src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cu

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2019,2020,2021, 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 class which recieves coordinates to GPU memory on PME task using CUDA
  38  *
  39  *
  40  * \author Alan Gray <alang@nvidia.com>
  41  *
  42  * \ingroup module_ewald
  43  */
  44 #include "gmxpre.h"
  45
  46 #include "gromacs/ewald/pme_pp_communication.h"
  47 #include "pme_coordinate_receiver_gpu_impl.h"
  48
  49 #include "config.h"
  50
  51 #include "gromacs/ewald/pme_force_sender_gpu.h"
  52 #include "gromacs/gpu_utils/cudautils.cuh"
  53 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
  54 #include "gromacs/utility/gmxmpi.h"
  55
  56 namespace gmx
  57 {
  58
  59 PmeCoordinateReceiverGpu::Impl::Impl(const DeviceStream&    pmeStream,
  60                                      MPI_Comm               comm,
  61                                      gmx::ArrayRef<PpRanks> ppRanks) :
  62     pmeStream_(pmeStream), comm_(comm), ppRanks_(ppRanks)
  63 {
  64     request_.resize(ppRanks.size());
  65     ppSync_.resize(ppRanks.size());
  66 }
  67
  68 PmeCoordinateReceiverGpu::Impl::~Impl() = default;
  69
  70 void PmeCoordinateReceiverGpu::Impl::sendCoordinateBufferAddressToPpRanks(DeviceBuffer<RVec> d_x)
  71 {
  72     // Need to send address to PP rank only for thread-MPI as PP rank pushes data using cudamemcpy
  73     if (GMX_THREAD_MPI)
  74     {
  75         int ind_start = 0;
  76         int ind_end   = 0;
  77         for (const auto& receiver : ppRanks_)
  78         {
  79             ind_start = ind_end;
  80             ind_end   = ind_start + receiver.numAtoms;
  81
  82             // Data will be transferred directly from GPU.
  83             void* sendBuf = reinterpret_cast<void*>(&d_x[ind_start]);
  84 #if GMX_MPI
  85             MPI_Send(&sendBuf, sizeof(void**), MPI_BYTE, receiver.rankId, 0, comm_);
  86 #else
  87             GMX_UNUSED_VALUE(sendBuf);
  88 #endif
  89         }
  90     }
  91 }
  92
  93 /*! \brief Receive coordinate synchronizer pointer from the PP ranks. */
  94 void PmeCoordinateReceiverGpu::Impl::receiveCoordinatesSynchronizerFromPpCudaDirect(int ppRank)
  95 {
  96     GMX_ASSERT(GMX_THREAD_MPI,
  97                "receiveCoordinatesSynchronizerFromPpCudaDirect is expected to be called only for "
  98                "Thread-MPI");
  99
 100     // Data will be pushed directly from PP task
 101
 102 #if GMX_MPI
 103     // Receive event from PP task
 104     MPI_Irecv(&ppSync_[recvCount_], sizeof(GpuEventSynchronizer*), MPI_BYTE, ppRank, 0, comm_, &request_[recvCount_]);
 105     recvCount_++;
 106 #else
 107     GMX_UNUSED_VALUE(ppRank);
 108 #endif
 109 }
 110
 111 /*! \brief Receive coordinate data using CUDA-aware MPI */
 112 void PmeCoordinateReceiverGpu::Impl::launchReceiveCoordinatesFromPpCudaMpi(DeviceBuffer<RVec> recvbuf,
 113                                                                            int numAtoms,
 114                                                                            int numBytes,
 115                                                                            int ppRank)
 116 {
 117     GMX_ASSERT(GMX_LIB_MPI,
 118                "launchReceiveCoordinatesFromPpCudaMpi is expected to be called only for Lib-MPI");
 119
 120 #if GMX_MPI
 121     MPI_Irecv(&recvbuf[numAtoms], numBytes, MPI_BYTE, ppRank, eCommType_COORD_GPU, comm_, &request_[recvCount_++]);
 122 #else
 123     GMX_UNUSED_VALUE(recvbuf);
 124     GMX_UNUSED_VALUE(numAtoms);
 125     GMX_UNUSED_VALUE(numBytes);
 126     GMX_UNUSED_VALUE(ppRank);
 127 #endif
 128 }
 129
 130 void PmeCoordinateReceiverGpu::Impl::synchronizeOnCoordinatesFromPpRanks()
 131 {
 132     if (recvCount_ > 0)
 133     {
 134         // ensure PME calculation doesn't commence until coordinate data/remote events
 135         // has been transferred
 136 #if GMX_MPI
 137         MPI_Waitall(recvCount_, request_.data(), MPI_STATUS_IGNORE);
 138 #endif
 139
 140         // Make PME stream wait on PP to PME data trasnfer events
 141         if (GMX_THREAD_MPI)
 142         {
 143             for (int i = 0; i < recvCount_; i++)
 144             {
 145                 ppSync_[i]->enqueueWaitEvent(pmeStream_);
 146             }
 147         }
 148
 149         // reset receive counter
 150         recvCount_ = 0;
 151     }
 152 }
 153
 154 PmeCoordinateReceiverGpu::PmeCoordinateReceiverGpu(const DeviceStream&    pmeStream,
 155                                                    MPI_Comm               comm,
 156                                                    gmx::ArrayRef<PpRanks> ppRanks) :
 157     impl_(new Impl(pmeStream, comm, ppRanks))
 158 {
 159 }
 160
 161 PmeCoordinateReceiverGpu::~PmeCoordinateReceiverGpu() = default;
 162
 163 void PmeCoordinateReceiverGpu::sendCoordinateBufferAddressToPpRanks(DeviceBuffer<RVec> d_x)
 164 {
 165     impl_->sendCoordinateBufferAddressToPpRanks(d_x);
 166 }
 167
 168 void PmeCoordinateReceiverGpu::receiveCoordinatesSynchronizerFromPpCudaDirect(int ppRank)
 169 {
 170     impl_->receiveCoordinatesSynchronizerFromPpCudaDirect(ppRank);
 171 }
 172
 173 void PmeCoordinateReceiverGpu::launchReceiveCoordinatesFromPpCudaMpi(DeviceBuffer<RVec> recvbuf,
 174                                                                      int                numAtoms,
 175                                                                      int                numBytes,
 176                                                                      int                ppRank)
 177 {
 178     impl_->launchReceiveCoordinatesFromPpCudaMpi(recvbuf, numAtoms, numBytes, ppRank);
 179 }
 180
 181 void PmeCoordinateReceiverGpu::synchronizeOnCoordinatesFromPpRanks()
 182 {
 183     impl_->synchronizeOnCoordinatesFromPpRanks();
 184 }
 185
 186 } // namespace gmx