src/gromacs/ewald/pme_force_sender_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 PME-PP communication using CUDA
  38  *
  39  *
  40  * \author Alan Gray <alang@nvidia.com>
  41  *
  42  * \ingroup module_ewald
  43  */
  44 #include "gmxpre.h"
  45
  46 #include "pme_force_sender_gpu_impl.h"
  47
  48 #include "config.h"
  49
  50 #include "gromacs/gpu_utils/cudautils.cuh"
  51 #include "gromacs/gpu_utils/gpueventsynchronizer.h"
  52 #include "gromacs/utility/gmxmpi.h"
  53
  54 namespace gmx
  55 {
  56
  57 /*! \brief Create PME-PP GPU communication object */
  58 PmeForceSenderGpu::Impl::Impl(GpuEventSynchronizer*  pmeForcesReady,
  59                               MPI_Comm               comm,
  60                               const DeviceContext&   deviceContext,
  61                               gmx::ArrayRef<PpRanks> ppRanks) :
  62     pmeForcesReady_(pmeForcesReady),
  63     comm_(comm),
  64     ppRanks_(ppRanks),
  65     ppCommStream_(ppRanks.size()),
  66     ppCommEvent_(ppRanks.size()),
  67     ppCommEventRecorded_(ppRanks.size()),
  68     deviceContext_(deviceContext),
  69     pmeRemoteCpuForcePtr_(ppRanks.size()),
  70     pmeRemoteGpuForcePtr_(ppRanks.size())
  71 {
  72     // Create streams and events to manage pushing of force buffers to remote PP ranks
  73     std::unique_ptr<DeviceStream>         stream;
  74     std::unique_ptr<GpuEventSynchronizer> event;
  75     size_t                                i = 0;
  76     for (i = 0; i < ppRanks_.size(); i++)
  77     {
  78         stream = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::High, false);
  79         ppCommStream_[i] = std::move(stream);
  80         event            = std::make_unique<GpuEventSynchronizer>();
  81         ppCommEvent_[i]  = std::move(event);
  82     }
  83 }
  84
  85 PmeForceSenderGpu::Impl::~Impl() = default;
  86
  87 /*! \brief Sets location of force to be sent to each PP rank  */
  88 void PmeForceSenderGpu::Impl::setForceSendBuffer(DeviceBuffer<Float3> d_f)
  89 {
  90
  91     // Need to send address to PP rank only for thread-MPI as PP rank pulls
  92     // data using cudamemcpy
  93     if (!GMX_THREAD_MPI)
  94     {
  95         return;
  96     }
  97
  98 #if GMX_MPI
  99
 100     if (localForcePtr_.empty())
 101     {
 102         localForcePtr_.resize(ppRanks_.size());
 103     }
 104     int ind_start = 0;
 105     int ind_end   = 0;
 106     int i         = 0;
 107     for (const auto& receiver : ppRanks_)
 108     {
 109         ind_start = ind_end;
 110         ind_end   = ind_start + receiver.numAtoms;
 111
 112         localForcePtr_[i] = &d_f[ind_start];
 113         // NOLINTNEXTLINE(bugprone-sizeof-expression)
 114         MPI_Recv(&pmeRemoteGpuForcePtr_[i], sizeof(float3*), MPI_BYTE, receiver.rankId, 0, comm_, MPI_STATUS_IGNORE);
 115         // NOLINTNEXTLINE(bugprone-sizeof-expression)
 116         MPI_Recv(&pmeRemoteCpuForcePtr_[i], sizeof(float3*), MPI_BYTE, receiver.rankId, 0, comm_, MPI_STATUS_IGNORE);
 117         // Send address of event and associated flag to PP rank, to allow remote enqueueing
 118         // NOLINTNEXTLINE(bugprone-sizeof-expression)
 119         MPI_Send(&ppCommEvent_[i], sizeof(GpuEventSynchronizer*), MPI_BYTE, receiver.rankId, 0, comm_);
 120
 121         std::atomic<bool>* tmpPpCommEventRecordedPtr =
 122                 reinterpret_cast<std::atomic<bool>*>(&(ppCommEventRecorded_[i]));
 123         tmpPpCommEventRecordedPtr->store(false, std::memory_order_release);
 124         // NOLINTNEXTLINE(bugprone-sizeof-expression)
 125         MPI_Send(&tmpPpCommEventRecordedPtr, sizeof(std::atomic<bool>*), MPI_BYTE, receiver.rankId, 0, comm_);
 126         i++;
 127     }
 128
 129 #else
 130     GMX_UNUSED_VALUE(d_f);
 131 #endif
