?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Avoid FP exception
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_free_energy.cpp
blobf23eefbf378ce20f4640701ea1e8630b19bef7f2
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, 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 = NULL;
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 rinv = gmx::invsqrt(rsq);
394 r = rsq*rinv;
395 }
396 else
397 {
398 /* The force at r=0 is zero, because of symmetry.
399 * But note that the potential is in general non-zero,
400 * since the soft-cored r will be non-zero.
401 */
402 rinv = 0;
403 r = 0;
404 }
405
406 if (sc_r_power == six)
407 {
408 rpm2 = rsq*rsq; /* r4 */
409 rp = rpm2*rsq; /* r6 */
410 }
411 else if (sc_r_power == fourtyeight)
412 {
413 rp = rsq*rsq*rsq; /* r6 */
414 rp = rp*rp; /* r12 */
415 rp = rp*rp; /* r24 */
416 rp = rp*rp; /* r48 */
417 rpm2 = rp/rsq; /* r46 */
418 }
419 else
420 {
421 rp = std::pow(r, sc_r_power); /* not currently supported as input, but can handle it */
422 rpm2 = rp/rsq;
423 }
424
425 Fscal = 0;
426
427 qq[STATE_A] = iqA*chargeA[jnr];
428 qq[STATE_B] = iqB*chargeB[jnr];
429
430 tj[STATE_A] = ntiA+2*typeA[jnr];
431 tj[STATE_B] = ntiB+2*typeB[jnr];
432
433 if (nlist->excl_fep == NULL || nlist->excl_fep[k])
434 {
435 c6[STATE_A] = nbfp[tj[STATE_A]];
436 c6[STATE_B] = nbfp[tj[STATE_B]];
437
438 for (i = 0; i < NSTATES; i++)
439 {
440 c12[i] = nbfp[tj[i]+1];
441 if ((c6[i] > 0) && (c12[i] > 0))
442 {
443 /* c12 is stored scaled with 12.0 and c6 is scaled with 6.0 - correct for this */
444 sigma6[i] = half*c12[i]/c6[i];
445 sigma2[i] = std::cbrt(sigma6[i]);
446 /* should be able to get rid of cbrt call eventually. Will require agreement on
447 what data to store externally. Can't be fixed without larger scale changes, so not 4.6 */
448 if (sigma6[i] < sigma6_min) /* for disappearing coul and vdw with soft core at the same time */
449 {
450 sigma6[i] = sigma6_min;
451 sigma2[i] = sigma2_min;
452 }
453 }
454 else
455 {
456 sigma6[i] = sigma6_def;
457 sigma2[i] = sigma2_def;
458 }
459 if (sc_r_power == six)
460 {
461 sigma_pow[i] = sigma6[i];
462 }
463 else if (sc_r_power == fourtyeight)
464 {
465 sigma_pow[i] = sigma6[i]*sigma6[i]; /* sigma^12 */
466 sigma_pow[i] = sigma_pow[i]*sigma_pow[i]; /* sigma^24 */
467 sigma_pow[i] = sigma_pow[i]*sigma_pow[i]; /* sigma^48 */
468 }
469 else
470 { /* not really supported as input, but in here for testing the general case*/
471 sigma_pow[i] = std::pow(sigma2[i], sc_r_power/2);
472 }
473 }
474
475 /* only use softcore if one of the states has a zero endstate - softcore is for avoiding infinities!*/
476 if ((c12[STATE_A] > 0) && (c12[STATE_B] > 0))
477 {
478 alpha_vdw_eff = 0;
479 alpha_coul_eff = 0;
480 }
481 else
482 {
483 alpha_vdw_eff = alpha_vdw;
484 alpha_coul_eff = alpha_coul;
485 }
486
487 for (i = 0; i < NSTATES; i++)
488 {
489 FscalC[i] = 0;
490 FscalV[i] = 0;
491 Vcoul[i] = 0;
492 Vvdw[i] = 0;
493
494 /* Only spend time on A or B state if it is non-zero */
495 if ( (qq[i] != 0) || (c6[i] != 0) || (c12[i] != 0) )
496 {
497 /* this section has to be inside the loop because of the dependence on sigma_pow */
498 rpinvC = one/(alpha_coul_eff*lfac_coul[i]*sigma_pow[i]+rp);
499 rinvC = std::pow(rpinvC, one/sc_r_power);
500 rC = one/rinvC;
501
502 rpinvV = one/(alpha_vdw_eff*lfac_vdw[i]*sigma_pow[i]+rp);
503 rinvV = std::pow(rpinvV, one/sc_r_power);
504 rV = one/rinvV;
505
506 if (do_tab)
507 {
508 rtC = rC*tabscale;
509 n0 = rtC;
510 epsC = rtC-n0;
511 eps2C = epsC*epsC;
512 n1C = tab_elemsize*n0;
513
514 rtV = rV*tabscale;
515 n0 = rtV;
516 epsV = rtV-n0;
517 eps2V = epsV*epsV;
518 n1V = tab_elemsize*n0;
519 }
520
521 /* Only process the coulomb interactions if we have charges,
522 * and if we either include all entries in the list (no cutoff
523 * used in the kernel), or if we are within the cutoff.
