src/gromacs/gmxlib/nonbonded/nb_free_energy.cpp

   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) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017, 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 #include "gmxpre.h"
  38
  39 #include "nb_free_energy.h"
  40
  41 #include <cmath>
  42
  43 #include <algorithm>
  44
  45 #include "gromacs/gmxlib/nrnb.h"
  46 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  47 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  48 #include "gromacs/math/functions.h"
  49 #include "gromacs/math/vec.h"
  50 #include "gromacs/mdtypes/forcerec.h"
  51 #include "gromacs/mdtypes/md_enums.h"
  52 #include "gromacs/utility/fatalerror.h"
  53
  54 void
  55 gmx_nb_free_energy_kernel(const t_nblist * gmx_restrict    nlist,
  56                           rvec * gmx_restrict              xx,
  57                           rvec * gmx_restrict              ff,
  58                           t_forcerec * gmx_restrict        fr,
  59                           const t_mdatoms * gmx_restrict   mdatoms,
  60                           nb_kernel_data_t * gmx_restrict  kernel_data,
  61                           t_nrnb * gmx_restrict            nrnb)
  62 {
  63
  64 #define  STATE_A  0
  65 #define  STATE_B  1
  66 #define  NSTATES  2
  67     int           i, n, ii, is3, ii3, k, nj0, nj1, jnr, j3, ggid;
  68     real          shX, shY, shZ;
  69     real          tx, ty, tz, Fscal;
  70     double        FscalC[NSTATES], FscalV[NSTATES];  /* Needs double for sc_power==48 */
  71     double        Vcoul[NSTATES], Vvdw[NSTATES];     /* Needs double for sc_power==48 */
  72     real          rinv6, r, rtC, rtV;
  73     real          iqA, iqB;
  74     real          qq[NSTATES], vctot;
  75     int           ntiA, ntiB, tj[NSTATES];
  76     real          Vvdw6, Vvdw12, vvtot;
  77     real          ix, iy, iz, fix, fiy, fiz;
  78     real          dx, dy, dz, rsq, rinv;
  79     real          c6[NSTATES], c12[NSTATES], c6grid;
  80     real          LFC[NSTATES], LFV[NSTATES], DLF[NSTATES];
  81     double        dvdl_coul, dvdl_vdw;
  82     real          lfac_coul[NSTATES], dlfac_coul[NSTATES], lfac_vdw[NSTATES], dlfac_vdw[NSTATES];
  83     real          sigma6[NSTATES], alpha_vdw_eff, alpha_coul_eff, sigma2_def, sigma2_min;
  84     double        rp, rpm2, rC, rV, rinvC, rpinvC, rinvV, rpinvV; /* Needs double for sc_power==48 */
  85     real          sigma2[NSTATES], sigma_pow[NSTATES];
  86     int           do_tab, tab_elemsize = 0;
  87     int           n0, n1C, n1V, nnn;
  88     real          Y, F, Fp, Geps, Heps2, epsC, eps2C, epsV, eps2V, VV, FF;
  89     int           icoul, ivdw;
  90     int           nri;
  91     const int *   iinr;
  92     const int *   jindex;
  93     const int *   jjnr;
  94     const int *   shift;
  95     const int *   gid;
  96     const int *   typeA;
  97     const int *   typeB;
  98     int           ntype;
  99     const real *  shiftvec;
 100     real *        fshift;
 101     real          tabscale = 0;
 102     const real *  VFtab    = nullptr;
 103     const real *  x;
 104     real *        f;
 105     real          facel, krf, crf;
 106     const real *  chargeA;
 107     const real *  chargeB;
 108     real          sigma6_min, sigma6_def, lam_power, sc_r_power;
 109     real          alpha_coul, alpha_vdw, lambda_coul, lambda_vdw;
 110     real          ewcljrsq, ewclj, ewclj2, exponent, poly, vvdw_disp, vvdw_rep, sh_lj_ewald;
 111     real          ewclj6;
 112     const real *  nbfp, *nbfp_grid;
 113     real *        dvdl;
 114     real *        Vv;
 115     real *        Vc;
 116     gmx_bool      bDoForces, bDoShiftForces, bDoPotential;
 117     real          rcoulomb, rvdw, sh_invrc6;
 118     gmx_bool      bExactElecCutoff, bExactVdwCutoff, bExactCutoffAll;
 119     gmx_bool      bEwald, bEwaldLJ;
 120     real          rcutoff_max2;
 121     const real *  tab_ewald_F_lj = nullptr;
 122     const real *  tab_ewald_V_lj = nullptr;
 123     real          d, d2, sw, dsw, rinvcorr;
 124     real          elec_swV3, elec_swV4, elec_swV5, elec_swF2, elec_swF3, elec_swF4;
 125     real          vdw_swV3, vdw_swV4, vdw_swV5, vdw_swF2, vdw_swF3, vdw_swF4;
 126     gmx_bool      bConvertEwaldToCoulomb, bConvertLJEwaldToLJ6;
 127     gmx_bool      bComputeVdwInteraction, bComputeElecInteraction;
 128     const real *  ewtab = nullptr;
 129     int           ewitab;
 130     real          ewrt, eweps, ewtabscale = 0, ewtabhalfspace = 0, sh_ewald = 0;
 131
 132     const real    onetwelfth  = 1.0/12.0;
 133     const real    onesixth    = 1.0/6.0;
 134     const real    zero        = 0.0;
 135     const real    half        = 0.5;
 136     const real    one         = 1.0;
 137     const real    two         = 2.0;
 138     const real    six         = 6.0;
 139     const real    fourtyeight = 48.0;
 140
 141     x                   = xx[0];
 142     f                   = ff[0];
 143
 144     fshift              = fr->fshift[0];
 145
 146     nri                 = nlist->nri;
 147     iinr                = nlist->iinr;
 148     jindex              = nlist->jindex;
 149     jjnr                = nlist->jjnr;
 150     icoul               = nlist->ielec;
