(float3 * __restrict__       data,
 106                                     const float3 * __restrict__ dataPacked,
 107                                     const int * __restrict__    map,
 108                                     const int                   mapSize)
 109 {
 110
 111     int           threadIndex           = blockIdx.x*blockDim.x+threadIdx.x;
 112     const float3 *gm_dataSrc            = &dataPacked[threadIndex];
 113     float3       *gm_dataDest           = &data[map[threadIndex]];
 114
 115     if (threadIndex < mapSize)
 116     {
 117         if (accumulate)
 118         {
 119             *gm_dataDest += *gm_dataSrc;
 120         }
 121         else
 122         {
 123             *gm_dataDest = *gm_dataSrc;
 124         }
 125     }
 126
 127     return;
 128 }
 129
 130 void GpuHaloExchange::Impl::reinitHalo(float3      *d_coordinatesBuffer,
 131                                        float3      *d_forcesBuffer)
 132 {
 133
 134     d_x_ = d_coordinatesBuffer;
 135     d_f_ = d_forcesBuffer;
 136
 137     cudaStream_t                  stream            = nonLocalStream_;
 138     int                           nzone             = 1;
 139     const gmx_domdec_comm_t      &comm              = *dd_->comm;
 140     const gmx_domdec_comm_dim_t  &cd                = comm.cd[0];
 141     const gmx_domdec_ind_t       &ind               = cd.ind[0];
 142     int                           newSize           = ind.nsend[nzone+1];
 143
 144     GMX_RELEASE_ASSERT(cd.numPulses() == 1, "Multiple pulses are not yet supported in GPU halo exchange");
 145     GMX_ASSERT(cd.receiveInPlace, "Out-of-place receive is not yet supported in GPU halo exchange");
 146
 147     // reallocates only if needed
 148     h_indexMap_.resize(newSize);
 149     // reallocate on device only if needed
 150     if (newSize > maxPackedBufferSize_)
 151     {
 152         reallocateDeviceBuffer(&d_indexMap_, newSize, &indexMapSize_, &indexMapSizeAlloc_, nullptr);
 153         reallocateDeviceBuffer(&d_sendBuf_, newSize, &sendBufSize_, &sendBufSizeAlloc_, nullptr);
 154         reallocateDeviceBuffer(&d_recvBuf_, newSize, &recvBufSize_, &recvBufSizeAlloc_, nullptr);
 155         maxPackedBufferSize_ = newSize;
 156     }
 157
 158     xSendSize_ = newSize;
 159 #if GMX_MPI
 160     MPI_Sendrecv(&xSendSize_, sizeof(int), MPI_BYTE, sendRankX_, 0,
 161                  &xRecvSize_, sizeof(int), MPI_BYTE, recvRankX_, 0,
 162                  mpi_comm_mysim_, MPI_STATUS_IGNORE);
 163 #endif
 164     fSendSize_ = xRecvSize_;
 165     fRecvSize_ = xSendSize_;
 166
 167     numHomeAtoms_ = comm.atomRanges.numHomeAtoms();  //offset for data recieved by this rank
 168
 169     GMX_ASSERT(ind.index.size() == h_indexMap_.size(), "Size mismatch");
 170     std::copy(ind.index.begin(), ind.index.end(), h_indexMap_.begin());
 171
 172     copyToDeviceBuffer(&d_indexMap_, h_indexMap_.data(), 0, newSize, stream, GpuApiCallBehavior::Async, nullptr);
 173
 174     // This rank will push data to its neighbor, so needs to know
 175     // the remote receive address and similarly send its receive
 176     // address to other neighbour. We can do this here in reinit fn
 177     // since the pointers will not change until the next NS step.