524 */
525 bComputeElecInteraction = !bExactElecCutoff ||
526 ( bConvertEwaldToCoulomb && r < rcoulomb) ||
527 (!bConvertEwaldToCoulomb && rC < rcoulomb);
528
529 if ( (qq[i] != 0) && bComputeElecInteraction)
530 {
531 switch (icoul)
532 {
533 case GMX_NBKERNEL_ELEC_COULOMB:
534 /* simple cutoff */
535 Vcoul[i] = qq[i]*rinvC;
536 FscalC[i] = Vcoul[i];
537 /* The shift for the Coulomb potential is stored in
538 * the RF parameter c_rf, which is 0 without shift.
539 */
540 Vcoul[i] -= qq[i]*fr->ic->c_rf;
541 break;
542
543 case GMX_NBKERNEL_ELEC_REACTIONFIELD:
544 /* reaction-field */
545 Vcoul[i] = qq[i]*(rinvC + krf*rC*rC-crf);
546 FscalC[i] = qq[i]*(rinvC - two*krf*rC*rC);
547 break;
548
549 case GMX_NBKERNEL_ELEC_CUBICSPLINETABLE:
550 /* non-Ewald tabulated coulomb */
551 nnn = n1C;
552 Y = VFtab[nnn];
553 F = VFtab[nnn+1];
554 Geps = epsC*VFtab[nnn+2];
555 Heps2 = eps2C*VFtab[nnn+3];
556 Fp = F+Geps+Heps2;
557 VV = Y+epsC*Fp;
558 FF = Fp+Geps+two*Heps2;
559 Vcoul[i] = qq[i]*VV;
560 FscalC[i] = -qq[i]*tabscale*FF*rC;
561 break;
562
563 case GMX_NBKERNEL_ELEC_GENERALIZEDBORN:
564 gmx_fatal(FARGS, "Free energy and GB not implemented.\n");
565 break;
566
567 case GMX_NBKERNEL_ELEC_EWALD:
568 if (bConvertEwaldToCoulomb)
569 {
570 /* Ewald FEP is done only on the 1/r part */
571 Vcoul[i] = qq[i]*(rinvC-sh_ewald);
572 FscalC[i] = qq[i]*rinvC;
573 }
574 else
575 {
576 ewrt = rC*ewtabscale;
577 ewitab = static_cast<int>(ewrt);
578 eweps = ewrt-ewitab;
579 ewitab = 4*ewitab;
580 FscalC[i] = ewtab[ewitab]+eweps*ewtab[ewitab+1];
581 rinvcorr = rinvC-sh_ewald;
582 Vcoul[i] = qq[i]*(rinvcorr-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+FscalC[i])));
583 FscalC[i] = qq[i]*(rinvC-rC*FscalC[i]);
584 }
585 break;
586
587 case GMX_NBKERNEL_ELEC_NONE:
588 FscalC[i] = zero;
589 Vcoul[i] = zero;
590 break;
591
592 default:
593 gmx_incons("Invalid icoul in free energy kernel");
594 break;
595 }
596
597 if (fr->coulomb_modifier == eintmodPOTSWITCH)
598 {
599 d = rC-fr->rcoulomb_switch;
600 d = (d > zero) ? d : zero;
601 d2 = d*d;
602 sw = one+d2*d*(elec_swV3+d*(elec_swV4+d*elec_swV5));
603 dsw = d2*(elec_swF2+d*(elec_swF3+d*elec_swF4));
604
605 FscalC[i] = FscalC[i]*sw - rC*Vcoul[i]*dsw;
606 Vcoul[i] *= sw;
607
608 FscalC[i] = (rC < rcoulomb) ? FscalC[i] : zero;
609 Vcoul[i] = (rC < rcoulomb) ? Vcoul[i] : zero;
610 }
611 }
612
613 /* Only process the VDW interactions if we have
614 * some non-zero parameters, and if we either
615 * include all entries in the list (no cutoff used
616 * in the kernel), or if we are within the cutoff.
