src/gromacs/ewald/pme_pp_comm_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 "gromacs/ewald/pme_pp_communication.h"
  47 #include "pme_pp_comm_gpu_impl.h"
  48
  49 #include "config.h"
  50
  51 #include "gromacs/gpu_utils/cudautils.cuh"
  52 #include "gromacs/gpu_utils/device_context.h"
  53 #include "gromacs/gpu_utils/device_stream.h"
  54 #include "gromacs/gpu_utils/devicebuffer.h"
  55 #include "gromacs/gpu_utils/gpueventsynchronizer.h"
  56 #include "gromacs/gpu_utils/typecasts.cuh"
  57 #include "gromacs/utility/gmxmpi.h"
  58
  59 namespace gmx
  60 {
  61
  62 PmePpCommGpu::Impl::Impl(MPI_Comm                comm,
  63                          int                     pmeRank,
  64                          std::vector<gmx::RVec>* pmeCpuForceBuffer,
  65                          const DeviceContext&    deviceContext,
  66                          const DeviceStream&     deviceStream) :
  67     deviceContext_(deviceContext),
  68     pmePpCommStream_(deviceStream),
  69     comm_(comm),
  70     pmeRank_(pmeRank),
  71     pmeCpuForceBuffer_(pmeCpuForceBuffer),
  72     d_pmeForces_(nullptr)
  73 {
  74 }
  75
  76 PmePpCommGpu::Impl::~Impl() = default;
  77
  78 void PmePpCommGpu::Impl::reinit(int size)
  79 {
  80     // Reallocate device buffer used for staging PME force
  81     reallocateDeviceBuffer(&d_pmeForces_, size, &d_pmeForcesSize_, &d_pmeForcesSizeAlloc_, deviceContext_);
  82
  83     // This rank will access PME rank memory directly, so needs to receive the remote PME buffer addresses.
  84 #if GMX_MPI
  85
  86     if (GMX_THREAD_MPI)
  87     {
  88         // receive device coordinate buffer address from PME rank
  89         MPI_Recv(&remotePmeXBuffer_, sizeof(float3*), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
  90         // send host and device force buffer addresses to PME rank
  91         MPI_Send(&d_pmeForces_, sizeof(float3*), MPI_BYTE, pmeRank_, 0, comm_);
  92         RVec* pmeCpuForceBufferData = pmeCpuForceBuffer_->data();
  93         MPI_Send(&pmeCpuForceBufferData, sizeof(RVec*), MPI_BYTE, pmeRank_, 0, comm_);
  94         // Receive address of event and associated flag from PME rank, to allow sync to local stream after force transfer
  95         MPI_Recv(&remotePmeForceSendEvent_, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
  96         MPI_Recv(&remotePmeForceSendEventRecorded_, sizeof(std::atomic<bool>*), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
  97     }
  98
  99 #endif
 100 }
 101
 102 void PmePpCommGpu::Impl::receiveForceFromPmeCudaDirect(bool receivePmeForceToGpu)
 103 {
 104 #if GMX_MPI
 105     // Wait until remote PME task has pushed data, and then enqueue remote event to local stream.
 106
 107     // Spin until PME rank sets flag
 108     while (!(remotePmeForceSendEventRecorded_->load(std::memory_order_acquire))) {};
 109
 110     // Enqueue remote event
 111     remotePmeForceSendEvent_->enqueueWaitEvent(pmePpCommStream_);
 112
 113     // Reset the flag
 114     remotePmeForceSendEventRecorded_->store(false, std::memory_order_release);
 115
 116     if (receivePmeForceToGpu)
 117     {
 118         // Record event to be enqueued in the GPU local buffer operations, to
 119         // satisfy dependency on receiving the PME force data before
 120         // reducing it with the other force contributions.
