2 * This file is part of the GROMACS molecular simulation package.
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.
7 * Copyright (c) 2018,2019,2020,2021, by the GROMACS development team, led by
8 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
9 * and including many others, as listed in the AUTHORS file in the
10 * top-level source directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
40 #include "nb_free_energy.h"
48 #include "gromacs/gmxlib/nrnb.h"
49 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
50 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
51 #include "gromacs/math/functions.h"
52 #include "gromacs/math/vec.h"
53 #include "gromacs/mdtypes/forceoutput.h"
54 #include "gromacs/mdtypes/forcerec.h"
55 #include "gromacs/mdtypes/interaction_const.h"
56 #include "gromacs/mdtypes/md_enums.h"
57 #include "gromacs/mdtypes/mdatom.h"
58 #include "gromacs/simd/simd.h"
59 #include "gromacs/utility/fatalerror.h"
60 #include "gromacs/utility/arrayref.h"
63 //! Scalar (non-SIMD) data types.
64 struct ScalarDataTypes
66 using RealType = real; //!< The data type to use as real.
67 using IntType = int; //!< The data type to use as int.
68 static constexpr int simdRealWidth = 1; //!< The width of the RealType.
69 static constexpr int simdIntWidth = 1; //!< The width of the IntType.
72 #if GMX_SIMD_HAVE_REAL && GMX_SIMD_HAVE_INT32_ARITHMETICS
76 using RealType = gmx::SimdReal; //!< The data type to use as real.
77 using IntType = gmx::SimdInt32; //!< The data type to use as int.
78 static constexpr int simdRealWidth = GMX_SIMD_REAL_WIDTH; //!< The width of the RealType.
79 static constexpr int simdIntWidth = GMX_SIMD_FINT32_WIDTH; //!< The width of the IntType.
83 //! Computes r^(1/p) and 1/r^(1/p) for the standard p=6
84 template<class RealType>
85 static inline void pthRoot(const RealType r, RealType* pthRoot, RealType* invPthRoot)
87 *invPthRoot = gmx::invsqrt(std::cbrt(r));
88 *pthRoot = 1 / (*invPthRoot);
91 template<class RealType>
92 static inline RealType calculateRinv6(const RealType rInvV)
94 RealType rInv6 = rInvV * rInvV;
95 return (rInv6 * rInv6 * rInv6);
98 template<class RealType>
99 static inline RealType calculateVdw6(const RealType c6, const RealType rInv6)
104 template<class RealType>
105 static inline RealType calculateVdw12(const RealType c12, const RealType rInv6)
107 return (c12 * rInv6 * rInv6);
110 /* reaction-field electrostatics */
111 template<class RealType>
112 static inline RealType reactionFieldScalarForce(const RealType qq,
118 return (qq * (rInv - two * krf * r * r));
120 template<class RealType>
121 static inline RealType reactionFieldPotential(const RealType qq,
125 const real potentialShift)
127 return (qq * (rInv + krf * r * r - potentialShift));
130 /* Ewald electrostatics */
131 template<class RealType>
132 static inline RealType ewaldScalarForce(const RealType coulomb, const RealType rInv)
134 return (coulomb * rInv);
136 template<class RealType>
137 static inline RealType ewaldPotential(const RealType coulomb, const RealType rInv, const real potentialShift)
139 return (coulomb * (rInv - potentialShift));
143 template<class RealType>
144 static inline RealType lennardJonesScalarForce(const RealType v6, const RealType v12)
148 template<class RealType>