617 */
618 bComputeVdwInteraction = !bExactVdwCutoff ||
619 ( bConvertLJEwaldToLJ6 && r < rvdw) ||
620 (!bConvertLJEwaldToLJ6 && rV < rvdw);
621 if ((c6[i] != 0 || c12[i] != 0) && bComputeVdwInteraction)
622 {
623 switch (ivdw)
624 {
625 case GMX_NBKERNEL_VDW_LENNARDJONES:
626 /* cutoff LJ */
627 if (sc_r_power == six)
628 {
629 rinv6 = rpinvV;
630 }
631 else
632 {
633 rinv6 = rinvV*rinvV;
634 rinv6 = rinv6*rinv6*rinv6;
635 }
636 Vvdw6 = c6[i]*rinv6;
637 Vvdw12 = c12[i]*rinv6*rinv6;
638
639 Vvdw[i] = ( (Vvdw12 - c12[i]*sh_invrc6*sh_invrc6)*onetwelfth
640 - (Vvdw6 - c6[i]*sh_invrc6)*onesixth);
641 FscalV[i] = Vvdw12 - Vvdw6;
642 break;
643
644 case GMX_NBKERNEL_VDW_BUCKINGHAM:
645 gmx_fatal(FARGS, "Buckingham free energy not supported.");
646 break;
647
648 case GMX_NBKERNEL_VDW_CUBICSPLINETABLE:
649 /* Table LJ */
650 nnn = n1V+4;
651 /* dispersion */
652 Y = VFtab[nnn];
653 F = VFtab[nnn+1];
654 Geps = epsV*VFtab[nnn+2];
655 Heps2 = eps2V*VFtab[nnn+3];
656 Fp = F+Geps+Heps2;
657 VV = Y+epsV*Fp;
658 FF = Fp+Geps+two*Heps2;
659 Vvdw[i] += c6[i]*VV;
660 FscalV[i] -= c6[i]*tabscale*FF*rV;
661
662 /* repulsion */
663 Y = VFtab[nnn+4];
664 F = VFtab[nnn+5];
665 Geps = epsV*VFtab[nnn+6];
666 Heps2 = eps2V*VFtab[nnn+7];
667 Fp = F+Geps+Heps2;
668 VV = Y+epsV*Fp;
669 FF = Fp+Geps+two*Heps2;
670 Vvdw[i] += c12[i]*VV;
671 FscalV[i] -= c12[i]*tabscale*FF*rV;
672 break;
673
674 case GMX_NBKERNEL_VDW_LJEWALD:
675 if (sc_r_power == six)
676 {
677 rinv6 = rpinvV;
678 }
679 else
680 {
681 rinv6 = rinvV*rinvV;
682 rinv6 = rinv6*rinv6*rinv6;
683 }
684 c6grid = nbfp_grid[tj[i]];
685
686 if (bConvertLJEwaldToLJ6)
687 {
688 /* cutoff LJ */
689 Vvdw6 = c6[i]*rinv6;
690 Vvdw12 = c12[i]*rinv6*rinv6;
691
692 Vvdw[i] = ( (Vvdw12 - c12[i]*sh_invrc6*sh_invrc6)*onetwelfth
693 - (Vvdw6 - c6[i]*sh_invrc6 - c6grid*sh_lj_ewald)*onesixth);
694 FscalV[i] = Vvdw12 - Vvdw6;
695 }
696 else
697 {
698 /* Normal LJ-PME */
699 ewcljrsq = ewclj2*rV*rV;
700 exponent = std::exp(-ewcljrsq);
701 poly = exponent*(one + ewcljrsq + ewcljrsq*ewcljrsq*half);
702 vvdw_disp = (c6[i]-c6grid*(one-poly))*rinv6;
703 vvdw_rep = c12[i]*rinv6*rinv6;
704 FscalV[i] = vvdw_rep - vvdw_disp - c6grid*onesixth*exponent*ewclj6;
705 Vvdw[i] = (vvdw_rep - c12[i]*sh_invrc6*sh_invrc6)*onetwelfth - (vvdw_disp - c6[i]*sh_invrc6 - c6grid*sh_lj_ewald)/six;
706 }
707 break;
708
709 case GMX_NBKERNEL_VDW_NONE:
710 Vvdw[i] = zero;
711 FscalV[i] = zero;
712 break;
713
714 default:
715 gmx_incons("Invalid ivdw in free energy kernel");
716 break;
717 }
718
719 if (fr->vdw_modifier == eintmodPOTSWITCH)
720 {
721 d = rV-fr->rvdw_switch;
722 d = (d > zero) ? d : zero;
723 d2 = d*d;
724 sw = one+d2*d*(vdw_swV3+d*(vdw_swV4+d*vdw_swV5));