 121         forcesReadySynchronizer_.markEvent(pmePpCommStream_);
 122     }
 123     else
 124     {
 125         // Ensure CPU waits for PME forces to be copied before reducing
 126         // them with other forces on the CPU
 127         pmePpCommStream_.synchronize();
 128     }
 129 #endif
 130 }
 131
 132 void PmePpCommGpu::Impl::receiveForceFromPmeCudaMpi(float3* pmeForcePtr, int recvSize)
 133 {
 134 #if GMX_MPI
 135     MPI_Recv(pmeForcePtr, recvSize * DIM, MPI_FLOAT, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
 136 #else
 137     GMX_UNUSED_VALUE(pmeForcePtr);
 138     GMX_UNUSED_VALUE(recvSize);
 139 #endif
 140 }
 141
 142 void PmePpCommGpu::Impl::receiveForceFromPme(float3* recvPtr, int recvSize, bool receivePmeForceToGpu)
 143 {
 144     float3* pmeForcePtr = receivePmeForceToGpu ? asFloat3(d_pmeForces_) : recvPtr;
 145     if (GMX_THREAD_MPI)
 146     {
 147         receiveForceFromPmeCudaDirect(receivePmeForceToGpu);
 148     }
 149     else
 150     {
 151         receiveForceFromPmeCudaMpi(pmeForcePtr, recvSize);
 152     }
 153 }
 154
 155 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaDirect(float3*               sendPtr,
 156                                                         int                   sendSize,
 157                                                         GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 158 {
 159     // ensure stream waits until coordinate data is available on device
 160     coordinatesReadyOnDeviceEvent->enqueueWaitEvent(pmePpCommStream_);
 161
 162     cudaError_t stat = cudaMemcpyAsync(remotePmeXBuffer_,
 163                                        sendPtr,
 164                                        sendSize * DIM * sizeof(float),
 165                                        cudaMemcpyDefault,
 166                                        pmePpCommStream_.stream());
 167     CU_RET_ERR(stat, "cudaMemcpyAsync on Send to PME CUDA direct data transfer failed");
 168
 169 #if GMX_MPI
 170     // Record and send event to allow PME task to sync to above transfer before commencing force calculations
 171     pmeCoordinatesSynchronizer_.markEvent(pmePpCommStream_);
 172     GpuEventSynchronizer* pmeSync = &pmeCoordinatesSynchronizer_;
 173     // NOLINTNEXTLINE(bugprone-sizeof-expression)
 174     MPI_Send(&pmeSync, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_);
 175 #endif
 176 }
 177
 178 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaMpi(float3*               sendPtr,
 179                                                      int                   sendSize,
 180                                                      GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 181 {
 182     // ensure coordinate data is available on device before we start transfer
 183     coordinatesReadyOnDeviceEvent->waitForEvent();
 184
 185 #if GMX_MPI
 186     float3* sendptr_x = sendPtr;
 187
 188     MPI_Send(sendptr_x, sendSize * DIM, MPI_FLOAT, pmeRank_, eCommType_COORD_GPU, comm_);
 189 #else
 190     GMX_UNUSED_VALUE(sendPtr);
 191     GMX_UNUSED_VALUE(sendSize);
 192 #endif
 193 }
 194
 195 void PmePpCommGpu::Impl::sendCoordinatesToPme(float3*               sendPtr,
 196                                               int                   sendSize,
 197                                               GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 198 {
 199     if (GMX_THREAD_MPI)
 200     {
 201         sendCoordinatesToPmeCudaDirect(sendPtr, sendSize, coordinatesReadyOnDeviceEvent);
 202     }
 203     else
 204     {
 205         sendCoordinatesToPmeCudaMpi(sendPtr, sendSize, coordinatesReadyOnDeviceEvent);
 206     }
 207 }
 208 DeviceBuffer<Float3> PmePpCommGpu::Impl::getGpuForceStagingPtr()
 209 {
 210     return d_pmeForces_;
 211 }
 212
 213 GpuEventSynchronizer* PmePpCommGpu::Impl::getForcesReadySynchronizer()
 214 {
 215     if (GMX_THREAD_MPI)
 216     {
 217         return &forcesReadySynchronizer_;
 218     }
 219     else
 220     {
 221         return nullptr;
 222     }
 223 }
 224
 225 PmePpCommGpu::PmePpCommGpu(MPI_Comm                comm,
 226                            int                     pmeRank,
 227                            std::vector<gmx::RVec>* pmeCpuForceBuffer,
 228                            const DeviceContext&    deviceContext,
 229                            const DeviceStream&     deviceStream) :
 230     impl_(new Impl(comm, pmeRank, pmeCpuForceBuffer, deviceContext, deviceStream))
 231 {
 232 }
 233
 234 PmePpCommGpu::~PmePpCommGpu() = default;
 235
 236 void PmePpCommGpu::reinit(int size)
 237 {
 238     impl_->reinit(size);
 239 }
 240
 241 void PmePpCommGpu::receiveForceFromPme(RVec* recvPtr, int recvSize, bool receivePmeForceToGpu)
 242 {
 243     impl_->receiveForceFromPme(asFloat3(recvPtr), recvSize, receivePmeForceToGpu);
 244 }
 245
 246 void PmePpCommGpu::sendCoordinatesToPmeFromGpu(DeviceBuffer<RVec>    sendPtr,
 247                                                int                   sendSize,
 248                                                GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 249 {
 250     impl_->sendCoordinatesToPme(asFloat3(sendPtr), sendSize, coordinatesReadyOnDeviceEvent);
 251 }
 252
 253 void PmePpCommGpu::sendCoordinatesToPmeFromCpu(RVec*                 sendPtr,
 254                                                int                   sendSize,
 255                                                GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 256 {
 257     impl_->sendCoordinatesToPme(asFloat3(sendPtr), sendSize, coordinatesReadyOnDeviceEvent);
 258 }
 259
 260 DeviceBuffer<Float3> PmePpCommGpu::getGpuForceStagingPtr()
 261 {
 262     return impl_->getGpuForceStagingPtr();
 263 }
 264
 265 GpuEventSynchronizer* PmePpCommGpu::getForcesReadySynchronizer()
 266 {
 267     return impl_->getForcesReadySynchronizer();
 268 }
 269
 270 } // namespace gmx