 178
 179     //Coordinates buffer:
 180 #if GMX_MPI
 181     void* recvPtr  = static_cast<void*> (&d_coordinatesBuffer[numHomeAtoms_]);
 182     MPI_Sendrecv(&recvPtr, sizeof(void*), MPI_BYTE, recvRankX_, 0,
 183                  &remoteXPtr_, sizeof(void*), MPI_BYTE, sendRankX_, 0,
 184                  mpi_comm_mysim_, MPI_STATUS_IGNORE);
 185
 186     //Force buffer:
 187     recvPtr  = static_cast<void*> (d_recvBuf_);
 188     MPI_Sendrecv(&recvPtr, sizeof(void*), MPI_BYTE, recvRankF_, 0,
 189                  &remoteFPtr_, sizeof(void*), MPI_BYTE, sendRankF_, 0,
 190                  mpi_comm_mysim_, MPI_STATUS_IGNORE);
 191 #endif
 192
 193
 194     return;
 195 }
 196
 197 void GpuHaloExchange::Impl::communicateHaloCoordinates(const matrix          box,
 198                                                        GpuEventSynchronizer *coordinatesReadyOnDeviceEvent)
 199 {
 200
 201     //ensure stream waits until coordinate data is available on device
 202     coordinatesReadyOnDeviceEvent->enqueueWaitEvent(nonLocalStream_);
 203
 204     // launch kernel to pack send buffer
 205     KernelLaunchConfig config;
 206     config.blockSize[0]     = c_threadsPerBlock;
 207     config.blockSize[1]     = 1;
 208     config.blockSize[2]     = 1;
 209     config.gridSize[0]      = (xSendSize_+c_threadsPerBlock-1)/c_threadsPerBlock;
 210     config.gridSize[1]      = 1;
 211     config.gridSize[2]      = 1;
 212     config.sharedMemorySize = 0;
 213     config.stream           = nonLocalStream_;
 214
 215     const float3     *sendBuf  = d_sendBuf_;
 216     const float3     *d_x      = d_x_;
 217     const int        *indexMap = d_indexMap_;
 218     const int         size     = xSendSize_;
 219     // The coordinateShift changes between steps when we have
 220     // performed a DD partition, or have updated the box e.g. when
 221     // performing pressure coupling. So, for simplicity, the the box
 222     // is used every step to pass the shift vector as an argument of
 223     // the packing kernel.
 224     //
 225     // Because only one-dimensional DD is supported, the coordinate
 226     // shift only needs to handle that dimension.
 227     const int         dimensionIndex = dd_->dim[0];
 228     const float3      coordinateShift {
 229         box[dimensionIndex][XX], box[dimensionIndex][YY], box[dimensionIndex][ZZ]
 230     };
 231
 232     // Avoid launching kernel when there is no work to do
 233     if (size > 0)
 234     {
 235         auto              kernelFn = usePBC_ ? packSendBufKernel<true> : packSendBufKernel<false>;
 236
 237         const auto        kernelArgs   = prepareGpuKernelArguments(kernelFn, config, &sendBuf, &d_x, &indexMap,
 238                                                                    &size, &coordinateShift);
 239
 240         launchGpuKernel(kernelFn, config, nullptr, "Domdec GPU Apply X Halo Exchange", kernelArgs);
 241     }
 242
 243     communicateHaloData(d_x_, HaloQuantity::HaloCoordinates);
 244
 245     return;
 246 }
 247
 248 // The following method should be called after non-local buffer operations,
 249 // and before the local buffer operations. It operates in the non-local stream.
 250 void GpuHaloExchange::Impl::communicateHaloForces(bool accumulateForces)
 251 {
 252
 253     // Communicate halo data (in non-local stream)
 254     communicateHaloData(d_f_, HaloQuantity::HaloForces);
 255
 256     float3            *d_f            = d_f_;
 257
 258     if (!accumulateForces)
 259     {
 260         //Clear local portion of force array (in local stream)
 261         cudaMemsetAsync(d_f, 0, numHomeAtoms_*sizeof(rvec), localStream_);
 262     }
 263
 264     // ensure non-local stream waits for local stream, due to dependence on
 265     // the previous H2D copy of CPU forces (if accumulateForces is true)
 266     // or the above clearing.