 151     ivdw                = nlist->ivdw;
 152     shift               = nlist->shift;
 153     gid                 = nlist->gid;
 154
 155     shiftvec            = fr->shift_vec[0];
 156     chargeA             = mdatoms->chargeA;
 157     chargeB             = mdatoms->chargeB;
 158     facel               = fr->epsfac;
 159     krf                 = fr->k_rf;
 160     crf                 = fr->c_rf;
 161     Vc                  = kernel_data->energygrp_elec;
 162     typeA               = mdatoms->typeA;
 163     typeB               = mdatoms->typeB;
 164     ntype               = fr->ntype;
 165     nbfp                = fr->nbfp;
 166     nbfp_grid           = fr->ljpme_c6grid;
 167     Vv                  = kernel_data->energygrp_vdw;
 168     lambda_coul         = kernel_data->lambda[efptCOUL];
 169     lambda_vdw          = kernel_data->lambda[efptVDW];
 170     dvdl                = kernel_data->dvdl;
 171     alpha_coul          = fr->sc_alphacoul;
 172     alpha_vdw           = fr->sc_alphavdw;
 173     lam_power           = fr->sc_power;
 174     sc_r_power          = fr->sc_r_power;
 175     sigma6_def          = fr->sc_sigma6_def;
 176     sigma6_min          = fr->sc_sigma6_min;
 177     bDoForces           = kernel_data->flags & GMX_NONBONDED_DO_FORCE;
 178     bDoShiftForces      = kernel_data->flags & GMX_NONBONDED_DO_SHIFTFORCE;
 179     bDoPotential        = kernel_data->flags & GMX_NONBONDED_DO_POTENTIAL;
 180
 181     rcoulomb            = fr->rcoulomb;
 182     rvdw                = fr->rvdw;
 183     sh_invrc6           = fr->ic->sh_invrc6;
 184     sh_lj_ewald         = fr->ic->sh_lj_ewald;
 185     ewclj               = fr->ewaldcoeff_lj;
 186     ewclj2              = ewclj*ewclj;
 187     ewclj6              = ewclj2*ewclj2*ewclj2;
 188
 189     if (fr->coulomb_modifier == eintmodPOTSWITCH)
 190     {
 191         d               = fr->rcoulomb-fr->rcoulomb_switch;
 192         elec_swV3       = -10.0/(d*d*d);
 193         elec_swV4       =  15.0/(d*d*d*d);
 194         elec_swV5       =  -6.0/(d*d*d*d*d);
 195         elec_swF2       = -30.0/(d*d*d);
 196         elec_swF3       =  60.0/(d*d*d*d);
 197         elec_swF4       = -30.0/(d*d*d*d*d);
 198     }
 199     else
 200     {
 201         /* Avoid warnings from stupid compilers (looking at you, Clang!) */
 202         elec_swV3 = elec_swV4 = elec_swV5 = elec_swF2 = elec_swF3 = elec_swF4 = 0.0;
 203     }
 204
 205     if (fr->vdw_modifier == eintmodPOTSWITCH)
 206     {
 207         d               = fr->rvdw-fr->rvdw_switch;
 208         vdw_swV3        = -10.0/(d*d*d);
 209         vdw_swV4        =  15.0/(d*d*d*d);
 210         vdw_swV5        =  -6.0/(d*d*d*d*d);
 211         vdw_swF2        = -30.0/(d*d*d);
 212         vdw_swF3        =  60.0/(d*d*d*d);
 213         vdw_swF4        = -30.0/(d*d*d*d*d);
 214     }
 215     else
 216     {
 217         /* Avoid warnings from stupid compilers (looking at you, Clang!) */
 218         vdw_swV3 = vdw_swV4 = vdw_swV5 = vdw_swF2 = vdw_swF3 = vdw_swF4 = 0.0;
 219     }
 220
 221     if (fr->cutoff_scheme == ecutsVERLET)
 222     {
 223         const interaction_const_t *ic;
 224
 225         ic = fr->ic;
 226         if (EVDW_PME(ic->vdwtype))
 227         {
 228             ivdw         = GMX_NBKERNEL_VDW_LJEWALD;
 229         }
 230         else
 231         {
 232             ivdw         = GMX_NBKERNEL_VDW_LENNARDJONES;
 233         }
 234
 235         if (ic->eeltype == eelCUT || EEL_RF(ic->eeltype))
 236         {
 237             icoul        = GMX_NBKERNEL_ELEC_REACTIONFIELD;
 238         }
 239         else if (EEL_PME_EWALD(ic->eeltype))
 240         {
 241             icoul        = GMX_NBKERNEL_ELEC_EWALD;
 242         }
 243         else
 244         {
 245             gmx_incons("Unsupported eeltype with Verlet and free-energy");
 246         }
 247
 248         bExactElecCutoff = TRUE;
 249         bExactVdwCutoff  = TRUE;
 250     }
 251     else
 252     {
 253         bExactElecCutoff = (fr->coulomb_modifier != eintmodNONE) || fr->eeltype == eelRF_ZERO;
 254         bExactVdwCutoff  = (fr->vdw_modifier != eintmodNONE);
 255     }
 256
 257     bExactCutoffAll = (bExactElecCutoff && bExactVdwCutoff);
 258     rcutoff_max2    = std::max(fr->rcoulomb, fr->rvdw);
 259     rcutoff_max2    = rcutoff_max2*rcutoff_max2;
 260
 261     bEwald          = (icoul == GMX_NBKERNEL_ELEC_EWALD);
 262     bEwaldLJ        = (ivdw == GMX_NBKERNEL_VDW_LJEWALD);
 263
 264     if (bEwald || bEwaldLJ)
 265     {
 266         sh_ewald       = fr->ic->sh_ewald;
 267         ewtab          = fr->ic->tabq_coul_FDV0;
 268         ewtabscale     = fr->ic->tabq_scale;
 269         ewtabhalfspace = half/ewtabscale;
 270         tab_ewald_F_lj = fr->ic->tabq_vdw_F;
 271         tab_ewald_V_lj = fr->ic->tabq_vdw_V;
 272     }
 273
 274     /* For Ewald/PME interactions we cannot easily apply the soft-core component to
 275      * reciprocal space. When we use vanilla (not switch/shift) Ewald interactions, we
 276      * can apply the small trick of subtracting the _reciprocal_ space contribution
 277      * in this kernel, and instead apply the free energy interaction to the 1/r
 278      * (standard coulomb) interaction.
