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, 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     gmx_bool            bEwald;
  70     const real*         Ftab = nullptr;
  71     real                rcut2, rvdw2, rlist2;
  72     int                 ntype;
  73     real                facel;
  74     int                 ish3;
  75     int                 sci;
  76     int                 cj4_ind0, cj4_ind1, cj4_ind;
  77     int                 ci, cj;
  78     int                 ic, jc, ia, ja, is, ifs, js, jfs, im, jm;
  79     int                 n0;
  80     int                 ggid;
  81     real                shX, shY, shZ;
  82     real                fscal, tx, ty, tz;
  83     real                rinvsq;
  84     real                iq;
  85     real                qq, vcoul = 0, krsq, vctot;
  86     int                 nti;
  87     int                 tj;
  88     real                rt, r, eps;
  89     real                rinvsix;
  90     real                Vvdwtot;
  91     real                Vvdw_rep, Vvdw_disp;
  92     real                ix, iy, iz, fix, fiy, fiz;
  93     real                jx, jy, jz;
  94     real                dx, dy, dz, rsq, rinv;
  95     real                int_bit;
  96     real                fexcl;
  97     real                c6, c12;
  98     const nbnxn_excl_t* excl[2];
  99
 100     int npair_tot, npair;
 101     int nhwu, nhwu_pruned;
 102
 103     if (nbl->na_ci != c_clSize)
 104     {
 105         gmx_fatal(FARGS,
 106                   "The neighborlist cluster size in the GPU reference kernel is %d, expected it to "
 107                   "be %d",
 108                   nbl->na_ci,
 109                   c_clSize);
 110     }
 111
 112     if (clearF == enbvClearFYes)
 113     {
 114         for (real& elem : f)
 115         {
 116             elem = 0;
 117         }
 118     }
 119
 120     bEwald = EEL_FULL(iconst->eeltype);
 121     if (bEwald)
 122     {
 123         Ftab = iconst->coulombEwaldTables->tableF.data();
 124     }
 125
 126     rcut2 = iconst->rcoulomb * iconst->rcoulomb;
 127     rvdw2 = iconst->rvdw * iconst->rvdw;
 128
 129     rlist2 = nbl->rlist * nbl->rlist;
 130
 131     const int* type      = nbat->params().type.data();
 132     facel                = iconst->epsfac;
 133     const real* shiftvec = shift_vec[0];
 134     const real* vdwparam = nbat->params().nbfp.data();
 135     ntype                = nbat->params().numTypes;
 136
 137     const real* x = nbat->x().data();
 138
 139     npair_tot   = 0;
 140     nhwu        = 0;
 141     nhwu_pruned = 0;
 142
 143     for (const nbnxn_sci_t& nbln : nbl->sci)
 144     {
 145         ish3     = 3 * nbln.shift;
 146         shX      = shiftvec[ish3];
 147         shY      = shiftvec[ish3 + 1];
 148         shZ      = shiftvec[ish3 + 2];
 149         cj4_ind0 = nbln.cj4_ind_start;
 150         cj4_ind1 = nbln.cj4_ind_end;
 151         sci      = nbln.sci;
 152         vctot    = 0;
 153         Vvdwtot  = 0;
 154
 155         if (nbln.shift == CENTRAL && nbl->cj4[cj4_ind0].cj[0] == sci * c_nbnxnGpuNumClusterPerSupercluster)
 156         {
 157             /* we have the diagonal:
 158              * add the charge self interaction energy term
 159              */
 160             for (im = 0; im < c_nbnxnGpuNumClusterPerSupercluster; im++)
 161             {
 162                 ci = sci * c_nbnxnGpuNumClusterPerSupercluster + im;
 163                 for (ic = 0; ic < c_clSize; ic++)
 164                 {
 165                     ia = ci * c_clSize + ic;
 166                     iq = x[ia * nbat->xstride + 3];
 167                     vctot += iq * iq;
 168                 }
 169             }
 170             if (!bEwald)
 171             {
 172                 vctot *= -facel * 0.5 * iconst->c_rf;
 173             }
 174             else
 175             {
 176                 /* last factor 1/sqrt(pi) */
 177                 vctot *= -facel * iconst->ewaldcoeff_q * M_1_SQRTPI;
 178             }
 179         }
 180
 181         for (cj4_ind = cj4_ind0; (cj4_ind < cj4_ind1); cj4_ind++)
 182         {
 183             excl[0] = &nbl->excl[nbl->cj4[cj4_ind].imei[0].excl_ind];
 184             excl[1] = &nbl->excl[nbl->cj4[cj4_ind].imei[1].excl_ind];
 185
 186             for (jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
 187             {
 188                 cj = nbl->cj4[cj4_ind].cj[jm];
 189
 190                 for (im = 0; im < c_nbnxnGpuNumClusterPerSupercluster; im++)
 191                 {
 192                     /* We're only using the first imask,
 193                      * but here imei[1].imask is identical.
