src/gromacs/nbnxm/nbnxm.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2019, 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.
   8  *
   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.
  13  *
  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.
  18  *
  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.
  23  *
  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.
  31  *
  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.
  34  */
  35
  36 /*! \internal \file
  37  * \brief
  38  * Implements the Nbnxm class
  39  *
  40  * \author Berk Hess <hess@kth.se>
  41  * \ingroup module_nbnxm
  42  */
  43
  44 #include "gmxpre.h"
  45
  46 #include "nbnxm.h"
  47
  48 #include "gromacs/domdec/domdec_struct.h"
  49 #include "gromacs/nbnxm/atomdata.h"
  50 #include "gromacs/timing/wallcycle.h"
  51
  52 #include "pairlistsets.h"
  53 #include "pairsearch.h"
  54
  55 /*! \cond INTERNAL */
  56
  57 void nbnxn_put_on_grid(nonbonded_verlet_t             *nb_verlet,
  58                        const matrix                    box,
  59                        int                             ddZone,
  60                        const rvec                      lowerCorner,
  61                        const rvec                      upperCorner,
  62                        const gmx::UpdateGroupsCog     *updateGroupsCog,
  63                        int                             atomStart,
  64                        int                             atomEnd,
  65                        real                            atomDensity,
  66                        gmx::ArrayRef<const int>        atomInfo,
  67                        gmx::ArrayRef<const gmx::RVec>  x,
  68                        int                             numAtomsMoved,
  69                        const int                      *move)
  70 {
  71     nb_verlet->pairSearch_->putOnGrid(box, ddZone, lowerCorner, upperCorner,
  72                                       updateGroupsCog, atomStart, atomEnd, atomDensity,
  73                                       atomInfo, x, numAtomsMoved, move,
  74                                       nb_verlet->nbat.get());
  75 }
  76
  77 /* Calls nbnxn_put_on_grid for all non-local domains */
  78 void nbnxn_put_on_grid_nonlocal(nonbonded_verlet_t              *nbv,
  79                                 const struct gmx_domdec_zones_t *zones,
  80                                 gmx::ArrayRef<const int>         atomInfo,
  81                                 gmx::ArrayRef<const gmx::RVec>   x)
  82 {
  83     for (int zone = 1; zone < zones->n; zone++)
  84     {
  85         rvec c0, c1;
  86         for (int d = 0; d < DIM; d++)
  87         {
  88             c0[d] = zones->size[zone].bb_x0[d];
  89             c1[d] = zones->size[zone].bb_x1[d];
  90         }
  91
  92         nbnxn_put_on_grid(nbv, nullptr,
  93                           zone, c0, c1,
  94                           nullptr,
  95                           zones->cg_range[zone],
  96                           zones->cg_range[zone+1],
  97                           -1,
  98                           atomInfo,
  99                           x,
 100                           0, nullptr);
 101     }
 102 }
 103
 104 bool nonbonded_verlet_t::isDynamicPruningStepCpu(int64_t step) const
 105 {
 106     return pairlistSets_->isDynamicPruningStepCpu(step);
 107 }
 108
 109 bool nonbonded_verlet_t::isDynamicPruningStepGpu(int64_t step) const
 110 {
 111     return pairlistSets_->isDynamicPruningStepGpu(step);
 112 }
 113
 114 gmx::ArrayRef<const int> nonbonded_verlet_t::getLocalAtomOrder() const
 115 {
 116     /* Return the atom order for the home cell (index 0) */
 117     const Nbnxm::Grid &grid       = pairSearch_->gridSet().grids()[0];
 118
 119     const int          numIndices = grid.atomIndexEnd() - grid.firstAtomInColumn(0);
 120
 121     return gmx::constArrayRefFromArray(pairSearch_->gridSet().atomIndices().data(), numIndices);
 122 }
 123
 124 void nonbonded_verlet_t::setLocalAtomOrder()
 125 {
 126     pairSearch_->setLocalAtomOrder();
 127 }
 128
 129 void nonbonded_verlet_t::setAtomProperties(const t_mdatoms          &mdatoms,
 130                                            gmx::ArrayRef<const int>  atomInfo)
 131 {
 132     nbnxn_atomdata_set(nbat.get(), pairSearch_->gridSet(), &mdatoms, atomInfo.data());
 133 }
 134
 135 void nonbonded_verlet_t::setCoordinates(const Nbnxm::AtomLocality       locality,
 136                                         const bool                      fillLocal,
 137                                         gmx::ArrayRef<const gmx::RVec>  x,
 138                                         BufferOpsUseGpu                 useGpu,
 139                                         void                           *xPmeDevicePtr)
 140 {
 141     wallcycle_start(wcycle_, ewcNB_XF_BUF_OPS);
 142     wallcycle_sub_start(wcycle_, ewcsNB_X_BUF_OPS);
 143
 144     auto fnPtr = (useGpu == BufferOpsUseGpu::True) ?