 279      *
 280      * However, we cannot use this approach for switch-modified since we would then
 281      * effectively end up evaluating a significantly different interaction here compared to the
 282      * normal (non-free-energy) kernels, either by applying a cutoff at a different
 283      * position than what the user requested, or by switching different
 284      * things (1/r rather than short-range Ewald). For these settings, we just
 285      * use the traditional short-range Ewald interaction in that case.
 286      */
 287     bConvertEwaldToCoulomb = (bEwald && (fr->coulomb_modifier != eintmodPOTSWITCH));
 288     /* For now the below will always be true (since LJ-PME only works with Shift in Gromacs-5.0),
 289      * but writing it this way means we stay in sync with coulomb, and it avoids future bugs.
 290      */
 291     bConvertLJEwaldToLJ6   = (bEwaldLJ && (fr->vdw_modifier   != eintmodPOTSWITCH));
 292
 293     /* We currently don't implement exclusion correction, needed with the Verlet cut-off scheme, without conversion */
 294     if (fr->cutoff_scheme == ecutsVERLET &&
 295         ((bEwald   && !bConvertEwaldToCoulomb) ||
 296          (bEwaldLJ && !bConvertLJEwaldToLJ6)))
 297     {
 298         gmx_incons("Unimplemented non-bonded setup");
 299     }
 300
 301     /* fix compiler warnings */
 302     n1C   = n1V   = 0;
 303     epsC  = epsV  = 0;
 304     eps2C = eps2V = 0;
 305
 306     dvdl_coul  = 0;
 307     dvdl_vdw   = 0;
 308
 309     /* Lambda factor for state A, 1-lambda*/
 310     LFC[STATE_A] = one - lambda_coul;
 311     LFV[STATE_A] = one - lambda_vdw;
 312
 313     /* Lambda factor for state B, lambda*/
 314     LFC[STATE_B] = lambda_coul;
 315     LFV[STATE_B] = lambda_vdw;
 316
 317     /*derivative of the lambda factor for state A and B */
 318     DLF[STATE_A] = -1;
 319     DLF[STATE_B] = 1;
 320
 321     for (i = 0; i < NSTATES; i++)
 322     {
 323         lfac_coul[i]  = (lam_power == 2 ? (1-LFC[i])*(1-LFC[i]) : (1-LFC[i]));
 324         dlfac_coul[i] = DLF[i]*lam_power/sc_r_power*(lam_power == 2 ? (1-LFC[i]) : 1);
 325         lfac_vdw[i]   = (lam_power == 2 ? (1-LFV[i])*(1-LFV[i]) : (1-LFV[i]));
 326         dlfac_vdw[i]  = DLF[i]*lam_power/sc_r_power*(lam_power == 2 ? (1-LFV[i]) : 1);
 327     }
 328     /* precalculate */
 329     sigma2_def = std::cbrt(sigma6_def);
 330     sigma2_min = std::cbrt(sigma6_min);
 331
 332     /* Ewald (not PME) table is special (icoul==enbcoulFEWALD) */
 333
 334     do_tab = (icoul == GMX_NBKERNEL_ELEC_CUBICSPLINETABLE ||
 335               ivdw == GMX_NBKERNEL_VDW_CUBICSPLINETABLE);
 336     if (do_tab)
 337     {
 338         tabscale         = kernel_data->table_elec_vdw->scale;
 339         VFtab            = kernel_data->table_elec_vdw->data;
 340         /* we always use the combined table here */
 341         tab_elemsize     = kernel_data->table_elec_vdw->stride;
 342     }
 343
 344     for (n = 0; (n < nri); n++)
 345     {
 346         int npair_within_cutoff;
 347
 348         npair_within_cutoff = 0;
 349
 350         is3              = 3*shift[n];
 351         shX              = shiftvec[is3];
 352         shY              = shiftvec[is3+1];
 353         shZ              = shiftvec[is3+2];
 354         nj0              = jindex[n];
 355         nj1              = jindex[n+1];
 356         ii               = iinr[n];
 357         ii3              = 3*ii;
 358         ix               = shX + x[ii3+0];
 359         iy               = shY + x[ii3+1];
 360         iz               = shZ + x[ii3+2];
 361         iqA              = facel*chargeA[ii];
 362         iqB              = facel*chargeB[ii];
 363         ntiA             = 2*ntype*typeA[ii];
 364         ntiB             = 2*ntype*typeB[ii];
 365         vctot            = 0;
 366         vvtot            = 0;
 367         fix              = 0;
 368         fiy              = 0;
 369         fiz              = 0;
 370
 371         for (k = nj0; (k < nj1); k++)
 372         {
 373             jnr              = jjnr[k];
 374             j3               = 3*jnr;
 375             dx               = ix - x[j3];
 376             dy               = iy - x[j3+1];
 377             dz               = iz - x[j3+2];
 378             rsq              = dx*dx + dy*dy + dz*dz;
 379
 380             if (bExactCutoffAll && rsq >= rcutoff_max2)
 381             {
 382                 /* We save significant time by skipping all code below.
