src/gromacs/mdlib/nbnxn_kernels/nbnxn_kernel_gpu_ref.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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 #include "gmxpre.h"
  36
  37 #include "config.h"
  38
  39 #include <math.h>
  40
  41 #include "gromacs/legacyheaders/types/simple.h"
  42 #include "gromacs/legacyheaders/typedefs.h"
  43 #include "gromacs/legacyheaders/force.h"
  44
  45 #include "gromacs/math/utilities.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/pbcutil/ishift.h"
  48
  49 #include "nbnxn_kernel_gpu_ref.h"
  50 #include "../nbnxn_consts.h"
  51 #include "nbnxn_kernel_common.h"
  52 #include "gromacs/mdlib/nb_verlet.h"
  53
  54 #define NCL_PER_SUPERCL         (NBNXN_GPU_NCLUSTER_PER_SUPERCLUSTER)
  55 #define CL_SIZE                 (NBNXN_GPU_CLUSTER_SIZE)
  56
  57 void
  58 nbnxn_kernel_gpu_ref(const nbnxn_pairlist_t     *nbl,
  59                      const nbnxn_atomdata_t     *nbat,
  60                      const interaction_const_t  *iconst,
  61                      rvec                       *shift_vec,
  62                      int                         force_flags,
  63                      int                         clearF,
  64                      real  *                     f,
  65                      real  *                     fshift,
  66                      real  *                     Vc,
  67                      real  *                     Vvdw)
  68 {
  69     const nbnxn_sci_t  *nbln;
  70     const real         *x;
  71     gmx_bool            bEner;
  72     gmx_bool            bEwald;
  73     const real         *Ftab = NULL;
  74     real                rcut2, rvdw2, rlist2;
  75     int                 ntype;
  76     real                facel;
  77     int                 n;
  78     int                 ish3;
  79     int                 sci;
  80     int                 cj4_ind0, cj4_ind1, cj4_ind;
  81     int                 ci, cj;
  82     int                 ic, jc, ia, ja, is, ifs, js, jfs, im, jm;
  83     int                 n0;
  84     int                 ggid;
  85     real                shX, shY, shZ;
  86     real                fscal, tx, ty, tz;
  87     real                rinvsq;
  88     real                iq;
  89     real                qq, vcoul = 0, krsq, vctot;
  90     int                 nti;
  91     int                 tj;
  92     real                rt, r, eps;
  93     real                rinvsix;
  94     real                Vvdwtot;
  95     real                Vvdw_rep, Vvdw_disp;
  96     real                ix, iy, iz, fix, fiy, fiz;
  97     real                jx, jy, jz;
  98     real                dx, dy, dz, rsq, rinv;
  99     int                 int_bit;
 100     real                fexcl;
 101     real                c6, c12, cexp1, cexp2, br;
 102     const real       *  shiftvec;
 103     real       *        vdwparam;
 104     int       *         shift;
 105     int       *         type;
 106     const nbnxn_excl_t *excl[2];
 107
 108     int                 npair_tot, npair;
 109     int                 nhwu, nhwu_pruned;
 110
 111     if (nbl->na_ci != CL_SIZE)
 112     {
 113         gmx_fatal(FARGS, "The neighborlist cluster size in the GPU reference kernel is %d, expected it to be %d", nbl->na_ci, CL_SIZE);
 114     }
 115
 116     if (clearF == enbvClearFYes)
 117     {
 118         clear_f(nbat, 0, f);
 119     }
 120
 121     bEner = (force_flags & GMX_FORCE_ENERGY);
 122
 123     bEwald = EEL_FULL(iconst->eeltype);
 124     if (bEwald)
 125     {
 126         Ftab = iconst->tabq_coul_F;
 127     }
 128
 129     rcut2               = iconst->rcoulomb*iconst->rcoulomb;
 130     rvdw2               = iconst->rvdw*iconst->rvdw;
 131
 132     rlist2              = nbl->rlist*nbl->rlist;
 133
 134     type                = nbat->type;
 135     facel               = iconst->epsfac;
 136     shiftvec            = shift_vec[0];
 137     vdwparam            = nbat->nbfp;
 138     ntype               = nbat->ntype;
 139
 140     x = nbat->x;
 141
 142     npair_tot   = 0;
 143     nhwu        = 0;
 144     nhwu_pruned = 0;
 145
 146     for (n = 0; n < nbl->nsci; n++)
 147     {
 148         nbln = &nbl->sci[n];
 149
 150         ish3             = 3*nbln->shift;
 151         shX              = shiftvec[ish3];
 152         shY              = shiftvec[ish3+1];
 153         shZ              = shiftvec[ish3+2];
 154         cj4_ind0         = nbln->cj4_ind_start;
 155         cj4_ind1         = nbln->cj4_ind_end;
 156         sci              = nbln->sci;
 157         vctot            = 0;
 158         Vvdwtot          = 0;
 159
 160         if (nbln->shift == CENTRAL &&
 161             nbl->cj4[cj4_ind0].cj[0] == sci*NCL_PER_SUPERCL)
 162         {
 163             /* we have the diagonal:
 164              * add the charge self interaction energy term
 165              */
 166             for (im = 0; im < NCL_PER_SUPERCL; im++)
 167             {
 168                 ci = sci*NCL_PER_SUPERCL + im;
 169                 for (ic = 0; ic < CL_SIZE; ic++)
 170                 {
 171                     ia     = ci*CL_SIZE + ic;
 172                     iq     = x[ia*nbat->xstride+3];
 173                     vctot += iq*iq;
 174                 }
 175             }
 176             if (!bEwald)
 177             {
 178                 vctot *= -facel*0.5*iconst->c_rf;
 179             }
 180             else
 181             {
 182                 /* last factor 1/sqrt(pi) */
 183                 vctot *= -facel*iconst->ewaldcoeff_q*M_1_SQRTPI;
 184             }
 185         }
 186
 187         for (cj4_ind = cj4_ind0; (cj4_ind < cj4_ind1); cj4_ind++)
 188         {
 189             excl[0]           = &nbl->excl[nbl->cj4[cj4_ind].imei[0].excl_ind];
 190             excl[1]           = &nbl->excl[nbl->cj4[cj4_ind].imei[1].excl_ind];
 191
 192             for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
 193             {
 194                 cj               = nbl->cj4[cj4_ind].cj[jm];
 195
 196                 for (im = 0; im < NCL_PER_SUPERCL; im++)
 197                 {
 198                     /* We're only using the first imask,
 199                      * but here imei[1].imask is identical.
 200                      */
 201                     if ((nbl->cj4[cj4_ind].imei[0].imask >> (jm*NCL_PER_SUPERCL+im)) & 1)
 202                     {
 203                         gmx_bool within_rlist;
 204
 205                         ci               = sci*NCL_PER_SUPERCL + im;
 206
 207                         within_rlist     = FALSE;
 208                         npair            = 0;
 209                         for (ic = 0; ic < CL_SIZE; ic++)
 210                         {
 211                             ia               = ci*CL_SIZE + ic;
 212
 213                             is               = ia*nbat->xstride;
 214                             ifs              = ia*nbat->fstride;
 215                             ix               = shX + x[is+0];
 216                             iy               = shY + x[is+1];
 217                             iz               = shZ + x[is+2];
 218                             iq               = facel*x[is+3];
 219                             nti              = ntype*2*type[ia];
 220
 221                             fix              = 0;
 222                             fiy              = 0;
 223                             fiz              = 0;
 224
 225                             for (jc = 0; jc < CL_SIZE; jc++)
 226                             {
 227                                 ja               = cj*CL_SIZE + jc;
 228
 229                                 if (nbln->shift == CENTRAL &&
 230                                     ci == cj && ja <= ia)
 231                                 {
 232                                     continue;
 233                                 }
 234
 235                                 int_bit = ((excl[jc>>2]->pair[(jc & 3)*CL_SIZE+ic] >> (jm*NCL_PER_SUPERCL+im)) & 1);
 236
 237                                 js               = ja*nbat->xstride;
 238                                 jfs              = ja*nbat->fstride;
 239                                 jx               = x[js+0];
 240                                 jy               = x[js+1];
 241                                 jz               = x[js+2];
 242                                 dx               = ix - jx;
 243                                 dy               = iy - jy;
 244                                 dz               = iz - jz;
 245                                 rsq              = dx*dx + dy*dy + dz*dz;
 246                                 if (rsq < rlist2)
 247                                 {
 248                                     within_rlist = TRUE;
 249                                 }
 250                                 if (rsq >= rcut2)
 251                                 {
 252                                     continue;
 253                                 }
 254
 255                                 if (type[ia] != ntype-1 && type[ja] != ntype-1)
 256                                 {
 257                                     npair++;
 258                                 }
 259
 260                                 /* avoid NaN for excluded pairs at r=0 */
 261                                 rsq             += (1.0 - int_bit)*NBNXN_AVOID_SING_R2_INC;
 262
 263                                 rinv             = gmx_invsqrt(rsq);
 264                                 rinvsq           = rinv*rinv;
