src/gromacs/nbnxm/kernels_reference/kernel_gpu_ref.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
   5  * Copyright (c) 2017,2018,2019,2020,2021, by the GROMACS development team, led by
   6  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   7  * and including many others, as listed in the AUTHORS file in the
   8  * top-level source directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36 #include "gmxpre.h"
  37
  38 #include "kernel_gpu_ref.h"
  39
  40 #include <cmath>
  41
  42 #include <algorithm>
  43
  44 #include "gromacs/math/functions.h"
  45 #include "gromacs/math/utilities.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/mdtypes/interaction_const.h"
  48 #include "gromacs/mdtypes/md_enums.h"
  49 #include "gromacs/mdtypes/simulation_workload.h"
  50 #include "gromacs/nbnxm/atomdata.h"
  51 #include "gromacs/nbnxm/nbnxm.h"
  52 #include "gromacs/nbnxm/pairlist.h"
  53 #include "gromacs/pbcutil/ishift.h"
  54 #include "gromacs/utility/fatalerror.h"
  55
  56 static constexpr int c_clSize = c_nbnxnGpuClusterSize;
  57
  58 void nbnxn_kernel_gpu_ref(const NbnxnPairlistGpu*    nbl,
  59                           const nbnxn_atomdata_t*    nbat,
  60                           const interaction_const_t* iconst,
  61                           rvec*                      shift_vec,
  62                           const gmx::StepWorkload&   stepWork,
  63                           int                        clearF,
  64                           gmx::ArrayRef<real>        f,
  65                           real*                      fshift,
  66                           real*                      Vc,
  67                           real*                      Vvdw)
  68 {
  69     real                fscal = NAN;
  70     real                vcoul = 0;
  71     const nbnxn_excl_t* excl[2];
  72
  73     if (nbl->na_ci != c_clSize)
  74     {
  75         gmx_fatal(FARGS,
  76                   "The neighborlist cluster size in the GPU reference kernel is %d, expected it to "
  77                   "be %d",
  78                   nbl->na_ci,
  79                   c_clSize);
  80     }
  81
  82     if (clearF == enbvClearFYes)
  83     {
  84         for (real& elem : f)
  85         {
  86             elem = 0;
  87         }
  88     }
  89
  90     const bool bEwald = EEL_FULL(iconst->eeltype);
  91
  92     const real rcut2 = iconst->rcoulomb * iconst->rcoulomb;
  93     const real rvdw2 = iconst->rvdw * iconst->rvdw;
  94
  95     const real rlist2 = nbl->rlist * nbl->rlist;
  96
  97     const int*  type     = nbat->params().type.data();
  98     const real  facel    = iconst->epsfac;
  99     const real* shiftvec = shift_vec[0];
 100     const real* vdwparam = nbat->params().nbfp.data();
 101     const int   ntype    = nbat->params().numTypes;
 102
 103     const real* x = nbat->x().data();
 104
 105     int npair_tot   = 0;
 106     int nhwu        = 0;
 107     int nhwu_pruned = 0;
 108
 109     for (const nbnxn_sci_t& nbln : nbl->sci)
 110     {
 111         const int  ish3     = 3 * nbln.shift;
 112         const real shX      = shiftvec[ish3];
 113         const real shY      = shiftvec[ish3 + 1];
 114         const real shZ      = shiftvec[ish3 + 2];
 115         const int  cj4_ind0 = nbln.cj4_ind_start;
 116         const int  cj4_ind1 = nbln.cj4_ind_end;
 117         const int  sci      = nbln.sci;
 118         real       vctot    = 0;
 119         real       Vvdwtot  = 0;
 120
 121         if (nbln.shift == CENTRAL && nbl->cj4[cj4_ind0].cj[0] == sci * c_nbnxnGpuNumClusterPerSupercluster)
 122         {
 123             /* we have the diagonal:
 124              * add the charge self interaction energy term
 125              */
 126             for (int im = 0; im < c_nbnxnGpuNumClusterPerSupercluster; im++)
 127             {
 128                 const int ci = sci * c_nbnxnGpuNumClusterPerSupercluster + im;
 129                 for (int ic = 0; ic < c_clSize; ic++)
 130                 {
 131                     const int ia = ci * c_clSize + ic;
 132                     real      iq = x[ia * nbat->xstride + 3];
 133                     vctot += iq * iq;
 134                 }
 135             }
 136             if (!bEwald)
 137             {
 138                 vctot *= -facel * 0.5 * iconst->c_rf;
 139             }
 140             else
 141             {
 142                 /* last factor 1/sqrt(pi) */
 143                 vctot *= -facel * iconst->ewaldcoeff_q * M_1_SQRTPI;
 144             }
 145         }
 146
 147         for (int cj4_ind = cj4_ind0; (cj4_ind < cj4_ind1); cj4_ind++)
 148         {
 149             excl[0] = &nbl->excl[nbl->cj4[cj4_ind].imei[0].excl_ind];
 150             excl[1] = &nbl->excl[nbl->cj4[cj4_ind].imei[1].excl_ind];
 151
 152             for (int jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
 153             {
 154                 const int cj = nbl->cj4[cj4_ind].cj[jm];
 155
 156                 for (int im = 0; im < c_nbnxnGpuNumClusterPerSupercluster; im++)
 157                 {
 158                     /* We're only using the first imask,
 159                      * but here imei[1].imask is identical.