 383                  * Note that with soft-core interactions, the actual cut-off
 384                  * check might be different. But since the soft-core distance
 385                  * is always larger than r, checking on r here is safe.
 386                  */
 387                 continue;
 388             }
 389             npair_within_cutoff++;
 390
 391             if (rsq > 0)
 392             {
 393                 /* Note that unlike in the nbnxn kernels, we do not need
 394                  * to clamp the value of rsq before taking the invsqrt
 395                  * to avoid NaN in the LJ calculation, since here we do
 396                  * not calculate LJ interactions when C6 and C12 are zero.
 397                  */
 398
 399                 rinv         = gmx::invsqrt(rsq);
 400                 r            = rsq*rinv;
 401             }
 402             else
 403             {
 404                 /* The force at r=0 is zero, because of symmetry.
 405                  * But note that the potential is in general non-zero,
 406                  * since the soft-cored r will be non-zero.
 407                  */
 408                 rinv         = 0;
 409                 r            = 0;
 410             }
 411
 412             if (sc_r_power == six)
 413             {
 414                 rpm2             = rsq*rsq;  /* r4 */
 415                 rp               = rpm2*rsq; /* r6 */
 416             }
 417             else if (sc_r_power == fourtyeight)
 418             {
 419                 rp               = rsq*rsq*rsq; /* r6 */
 420                 rp               = rp*rp;       /* r12 */
 421                 rp               = rp*rp;       /* r24 */
 422                 rp               = rp*rp;       /* r48 */
 423                 rpm2             = rp/rsq;      /* r46 */
 424             }
 425             else
 426             {
 427                 rp             = std::pow(r, sc_r_power);  /* not currently supported as input, but can handle it */
 428                 rpm2           = rp/rsq;
 429             }
 430
 431             Fscal = 0;
 432
 433             qq[STATE_A]      = iqA*chargeA[jnr];
 434             qq[STATE_B]      = iqB*chargeB[jnr];
 435
 436             tj[STATE_A]      = ntiA+2*typeA[jnr];
 437             tj[STATE_B]      = ntiB+2*typeB[jnr];
 438
 439             if (nlist->excl_fep == nullptr || nlist->excl_fep[k])
 440             {
 441                 c6[STATE_A]      = nbfp[tj[STATE_A]];
 442                 c6[STATE_B]      = nbfp[tj[STATE_B]];
 443
 444                 for (i = 0; i < NSTATES; i++)
 445                 {
 446                     c12[i]             = nbfp[tj[i]+1];
 447                     if ((c6[i] > 0) && (c12[i] > 0))
 448                     {
 449                         /* c12 is stored scaled with 12.0 and c6 is scaled with 6.0 - correct for this */
 450                         sigma6[i]       = half*c12[i]/c6[i];
 451                         sigma2[i]       = std::cbrt(sigma6[i]);
 452                         /* should be able to get rid of cbrt call eventually.  Will require agreement on
 453                            what data to store externally.  Can't be fixed without larger scale changes, so not 4.6 */
 454                         if (sigma6[i] < sigma6_min)   /* for disappearing coul and vdw with soft core at the same time */
 455                         {
 456                             sigma6[i] = sigma6_min;
 457                             sigma2[i] = sigma2_min;
 458                         }
 459                     }
 460                     else
 461                     {
 462                         sigma6[i]       = sigma6_def;
 463                         sigma2[i]       = sigma2_def;
 464                     }
 465                     if (sc_r_power == six)
 466                     {
 467                         sigma_pow[i]    = sigma6[i];
 468                     }
 469                     else if (sc_r_power == fourtyeight)
 470                     {
 471                         sigma_pow[i]    = sigma6[i]*sigma6[i];       /* sigma^12 */
 472                         sigma_pow[i]    = sigma_pow[i]*sigma_pow[i]; /* sigma^24 */
 473                         sigma_pow[i]    = sigma_pow[i]*sigma_pow[i]; /* sigma^48 */
 474                     }
 475                     else
 476                     {    /* not really supported as input, but in here for testing the general case*/
 477                         sigma_pow[i]    = std::pow(sigma2[i], sc_r_power/2);
 478                     }
 479                 }
 480
 481                 /* only use softcore if one of the states has a zero endstate - softcore is for avoiding infinities!*/
 482                 if ((c12[STATE_A] > 0) && (c12[STATE_B] > 0))
 483                 {
 484                     alpha_vdw_eff    = 0;
 485                     alpha_coul_eff   = 0;
 486                 }
 487                 else
 488                 {
 489                     alpha_vdw_eff    = alpha_vdw;
 490                     alpha_coul_eff   = alpha_coul;
 491                 }
 492
 493                 for (i = 0; i < NSTATES; i++)
 494                 {
 495                     FscalC[i]    = 0;
 496                     FscalV[i]    = 0;
 497                     Vcoul[i]     = 0;
 498                     Vvdw[i]      = 0;
 499
 500                     /* Only spend time on A or B state if it is non-zero */
 501                     if ( (qq[i] != 0) || (c6[i] != 0) || (c12[i] != 0) )
 502                     {
 503                         /* this section has to be inside the loop because of the dependence on sigma_pow */
 504                         rpinvC         = one/(alpha_coul_eff*lfac_coul[i]*sigma_pow[i]+rp);
 505                         rinvC          = std::pow(rpinvC, one/sc_r_power);
 506                         rC             = one/rinvC;
 507
 508                         rpinvV         = one/(alpha_vdw_eff*lfac_vdw[i]*sigma_pow[i]+rp);
 509                         rinvV          = std::pow(rpinvV, one/sc_r_power);
 510                         rV             = one/rinvV;
 511
 512                         if (do_tab)
 513                         {
 514                             rtC        = rC*tabscale;
 515                             n0         = rtC;
 516                             epsC       = rtC-n0;
