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 Declares CUDA implementation class for PME-PP communications
39 * \author Alan Gray <alang@nvidia.com>
41 * \ingroup module_ewald
43 #ifndef GMX_PME_PP_COMM_GPU_IMPL_H
44 #define GMX_PME_PP_COMM_GPU_IMPL_H
46 #include "gromacs/ewald/pme_pp_comm_gpu.h"
47 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
48 #include "gromacs/math/vectypes.h"
49 #include "gromacs/utility/gmxmpi.h"
54 /*! \internal \brief Class with interfaces and data for CUDA version of PME-PP Communication */
55 class PmePpCommGpu::Impl
59 /*! \brief Creates PME-PP GPU communication object.
61 * \param[in] comm Communicator used for simulation
62 * \param[in] pmeRank Rank of PME task
63 * \param[in] pmeCpuForceBuffer Buffer for PME force in CPU memory
64 * \param[in] deviceContext GPU context.
65 * \param[in] deviceStream GPU stream.
69 std::vector<gmx::RVec>* pmeCpuForceBuffer,
70 const DeviceContext& deviceContext,
71 const DeviceStream& deviceStream);
74 /*! \brief Perform steps required when buffer size changes
75 * \param[in] size Number of elements in buffer
77 void reinit(int size);
79 /*! \brief Pull force buffer directly from GPU memory on PME
80 * rank to either GPU or CPU memory on PP task using CUDA
81 * Memory copy or CUDA-aware MPI.
83 * recvPtr should be in GPU or CPU memory if recvPmeForceToGpu
84 * is true or false, respectively. If receiving to GPU, this
85 * method should be called before the local GPU buffer
86 * operations. If receiving to CPU it should be called
87 * before forces are reduced with the other force
88 * contributions on the CPU. It will automatically wait for
89 * remote PME force data to be ready.
91 * \param[out] recvPtr CPU buffer to receive PME force data
92 * \param[in] recvSize Number of elements to receive
93 * \param[in] receivePmeForceToGpu Whether receive is to GPU, otherwise CPU
95 void receiveForceFromPme(float3* recvPtr, int recvSize, bool receivePmeForceToGpu);
97 /*! \brief Push coordinates buffer directly to GPU memory on PME
98 * task, from either GPU or CPU memory on PP task using CUDA
99 * Memory copy or CUDA-aware MPI. If sending from GPU, this method should
100 * be called after the local GPU coordinate buffer operations.
101 * The remote PME task will automatically wait for data to be copied
102 * before commencing PME force calculations.
103 * \param[in] sendPtr Buffer with coordinate data
104 * \param[in] sendSize Number of elements to send
105 * \param[in] coordinatesReadyOnDeviceEvent Event recorded when coordinates are available on device
107 void sendCoordinatesToPme(float3* sendPtr, int sendSize, GpuEventSynchronizer* coordinatesReadyOnDeviceEvent);
110 * Return pointer to buffer used for staging PME force on GPU
112 DeviceBuffer<Float3> getGpuForceStagingPtr();
115 * Return pointer to event recorded when forces are ready
117 GpuEventSynchronizer* getForcesReadySynchronizer();
120 /*! \brief Pull force buffer directly from GPU memory on PME
121 * rank to either GPU or CPU memory on PP task using CUDA
122 * Memory copy. This method is used with Thread-MPI.
123 * \param[in] receivePmeForceToGpu Whether receive is to GPU, otherwise CPU
125 void receiveForceFromPmeCudaDirect(bool receivePmeForceToGpu);
127 /*! \brief Pull force buffer directly from GPU memory on PME
128 * rank to either GPU or CPU memory on PP task using CUDA-aware
129 * MPI. This method is used with process-MPI.
130 * \param[out] recvPtr CPU buffer to receive PME force data
131 * \param[in] recvSize Number of elements to receive
133 void receiveForceFromPmeCudaMpi(float3* recvPtr, int recvSize);
135 /*! \brief Push coordinates buffer directly to GPU memory on PME
136 * task, from either GPU or CPU memory on PP task using CUDA Memory copy.
137 * This method is used with Thread-MPI.
138 * \param[in] sendPtr Buffer with coordinate data
139 * \param[in] sendSize Number of elements to send
140 * \param[in] coordinatesReadyOnDeviceEvent Event recorded when coordinates are available on device
142 void sendCoordinatesToPmeCudaDirect(float3* sendPtr,
144 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent);
146 /*! \brief Push coordinates buffer directly to GPU memory on PME
147 * task, from either GPU or CPU memory on PP task using CUDA-aware MPI.
148 * This method is used with process-MPI.
149 * \param[in] sendPtr Buffer with coordinate data
150 * \param[in] sendSize Number of elements to send
151 * \param[in] coordinatesReadyOnDeviceEvent Event recorded when coordinates are available on device
153 void sendCoordinatesToPmeCudaMpi(float3* sendPtr,
155 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent);
157 //! GPU context handle (not used in CUDA)
158 const DeviceContext& deviceContext_;
159 //! Handle for CUDA stream used for the communication operations in this class
160 const DeviceStream& pmePpCommStream_;
161 //! Remote location of PME coordinate data buffer
162 float3* remotePmeXBuffer_ = nullptr;
163 //! communicator for simulation
167 //! Buffer for PME force on CPU
168 std::vector<gmx::RVec>* pmeCpuForceBuffer_;
169 //! Buffer for staging PME force on GPU
170 DeviceBuffer<gmx::RVec> d_pmeForces_;
171 //! number of atoms in PME force staging array
172 int d_pmeForcesSize_ = -1;
173 //! number of atoms allocated in recvbuf array
174 int d_pmeForcesSizeAlloc_ = -1;
175 //! Event recorded when PME forces are ready on PME task
176 GpuEventSynchronizer forcesReadySynchronizer_;
177 //! Event recorded when coordinates have been transferred to PME task
178 GpuEventSynchronizer pmeCoordinatesSynchronizer_;