2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019, 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.
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.
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.
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.
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.
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.
37 * \brief Implements PME-PP communication using CUDA
40 * \author Alan Gray <alang@nvidia.com>
42 * \ingroup module_ewald
46 #include "pme_pp_comm_gpu_impl.h"
50 #include "gromacs/gpu_utils/cudautils.cuh"
51 #include "gromacs/gpu_utils/devicebuffer.h"
52 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
53 #include "gromacs/utility/gmxmpi.h"
58 PmePpCommGpu::Impl::Impl(MPI_Comm comm, int pmeRank) : comm_(comm), pmeRank_(pmeRank)
62 "PME-PP GPU Communication is currently only supported with thread-MPI enabled");
63 cudaStreamCreate(&pmePpCommStream_);
66 PmePpCommGpu::Impl::~Impl() = default;
68 void PmePpCommGpu::Impl::reinit(int size)
70 // This rank will access PME rank memory directly, so needs to receive the remote PME buffer addresses.
71 MPI_Recv(&remotePmeXBuffer_, sizeof(void**), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
72 MPI_Recv(&remotePmeFBuffer_, sizeof(void**), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
74 // Reallocate buffer used for staging PME force on GPU
75 reallocateDeviceBuffer(&d_pmeForces_, size, &d_pmeForcesSize_, &d_pmeForcesSizeAlloc_, nullptr);
80 // TODO make this asynchronous by splitting into this into
81 // launchRecvForceFromPmeCudaDirect() and sycnRecvForceFromPmeCudaDirect()
82 void PmePpCommGpu::Impl::receiveForceFromPmeCudaDirect(void* recvPtr, int recvSize, bool receivePmeForceToGpu)
85 // Receive event from PME task and add to stream, to ensure pull of data doesn't
86 // occur before PME force calc is completed
87 GpuEventSynchronizer* pmeSync;
88 MPI_Recv(&pmeSync, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
89 pmeSync->enqueueWaitEvent(pmePpCommStream_);
91 // Pull force data from remote GPU
92 void* pmeForcePtr = receivePmeForceToGpu ? static_cast<void*>(d_pmeForces_) : recvPtr;
93 cudaError_t stat = cudaMemcpyAsync(pmeForcePtr, remotePmeFBuffer_, recvSize * DIM * sizeof(float),
94 cudaMemcpyDefault, pmePpCommStream_);
95 CU_RET_ERR(stat, "cudaMemcpyAsync on Recv from PME CUDA direct data transfer failed");
97 if (receivePmeForceToGpu)
99 // Record event to be enqueued in the GPU local buffer operations, to
100 // satisfy dependency on receiving the PME force data before
101 // reducing it with the other force contributions.
102 forcesReadySynchronizer_.markEvent(pmePpCommStream_);
106 // Ensure CPU waits for PME forces to be copied before reducing
107 // them with other forces on the CPU
108 cudaStreamSynchronize(pmePpCommStream_);
112 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaDirect(void* sendPtr,
114 bool gmx_unused sendPmeCoordinatesFromGpu,
115 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
118 // ensure stream waits until coordinate data is available on device
119 coordinatesReadyOnDeviceEvent->enqueueWaitEvent(pmePpCommStream_);
121 cudaError_t stat = cudaMemcpyAsync(remotePmeXBuffer_, sendPtr, sendSize * DIM * sizeof(float),
122 cudaMemcpyDefault, pmePpCommStream_);
123 CU_RET_ERR(stat, "cudaMemcpyAsync on Send to PME CUDA direct data transfer failed");
125 // Record and send event to allow PME task to sync to above transfer before commencing force calculations
126 pmeCoordinatesSynchronizer_.markEvent(pmePpCommStream_);
127 GpuEventSynchronizer* pmeSync = &pmeCoordinatesSynchronizer_;
128 MPI_Send(&pmeSync, sizeof(GpuEventSynchronizer*), MPI_BYTE, pmeRank_, 0, comm_);
130 void* PmePpCommGpu::Impl::getGpuForceStagingPtr()
132 return static_cast<void*>(d_pmeForces_);
135 void* PmePpCommGpu::Impl::getForcesReadySynchronizer()
137 return static_cast<void*>(&forcesReadySynchronizer_);
140 PmePpCommGpu::PmePpCommGpu(MPI_Comm comm, int pmeRank) : impl_(new Impl(comm, pmeRank)) {}
142 PmePpCommGpu::~PmePpCommGpu() = default;
144 void PmePpCommGpu::reinit(int size)
149 void PmePpCommGpu::receiveForceFromPmeCudaDirect(void* recvPtr, int recvSize, bool receivePmeForceToGpu)
151 impl_->receiveForceFromPmeCudaDirect(recvPtr, recvSize, receivePmeForceToGpu);
154 void PmePpCommGpu::sendCoordinatesToPmeCudaDirect(void* sendPtr,
156 bool sendPmeCoordinatesFromGpu,
157 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
159 impl_->sendCoordinatesToPmeCudaDirect(sendPtr, sendSize, sendPmeCoordinatesFromGpu,
160 coordinatesReadyOnDeviceEvent);
163 void* PmePpCommGpu::getGpuForceStagingPtr()
165 return impl_->getGpuForceStagingPtr();
168 void* PmePpCommGpu::getForcesReadySynchronizer()
170 return impl_->getForcesReadySynchronizer();