 517                             eps2C      = epsC*epsC;
 518                             n1C        = tab_elemsize*n0;
 519
 520                             rtV        = rV*tabscale;
 521                             n0         = rtV;
 522                             epsV       = rtV-n0;
 523                             eps2V      = epsV*epsV;
 524                             n1V        = tab_elemsize*n0;
 525                         }
 526
 527                         /* Only process the coulomb interactions if we have charges,
 528                          * and if we either include all entries in the list (no cutoff
 529                          * used in the kernel), or if we are within the cutoff.
 530                          */
 531                         bComputeElecInteraction = !bExactElecCutoff ||
 532                             ( bConvertEwaldToCoulomb && r < rcoulomb) ||
 533                             (!bConvertEwaldToCoulomb && rC < rcoulomb);
 534
 535                         if ( (qq[i] != 0) && bComputeElecInteraction)
 536                         {
 537                             switch (icoul)
 538                             {
 539                                 case GMX_NBKERNEL_ELEC_COULOMB:
 540                                     /* simple cutoff */
 541                                     Vcoul[i]   = qq[i]*rinvC;
 542                                     FscalC[i]  = Vcoul[i];
 543                                     /* The shift for the Coulomb potential is stored in
 544                                      * the RF parameter c_rf, which is 0 without shift.
 545                                      */
 546                                     Vcoul[i]  -= qq[i]*fr->ic->c_rf;
 547                                     break;
 548
 549                                 case GMX_NBKERNEL_ELEC_REACTIONFIELD:
 550                                     /* reaction-field */
 551                                     Vcoul[i]   = qq[i]*(rinvC + krf*rC*rC-crf);
 552                                     FscalC[i]  = qq[i]*(rinvC - two*krf*rC*rC);
 553                                     break;
 554
 555                                 case GMX_NBKERNEL_ELEC_CUBICSPLINETABLE:
 556                                     /* non-Ewald tabulated coulomb */
 557                                     nnn        = n1C;
 558                                     Y          = VFtab[nnn];
 559                                     F          = VFtab[nnn+1];
 560                                     Geps       = epsC*VFtab[nnn+2];
 561                                     Heps2      = eps2C*VFtab[nnn+3];
 562                                     Fp         = F+Geps+Heps2;
 563                                     VV         = Y+epsC*Fp;
 564                                     FF         = Fp+Geps+two*Heps2;
 565                                     Vcoul[i]   = qq[i]*VV;
 566                                     FscalC[i]  = -qq[i]*tabscale*FF*rC;
 567                                     break;
 568
 569                                 case GMX_NBKERNEL_ELEC_GENERALIZEDBORN:
 570                                     gmx_fatal(FARGS, "Free energy and GB not implemented.\n");
 571                                     break;
 572
 573                                 case GMX_NBKERNEL_ELEC_EWALD:
 574                                     if (bConvertEwaldToCoulomb)
 575                                     {
 576                                         /* Ewald FEP is done only on the 1/r part */
 577                                         Vcoul[i]   = qq[i]*(rinvC-sh_ewald);
 578                                         FscalC[i]  = qq[i]*rinvC;
 579                                     }
 580                                     else
 581                                     {
 582                                         ewrt      = rC*ewtabscale;
 583                                         ewitab    = static_cast<int>(ewrt);
 584                                         eweps     = ewrt-ewitab;
 585                                         ewitab    = 4*ewitab;
 586                                         FscalC[i] = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 587                                         rinvcorr  = rinvC-sh_ewald;
 588                                         Vcoul[i]  = qq[i]*(rinvcorr-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+FscalC[i])));
 589                                         FscalC[i] = qq[i]*(rinvC-rC*FscalC[i]);
 590                                     }
 591                                     break;
 592
 593                                 case GMX_NBKERNEL_ELEC_NONE:
 594                                     FscalC[i]  = zero;
 595                                     Vcoul[i]   = zero;
 596                                     break;
 597
 598                                 default:
 599                                     gmx_incons("Invalid icoul in free energy kernel");
 600                                     break;
 601                             }
 602
 603                             if (fr->coulomb_modifier == eintmodPOTSWITCH)
 604                             {
 605                                 d                = rC-fr->rcoulomb_switch;
 606                                 d                = (d > zero) ? d : zero;
 607                                 d2               = d*d;
 608                                 sw               = one+d2*d*(elec_swV3+d*(elec_swV4+d*elec_swV5));
 609                                 dsw              = d2*(elec_swF2+d*(elec_swF3+d*elec_swF4));
 610
 611                                 FscalC[i]        = FscalC[i]*sw - rC*Vcoul[i]*dsw;
 612                                 Vcoul[i]        *= sw;
 613
 614                                 FscalC[i]        = (rC < rcoulomb) ? FscalC[i] : zero;
 615                                 Vcoul[i]         = (rC < rcoulomb) ? Vcoul[i] : zero;
 616                             }
 617                         }
 618
 619                         /* Only process the VDW interactions if we have
 620                          * some non-zero parameters, and if we either
 621                          * include all entries in the list (no cutoff used
 622                          * in the kernel), or if we are within the cutoff.
