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 "gromacs/ewald/pme_pp_communication.h"
47 #include "pme_pp_comm_gpu_impl.h"
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"
62 PmePpCommGpu::Impl::Impl(MPI_Comm comm,
64 std::vector<gmx::RVec>* pmeCpuForceBuffer,
65 const DeviceContext& deviceContext,
66 const DeviceStream& deviceStream) :
67 deviceContext_(deviceContext),
68 pmePpCommStream_(deviceStream),
71 pmeCpuForceBuffer_(pmeCpuForceBuffer),
76 PmePpCommGpu::Impl::~Impl() = default;
78 void PmePpCommGpu::Impl::reinit(int size)
80 // Reallocate device buffer used for staging PME force
81 reallocateDeviceBuffer(&d_pmeForces_, size, &d_pmeForcesSize_, &d_pmeForcesSizeAlloc_, deviceContext_);
83 // This rank will access PME rank memory directly, so needs to receive the remote PME buffer addresses.
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);
102 void PmePpCommGpu::Impl::receiveForceFromPmeCudaDirect(bool receivePmeForceToGpu)
105 // Wait until remote PME task has pushed data, and then enqueue remote event to local stream.
107 // Spin until PME rank sets flag
108 while (!(remotePmeForceSendEventRecorded_->load(std::memory_order_acquire))) {};
110 // Enqueue remote event
111 remotePmeForceSendEvent_->enqueueWaitEvent(pmePpCommStream_);
114 remotePmeForceSendEventRecorded_->store(false, std::memory_order_release);
116 if (receivePmeForceToGpu)
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_);
125 // Ensure CPU waits for PME forces to be copied before reducing
126 // them with other forces on the CPU
127 pmePpCommStream_.synchronize();
132 void PmePpCommGpu::Impl::receiveForceFromPmeCudaMpi(float3* pmeForcePtr, int recvSize)
135 MPI_Recv(pmeForcePtr, recvSize * DIM, MPI_FLOAT, pmeRank_, 0, comm_, MPI_STATUS_IGNORE);
137 GMX_UNUSED_VALUE(pmeForcePtr);
138 GMX_UNUSED_VALUE(recvSize);
142 void PmePpCommGpu::Impl::receiveForceFromPme(float3* recvPtr, int recvSize, bool receivePmeForceToGpu)
144 float3* pmeForcePtr = receivePmeForceToGpu ? asFloat3(d_pmeForces_) : recvPtr;
147 receiveForceFromPmeCudaDirect(receivePmeForceToGpu);
151 receiveForceFromPmeCudaMpi(pmeForcePtr, recvSize);
155 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaDirect(float3* sendPtr,
157 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
159 // ensure stream waits until coordinate data is available on device
160 coordinatesReadyOnDeviceEvent->enqueueWaitEvent(pmePpCommStream_);
162 cudaError_t stat = cudaMemcpyAsync(remotePmeXBuffer_,
164 sendSize * DIM * sizeof(float),
166 pmePpCommStream_.stream());
167 CU_RET_ERR(stat, "cudaMemcpyAsync on Send to PME CUDA direct data transfer failed");
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_);
178 void PmePpCommGpu::Impl::sendCoordinatesToPmeCudaMpi(float3* sendPtr,
180 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
182 // ensure coordinate data is available on device before we start transfer
183 coordinatesReadyOnDeviceEvent->waitForEvent();
186 float3* sendptr_x = sendPtr;
188 MPI_Send(sendptr_x, sendSize * DIM, MPI_FLOAT, pmeRank_, eCommType_COORD_GPU, comm_);
190 GMX_UNUSED_VALUE(sendPtr);
191 GMX_UNUSED_VALUE(sendSize);
195 void PmePpCommGpu::Impl::sendCoordinatesToPme(float3* sendPtr,
197 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
201 sendCoordinatesToPmeCudaDirect(sendPtr, sendSize, coordinatesReadyOnDeviceEvent);
205 sendCoordinatesToPmeCudaMpi(sendPtr, sendSize, coordinatesReadyOnDeviceEvent);
208 DeviceBuffer<Float3> PmePpCommGpu::Impl::getGpuForceStagingPtr()
213 GpuEventSynchronizer* PmePpCommGpu::Impl::getForcesReadySynchronizer()
217 return &forcesReadySynchronizer_;
225 PmePpCommGpu::PmePpCommGpu(MPI_Comm comm,
227 std::vector<gmx::RVec>* pmeCpuForceBuffer,
228 const DeviceContext& deviceContext,
229 const DeviceStream& deviceStream) :
230 impl_(new Impl(comm, pmeRank, pmeCpuForceBuffer, deviceContext, deviceStream))
234 PmePpCommGpu::~PmePpCommGpu() = default;
236 void PmePpCommGpu::reinit(int size)
241 void PmePpCommGpu::receiveForceFromPme(RVec* recvPtr, int recvSize, bool receivePmeForceToGpu)
243 impl_->receiveForceFromPme(asFloat3(recvPtr), recvSize, receivePmeForceToGpu);
246 void PmePpCommGpu::sendCoordinatesToPmeFromGpu(DeviceBuffer<RVec> sendPtr,
248 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
250 impl_->sendCoordinatesToPme(asFloat3(sendPtr), sendSize, coordinatesReadyOnDeviceEvent);
253 void PmePpCommGpu::sendCoordinatesToPmeFromCpu(RVec* sendPtr,
255 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
257 impl_->sendCoordinatesToPme(asFloat3(sendPtr), sendSize, coordinatesReadyOnDeviceEvent);
260 DeviceBuffer<Float3> PmePpCommGpu::getGpuForceStagingPtr()
262 return impl_->getGpuForceStagingPtr();
265 GpuEventSynchronizer* PmePpCommGpu::getForcesReadySynchronizer()
267 return impl_->getForcesReadySynchronizer();