 160                      */
 161                     if ((nbl->cj4[cj4_ind].imei[0].imask >> (jm * c_nbnxnGpuNumClusterPerSupercluster + im))
 162                         & 1)
 163                     {
 164                         const int ci = sci * c_nbnxnGpuNumClusterPerSupercluster + im;
 165
 166                         bool within_rlist = false;
 167                         int  npair        = 0;
 168                         for (int ic = 0; ic < c_clSize; ic++)
 169                         {
 170                             const int ia = ci * c_clSize + ic;
 171
 172                             const int  is  = ia * nbat->xstride;
 173                             const int  ifs = ia * nbat->fstride;
 174                             const real ix  = shX + x[is + 0];
 175                             const real iy  = shY + x[is + 1];
 176                             const real iz  = shZ + x[is + 2];
 177                             const real iq  = facel * x[is + 3];
 178                             const int  nti = ntype * 2 * type[ia];
 179
 180                             real fix = 0;
 181                             real fiy = 0;
 182                             real fiz = 0;
 183
 184                             for (int jc = 0; jc < c_clSize; jc++)
 185                             {
 186                                 const int ja = cj * c_clSize + jc;
 187
 188                                 if (nbln.shift == CENTRAL && ci == cj && ja <= ia)
 189                                 {
 190                                     continue;
 191                                 }
 192
 193                                 constexpr int clusterPerSplit =
 194                                         c_nbnxnGpuClusterSize / c_nbnxnGpuClusterpairSplit;
 195                                 const real int_bit = static_cast<real>(
 196                                         (excl[jc / clusterPerSplit]->pair[(jc & (clusterPerSplit - 1)) * c_clSize + ic]
 197                                          >> (jm * c_nbnxnGpuNumClusterPerSupercluster + im))
 198                                         & 1);
 199
 200                                 int        js  = ja * nbat->xstride;
 201                                 int        jfs = ja * nbat->fstride;
 202                                 const real jx  = x[js + 0];
 203                                 const real jy  = x[js + 1];
 204                                 const real jz  = x[js + 2];
 205                                 const real dx  = ix - jx;
 206                                 const real dy  = iy - jy;
 207                                 const real dz  = iz - jz;
 208                                 real       rsq = dx * dx + dy * dy + dz * dz;
 209                                 if (rsq < rlist2)
 210                                 {
 211                                     within_rlist = true;
 212                                 }
 213                                 if (rsq >= rcut2)
 214                                 {
 215                                     continue;
 216                                 }
 217
 218                                 if (type[ia] != ntype - 1 && type[ja] != ntype - 1)
 219                                 {
 220                                     npair++;
 221                                 }
 222
 223                                 // Ensure distance do not become so small that r^-12 overflows
 224                                 rsq = std::max(rsq, c_nbnxnMinDistanceSquared);
 225
 226                                 const real rinv   = gmx::invsqrt(rsq);
 227                                 const real rinvsq = rinv * rinv;
 228
 229                                 const real qq = iq * x[js + 3];
 230                                 if (!bEwald)
 231                                 {
 232                                     /* Reaction-field */
 233                                     const real krsq = iconst->k_rf * rsq;
 234                                     fscal           = qq * (int_bit * rinv - 2 * krsq) * rinvsq;
 235                                     if (stepWork.computeEnergy)
 236                                     {
 237                                         vcoul = qq * (int_bit * rinv + krsq - iconst->c_rf);
 238                                     }
 239                                 }
 240                                 else
 241                                 {
 242                                     const real  r    = rsq * rinv;
 243                                     const real  rt   = r * iconst->coulombEwaldTables->scale;
 244                                     const int   n0   = static_cast<int>(rt);
 245                                     const real  eps  = rt - static_cast<real>(n0);
 246                                     const real* Ftab = iconst->coulombEwaldTables->tableF.data();
 247
 248                                     const real fexcl = (1 - eps) * Ftab[n0] + eps * Ftab[n0 + 1];