149 static inline RealType lennardJonesPotential(const RealType v6,
153 const real repulsionShift,
154 const real dispersionShift,
156 const real oneTwelfth)
158 return ((v12 + c12 * repulsionShift) * oneTwelfth - (v6 + c6 * dispersionShift) * oneSixth);
162 static inline real ewaldLennardJonesGridSubtract(const real c6grid, const real potentialShift, const real oneSixth)
164 return (c6grid * potentialShift * oneSixth);
167 /* LJ Potential switch */
168 template<class RealType>
169 static inline RealType potSwitchScalarForceMod(const RealType fScalarInp,
170 const RealType potential,
179 real fScalar = fScalarInp * sw - r * potential * dsw;
184 template<class RealType>
185 static inline RealType potSwitchPotentialMod(const RealType potentialInp,
193 real potential = potentialInp * sw;
200 //! Templated free-energy non-bonded kernel
201 template<typename DataTypes, bool useSoftCore, bool scLambdasOrAlphasDiffer, bool vdwInteractionTypeIsEwald, bool elecInteractionTypeIsEwald, bool vdwModifierIsPotSwitch>
202 static void nb_free_energy_kernel(const t_nblist* gmx_restrict nlist,
203 rvec* gmx_restrict xx,
204 gmx::ForceWithShiftForces* forceWithShiftForces,
205 const t_forcerec* gmx_restrict fr,
206 const t_mdatoms* gmx_restrict mdatoms,
207 nb_kernel_data_t* gmx_restrict kernel_data,
208 t_nrnb* gmx_restrict nrnb)
214 using RealType = typename DataTypes::RealType;
215 using IntType = typename DataTypes::IntType;
217 /* FIXME: How should these be handled with SIMD? */
218 constexpr real oneTwelfth = 1.0 / 12.0;
219 constexpr real oneSixth = 1.0 / 6.0;
220 constexpr real zero = 0.0;
221 constexpr real half = 0.5;
222 constexpr real one = 1.0;
223 constexpr real two = 2.0;
224 constexpr real six = 6.0;
226 /* Extract pointer to non-bonded interaction constants */
227 const interaction_const_t* ic = fr->ic.get();
229 // Extract pair list data
230 const int nri = nlist->nri;
231 gmx::ArrayRef<const int> iinr = nlist->iinr;
232 gmx::ArrayRef<const int> jindex = nlist->jindex;
233 gmx::ArrayRef<const int> jjnr = nlist->jjnr;
234 gmx::ArrayRef<const int> shift = nlist->shift;
235 gmx::ArrayRef<const int> gid = nlist->gid;
237 const real* shiftvec = fr->shift_vec[0];
238 const real* chargeA = mdatoms->chargeA;
239 const real* chargeB = mdatoms->chargeB;
240 real* Vc = kernel_data->energygrp_elec;
241 const int* typeA = mdatoms->typeA;
242 const int* typeB = mdatoms->typeB;
243 const int ntype = fr->ntype;
244 gmx::ArrayRef<const real> nbfp = fr->nbfp;
245 gmx::ArrayRef<const real> nbfp_grid = fr->ljpme_c6grid;
247 real* Vv = kernel_data->energygrp_vdw;
248 const real lambda_coul =
249 kernel_data->lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)];
250 const real lambda_vdw =
251 kernel_data->lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)];
252 gmx::ArrayRef<real> dvdl = kernel_data->dvdl;
253 const auto& scParams = *ic->softCoreParameters;
254 const real alpha_coul = scParams.alphaCoulomb;
255 const real alpha_vdw = scParams.alphaVdw;
256 const real lam_power = scParams.lambdaPower;
257 const real sigma6_def = scParams.sigma6WithInvalidSigma;
258 const real sigma6_min = scParams.sigma6Minimum;
259 const bool doForces = ((kernel_data->flags & GMX_NONBONDED_DO_FORCE) != 0);