 267     // TODO remove this dependency on localStream - edmine issue #3093
 268     GpuEventSynchronizer eventLocal;
 269     eventLocal.markEvent(localStream_);
 270     eventLocal.enqueueWaitEvent(nonLocalStream_);
 271
 272     //Unpack halo buffer into force array
 273
 274     KernelLaunchConfig config;
 275     config.blockSize[0]     = c_threadsPerBlock;
 276     config.blockSize[1]     = 1;
 277     config.blockSize[2]     = 1;
 278     config.gridSize[0]      = (fRecvSize_+c_threadsPerBlock-1)/c_threadsPerBlock;
 279     config.gridSize[1]      = 1;
 280     config.gridSize[2]      = 1;
 281     config.sharedMemorySize = 0;
 282     config.stream           = nonLocalStream_;
 283
 284     const float3    *recvBuf    = d_recvBuf_;
 285     const int       *indexMap   = d_indexMap_;
 286     const int        size       = fRecvSize_;
 287
 288     if (size > 0)
 289     {
 290         auto             kernelFn = accumulateForces ? unpackRecvBufKernel<true> : unpackRecvBufKernel<false>;
 291
 292         const auto       kernelArgs   = prepareGpuKernelArguments(kernelFn, config, &d_f,
 293                                                                   &recvBuf, &indexMap,
 294                                                                   &size);
 295
 296         launchGpuKernel(kernelFn, config, nullptr, "Domdec GPU Apply F Halo Exchange", kernelArgs);
 297     }
 298 }
 299
 300
 301 void GpuHaloExchange::Impl::communicateHaloData(float3     * d_ptr,
 302                                                 HaloQuantity haloQuantity)
 303 {
 304
 305     void * sendPtr;
 306     int    sendSize;
 307     void * remotePtr;
 308     int    sendRank;
 309     int    recvRank;
 310
 311     if (haloQuantity == HaloQuantity::HaloCoordinates)
 312     {
 313         sendPtr   = static_cast<void*> (d_sendBuf_);
 314         sendSize  = xSendSize_;
 315         remotePtr = remoteXPtr_;
 316         sendRank  = sendRankX_;
 317         recvRank  = recvRankX_;
 318
 319 #if GMX_MPI
 320         //Wait for signal from receiving task that it is ready, and similarly send signal to task that will push data to this task
 321         char thisTaskIsReady, remoteTaskIsReady;
 322         MPI_Sendrecv(&thisTaskIsReady, sizeof(char), MPI_BYTE, recvRank, 0,
 323                      &remoteTaskIsReady, sizeof(char), MPI_BYTE, sendRank, 0,
 324                      mpi_comm_mysim_, MPI_STATUS_IGNORE);
 325 #endif
 326     }
 327     else
 328     {
 329         sendPtr   = static_cast<void*> (&(d_ptr[numHomeAtoms_]));
 330         sendSize  = fSendSize_;
 331         remotePtr = remoteFPtr_;
 332         sendRank  = sendRankF_;
 333         recvRank  = recvRankF_;
 334     }
 335
 336     communicateHaloDataWithCudaDirect(sendPtr, sendSize, sendRank, remotePtr, recvRank);
 337 }
 338
 339 void GpuHaloExchange::Impl::communicateHaloDataWithCudaDirect(void *sendPtr,
 340                                                               int   sendSize,
 341                                                               int   sendRank,
 342                                                               void *remotePtr,
 343                                                               int   recvRank)
 344 {
 345
 346     cudaError_t  stat;
 347     cudaStream_t stream = nonLocalStream_;
 348
 349     // We asynchronously push data to remote rank. The remote
 350     // destination pointer has already been set in the init fn.  We
 351     // don't need to worry about overwriting data the remote ranks
 352     // still needs since the halo exchange is just done once per
 353     // timestep, for each of X and F.
 354
 355     // send data to neighbor, if any data exists to send
 356     if (sendSize > 0)
 357     {
 358         stat = cudaMemcpyAsync(remotePtr, sendPtr, sendSize*DIM*sizeof(float), cudaMemcpyDeviceToDevice, stream);
 359         CU_RET_ERR(stat, "cudaMemcpyAsync on GPU Domdec CUDA direct data transfer failed");
 360     }
 361
 362 #if GMX_MPI
 363     //ensure pushed data has arrived before remote rank progresses
 364     // This rank records an event and sends it to the remote rank which has just been pushed data.
 365     // This rank recieves event from remote rank which has pushed data here, and enqueues that event to
 366     // its stream.