 623                          */
 624                         bComputeVdwInteraction = !bExactVdwCutoff ||
 625                             ( bConvertLJEwaldToLJ6 && r < rvdw) ||
 626                             (!bConvertLJEwaldToLJ6 && rV < rvdw);
 627                         if ((c6[i] != 0 || c12[i] != 0) && bComputeVdwInteraction)
 628                         {
 629                             switch (ivdw)
 630                             {
 631                                 case GMX_NBKERNEL_VDW_LENNARDJONES:
 632                                     /* cutoff LJ */
 633                                     if (sc_r_power == six)
 634                                     {
 635                                         rinv6            = rpinvV;
 636                                     }
 637                                     else
 638                                     {
 639                                         rinv6            = rinvV*rinvV;
 640                                         rinv6            = rinv6*rinv6*rinv6;
 641                                     }
 642                                     Vvdw6            = c6[i]*rinv6;
 643                                     Vvdw12           = c12[i]*rinv6*rinv6;
 644
 645                                     Vvdw[i]          = ( (Vvdw12 - c12[i]*sh_invrc6*sh_invrc6)*onetwelfth
 646                                                          - (Vvdw6 - c6[i]*sh_invrc6)*onesixth);
 647                                     FscalV[i]        = Vvdw12 - Vvdw6;
 648                                     break;
 649
 650                                 case GMX_NBKERNEL_VDW_BUCKINGHAM:
 651                                     gmx_fatal(FARGS, "Buckingham free energy not supported.");
 652                                     break;
 653
 654                                 case GMX_NBKERNEL_VDW_CUBICSPLINETABLE:
 655                                     /* Table LJ */
 656                                     nnn = n1V+4;
 657                                     /* dispersion */
 658                                     Y          = VFtab[nnn];
 659                                     F          = VFtab[nnn+1];
 660                                     Geps       = epsV*VFtab[nnn+2];
 661                                     Heps2      = eps2V*VFtab[nnn+3];
 662                                     Fp         = F+Geps+Heps2;
 663                                     VV         = Y+epsV*Fp;
 664                                     FF         = Fp+Geps+two*Heps2;
 665                                     Vvdw[i]   += c6[i]*VV;
 666                                     FscalV[i] -= c6[i]*tabscale*FF*rV;
 667
 668                                     /* repulsion */
 669                                     Y          = VFtab[nnn+4];
 670                                     F          = VFtab[nnn+5];
 671                                     Geps       = epsV*VFtab[nnn+6];
 672                                     Heps2      = eps2V*VFtab[nnn+7];
 673                                     Fp         = F+Geps+Heps2;
 674                                     VV         = Y+epsV*Fp;
 675                                     FF         = Fp+Geps+two*Heps2;
 676                                     Vvdw[i]   += c12[i]*VV;
 677                                     FscalV[i] -= c12[i]*tabscale*FF*rV;
 678                                     break;
 679
 680                                 case GMX_NBKERNEL_VDW_LJEWALD:
 681                                     if (sc_r_power == six)
 682                                     {
 683                                         rinv6            = rpinvV;
 684                                     }
 685                                     else
 686                                     {
 687                                         rinv6            = rinvV*rinvV;
 688                                         rinv6            = rinv6*rinv6*rinv6;
 689                                     }
 690                                     c6grid           = nbfp_grid[tj[i]];
 691
 692                                     if (bConvertLJEwaldToLJ6)
 693                                     {
 694                                         /* cutoff LJ */
 695                                         Vvdw6            = c6[i]*rinv6;
 696                                         Vvdw12           = c12[i]*rinv6*rinv6;
 697
 698                                         Vvdw[i]          = ( (Vvdw12 - c12[i]*sh_invrc6*sh_invrc6)*onetwelfth
 699                                                              - (Vvdw6 - c6[i]*sh_invrc6 - c6grid*sh_lj_ewald)*onesixth);
 700                                         FscalV[i]        = Vvdw12 - Vvdw6;
 701                                     }
 702                                     else
 703                                     {
 704                                         /* Normal LJ-PME */
 705                                         ewcljrsq         = ewclj2*rV*rV;
 706                                         exponent         = std::exp(-ewcljrsq);
 707                                         poly             = exponent*(one + ewcljrsq + ewcljrsq*ewcljrsq*half);
 708                                         vvdw_disp        = (c6[i]-c6grid*(one-poly))*rinv6;