 194                      */
 195                     if ((nbl->cj4[cj4_ind].imei[0].imask >> (jm * c_nbnxnGpuNumClusterPerSupercluster + im))
 196                         & 1)
 197                     {
 198                         gmx_bool within_rlist;
 199
 200                         ci = sci * c_nbnxnGpuNumClusterPerSupercluster + im;
 201
 202                         within_rlist = FALSE;
 203                         npair        = 0;
 204                         for (ic = 0; ic < c_clSize; ic++)
 205                         {
 206                             ia = ci * c_clSize + ic;
 207
 208                             is  = ia * nbat->xstride;
 209                             ifs = ia * nbat->fstride;
 210                             ix  = shX + x[is + 0];
 211                             iy  = shY + x[is + 1];
 212                             iz  = shZ + x[is + 2];
 213                             iq  = facel * x[is + 3];
 214                             nti = ntype * 2 * type[ia];
 215
 216                             fix = 0;
 217                             fiy = 0;
 218                             fiz = 0;
 219
 220                             for (jc = 0; jc < c_clSize; jc++)
 221                             {
 222                                 ja = cj * c_clSize + jc;
 223
 224                                 if (nbln.shift == CENTRAL && ci == cj && ja <= ia)
 225                                 {
 226                                     continue;
 227                                 }
 228
 229                                 constexpr int clusterPerSplit =
 230                                         c_nbnxnGpuClusterSize / c_nbnxnGpuClusterpairSplit;
 231                                 int_bit = static_cast<real>(
 232                                         (excl[jc / clusterPerSplit]->pair[(jc & (clusterPerSplit - 1)) * c_clSize + ic]
 233                                          >> (jm * c_nbnxnGpuNumClusterPerSupercluster + im))
 234                                         & 1);
 235
 236                                 js  = ja * nbat->xstride;
 237                                 jfs = ja * nbat->fstride;
 238                                 jx  = x[js + 0];
 239                                 jy  = x[js + 1];
 240                                 jz  = x[js + 2];
 241                                 dx  = ix - jx;
 242                                 dy  = iy - jy;
 243                                 dz  = iz - jz;
 244                                 rsq = dx * dx + dy * dy + dz * dz;
 245                                 if (rsq < rlist2)
 246                                 {
 247                                     within_rlist = TRUE;
 248                                 }
 249                                 if (rsq >= rcut2)
 250                                 {
 251                                     continue;
 252                                 }
 253
 254                                 if (type[ia] != ntype - 1 && type[ja] != ntype - 1)
 255                                 {
 256                                     npair++;
 257                                 }
 258
 259                                 // Ensure distance do not become so small that r^-12 overflows
 260                                 rsq = std::max(rsq, c_nbnxnMinDistanceSquared);
 261
 262                                 rinv   = gmx::invsqrt(rsq);
 263                                 rinvsq = rinv * rinv;
 264
 265                                 qq = iq * x[js + 3];
 266                                 if (!bEwald)
 267                                 {
 268                                     /* Reaction-field */
 269                                     krsq  = iconst->k_rf * rsq;
 270                                     fscal = qq * (int_bit * rinv - 2 * krsq) * rinvsq;
 271                                     if (stepWork.computeEnergy)
 272                                     {
 273                                         vcoul = qq * (int_bit * rinv + krsq - iconst->c_rf);
 274                                     }
 275                                 }
 276                                 else
 277                                 {
 278                                     r   = rsq * rinv;
 279                                     rt  = r * iconst->coulombEwaldTables->scale;
 280                                     n0  = static_cast<int>(rt);
 281                                     eps = rt - static_cast<real>(n0);
 282
 283                                     fexcl = (1 - eps) * Ftab[n0] + eps * Ftab[n0 + 1];
 284
 285                                     fscal = qq * (int_bit * rinvsq - fexcl) * rinv;
 286