 265                                 fscal            = 0;
 266
 267                                 qq               = iq*x[js+3];
 268                                 if (!bEwald)
 269                                 {
 270                                     /* Reaction-field */
 271                                     krsq  = iconst->k_rf*rsq;
 272                                     fscal = qq*(int_bit*rinv - 2*krsq)*rinvsq;
 273                                     if (bEner)
 274                                     {
 275                                         vcoul = qq*(int_bit*rinv + krsq - iconst->c_rf);
 276                                     }
 277                                 }
 278                                 else
 279                                 {
 280                                     r     = rsq*rinv;
 281                                     rt    = r*iconst->tabq_scale;
 282                                     n0    = rt;
 283                                     eps   = rt - n0;
 284
 285                                     fexcl = (1 - eps)*Ftab[n0] + eps*Ftab[n0+1];
 286
 287                                     fscal = qq*(int_bit*rinvsq - fexcl)*rinv;
 288
 289                                     if (bEner)
 290                                     {
 291                                         vcoul = qq*((int_bit - gmx_erf(iconst->ewaldcoeff_q*r))*rinv - 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 (bEner)
 309                                     {
 310                                         vctot   += vcoul;
 311
 312                                         Vvdwtot +=
 313                                             (Vvdw_rep - int_bit*c12*iconst->sh_invrc6*iconst->sh_invrc6)/12 -
 314                                             (Vvdw_disp - int_bit*c6*iconst->sh_invrc6)/6;
 315                                     }
 316                                 }
 317
 318                                 tx        = fscal*dx;
 319                                 ty        = fscal*dy;
 320                                 tz        = fscal*dz;
 321                                 fix       = fix + tx;
 322                                 fiy       = fiy + ty;
 323                                 fiz       = fiz + tz;
 324                                 f[jfs+0] -= tx;
 325                                 f[jfs+1] -= ty;
 326                                 f[jfs+2] -= tz;
 327                             }
 328
 329                             f[ifs+0]        += fix;
 330                             f[ifs+1]        += fiy;
 331                             f[ifs+2]        += fiz;
 332                             fshift[ish3]     = fshift[ish3]   + fix;
 333                             fshift[ish3+1]   = fshift[ish3+1] + fiy;
 334                             fshift[ish3+2]   = fshift[ish3+2] + fiz;
 335
 336                             /* Count in half work-units.
 337                              * In CUDA one work-unit is 2 warps.
 338                              */
 339                             if ((ic+1) % (CL_SIZE/2) == 0)
 340                             {
 341                                 npair_tot += npair;
 342
 343                                 nhwu++;
 344                                 if (within_rlist)
 345                                 {
 346                                     nhwu_pruned++;
 347                                 }
 348
 349                                 within_rlist = FALSE;
 350                                 npair        = 0;
 351                             }
 352                         }
 353                     }
 354                 }
 355             }
 356         }
 357
 358         if (bEner)
 359         {
 360             ggid             = 0;
 361             Vc[ggid]         = Vc[ggid]   + vctot;
 362             Vvdw[ggid]       = Vvdw[ggid] + Vvdwtot;
 363         }
 364     }
 365
 366     if (debug)
 367     {
 368         fprintf(debug, "number of half %dx%d atom pairs: %d after pruning: %d fraction %4.2f\n",
 369                 nbl->na_ci, nbl->na_ci,
 370                 nhwu, nhwu_pruned, nhwu_pruned/(double)nhwu);
 371         fprintf(debug, "generic kernel pair interactions:            %d\n",
 372                 nhwu*nbl->na_ci/2*nbl->na_ci);
 373         fprintf(debug, "generic kernel post-prune pair interactions: %d\n",
 374                 nhwu_pruned*nbl->na_ci/2*nbl->na_ci);
 375         fprintf(debug, "generic kernel non-zero pair interactions:   %d\n",
 376                 npair_tot);
 377         fprintf(debug, "ratio non-zero/post-prune pair interactions: %4.2f\n",
 378                 npair_tot/(double)(nhwu_pruned*nbl->na_ci/2*nbl->na_ci));
 379     }
 380 }