 132 }
 133
 134
 135 /*! \brief Send PME synchronizer directly using CUDA memory copy */
 136 void PmeForceSenderGpu::Impl::sendFToPpCudaDirect(int ppRank, int numAtoms, bool sendForcesDirectToPpGpu)
 137 {
 138
 139     GMX_ASSERT(GMX_THREAD_MPI, "sendFToPpCudaDirect is expected to be called only for Thread-MPI");
 140
 141
 142 #if GMX_MPI
 143     float3* pmeRemoteForcePtr =
 144             sendForcesDirectToPpGpu ? pmeRemoteGpuForcePtr_[ppRank] : pmeRemoteCpuForcePtr_[ppRank];
 145
 146     pmeForcesReady_->enqueueWaitEvent(*ppCommStream_[ppRank]);
 147
 148     cudaError_t stat = cudaMemcpyAsync(pmeRemoteForcePtr,
 149                                        localForcePtr_[ppRank],
 150                                        numAtoms * sizeof(rvec),
 151                                        cudaMemcpyDefault,
 152                                        ppCommStream_[ppRank]->stream());
 153     CU_RET_ERR(stat, "cudaMemcpyAsync on Recv from PME CUDA direct data transfer failed");
 154     ppCommEvent_[ppRank]->markEvent(*ppCommStream_[ppRank]);
 155     std::atomic<bool>* tmpPpCommEventRecordedPtr =
 156             reinterpret_cast<std::atomic<bool>*>(&(ppCommEventRecorded_[ppRank]));
 157     tmpPpCommEventRecordedPtr->store(true, std::memory_order_release);
 158 #else
 159     GMX_UNUSED_VALUE(ppRank);
 160     GMX_UNUSED_VALUE(numAtoms);
 161 #endif
 162 }
 163
 164 /*! \brief Send PME data directly using CUDA-aware MPI */
 165 void PmeForceSenderGpu::Impl::sendFToPpCudaMpi(DeviceBuffer<RVec> sendbuf,
 166                                                int                offset,
 167                                                int                numBytes,
 168                                                int                ppRank,
 169                                                MPI_Request*       request)
 170 {
 171     GMX_ASSERT(GMX_LIB_MPI, "sendFToPpCudaMpi is expected to be called only for Lib-MPI");
 172
 173 #if GMX_MPI
 174     // if using GPU direct comm with CUDA-aware MPI, make sure forces are ready on device
 175     // before sending it to PP ranks
 176     pmeForcesReady_->waitForEvent();
 177
 178     MPI_Isend(sendbuf[offset], numBytes, MPI_BYTE, ppRank, 0, comm_, request);
 179
 180 #else
 181     GMX_UNUSED_VALUE(sendbuf);
 182     GMX_UNUSED_VALUE(offset);
 183     GMX_UNUSED_VALUE(numBytes);
 184     GMX_UNUSED_VALUE(ppRank);
 185     GMX_UNUSED_VALUE(request);
 186 #endif
 187 }
 188
 189 PmeForceSenderGpu::PmeForceSenderGpu(GpuEventSynchronizer*  pmeForcesReady,
 190                                      MPI_Comm               comm,
 191                                      const DeviceContext&   deviceContext,
 192                                      gmx::ArrayRef<PpRanks> ppRanks) :
 193     impl_(new Impl(pmeForcesReady, comm, deviceContext, ppRanks))
 194 {
 195 }
 196
 197 PmeForceSenderGpu::~PmeForceSenderGpu() = default;
 198
 199
 200 void PmeForceSenderGpu::setForceSendBuffer(DeviceBuffer<RVec> d_f)
 201 {
 202     impl_->setForceSendBuffer(d_f);
 203 }
 204
 205 void PmeForceSenderGpu::sendFToPpCudaMpi(DeviceBuffer<RVec> sendbuf,
 206                                          int                offset,
 207                                          int                numBytes,
 208                                          int                ppRank,
 209                                          MPI_Request*       request)
 210 {
 211     impl_->sendFToPpCudaMpi(sendbuf, offset, numBytes, ppRank, request);
 212 }
 213
 214 void PmeForceSenderGpu::sendFToPpCudaDirect(int ppRank, int numAtoms, bool sendForcesDirectToPpGpu)
 215 {
 216     impl_->sendFToPpCudaDirect(ppRank, numAtoms, sendForcesDirectToPpGpu);
 217 }
 218
 219
 220 } // namespace gmx