260 const bool doShiftForces = ((kernel_data->flags & GMX_NONBONDED_DO_SHIFTFORCE) != 0);
261 const bool doPotential = ((kernel_data->flags & GMX_NONBONDED_DO_POTENTIAL) != 0);
263 // Extract data from interaction_const_t
264 const real facel = ic->epsfac;
265 const real rCoulomb = ic->rcoulomb;
266 const real krf = ic->reactionFieldCoefficient;
267 const real crf = ic->reactionFieldShift;
268 const real shLjEwald = ic->sh_lj_ewald;
269 const real rVdw = ic->rvdw;
270 const real dispersionShift = ic->dispersion_shift.cpot;
271 const real repulsionShift = ic->repulsion_shift.cpot;
273 // Note that the nbnxm kernels do not support Coulomb potential switching at all
274 GMX_ASSERT(ic->coulomb_modifier != InteractionModifiers::PotSwitch,
275 "Potential switching is not supported for Coulomb with FEP");
277 real vdw_swV3, vdw_swV4, vdw_swV5, vdw_swF2, vdw_swF3, vdw_swF4;
278 if (vdwModifierIsPotSwitch)
280 const real d = ic->rvdw - ic->rvdw_switch;
281 vdw_swV3 = -10.0 / (d * d * d);
282 vdw_swV4 = 15.0 / (d * d * d * d);
283 vdw_swV5 = -6.0 / (d * d * d * d * d);
284 vdw_swF2 = -30.0 / (d * d * d);
285 vdw_swF3 = 60.0 / (d * d * d * d);
286 vdw_swF4 = -30.0 / (d * d * d * d * d);
290 /* Avoid warnings from stupid compilers (looking at you, Clang!) */
291 vdw_swV3 = vdw_swV4 = vdw_swV5 = vdw_swF2 = vdw_swF3 = vdw_swF4 = 0.0;
294 NbkernelElecType icoul;
295 if (ic->eeltype == CoulombInteractionType::Cut || EEL_RF(ic->eeltype))
297 icoul = NbkernelElecType::ReactionField;
301 icoul = NbkernelElecType::None;
304 real rcutoff_max2 = std::max(ic->rcoulomb, ic->rvdw);
305 rcutoff_max2 = rcutoff_max2 * rcutoff_max2;
307 const real* tab_ewald_F_lj = nullptr;
308 const real* tab_ewald_V_lj = nullptr;
309 const real* ewtab = nullptr;
310 real coulombTableScale = 0;
311 real coulombTableScaleInvHalf = 0;
312 real vdwTableScale = 0;
313 real vdwTableScaleInvHalf = 0;
315 if (elecInteractionTypeIsEwald || vdwInteractionTypeIsEwald)
317 sh_ewald = ic->sh_ewald;
319 if (elecInteractionTypeIsEwald)
321 const auto& coulombTables = *ic->coulombEwaldTables;
322 ewtab = coulombTables.tableFDV0.data();
323 coulombTableScale = coulombTables.scale;
324 coulombTableScaleInvHalf = half / coulombTableScale;
326 if (vdwInteractionTypeIsEwald)
328 const auto& vdwTables = *ic->vdwEwaldTables;
329 tab_ewald_F_lj = vdwTables.tableF.data();
330 tab_ewald_V_lj = vdwTables.tableV.data();
331 vdwTableScale = vdwTables.scale;
332 vdwTableScaleInvHalf = half / vdwTableScale;
335 /* For Ewald/PME interactions we cannot easily apply the soft-core component to
336 * reciprocal space. When we use non-switched Ewald interactions, we
337 * assume the soft-coring does not significantly affect the grid contribution
338 * and apply the soft-core only to the full 1/r (- shift) pair contribution.
340 * However, we cannot use this approach for switch-modified since we would then
341 * effectively end up evaluating a significantly different interaction here compared to the
342 * normal (non-free-energy) kernels, either by applying a cutoff at a different
343 * position than what the user requested, or by switching different
344 * things (1/r rather than short-range Ewald). For these settings, we just
345 * use the traditional short-range Ewald interaction in that case.
