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.cuh"
  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     }
  95
  96 #endif
  97 }
  98
  99 // TODO make this asynchronous by splitting into this into
 100 // launchRecvForceFromPmeCudaDirect() and sycnRecvForceFromPmeCudaDirect()
 101 void PmePpCommGpu::Impl::receiveForceFromPmeCudaDirect(bool receivePmeForceToGpu)
 102 {
 103 #if GMX_MPI
 104     // Remote PME task pushes GPU data directly data to this PP task.
 105
 106     // Recieve event from PME task after PME->PP force data push has
 107     // been scheduled and enqueue this to PP stream.
 108     GpuEventSynchronizer* eventptr;
 109     // NOLINTNEXTLINE(bugprone-sizeof-expression)
 110     MPI_Recv(&eventptr, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
 111     eventptr->enqueueWaitEvent(pmePpCommStream_);
 112
 113     if (receivePmeForceToGpu)
 114     {
 115         // Record event to be enqueued in the GPU local buffer operations, to
 116         // satisfy dependency on receiving the PME force data before
 117         // reducing it with the other force contributions.
 118         forcesReadySynchronizer_.markEvent(pmePpCommStream_);
 119     }
 120     else
 121     {
 122         // Ensure CPU waits for PME forces to be copied before reducing
 123         // them with other forces on the CPU
 124         pmePpCommStream_.synchronize();
 125     }
 126 #endif
 127 }
 128
 129 void PmePpCommGpu::Impl::receiveForceFromPmeCudaMpi(float3* pmeForcePtr, int recvSize)
 130 {
 131 #if GMX_MPI
 132     MPI_Recv(pmeForcePtr, recvSize * DIM, MPI_FLOAT, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
 133 #else
 134     GMX_UNUSED_VALUE(pmeForcePtr);
 135     GMX_UNUSED_VALUE(recvSize);
 136 #endif
 137 }
 138
 139 void PmePpCommGpu::Impl::receiveForceFromPme(float3* recvPtr, int recvSize, bool receivePmeForceToGpu)
 140 {
 141     float3* pmeForcePtr = receivePmeForceToGpu ? asFloat3(d_pmeForces_) : recvPtr;
 142     if (GMX_THREAD_MPI)
 143     {
 144         receiveForceFromPmeCudaDirect(receivePmeForceToGpu);
 145     }
 146     else
 147     {
 148         receiveForceFromPmeCudaMpi(pmeForcePtr, recvSize);
 149     }
 150 }
 151
 152 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaDirect(float3*               sendPtr,
 153                                                         int                   sendSize,
 154                                                         GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 155 {
 156     // ensure stream waits until coordinate data is available on device
 157     coordinatesReadyOnDeviceEvent->enqueueWaitEvent(pmePpCommStream_);
 158
 159     cudaError_t stat = cudaMemcpyAsync(remotePmeXBuffer_,
 160                                        sendPtr,
 161                                        sendSize * DIM * sizeof(float),
 162                                        cudaMemcpyDefault,
 163                                        pmePpCommStream_.stream());
 164     CU_RET_ERR(stat, "cudaMemcpyAsync on Send to PME CUDA direct data transfer failed");
 165
 166 #if GMX_MPI
 167     // Record and send event to allow PME task to sync to above transfer before commencing force calculations
 168     pmeCoordinatesSynchronizer_.markEvent(pmePpCommStream_);
 169     GpuEventSynchronizer* pmeSync = &pmeCoordinatesSynchronizer_;
 170     // NOLINTNEXTLINE(bugprone-sizeof-expression)
 171     MPI_Send(&pmeSync, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_);
 172 #endif
 173 }
 174
 175 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaMpi(float3*               sendPtr,
 176                                                      int                   sendSize,
 177                                                      GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 178 {
 179     // ensure coordinate data is available on device before we start transfer
 180     coordinatesReadyOnDeviceEvent->waitForEvent();
 181
 182 #if GMX_MPI
 183     float3* sendptr_x = sendPtr;
 184
 185     MPI_Send(sendptr_x, sendSize * DIM, MPI_FLOAT, pmeRank_, eCommType_COORD_GPU, comm_);
 186 #else
 187     GMX_UNUSED_VALUE(sendPtr);
 188     GMX_UNUSED_VALUE(sendSize);
 189 #endif
 190 }
 191
 192 void PmePpCommGpu::Impl::sendCoordinatesToPme(float3*               sendPtr,
 193                                               int                   sendSize,
 194                                               GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 195 {
 196     if (GMX_THREAD_MPI)
 197     {
 198         sendCoordinatesToPmeCudaDirect(sendPtr, sendSize, coordinatesReadyOnDeviceEvent);
 199     }
 200     else
 201     {
 202         sendCoordinatesToPmeCudaMpi(sendPtr, sendSize, coordinatesReadyOnDeviceEvent);
 203     }
 204 }
 205 DeviceBuffer<Float3> PmePpCommGpu::Impl::getGpuForceStagingPtr()
 206 {
 207     return d_pmeForces_;
 208 }
 209
 210 GpuEventSynchronizer* PmePpCommGpu::Impl::getForcesReadySynchronizer()
 211 {
 212     if (GMX_THREAD_MPI)
 213     {
 214         return &forcesReadySynchronizer_;
 215     }
 216     else
 217     {
 218         return nullptr;
 219     }
 220 }
 221
 222 PmePpCommGpu::PmePpCommGpu(MPI_Comm                comm,
 223                            int                     pmeRank,
 224                            std::vector<gmx::RVec>* pmeCpuForceBuffer,
 225                            const DeviceContext&    deviceContext,
 226                            const DeviceStream&     deviceStream) :
 227     impl_(new Impl(comm, pmeRank, pmeCpuForceBuffer, deviceContext, deviceStream))
 228 {
 229 }
 230
 231 PmePpCommGpu::~PmePpCommGpu() = default;
 232
 233 void PmePpCommGpu::reinit(int size)
 234 {
 235     impl_->reinit(size);
 236 }
 237
 238 void PmePpCommGpu::receiveForceFromPme(RVec* recvPtr, int recvSize, bool receivePmeForceToGpu)
 239 {
 240     impl_->receiveForceFromPme(asFloat3(recvPtr), recvSize, receivePmeForceToGpu);
 241 }
 242
 243 void PmePpCommGpu::sendCoordinatesToPmeFromGpu(DeviceBuffer<RVec>    sendPtr,
 244                                                int                   sendSize,
 245                                                GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 246 {
 247     impl_->sendCoordinatesToPme(asFloat3(sendPtr), sendSize, coordinatesReadyOnDeviceEvent);
 248 }
 249
 250 void PmePpCommGpu::sendCoordinatesToPmeFromCpu(RVec*                 sendPtr,
 251                                                int                   sendSize,
 252                                                GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 253 {
 254     impl_->sendCoordinatesToPme(asFloat3(sendPtr), sendSize, coordinatesReadyOnDeviceEvent);
 255 }
 256
 257 DeviceBuffer<Float3> PmePpCommGpu::getGpuForceStagingPtr()
 258 {
 259     return impl_->getGpuForceStagingPtr();
 260 }
 261
 262 GpuEventSynchronizer* PmePpCommGpu::getForcesReadySynchronizer()
 263 {
 264     return impl_->getForcesReadySynchronizer();
 265 }
 266
 267 } // namespace gmx