 287                                     if (stepWork.computeEnergy)
 288                                     {
 289                                         vcoul = qq
 290                                                 * ((int_bit - std::erf(iconst->ewaldcoeff_q * r)) * rinv
 291                                                    - int_bit * iconst->sh_ewald);
 292                                     }
 293                                 }
 294
 295                                 if (rsq < rvdw2)
 296                                 {
 297                                     tj = nti + 2 * type[ja];
 298
 299                                     /* Vanilla Lennard-Jones cutoff */
 300                                     c6  = vdwparam[tj];
 301                                     c12 = vdwparam[tj + 1];
 302
 303                                     rinvsix   = int_bit * rinvsq * rinvsq * rinvsq;
 304                                     Vvdw_disp = c6 * rinvsix;
 305                                     Vvdw_rep  = c12 * rinvsix * rinvsix;
 306                                     fscal += (Vvdw_rep - Vvdw_disp) * rinvsq;
 307
 308                                     if (stepWork.computeEnergy)
 309                                     {
 310                                         vctot += vcoul;
 311
 312                                         Vvdwtot +=
 313                                                 (Vvdw_rep + int_bit * c12 * iconst->repulsion_shift.cpot) / 12
 314                                                 - (Vvdw_disp
 315                                                    + int_bit * c6 * iconst->dispersion_shift.cpot)
 316                                                           / 6;
 317                                     }
 318                                 }
 319
 320                                 tx  = fscal * dx;
 321                                 ty  = fscal * dy;
 322                                 tz  = fscal * dz;
 323                                 fix = fix + tx;
 324                                 fiy = fiy + ty;
 325                                 fiz = fiz + tz;
 326                                 f[jfs + 0] -= tx;
 327                                 f[jfs + 1] -= ty;
 328                                 f[jfs + 2] -= tz;
 329                             }
 330
 331                             f[ifs + 0] += fix;
 332                             f[ifs + 1] += fiy;
 333                             f[ifs + 2] += fiz;
 334                             fshift[ish3]     = fshift[ish3] + fix;
 335                             fshift[ish3 + 1] = fshift[ish3 + 1] + fiy;
 336                             fshift[ish3 + 2] = fshift[ish3 + 2] + fiz;
 337
 338                             /* Count in half work-units.
 339                              * In CUDA one work-unit is 2 warps.
 340                              */
 341                             if ((ic + 1) % (c_clSize / c_nbnxnGpuClusterpairSplit) == 0)
 342                             {
 343                                 npair_tot += npair;
 344
 345                                 nhwu++;
 346                                 if (within_rlist)
 347                                 {
 348                                     nhwu_pruned++;
 349                                 }
 350
 351                                 within_rlist = FALSE;
 352                                 npair        = 0;
 353                             }
 354                         }
 355                     }
 356                 }
 357             }
 358         }
 359
 360         if (stepWork.computeEnergy)
 361         {
 362             ggid       = 0;
 363             Vc[ggid]   = Vc[ggid] + vctot;
 364             Vvdw[ggid] = Vvdw[ggid] + Vvdwtot;
 365         }
 366     }
 367
 368     if (debug)
 369     {
 370         fprintf(debug,
 371                 "number of half %dx%d atom pairs: %d after pruning: %d fraction %4.2f\n",
 372                 nbl->na_ci,
 373                 nbl->na_ci,
 374                 nhwu,
 375                 nhwu_pruned,
 376                 nhwu_pruned / static_cast<double>(nhwu));
 377         fprintf(debug, "generic kernel pair interactions:            %d\n", nhwu * nbl->na_ci / 2 * nbl->na_ci);
 378         fprintf(debug,
 379                 "generic kernel post-prune pair interactions: %d\n",
 380                 nhwu_pruned * nbl->na_ci / 2 * nbl->na_ci);
 381         fprintf(debug, "generic kernel non-zero pair interactions:   %d\n", npair_tot);
 382         fprintf(debug,
 383                 "ratio non-zero/post-prune pair interactions: %4.2f\n",
 384                 npair_tot / static_cast<double>(nhwu_pruned * gmx::exactDiv(nbl->na_ci, 2) * nbl->na_ci));
 385     }
 386 }