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 "pme_pp_comm_gpu_impl.h"
  47
  48 #include "config.h"
  49
  50 #include "gromacs/gpu_utils/cudautils.cuh"
  51 #include "gromacs/gpu_utils/device_context.h"
  52 #include "gromacs/gpu_utils/device_stream.h"
  53 #include "gromacs/gpu_utils/devicebuffer.h"
  54 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
  55 #include "gromacs/utility/gmxmpi.h"
  56
  57 namespace gmx
  58 {
  59
  60 PmePpCommGpu::Impl::Impl(MPI_Comm             comm,
  61                          int                  pmeRank,
  62                          const DeviceContext& deviceContext,
  63                          const DeviceStream&  deviceStream) :
  64     deviceContext_(deviceContext),
  65     pmePpCommStream_(deviceStream),
  66     comm_(comm),
  67     pmeRank_(pmeRank),
  68     d_pmeForces_(nullptr)
  69 {
  70     GMX_RELEASE_ASSERT(
  71             GMX_THREAD_MPI,
  72             "PME-PP GPU Communication is currently only supported with thread-MPI enabled");
  73 }
  74
  75 PmePpCommGpu::Impl::~Impl() = default;
  76
  77 void PmePpCommGpu::Impl::reinit(int size)
  78 {
  79     // This rank will access PME rank memory directly, so needs to receive the remote PME buffer addresses.
  80 #if GMX_MPI
  81     MPI_Recv(&remotePmeXBuffer_, sizeof(void**), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
  82     MPI_Recv(&remotePmeFBuffer_, sizeof(void**), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
  83
  84     // Reallocate buffer used for staging PME force on GPU
  85     reallocateDeviceBuffer(&d_pmeForces_, size, &d_pmeForcesSize_, &d_pmeForcesSizeAlloc_, deviceContext_);
  86 #else
  87     GMX_UNUSED_VALUE(size);
  88 #endif
  89     return;
  90 }
  91
  92 // TODO make this asynchronous by splitting into this into
  93 // launchRecvForceFromPmeCudaDirect() and sycnRecvForceFromPmeCudaDirect()
  94 void PmePpCommGpu::Impl::receiveForceFromPmeCudaDirect(void* recvPtr, int recvSize, bool receivePmeForceToGpu)
  95 {
  96 #if GMX_MPI
  97     // Receive event from PME task and add to stream, to ensure pull of data doesn't
  98     // occur before PME force calc is completed
  99     GpuEventSynchronizer* pmeSync;
 100     MPI_Recv(&pmeSync, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
 101     pmeSync->enqueueWaitEvent(pmePpCommStream_);
 102
 103     // Pull force data from remote GPU
 104     void*       pmeForcePtr = receivePmeForceToGpu ? static_cast<void*>(d_pmeForces_) : recvPtr;
 105     cudaError_t stat        = cudaMemcpyAsync(pmeForcePtr,
 106                                        remotePmeFBuffer_,
 107                                        recvSize * DIM * sizeof(float),
 108                                        cudaMemcpyDefault,
 109                                        pmePpCommStream_.stream());
 110     CU_RET_ERR(stat, "cudaMemcpyAsync on Recv from PME CUDA direct data transfer failed");
 111
 112     if (receivePmeForceToGpu)
 113     {
 114         // Record event to be enqueued in the GPU local buffer operations, to
 115         // satisfy dependency on receiving the PME force data before
 116         // reducing it with the other force contributions.
 117         forcesReadySynchronizer_.markEvent(pmePpCommStream_);
 118     }
 119     else
 120     {
 121         // Ensure CPU waits for PME forces to be copied before reducing
 122         // them with other forces on the CPU
 123         cudaStreamSynchronize(pmePpCommStream_.stream());
 124     }
 125 #else
 126     GMX_UNUSED_VALUE(recvPtr);
 127     GMX_UNUSED_VALUE(recvSize);
 128     GMX_UNUSED_VALUE(receivePmeForceToGpu);
 129 #endif
 130 }
 131
 132 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaDirect(void* sendPtr,
 133                                                         int   sendSize,
 134                                                         bool gmx_unused sendPmeCoordinatesFromGpu,
 135                                                         GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 136 {
 137 #if GMX_MPI
 138     // ensure stream waits until coordinate data is available on device
 139     coordinatesReadyOnDeviceEvent->enqueueWaitEvent(pmePpCommStream_);
 140
 141     cudaError_t stat = cudaMemcpyAsync(remotePmeXBuffer_,
 142                                        sendPtr,
 143                                        sendSize * DIM * sizeof(float),
 144                                        cudaMemcpyDefault,
 145                                        pmePpCommStream_.stream());
 146     CU_RET_ERR(stat, "cudaMemcpyAsync on Send to PME CUDA direct data transfer failed");
 147
 148     // Record and send event to allow PME task to sync to above transfer before commencing force calculations
 149     pmeCoordinatesSynchronizer_.markEvent(pmePpCommStream_);
 150     GpuEventSynchronizer* pmeSync = &pmeCoordinatesSynchronizer_;
 151     MPI_Send(&pmeSync, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_);
 152 #else
 153     GMX_UNUSED_VALUE(sendPtr);
 154     GMX_UNUSED_VALUE(sendSize);
 155     GMX_UNUSED_VALUE(sendPmeCoordinatesFromGpu);
 156     GMX_UNUSED_VALUE(coordinatesReadyOnDeviceEvent);
 157 #endif
 158 }
 159
 160 DeviceBuffer<Float3> PmePpCommGpu::Impl::getGpuForceStagingPtr()
 161 {
 162     return d_pmeForces_;
 163 }
 164
 165 GpuEventSynchronizer* PmePpCommGpu::Impl::getForcesReadySynchronizer()
 166 {
 167     return &forcesReadySynchronizer_;
 168 }
 169
 170 PmePpCommGpu::PmePpCommGpu(MPI_Comm             comm,
 171                            int                  pmeRank,
 172                            const DeviceContext& deviceContext,
 173                            const DeviceStream&  deviceStream) :
 174     impl_(new Impl(comm, pmeRank, deviceContext, deviceStream))
 175 {
 176 }
 177
 178 PmePpCommGpu::~PmePpCommGpu() = default;
 179
 180 void PmePpCommGpu::reinit(int size)
 181 {
 182     impl_->reinit(size);
 183 }
 184
 185 void PmePpCommGpu::receiveForceFromPmeCudaDirect(void* recvPtr, int recvSize, bool receivePmeForceToGpu)
 186 {
 187     impl_->receiveForceFromPmeCudaDirect(recvPtr, recvSize, receivePmeForceToGpu);
 188 }
 189
 190 void PmePpCommGpu::sendCoordinatesToPmeCudaDirect(void*                 sendPtr,
 191                                                   int                   sendSize,
 192                                                   bool                  sendPmeCoordinatesFromGpu,
 193                                                   GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
 194 {
 195     impl_->sendCoordinatesToPmeCudaDirect(
 196             sendPtr, sendSize, sendPmeCoordinatesFromGpu, coordinatesReadyOnDeviceEvent);
 197 }
 198
 199 DeviceBuffer<gmx::RVec> PmePpCommGpu::getGpuForceStagingPtr()
 200 {
 201     return impl_->getGpuForceStagingPtr();
 202 }
 203
 204 GpuEventSynchronizer* PmePpCommGpu::getForcesReadySynchronizer()
 205 {
 206     return impl_->getForcesReadySynchronizer();
 207 }
 208
 209 } // namespace gmx