src/gromacs/nbnxm/cuda/nbnxm_cuda_types.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   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, 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.
  10  *
  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.
  15  *
  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.
  20  *
  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.
  25  *
  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.
  33  *
  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.
  36  */
  37
  38 /*! \internal \file
  39  *  \brief
  40  *  Data types used internally in the nbnxn_cuda module.
  41  *
  42  *  \author Szilárd Páll <pall.szilard@gmail.com>
  43  *  \ingroup module_nbnxm
  44  */
  45
  46 #ifndef NBNXM_CUDA_TYPES_H
  47 #define NBNXM_CUDA_TYPES_H
  48
  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/gputraits.cuh"
  54 #include "gromacs/mdtypes/interaction_const.h"
  55 #include "gromacs/nbnxm/gpu_types_common.h"
  56 #include "gromacs/nbnxm/nbnxm.h"
  57 #include "gromacs/nbnxm/pairlist.h"
  58 #include "gromacs/timing/gpu_timing.h"
  59 #include "gromacs/utility/enumerationhelpers.h"
  60
  61 /*! \brief Macro definining default for the prune kernel's j4 processing concurrency.
  62  *
  63  *  The GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY macro allows compile-time override.
  64  */
  65 #ifndef GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY
  66 #    define GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY 4
  67 #endif
  68 /*! \brief Default for the prune kernel's j4 processing concurrency.
  69  *
  70  *  Initialized using the #GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY macro which allows compile-time override.
  71  */
  72 const int c_cudaPruneKernelJ4Concurrency = GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY;
  73
  74 /* TODO: consider moving this to kernel_utils */
  75 /* Convenience defines */
  76 /*! \brief cluster size = number of atoms per cluster. */
  77 static constexpr int c_clSize = c_nbnxnGpuClusterSize;
  78
  79 /* All structs prefixed with "cu_" hold data used in GPU calculations and
  80  * are passed to the kernels, except cu_timers_t. */
  81 /*! \cond */
  82 typedef struct cu_atomdata cu_atomdata_t;
  83 /*! \endcond */
  84
  85
  86 /** \internal
  87  * \brief Staging area for temporary data downloaded from the GPU.
  88  *
  89  *  The energies/shift forces get downloaded here first, before getting added
  90  *  to the CPU-side aggregate values.
  91  */
  92 struct nb_staging_t
  93 {
  94     //! LJ energy
  95     float* e_lj = nullptr;
  96     //! electrostatic energy
  97     float* e_el = nullptr;
  98     //! shift forces
  99     float3* fshift = nullptr;
 100 };
 101
 102 /** \internal
 103  * \brief Nonbonded atom data - both inputs and outputs.
 104  */
 105 struct cu_atomdata
 106 {
 107     //! number of atoms
 108     int natoms;
 109     //! number of local atoms
 110     int natoms_local;
 111     //! allocation size for the atom data (xq, f)
 112     int nalloc;
 113
 114     //! atom coordinates + charges, size natoms
 115     DeviceBuffer<float4> xq;
 116     //! force output array, size natoms
 117     DeviceBuffer<float3> f;
 118
 119     //! LJ energy output, size 1
 120     DeviceBuffer<float> e_lj;
 121     //! Electrostatics energy input, size 1
 122     DeviceBuffer<float> e_el;
 123
 124     //! shift forces
 125     DeviceBuffer<float3> fshift;
 126
 127     //! number of atom types
 128     int ntypes;
 129     //! atom type indices, size natoms
 130     DeviceBuffer<int> atom_types;
 131     //! sqrt(c6),sqrt(c12) size natoms
 132     DeviceBuffer<float2> lj_comb;
 133
 134     //! shifts
 135     DeviceBuffer<float3> shift_vec;
 136     //! true if the shift vector has been uploaded
 137     bool bShiftVecUploaded;
 138 };
 139
 140 /** \internal
 141  * \brief Pair list data.
 142  */
 143 using cu_plist_t = Nbnxm::gpu_plist;
 144
 145 /** \internal
 146  * \brief Typedef of actual timer type.
 147  */
 148 typedef struct Nbnxm::gpu_timers_t cu_timers_t;
 149
 150 class GpuEventSynchronizer;
 151
 152 /*! \internal
 153  * \brief Main data structure for CUDA nonbonded force calculations.
 154  */
 155 struct NbnxmGpu
 156 {
 157     /*! \brief GPU device context.
 158      *
 159      * \todo Make it constant reference, once NbnxmGpu is a proper class.