725 dsw = d2*(vdw_swF2+d*(vdw_swF3+d*vdw_swF4));
726
727 FscalV[i] = FscalV[i]*sw - rV*Vvdw[i]*dsw;
728 Vvdw[i] *= sw;
729
730 FscalV[i] = (rV < rvdw) ? FscalV[i] : zero;
731 Vvdw[i] = (rV < rvdw) ? Vvdw[i] : zero;
732 }
733 }
734
735 /* FscalC (and FscalV) now contain: dV/drC * rC
736 * Now we multiply by rC^-p, so it will be: dV/drC * rC^1-p
737 * Further down we first multiply by r^p-2 and then by
738 * the vector r, which in total gives: dV/drC * (r/rC)^1-p
739 */
740 FscalC[i] *= rpinvC;
741 FscalV[i] *= rpinvV;
742 }
743 }
744
745 /* Assemble A and B states */
746 for (i = 0; i < NSTATES; i++)
747 {
748 vctot += LFC[i]*Vcoul[i];
749 vvtot += LFV[i]*Vvdw[i];
750
751 Fscal += LFC[i]*FscalC[i]*rpm2;
752 Fscal += LFV[i]*FscalV[i]*rpm2;
753
754 dvdl_coul += Vcoul[i]*DLF[i] + LFC[i]*alpha_coul_eff*dlfac_coul[i]*FscalC[i]*sigma_pow[i];
755 dvdl_vdw += Vvdw[i]*DLF[i] + LFV[i]*alpha_vdw_eff*dlfac_vdw[i]*FscalV[i]*sigma_pow[i];
756 }
757 }
758 else if (icoul == GMX_NBKERNEL_ELEC_REACTIONFIELD)
759 {
760 /* For excluded pairs, which are only in this pair list when
761 * using the Verlet scheme, we don't use soft-core.
762 * The group scheme also doesn't soft-core for these.
763 * As there is no singularity, there is no need for soft-core.
764 */
765 VV = krf*rsq - crf;
766 FF = -two*krf;
767
768 if (ii == jnr)
769 {
770 VV *= half;
771 }
772
773 for (i = 0; i < NSTATES; i++)
774 {
775 vctot += LFC[i]*qq[i]*VV;
776 Fscal += LFC[i]*qq[i]*FF;
777 dvdl_coul += DLF[i]*qq[i]*VV;
778 }
779 }
780
781 if (bConvertEwaldToCoulomb && ( !bExactElecCutoff || r < rcoulomb ) )
782 {
783 /* See comment in the preamble. When using Ewald interactions
784 * (unless we use a switch modifier) we subtract the reciprocal-space
785 * Ewald component here which made it possible to apply the free
786 * energy interaction to 1/r (vanilla coulomb short-range part)
787 * above. This gets us closer to the ideal case of applying
788 * the softcore to the entire electrostatic interaction,
789 * including the reciprocal-space component.
790 */
791 real v_lr, f_lr;
792
793 ewrt = r*ewtabscale;
794 ewitab = static_cast<int>(ewrt);
795 eweps = ewrt-ewitab;
796 ewitab = 4*ewitab;
797 f_lr = ewtab[ewitab]+eweps*ewtab[ewitab+1];
798 v_lr = (ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+f_lr));
799 f_lr *= rinv;
800
801 /* Note that any possible Ewald shift has already been applied in
802 * the normal interaction part above.
803 */
804
805 if (ii == jnr)
806 {
807 /* If we get here, the i particle (ii) has itself (jnr)
808 * in its neighborlist. This can only happen with the Verlet
809 * scheme, and corresponds to a self-interaction that will
810 * occur twice. Scale it down by 50% to only include it once.