347 GMX_RELEASE_ASSERT(!(vdwInteractionTypeIsEwald && vdwModifierIsPotSwitch),
348 "Can not apply soft-core to switched Ewald potentials");
353 /* Lambda factor for state A, 1-lambda*/
354 real LFC[NSTATES], LFV[NSTATES];
355 LFC[STATE_A] = one - lambda_coul;
356 LFV[STATE_A] = one - lambda_vdw;
358 /* Lambda factor for state B, lambda*/
359 LFC[STATE_B] = lambda_coul;
360 LFV[STATE_B] = lambda_vdw;
362 /*derivative of the lambda factor for state A and B */
367 real lFacCoul[NSTATES], dlFacCoul[NSTATES], lFacVdw[NSTATES], dlFacVdw[NSTATES];
368 constexpr real sc_r_power = 6.0_real;
369 for (int i = 0; i < NSTATES; i++)
371 lFacCoul[i] = (lam_power == 2 ? (1 - LFC[i]) * (1 - LFC[i]) : (1 - LFC[i]));
372 dlFacCoul[i] = DLF[i] * lam_power / sc_r_power * (lam_power == 2 ? (1 - LFC[i]) : 1);
373 lFacVdw[i] = (lam_power == 2 ? (1 - LFV[i]) * (1 - LFV[i]) : (1 - LFV[i]));
374 dlFacVdw[i] = DLF[i] * lam_power / sc_r_power * (lam_power == 2 ? (1 - LFV[i]) : 1);
377 // TODO: We should get rid of using pointers to real
378 const real* x = xx[0];
379 real* gmx_restrict f = &(forceWithShiftForces->force()[0][0]);
380 real* gmx_restrict fshift = &(forceWithShiftForces->shiftForces()[0][0]);
382 const real rlistSquared = gmx::square(fr->rlist);
384 int numExcludedPairsBeyondRlist = 0;
386 for (int n = 0; n < nri; n++)
388 int npair_within_cutoff = 0;
390 const int is3 = 3 * shift[n];
391 const real shX = shiftvec[is3];
392 const real shY = shiftvec[is3 + 1];
393 const real shZ = shiftvec[is3 + 2];
394 const int nj0 = jindex[n];
395 const int nj1 = jindex[n + 1];
396 const int ii = iinr[n];
397 const int ii3 = 3 * ii;
398 const real ix = shX + x[ii3 + 0];
399 const real iy = shY + x[ii3 + 1];
400 const real iz = shZ + x[ii3 + 2];
401 const real iqA = facel * chargeA[ii];
402 const real iqB = facel * chargeB[ii];
403 const int ntiA = 2 * ntype * typeA[ii];
404 const int ntiB = 2 * ntype * typeB[ii];
411 for (int k = nj0; k < nj1; k++)
414 const int jnr = jjnr[k];
415 const int j3 = 3 * jnr;
416 RealType c6[NSTATES], c12[NSTATES], qq[NSTATES], vCoul[NSTATES], vVdw[NSTATES];
417 RealType r, rInv, rp, rpm2;
418 RealType alphaVdwEff, alphaCoulEff, sigma6[NSTATES];
419 const RealType dX = ix - x[j3];
420 const RealType dY = iy - x[j3 + 1];
421 const RealType dZ = iz - x[j3 + 2];
422 const RealType rSq = dX * dX + dY * dY + dZ * dZ;
423 RealType fScalC[NSTATES], fScalV[NSTATES];
424 /* Check if this pair on the exlusions list.*/
425 const bool bPairIncluded = nlist->excl_fep.empty() || nlist->excl_fep[k];
427 if (rSq >= rcutoff_max2 && bPairIncluded)
429 /* We save significant time by skipping all code below.
430 * Note that with soft-core interactions, the actual cut-off
431 * check might be different. But since the soft-core distance
432 * is always larger than r, checking on r here is safe.
433 * Exclusions outside the cutoff can not be skipped as
434 * when using Ewald: the reciprocal-space
435 * Ewald component still needs to be subtracted.
440 npair_within_cutoff++;
442 if (rSq > rlistSquared)
444 numExcludedPairsBeyondRlist++;
449 /* Note that unlike in the nbnxn kernels, we do not need
450 * to clamp the value of rSq before taking the invsqrt
451 * to avoid NaN in the LJ calculation, since here we do
452 * not calculate LJ interactions when C6 and C12 are zero.
455 rInv = gmx::invsqrt(rSq);
460 /* The force at r=0 is zero, because of symmetry.
461 * But note that the potential is in general non-zero,
462 * since the soft-cored r will be non-zero.
470 rpm2 = rSq * rSq; /* r4 */
471 rp = rpm2 * rSq; /* r6 */
475 /* The soft-core power p will not affect the results
476 * with not using soft-core, so we use power of 0 which gives
477 * the simplest math and cheapest code.