 145         nbnxn_atomdata_copy_x_to_nbat_x<true> :
 146         nbnxn_atomdata_copy_x_to_nbat_x<false>;
 147
 148     fnPtr(pairSearch_->gridSet(), locality, fillLocal,
 149           as_rvec_array(x.data()),
 150           nbat.get(), gpu_nbv, xPmeDevicePtr);
 151
 152     wallcycle_sub_stop(wcycle_, ewcsNB_X_BUF_OPS);
 153     wallcycle_stop(wcycle_, ewcNB_XF_BUF_OPS);
 154 }
 155
 156 gmx::ArrayRef<const int> nonbonded_verlet_t::getGridIndices() const
 157 {
 158     return pairSearch_->gridSet().cells();
 159 }
 160
 161 void
 162 nonbonded_verlet_t::atomdata_add_nbat_f_to_f(const Nbnxm::AtomLocality           locality,
 163                                              gmx::ArrayRef<gmx::RVec>            force)
 164 {
 165
 166     /* Skip the reduction if there was no short-range GPU work to do
 167      * (either NB or both NB and bonded work). */
 168     if (!pairlistIsSimple() && !haveGpuShortRangeWork(locality))
 169     {
 170         return;
 171     }
 172
 173     wallcycle_start(wcycle_, ewcNB_XF_BUF_OPS);
 174     wallcycle_sub_start(wcycle_, ewcsNB_F_BUF_OPS);
 175
 176     reduceForces<false>(nbat.get(), locality, pairSearch_->gridSet(), as_rvec_array(force.data()), nullptr, nullptr, gpu_nbv, false, false);
 177
 178     wallcycle_sub_stop(wcycle_, ewcsNB_F_BUF_OPS);
 179     wallcycle_stop(wcycle_, ewcNB_XF_BUF_OPS);
 180 }
 181
 182 void
 183 nonbonded_verlet_t::atomdata_add_nbat_f_to_f(const Nbnxm::AtomLocality           locality,
 184                                              gmx::ArrayRef<gmx::RVec>            force,
 185                                              void                               *fPmeDeviceBuffer,
 186                                              GpuEventSynchronizer               *pmeForcesReady,
 187                                              BufferOpsUseGpu                     useGpu,
 188                                              bool                                useGpuFPmeReduction,
 189                                              bool                                accumulateForce)
 190 {
 191
 192     GMX_ASSERT(!((useGpu == BufferOpsUseGpu::False) && accumulateForce),
 193                "Accumulatation of force is only valid when GPU buffer ops are active");
 194
 195     GMX_ASSERT((useGpuFPmeReduction == (fPmeDeviceBuffer != nullptr)),
 196                "GPU PME force reduction is only valid when a non-null GPU PME force pointer is available");
 197
 198     /* Skip the reduction if there was no short-range GPU work to do
 199      * (either NB or both NB and bonded work). */
 200     if (!pairlistIsSimple() && !haveGpuShortRangeWork(locality))
 201     {
 202         return;
 203     }
 204
 205     wallcycle_start(wcycle_, ewcNB_XF_BUF_OPS);
 206     wallcycle_sub_start(wcycle_, ewcsNB_F_BUF_OPS);
 207
 208     auto fn = useGpu == BufferOpsUseGpu::True ? reduceForces<true> : reduceForces<false>;
 209     fn(nbat.get(), locality, pairSearch_->gridSet(), as_rvec_array(force.data()), fPmeDeviceBuffer, pmeForcesReady, gpu_nbv, useGpuFPmeReduction, accumulateForce);
 210
 211     wallcycle_sub_stop(wcycle_, ewcsNB_F_BUF_OPS);
 212     wallcycle_stop(wcycle_, ewcNB_XF_BUF_OPS);
 213 }
 214
 215 void
 216 nonbonded_verlet_t::atomdata_init_add_nbat_f_to_f_gpu()
 217 {
 218
 219     wallcycle_start(wcycle_, ewcNB_XF_BUF_OPS);
 220     wallcycle_sub_start(wcycle_, ewcsNB_F_BUF_OPS);
 221
 222     const Nbnxm::GridSet      &gridSet = pairSearch_->gridSet();
 223
 224     Nbnxm::nbnxn_gpu_init_add_nbat_f_to_f(gridSet.cells().data(),
 225                                           gpu_nbv,
 226                                           gridSet.numRealAtomsTotal());
 227
 228     wallcycle_sub_stop(wcycle_, ewcsNB_F_BUF_OPS);
 229     wallcycle_stop(wcycle_, ewcNB_XF_BUF_OPS);
 230 }
 231
 232 real nonbonded_verlet_t::pairlistInnerRadius() const
 233 {
 234     return pairlistSets_->params().rlistInner;
 235 }
 236
 237 real nonbonded_verlet_t::pairlistOuterRadius() const
 238 {
 239     return pairlistSets_->params().rlistOuter;
 240 }
 241
 242 void nonbonded_verlet_t::changePairlistRadii(real rlistOuter,
 243                                              real rlistInner)
 244 {
 245     pairlistSets_->changePairlistRadii(rlistOuter, rlistInner);
 246 }
 247
 248 void
 249 nonbonded_verlet_t::atomdata_init_copy_x_to_nbat_x_gpu()
 250 {
 251     Nbnxm::nbnxn_gpu_init_x_to_nbat_x(pairSearch_->gridSet(), gpu_nbv);
 252 }
 253
 254 void nonbonded_verlet_t::insertNonlocalGpuDependency(const Nbnxm::InteractionLocality interactionLocality)
 255 {
 256     Nbnxm::nbnxnInsertNonlocalGpuDependency(gpu_nbv, interactionLocality);
 257 }
 258
 259 void nonbonded_verlet_t::launch_copy_f_to_gpu(rvec *f, const Nbnxm::AtomLocality locality)
 260 {
 261     nbnxn_launch_copy_f_to_gpu(locality,
 262                                pairSearch_->gridSet(),
 263                                gpu_nbv,
 264                                f);
 265 }
 266
 267 void nonbonded_verlet_t::launch_copy_f_from_gpu(rvec *f, const Nbnxm::AtomLocality locality)
 268 {
 269     nbnxn_launch_copy_f_from_gpu(locality,
 270                                  pairSearch_->gridSet(),
 271                                  gpu_nbv,
 272                                  f);
 273 }
 274
 275 void nonbonded_verlet_t::wait_for_gpu_force_reduction(const Nbnxm::AtomLocality locality)
 276 {
 277     nbnxn_wait_for_gpu_force_reduction(locality, gpu_nbv);
 278 }
 279
 280 /*! \endcond */