2 * This file is part of the GROMACS molecular simulation package.
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.
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_force_sender_gpu_impl.h"
50 #include "gromacs/gpu_utils/cudautils.cuh"
51 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
52 #include "gromacs/utility/gmxmpi.h"
57 /*! \brief Create PME-PP GPU communication object */
58 PmeForceSenderGpu::Impl::Impl(GpuEventSynchronizer* pmeForcesReady,
60 const DeviceContext& deviceContext,
61 gmx::ArrayRef<PpRanks> ppRanks) :
62 pmeForcesReady_(pmeForcesReady),
65 ppCommStream_(ppRanks.size()),
66 ppCommEvent_(ppRanks.size()),
67 deviceContext_(deviceContext),
68 pmeRemoteCpuForcePtr_(ppRanks.size()),
69 pmeRemoteGpuForcePtr_(ppRanks.size())
71 // Create streams and events to manage pushing of force buffers to remote PP ranks
72 std::unique_ptr<DeviceStream> stream;
73 std::unique_ptr<GpuEventSynchronizer> event;
75 for (i = 0; i < ppRanks_.size(); i++)
77 stream = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::High, false);
78 ppCommStream_[i] = std::move(stream);
79 event = std::make_unique<GpuEventSynchronizer>();
80 ppCommEvent_[i] = std::move(event);
84 PmeForceSenderGpu::Impl::~Impl() = default;
86 /*! \brief Sets location of force to be sent to each PP rank */
87 void PmeForceSenderGpu::Impl::setForceSendBuffer(DeviceBuffer<Float3> d_f)
90 // Need to send address to PP rank only for thread-MPI as PP rank pulls
91 // data using cudamemcpy
99 if (localForcePtr_.empty())
101 localForcePtr_.resize(ppRanks_.size());
106 for (const auto& receiver : ppRanks_)
109 ind_end = ind_start + receiver.numAtoms;
111 localForcePtr_[i] = &d_f[ind_start];
112 // NOLINTNEXTLINE(bugprone-sizeof-expression)
113 MPI_Recv(&pmeRemoteGpuForcePtr_[i], sizeof(float3*), MPI_BYTE, receiver.rankId, 0, comm_, MPI_STATUS_IGNORE);
114 // NOLINTNEXTLINE(bugprone-sizeof-expression)
115 MPI_Recv(&pmeRemoteCpuForcePtr_[i], sizeof(float3*), MPI_BYTE, receiver.rankId, 0, comm_, MPI_STATUS_IGNORE);
120 GMX_UNUSED_VALUE(d_f);
125 /*! \brief Send PME synchronizer directly using CUDA memory copy */
126 void PmeForceSenderGpu::Impl::sendFToPpCudaDirect(int ppRank, int numAtoms, bool sendForcesDirectToPpGpu)
129 GMX_ASSERT(GMX_THREAD_MPI, "sendFToPpCudaDirect is expected to be called only for Thread-MPI");
133 float3* pmeRemoteForcePtr =
134 sendForcesDirectToPpGpu ? pmeRemoteGpuForcePtr_[ppRank] : pmeRemoteCpuForcePtr_[ppRank];
136 pmeForcesReady_->enqueueWaitEvent(*ppCommStream_[ppRank]);
138 cudaError_t stat = cudaMemcpyAsync(pmeRemoteForcePtr,
139 localForcePtr_[ppRank],
140 numAtoms * sizeof(rvec),
142 ppCommStream_[ppRank]->stream());
143 CU_RET_ERR(stat, "cudaMemcpyAsync on Recv from PME CUDA direct data transfer failed");
144 ppCommEvent_[ppRank]->markEvent(*ppCommStream_[ppRank]);
145 // NOLINTNEXTLINE(bugprone-sizeof-expression)
146 MPI_Send(&ppCommEvent_[ppRank], sizeof(GpuEventSynchronizer*), MPI_BYTE, ppRank, 0, comm_);
148 GMX_UNUSED_VALUE(ppRank);
149 GMX_UNUSED_VALUE(numAtoms);
153 /*! \brief Send PME data directly using CUDA-aware MPI */
154 void PmeForceSenderGpu::Impl::sendFToPpCudaMpi(DeviceBuffer<RVec> sendbuf,
158 MPI_Request* request)
160 GMX_ASSERT(GMX_LIB_MPI, "sendFToPpCudaMpi is expected to be called only for Lib-MPI");
163 // if using GPU direct comm with CUDA-aware MPI, make sure forces are ready on device
164 // before sending it to PP ranks
165 pmeForcesReady_->waitForEvent();
167 MPI_Isend(sendbuf[offset], numBytes, MPI_BYTE, ppRank, 0, comm_, request);
170 GMX_UNUSED_VALUE(sendbuf);
171 GMX_UNUSED_VALUE(offset);
172 GMX_UNUSED_VALUE(numBytes);
173 GMX_UNUSED_VALUE(ppRank);
174 GMX_UNUSED_VALUE(request);
178 PmeForceSenderGpu::PmeForceSenderGpu(GpuEventSynchronizer* pmeForcesReady,
180 const DeviceContext& deviceContext,
181 gmx::ArrayRef<PpRanks> ppRanks) :
182 impl_(new Impl(pmeForcesReady, comm, deviceContext, ppRanks))
186 PmeForceSenderGpu::~PmeForceSenderGpu() = default;
189 void PmeForceSenderGpu::setForceSendBuffer(DeviceBuffer<RVec> d_f)
191 impl_->setForceSendBuffer(d_f);
194 void PmeForceSenderGpu::sendFToPpCudaMpi(DeviceBuffer<RVec> sendbuf,
198 MPI_Request* request)
200 impl_->sendFToPpCudaMpi(sendbuf, offset, numBytes, ppRank, request);
203 void PmeForceSenderGpu::sendFToPpCudaDirect(int ppRank, int numAtoms, bool sendForcesDirectToPpGpu)
205 impl_->sendFToPpCudaDirect(ppRank, numAtoms, sendForcesDirectToPpGpu);