 249
 250                                     fscal = qq * (int_bit * rinvsq - fexcl) * rinv;
 251
 252                                     if (stepWork.computeEnergy)
 253                                     {
 254                                         vcoul = qq
 255                                                 * ((int_bit - std::erf(iconst->ewaldcoeff_q * r)) * rinv
 256                                                    - int_bit * iconst->sh_ewald);
 257                                     }
 258                                 }
 259
 260                                 if (rsq < rvdw2)
 261                                 {
 262                                     const int tj = nti + 2 * type[ja];
 263
 264                                     /* Vanilla Lennard-Jones cutoff */
 265                                     const real c6  = vdwparam[tj];
 266                                     const real c12 = vdwparam[tj + 1];
 267
 268                                     const real rinvsix   = int_bit * rinvsq * rinvsq * rinvsq;
 269                                     const real Vvdw_disp = c6 * rinvsix;
 270                                     const real Vvdw_rep  = c12 * rinvsix * rinvsix;
 271                                     fscal += (Vvdw_rep - Vvdw_disp) * rinvsq;
 272
 273                                     if (stepWork.computeEnergy)
 274                                     {
 275                                         vctot += vcoul;
 276
 277                                         Vvdwtot +=
 278                                                 (Vvdw_rep + int_bit * c12 * iconst->repulsion_shift.cpot) / 12
 279                                                 - (Vvdw_disp
 280                                                    + int_bit * c6 * iconst->dispersion_shift.cpot)
 281                                                           / 6;
 282                                     }
 283                                 }
 284
 285                                 real tx = fscal * dx;
 286                                 real ty = fscal * dy;
 287                                 real tz = fscal * dz;
 288                                 fix     = fix + tx;
 289                                 fiy     = fiy + ty;
 290                                 fiz     = fiz + tz;
 291                                 f[jfs + 0] -= tx;
 292                                 f[jfs + 1] -= ty;
 293                                 f[jfs + 2] -= tz;
 294                             }
 295
 296                             f[ifs + 0] += fix;
 297                             f[ifs + 1] += fiy;
 298                             f[ifs + 2] += fiz;
 299                             fshift[ish3]     = fshift[ish3] + fix;
 300                             fshift[ish3 + 1] = fshift[ish3 + 1] + fiy;
 301                             fshift[ish3 + 2] = fshift[ish3 + 2] + fiz;
 302
 303                             /* Count in half work-units.
 304                              * In CUDA one work-unit is 2 warps.
 305                              */
 306                             if ((ic + 1) % (c_clSize / c_nbnxnGpuClusterpairSplit) == 0)
 307                             {
 308                                 npair_tot += npair;
 309
 310                                 nhwu++;
 311                                 if (within_rlist)
 312                                 {
 313                                     nhwu_pruned++;
 314                                 }
 315
 316                                 within_rlist = false;
 317                                 npair        = 0;
 318                             }
 319                         }
 320                     }
 321                 }
 322             }
 323         }
 324
 325         if (stepWork.computeEnergy)
 326         {
 327             const int ggid = 0;
 328             Vc[ggid]       = Vc[ggid] + vctot;
 329             Vvdw[ggid]     = Vvdw[ggid] + Vvdwtot;
 330         }
 331     }
 332
 333     if (debug)
 334     {
 335         fprintf(debug,
 336                 "number of half %dx%d atom pairs: %d after pruning: %d fraction %4.2f\n",
 337                 nbl->na_ci,
 338                 nbl->na_ci,
 339                 nhwu,
 340                 nhwu_pruned,
 341                 nhwu_pruned / static_cast<double>(nhwu));
 342         fprintf(debug, "generic kernel pair interactions:            %d\n", nhwu * nbl->na_ci / 2 * nbl->na_ci);
 343         fprintf(debug,
 344                 "generic kernel post-prune pair interactions: %d\n",
 345                 nhwu_pruned * nbl->na_ci / 2 * nbl->na_ci);
 346         fprintf(debug, "generic kernel non-zero pair interactions:   %d\n", npair_tot);
 347         fprintf(debug,
 348                 "ratio non-zero/post-prune pair interactions: %4.2f\n",
 349                 npair_tot / static_cast<double>(nhwu_pruned * gmx::exactDiv(nbl->na_ci, 2) * nbl->na_ci));
 350     }
 351 }