 709                                         vvdw_rep         = c12[i]*rinv6*rinv6;
 710                                         FscalV[i]        = vvdw_rep - vvdw_disp - c6grid*onesixth*exponent*ewclj6;
 711                                         Vvdw[i]          = (vvdw_rep - c12[i]*sh_invrc6*sh_invrc6)*onetwelfth - (vvdw_disp - c6[i]*sh_invrc6 - c6grid*sh_lj_ewald)/six;
 712                                     }
 713                                     break;
 714
 715                                 case GMX_NBKERNEL_VDW_NONE:
 716                                     Vvdw[i]    = zero;
 717                                     FscalV[i]  = zero;
 718                                     break;
 719
 720                                 default:
 721                                     gmx_incons("Invalid ivdw in free energy kernel");
 722                                     break;
 723                             }
 724
 725                             if (fr->vdw_modifier == eintmodPOTSWITCH)
 726                             {
 727                                 d                = rV-fr->rvdw_switch;
 728                                 d                = (d > zero) ? d : zero;
 729                                 d2               = d*d;
 730                                 sw               = one+d2*d*(vdw_swV3+d*(vdw_swV4+d*vdw_swV5));
 731                                 dsw              = d2*(vdw_swF2+d*(vdw_swF3+d*vdw_swF4));
 732
 733                                 FscalV[i]        = FscalV[i]*sw - rV*Vvdw[i]*dsw;
 734                                 Vvdw[i]         *= sw;
 735
 736                                 FscalV[i]  = (rV < rvdw) ? FscalV[i] : zero;
 737                                 Vvdw[i]    = (rV < rvdw) ? Vvdw[i] : zero;
 738                             }
 739                         }
 740
 741                         /* FscalC (and FscalV) now contain: dV/drC * rC
 742                          * Now we multiply by rC^-p, so it will be: dV/drC * rC^1-p
 743                          * Further down we first multiply by r^p-2 and then by
 744                          * the vector r, which in total gives: dV/drC * (r/rC)^1-p
 745                          */
 746                         FscalC[i] *= rpinvC;
 747                         FscalV[i] *= rpinvV;
 748                     }
 749                 }
 750
 751                 /* Assemble A and B states */
 752                 for (i = 0; i < NSTATES; i++)
 753                 {
 754                     vctot         += LFC[i]*Vcoul[i];
 755                     vvtot         += LFV[i]*Vvdw[i];
 756
 757                     Fscal         += LFC[i]*FscalC[i]*rpm2;
 758                     Fscal         += LFV[i]*FscalV[i]*rpm2;
 759
 760                     dvdl_coul     += Vcoul[i]*DLF[i] + LFC[i]*alpha_coul_eff*dlfac_coul[i]*FscalC[i]*sigma_pow[i];
 761                     dvdl_vdw      += Vvdw[i]*DLF[i] + LFV[i]*alpha_vdw_eff*dlfac_vdw[i]*FscalV[i]*sigma_pow[i];
 762                 }
 763             }
 764             else if (icoul == GMX_NBKERNEL_ELEC_REACTIONFIELD)
 765             {
 766                 /* For excluded pairs, which are only in this pair list when
 767                  * using the Verlet scheme, we don't use soft-core.
 768                  * The group scheme also doesn't soft-core for these.
 769                  * As there is no singularity, there is no need for soft-core.
 770                  */
 771                 VV = krf*rsq - crf;
 772                 FF = -two*krf;
 773
 774                 if (ii == jnr)
 775                 {
 776                     VV *= half;
 777                 }
 778
 779                 for (i = 0; i < NSTATES; i++)
 780                 {
 781                     vctot      += LFC[i]*qq[i]*VV;
 782                     Fscal      += LFC[i]*qq[i]*FF;
 783                     dvdl_coul  += DLF[i]*qq[i]*VV;
 784                 }
 785             }
 786
 787             if (bConvertEwaldToCoulomb && ( !bExactElecCutoff || r < rcoulomb ) )
 788             {
 789                 /* See comment in the preamble. When using Ewald interactions
 790                  * (unless we use a switch modifier) we subtract the reciprocal-space
 791                  * Ewald component here which made it possible to apply the free
 792                  * energy interaction to 1/r (vanilla coulomb short-range part)
 793                  * above. This gets us closer to the ideal case of applying
 794                  * the softcore to the entire electrostatic interaction,
 795                  * including the reciprocal-space component.
 796                  */
 797                 real v_lr, f_lr;
 798
 799                 ewrt      = r*ewtabscale;
 800                 ewitab    = static_cast<int>(ewrt);
 801                 eweps     = ewrt-ewitab;
 802                 ewitab    = 4*ewitab;
 803                 f_lr      = ewtab[ewitab]+eweps*ewtab[ewitab+1];
 804                 v_lr      = (ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+f_lr));
 805                 f_lr     *= rinv;
 806
 807                 /* Note that any possible Ewald shift has already been applied in
 808                  * the normal interaction part above.