811 */
812 v_lr *= half;
813 }
814
815 for (i = 0; i < NSTATES; i++)
816 {
817 vctot -= LFC[i]*qq[i]*v_lr;
818 Fscal -= LFC[i]*qq[i]*f_lr;
819 dvdl_coul -= (DLF[i]*qq[i])*v_lr;
820 }
821 }
822
823 if (bConvertLJEwaldToLJ6 && (!bExactVdwCutoff || r < rvdw))
824 {
825 /* See comment in the preamble. When using LJ-Ewald interactions
826 * (unless we use a switch modifier) we subtract the reciprocal-space
827 * Ewald component here which made it possible to apply the free
828 * energy interaction to r^-6 (vanilla LJ6 short-range part)
829 * above. This gets us closer to the ideal case of applying
830 * the softcore to the entire VdW interaction,
831 * including the reciprocal-space component.
832 */
833 /* We could also use the analytical form here
834 * iso a table, but that can cause issues for
835 * r close to 0 for non-interacting pairs.
836 */
837 real rs, frac, f_lr;
838 int ri;
839
840 rs = rsq*rinv*ewtabscale;
841 ri = static_cast<int>(rs);
842 frac = rs - ri;
843 f_lr = (1 - frac)*tab_ewald_F_lj[ri] + frac*tab_ewald_F_lj[ri+1];
844 /* TODO: Currently the Ewald LJ table does not contain
845 * the factor 1/6, we should add this.
846 */
847 FF = f_lr*rinv/six;
848 VV = (tab_ewald_V_lj[ri] - ewtabhalfspace*frac*(tab_ewald_F_lj[ri] + f_lr))/six;
849
850 if (ii == jnr)
851 {
852 /* If we get here, the i particle (ii) has itself (jnr)
853 * in its neighborlist. This can only happen with the Verlet
854 * scheme, and corresponds to a self-interaction that will
855 * occur twice. Scale it down by 50% to only include it once.
856 */
857 VV *= half;
858 }
859
860 for (i = 0; i < NSTATES; i++)
861 {
862 c6grid = nbfp_grid[tj[i]];
863 vvtot += LFV[i]*c6grid*VV;
864 Fscal += LFV[i]*c6grid*FF;
865 dvdl_vdw += (DLF[i]*c6grid)*VV;
866 }
867 }
868
869 if (bDoForces)
870 {
871 tx = Fscal*dx;
872 ty = Fscal*dy;
873 tz = Fscal*dz;
874 fix = fix + tx;
875 fiy = fiy + ty;
876 fiz = fiz + tz;
877 /* OpenMP atomics are expensive, but this kernels is also
878 * expensive, so we can take this hit, instead of using
879 * thread-local output buffers and extra reduction.
880 *
881 * All the OpenMP regions in this file are trivial and should
882 * not throw, so no need for try/catch.
883 */
884 #pragma omp atomic
885 f[j3] -= tx;
886 #pragma omp atomic
887 f[j3+1] -= ty;
888 #pragma omp atomic
889 f[j3+2] -= tz;
890 }
891 }
892
893 /* The atomics below are expensive with many OpenMP threads.
894 * Here unperturbed i-particles will usually only have a few
895 * (perturbed) j-particles in the list. Thus with a buffered list
896 * we can skip a significant number of i-reductions with a check.
897 */
898 if (npair_within_cutoff > 0)
899 {
900 if (bDoForces)
901 {
902 #pragma omp atomic
903 f[ii3] += fix;
904 #pragma omp atomic
905 f[ii3+1] += fiy;
906 #pragma omp atomic
907 f[ii3+2] += fiz;
908 }
909 if (bDoShiftForces)
910 {
911 #pragma omp atomic
912 fshift[is3] += fix;
913 #pragma omp atomic
914 fshift[is3+1] += fiy;
915 #pragma omp atomic
916 fshift[is3+2] += fiz;
917 }
918 if (bDoPotential)
919 {
920 ggid = gid[n];
921 #pragma omp atomic
922 Vc[ggid] += vctot;
923 #pragma omp atomic
924 Vv[ggid] += vvtot;
925 }
926 }
927 }
928
929 #pragma omp atomic
930 dvdl[efptCOUL] += dvdl_coul;
931 #pragma omp atomic
932 dvdl[efptVDW] += dvdl_vdw;
933
934 /* Estimate flops, average for free energy stuff:
935 * 12 flops per outer iteration
936 * 150 flops per inner iteration
937 */
938 #pragma omp atomic
939 inc_nrnb(nrnb, eNR_NBKERNEL_FREE_ENERGY, nlist->nri*12 + nlist->jindex[n]*150);
940 }
The ultimate molecular dynamics simulation package