 367     GpuEventSynchronizer *haloDataTransferRemote;
 368
 369     haloDataTransferLaunched_->markEvent(stream);
 370
 371     MPI_Sendrecv(&haloDataTransferLaunched_, sizeof(GpuEventSynchronizer*), MPI_BYTE, sendRank, 0,
 372                  &haloDataTransferRemote, sizeof(GpuEventSynchronizer*), MPI_BYTE, recvRank, 0,
 373                  mpi_comm_mysim_, MPI_STATUS_IGNORE);
 374
 375     haloDataTransferRemote->enqueueWaitEvent(stream);
 376 #else
 377     GMX_UNUSED_VALUE(sendRank);
 378     GMX_UNUSED_VALUE(recvRank);
 379 #endif
 380
 381 }
 382
 383 /*! \brief Create Domdec GPU object */
 384 GpuHaloExchange::Impl::Impl(gmx_domdec_t *dd,
 385                             MPI_Comm      mpi_comm_mysim,
 386                             void        * localStream,
 387                             void        * nonLocalStream)
 388     : dd_(dd),
 389       sendRankX_(dd->neighbor[0][1]),
 390       recvRankX_(dd->neighbor[0][0]),
 391       sendRankF_(dd->neighbor[0][0]),
 392       recvRankF_(dd->neighbor[0][1]),
 393       usePBC_(dd->ci[dd->dim[0]] == 0),
 394       haloDataTransferLaunched_(new GpuEventSynchronizer()),
 395       mpi_comm_mysim_(mpi_comm_mysim),
 396       localStream_(*static_cast<cudaStream_t*> (localStream)),
 397       nonLocalStream_(*static_cast<cudaStream_t*> (nonLocalStream))
 398 {
 399
 400     GMX_RELEASE_ASSERT(GMX_THREAD_MPI, "GPU Halo exchange is currently only supported with thread-MPI enabled");
 401
 402     if (dd->ndim > 1)
 403     {
 404         gmx_fatal(FARGS, "Error: dd->ndim > 1 is not yet supported in GPU halo exchange");
 405     }
 406
 407     if (usePBC_ && dd->unitCellInfo.haveScrewPBC)
 408     {
 409         gmx_fatal(FARGS, "Error: screw is not yet supported in GPU halo exchange\n");
 410     }
 411
 412     changePinningPolicy(&h_indexMap_, gmx::PinningPolicy::PinnedIfSupported);
 413
 414     allocateDeviceBuffer(&d_fShift_, 1, nullptr);
 415
 416 }
 417
 418 GpuHaloExchange::Impl::~Impl()
 419 {
 420     freeDeviceBuffer(&d_indexMap_);
 421     freeDeviceBuffer(&d_sendBuf_);
 422     freeDeviceBuffer(&d_recvBuf_);
 423     freeDeviceBuffer(&d_fShift_);
 424     delete haloDataTransferLaunched_;
 425 }
 426
 427 GpuHaloExchange::GpuHaloExchange(gmx_domdec_t *dd,
 428                                  MPI_Comm      mpi_comm_mysim,
 429                                  void         *localStream,
 430                                  void         *nonLocalStream)
 431     : impl_(new Impl(dd, mpi_comm_mysim, localStream, nonLocalStream))
 432 {
 433 }
 434
 435 GpuHaloExchange::~GpuHaloExchange() = default;
 436
 437 void GpuHaloExchange::reinitHalo(DeviceBuffer<float>  d_coordinatesBuffer,
 438                                  DeviceBuffer<float>  d_forcesBuffer)
 439 {
 440     impl_->reinitHalo(reinterpret_cast<float3*>(d_coordinatesBuffer), reinterpret_cast<float3*>(d_forcesBuffer));
 441 }
 442
 443 void GpuHaloExchange::communicateHaloCoordinates(const matrix box, GpuEventSynchronizer *coordinatesReadyOnDeviceEvent)
 444 {
 445     impl_->communicateHaloCoordinates(box, coordinatesReadyOnDeviceEvent);
 446 }
 447
 448 void GpuHaloExchange::communicateHaloForces(bool accumulateForces)
 449 {
 450     impl_->communicateHaloForces(accumulateForces);
 451 }
 452
 453 } //namespace gmx