2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2012, The GROMACS development team.
6 * Copyright (c) 2013-2019,2020,2021, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
40 * Data types used internally in the nbnxn_cuda module.
42 * \author Szilárd Páll <pall.szilard@gmail.com>
43 * \ingroup module_nbnxm
46 #ifndef NBNXM_CUDA_TYPES_H
47 #define NBNXM_CUDA_TYPES_H
49 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
50 #include "gromacs/gpu_utils/cudautils.cuh"
51 #include "gromacs/gpu_utils/devicebuffer.h"
52 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
53 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
54 #include "gromacs/gpu_utils/gputraits.cuh"
55 #include "gromacs/mdtypes/interaction_const.h"
56 #include "gromacs/nbnxm/gpu_types_common.h"
57 #include "gromacs/nbnxm/nbnxm.h"
58 #include "gromacs/nbnxm/pairlist.h"
59 #include "gromacs/timing/gpu_timing.h"
60 #include "gromacs/utility/enumerationhelpers.h"
62 /* TODO: consider moving this to kernel_utils */
63 /* Convenience defines */
64 /*! \brief cluster size = number of atoms per cluster. */
65 static constexpr int c_clSize = c_nbnxnGpuClusterSize;
68 * \brief Main data structure for CUDA nonbonded force calculations.
72 /*! \brief GPU device context.
74 * \todo Make it constant reference, once NbnxmGpu is a proper class.
76 const DeviceContext* deviceContext_;
77 /*! \brief true if doing both local/non-local NB work on GPU */
78 bool bUseTwoStreams = false;
79 //! true indicates that the nonlocal_done event was marked
80 bool bNonLocalStreamDoneMarked = false;
82 /*! \brief atom data */
83 NBAtomData* atdat = nullptr;
84 /*! \brief array of atom indices */
85 int* atomIndices = nullptr;
86 /*! \brief size of atom indices */
87 int atomIndicesSize = 0;
88 /*! \brief size of atom indices allocated in device buffer */
89 int atomIndicesSize_alloc = 0;
90 /*! \brief x buf ops num of atoms */
91 int* cxy_na = nullptr;
92 /*! \brief number of elements in cxy_na */
94 /*! \brief number of elements allocated allocated in device buffer */
95 int ncxy_na_alloc = 0;
96 /*! \brief x buf ops cell index mapping */
97 int* cxy_ind = nullptr;
98 /*! \brief number of elements in cxy_ind */
100 /*! \brief number of elements allocated allocated in device buffer */
101 int ncxy_ind_alloc = 0;
102 /*! \brief parameters required for the non-bonded calc. */
103 NBParamGpu* nbparam = nullptr;
104 /*! \brief pair-list data structures (local and non-local) */
105 gmx::EnumerationArray<Nbnxm::InteractionLocality, Nbnxm::gpu_plist*> plist = { { nullptr } };
106 /*! \brief staging area where fshift/energies get downloaded */
108 /*! \brief local and non-local GPU streams */
109 gmx::EnumerationArray<Nbnxm::InteractionLocality, const DeviceStream*> deviceStreams;
111 /*! \brief Event triggered when the non-local non-bonded
112 * kernel is done (and the local transfer can proceed) */
113 GpuEventSynchronizer nonlocal_done;
114 /*! \brief Event triggered when the tasks issued in the local
115 * stream that need to precede the non-local force or buffer
116 * operation calculations are done (e.g. f buffer 0-ing, local
117 * x/q H2D, buffer op initialization in local stream that is
118 * required also by nonlocal stream ) */
119 GpuEventSynchronizer misc_ops_and_local_H2D_done;
121 /*! \brief True if there is work for the current domain in the
122 * respective locality.
124 * This includes local/nonlocal GPU work, either bonded or
125 * nonbonded, scheduled to be executed in the current
126 * domain. As long as bonded work is not split up into
127 * local/nonlocal, if there is bonded GPU work, both flags
129 gmx::EnumerationArray<Nbnxm::InteractionLocality, bool> haveWork = { { false } };
131 /* NOTE: With current CUDA versions (<=5.0) timing doesn't work with multiple
132 * concurrent streams, so we won't time if both l/nl work is done on GPUs.
133 * Timer init/uninit is still done even with timing off so only the condition
134 * setting bDoTime needs to be change if this CUDA "feature" gets fixed. */
135 /*! \brief True if event-based timing is enabled. */
136 bool bDoTime = false;
137 /*! \brief CUDA event-based timers. */
138 Nbnxm::GpuTimers* timers = nullptr;
139 /*! \brief Timing data. TODO: deprecate this and query timers for accumulated data instead */
140 gmx_wallclock_gpu_nbnxn_t* timings = nullptr;
143 #endif /* NBNXN_CUDA_TYPES_H */