 809                  */
 810
 811                 if (ii == jnr)
 812                 {
 813                     /* If we get here, the i particle (ii) has itself (jnr)
 814                      * in its neighborlist. This can only happen with the Verlet
 815                      * scheme, and corresponds to a self-interaction that will
 816                      * occur twice. Scale it down by 50% to only include it once.
 817                      */
 818                     v_lr *= half;
 819                 }
 820
 821                 for (i = 0; i < NSTATES; i++)
 822                 {
 823                     vctot      -= LFC[i]*qq[i]*v_lr;
 824                     Fscal      -= LFC[i]*qq[i]*f_lr;
 825                     dvdl_coul  -= (DLF[i]*qq[i])*v_lr;
 826                 }
 827             }
 828
 829             if (bConvertLJEwaldToLJ6 && (!bExactVdwCutoff || r < rvdw))
 830             {
 831                 /* See comment in the preamble. When using LJ-Ewald interactions
 832                  * (unless we use a switch modifier) we subtract the reciprocal-space
 833                  * Ewald component here which made it possible to apply the free
 834                  * energy interaction to r^-6 (vanilla LJ6 short-range part)
 835                  * above. This gets us closer to the ideal case of applying
 836                  * the softcore to the entire VdW interaction,
 837                  * including the reciprocal-space component.
 838                  */
 839                 /* We could also use the analytical form here
 840                  * iso a table, but that can cause issues for
 841                  * r close to 0 for non-interacting pairs.
 842                  */
 843                 real rs, frac, f_lr;
 844                 int  ri;
 845
 846                 rs     = rsq*rinv*ewtabscale;
 847                 ri     = static_cast<int>(rs);
 848                 frac   = rs - ri;
 849                 f_lr   = (1 - frac)*tab_ewald_F_lj[ri] + frac*tab_ewald_F_lj[ri+1];
 850                 /* TODO: Currently the Ewald LJ table does not contain
 851                  * the factor 1/6, we should add this.
 852                  */
 853                 FF     = f_lr*rinv/six;
 854                 VV     = (tab_ewald_V_lj[ri] - ewtabhalfspace*frac*(tab_ewald_F_lj[ri] + f_lr))/six;
 855
 856                 if (ii == jnr)
 857                 {
 858                     /* If we get here, the i particle (ii) has itself (jnr)
 859                      * in its neighborlist. This can only happen with the Verlet
 860                      * scheme, and corresponds to a self-interaction that will
 861                      * occur twice. Scale it down by 50% to only include it once.
 862                      */
 863                     VV *= half;
 864                 }
 865
 866                 for (i = 0; i < NSTATES; i++)
 867                 {
 868                     c6grid      = nbfp_grid[tj[i]];
 869                     vvtot      += LFV[i]*c6grid*VV;
 870                     Fscal      += LFV[i]*c6grid*FF;
 871                     dvdl_vdw   += (DLF[i]*c6grid)*VV;
 872                 }
 873             }
 874
 875             if (bDoForces)
 876             {
 877                 tx         = Fscal*dx;
 878                 ty         = Fscal*dy;
 879                 tz         = Fscal*dz;
 880                 fix        = fix + tx;
 881                 fiy        = fiy + ty;
 882                 fiz        = fiz + tz;
 883                 /* OpenMP atomics are expensive, but this kernels is also
 884                  * expensive, so we can take this hit, instead of using
 885                  * thread-local output buffers and extra reduction.
 886                  *
 887                  * All the OpenMP regions in this file are trivial and should
 888                  * not throw, so no need for try/catch.
 889                  */
 890 #pragma omp atomic
 891                 f[j3]     -= tx;
 892 #pragma omp atomic
 893                 f[j3+1]   -= ty;
 894 #pragma omp atomic
 895                 f[j3+2]   -= tz;
 896             }
 897         }
 898
 899         /* The atomics below are expensive with many OpenMP threads.
 900          * Here unperturbed i-particles will usually only have a few
 901          * (perturbed) j-particles in the list. Thus with a buffered list
 902          * we can skip a significant number of i-reductions with a check.
 903          */
 904         if (npair_within_cutoff > 0)
 905         {
 906             if (bDoForces)
 907             {
 908 #pragma omp atomic
 909                 f[ii3]        += fix;
 910 #pragma omp atomic
 911                 f[ii3+1]      += fiy;
 912 #pragma omp atomic
 913                 f[ii3+2]      += fiz;
 914             }
 915             if (bDoShiftForces)
 916             {
 917 #pragma omp atomic
 918                 fshift[is3]   += fix;
 919 #pragma omp atomic
 920                 fshift[is3+1] += fiy;
 921 #pragma omp atomic
 922                 fshift[is3+2] += fiz;
 923             }
 924             if (bDoPotential)
 925             {
 926                 ggid               = gid[n];
 927 #pragma omp atomic
 928                 Vc[ggid]          += vctot;
 929 #pragma omp atomic
 930                 Vv[ggid]          += vvtot;
 931             }
 932         }
 933     }
 934
 935 #pragma omp atomic
 936     dvdl[efptCOUL]     += dvdl_coul;
 937  #pragma omp atomic
 938     dvdl[efptVDW]      += dvdl_vdw;
 939
 940     /* Estimate flops, average for free energy stuff:
 941      * 12  flops per outer iteration
 942      * 150 flops per inner iteration
 943      */
 944 #pragma omp atomic
 945     inc_nrnb(nrnb, eNR_NBKERNEL_FREE_ENERGY, nlist->nri*12 + nlist->jindex[n]*150);
 946 }