 160      */
 161     const DeviceContext* deviceContext_;
 162     /*! \brief true if doing both local/non-local NB work on GPU */
 163     bool bUseTwoStreams = false;
 164     /*! \brief atom data */
 165     cu_atomdata_t* atdat = nullptr;
 166     /*! \brief f buf ops cell index mapping */
 167     int* cell = nullptr;
 168     /*! \brief number of indices in cell buffer */
 169     int ncell = 0;
 170     /*! \brief number of indices allocated in cell buffer */
 171     int ncell_alloc = 0;
 172     /*! \brief array of atom indices */
 173     int* atomIndices = nullptr;
 174     /*! \brief size of atom indices */
 175     int atomIndicesSize = 0;
 176     /*! \brief size of atom indices allocated in device buffer */
 177     int atomIndicesSize_alloc = 0;
 178     /*! \brief x buf ops num of atoms */
 179     int* cxy_na = nullptr;
 180     /*! \brief number of elements in cxy_na */
 181     int ncxy_na = 0;
 182     /*! \brief number of elements allocated allocated in device buffer */
 183     int ncxy_na_alloc = 0;
 184     /*! \brief x buf ops cell index mapping */
 185     int* cxy_ind = nullptr;
 186     /*! \brief number of elements in cxy_ind */
 187     int ncxy_ind = 0;
 188     /*! \brief number of elements allocated allocated in device buffer */
 189     int ncxy_ind_alloc = 0;
 190     /*! \brief parameters required for the non-bonded calc. */
 191     NBParamGpu* nbparam = nullptr;
 192     /*! \brief pair-list data structures (local and non-local) */
 193     gmx::EnumerationArray<Nbnxm::InteractionLocality, cu_plist_t*> plist = { { nullptr } };
 194     /*! \brief staging area where fshift/energies get downloaded */
 195     nb_staging_t nbst;
 196     /*! \brief local and non-local GPU streams */
 197     gmx::EnumerationArray<Nbnxm::InteractionLocality, const DeviceStream*> deviceStreams;
 198
 199     /*! \brief Events used for synchronization */
 200     /*! \{ */
 201     /*! \brief Event triggered when the non-local non-bonded
 202      * kernel is done (and the local transfer can proceed) */
 203     cudaEvent_t nonlocal_done = nullptr;
 204     /*! \brief Event triggered when the tasks issued in the local
 205      * stream that need to precede the non-local force or buffer
 206      * operation calculations are done (e.g. f buffer 0-ing, local
 207      * x/q H2D, buffer op initialization in local stream that is
 208      * required also by nonlocal stream ) */
 209     cudaEvent_t misc_ops_and_local_H2D_done = nullptr;
 210     /*! \} */
 211
 212     /*! \brief True if there is work for the current domain in the
 213      * respective locality.
 214      *
 215      * This includes local/nonlocal GPU work, either bonded or
 216      * nonbonded, scheduled to be executed in the current
 217      * domain. As long as bonded work is not split up into
 218      * local/nonlocal, if there is bonded GPU work, both flags
 219      * will be true. */
 220     gmx::EnumerationArray<Nbnxm::InteractionLocality, bool> haveWork = { { false } };
 221
 222     /*! \brief Pointer to event synchronizer triggered when the local
 223      * GPU buffer ops / reduction is complete
 224      *
 225      * \note That the synchronizer is managed outside of this module
 226      * in StatePropagatorDataGpu.
 227      */
 228     GpuEventSynchronizer* localFReductionDone = nullptr;
 229
 230     /*! \brief Event triggered when non-local coordinate buffer
 231      * has been copied from device to host. */
 232     GpuEventSynchronizer* xNonLocalCopyD2HDone = nullptr;
 233
 234     /* NOTE: With current CUDA versions (<=5.0) timing doesn't work with multiple
 235      * concurrent streams, so we won't time if both l/nl work is done on GPUs.
 236      * Timer init/uninit is still done even with timing off so only the condition
 237      * setting bDoTime needs to be change if this CUDA "feature" gets fixed. */
 238     /*! \brief True if event-based timing is enabled. */
 239     bool bDoTime = false;
 240     /*! \brief CUDA event-based timers. */
 241     cu_timers_t* timers = nullptr;
 242     /*! \brief Timing data. TODO: deprecate this and query timers for accumulated data instead */
 243     gmx_wallclock_gpu_nbnxn_t* timings = nullptr;
 244 };
 245
 246 #endif /* NBNXN_CUDA_TYPES_H */