485 qq[STATE_A] = iqA * chargeA[jnr];
486 qq[STATE_B] = iqB * chargeB[jnr];
488 tj[STATE_A] = ntiA + 2 * typeA[jnr];
489 tj[STATE_B] = ntiB + 2 * typeB[jnr];
493 c6[STATE_A] = nbfp[tj[STATE_A]];
494 c6[STATE_B] = nbfp[tj[STATE_B]];
496 for (int i = 0; i < NSTATES; i++)
498 c12[i] = nbfp[tj[i] + 1];
501 if ((c6[i] > 0) && (c12[i] > 0))
503 /* c12 is stored scaled with 12.0 and c6 is scaled with 6.0 - correct for this */
504 sigma6[i] = half * c12[i] / c6[i];
505 if (sigma6[i] < sigma6_min) /* for disappearing coul and vdw with soft core at the same time */
507 sigma6[i] = sigma6_min;
512 sigma6[i] = sigma6_def;
519 /* only use softcore if one of the states has a zero endstate - softcore is for avoiding infinities!*/
520 if ((c12[STATE_A] > 0) && (c12[STATE_B] > 0))
527 alphaVdwEff = alpha_vdw;
528 alphaCoulEff = alpha_coul;
532 for (int i = 0; i < NSTATES; i++)
539 RealType rInvC, rInvV, rC, rV, rPInvC, rPInvV;
541 /* Only spend time on A or B state if it is non-zero */
542 if ((qq[i] != 0) || (c6[i] != 0) || (c12[i] != 0))
544 /* this section has to be inside the loop because of the dependence on sigma6 */
547 rPInvC = one / (alphaCoulEff * lFacCoul[i] * sigma6[i] + rp);
548 pthRoot(rPInvC, &rInvC, &rC);
549 if (scLambdasOrAlphasDiffer)
551 rPInvV = one / (alphaVdwEff * lFacVdw[i] * sigma6[i] + rp);
552 pthRoot(rPInvV, &rInvV, &rV);
556 /* We can avoid one expensive pow and one / operation */
573 /* Only process the coulomb interactions if we have charges,
574 * and if we either include all entries in the list (no cutoff
575 * used in the kernel), or if we are within the cutoff.
577 bool computeElecInteraction = (elecInteractionTypeIsEwald && r < rCoulomb)
578 || (!elecInteractionTypeIsEwald && rC < rCoulomb);
580 if ((qq[i] != 0) && computeElecInteraction)
582 if (elecInteractionTypeIsEwald)
584 vCoul[i] = ewaldPotential(qq[i], rInvC, sh_ewald);
585 fScalC[i] = ewaldScalarForce(qq[i], rInvC);
589 vCoul[i] = reactionFieldPotential(qq[i], rInvC, rC, krf, crf);
590 fScalC[i] = reactionFieldScalarForce(qq[i], rInvC, rC, krf, two);
594 /* Only process the VDW interactions if we have
595 * some non-zero parameters, and if we either
596 * include all entries in the list (no cutoff used
597 * in the kernel), or if we are within the cutoff.
599 bool computeVdwInteraction = (vdwInteractionTypeIsEwald && r < rVdw)
600 || (!vdwInteractionTypeIsEwald && rV < rVdw);
601 if ((c6[i] != 0 || c12[i] != 0) && computeVdwInteraction)
610 rInv6 = calculateRinv6(rInvV);
612 RealType vVdw6 = calculateVdw6(c6[i], rInv6);
613 RealType vVdw12 = calculateVdw12(c12[i], rInv6);
615 vVdw[i] = lennardJonesPotential(
616 vVdw6, vVdw12, c6[i], c12[i], repulsionShift, dispersionShift, oneSixth, oneTwelfth);
617 fScalV[i] = lennardJonesScalarForce(vVdw6, vVdw12);
619 if (vdwInteractionTypeIsEwald)
621 /* Subtract the grid potential at the cut-off */
622 vVdw[i] += ewaldLennardJonesGridSubtract(
623 nbfp_grid[tj[i]], shLjEwald, oneSixth);
626 if (vdwModifierIsPotSwitch)
628 RealType d = rV - ic->rvdw_switch;
629 d = (d > zero) ? d : zero;
630 const RealType d2 = d * d;
632 one + d2 * d * (vdw_swV3 + d * (vdw_swV4 + d * vdw_swV5));
633 const RealType dsw = d2 * (vdw_swF2 + d * (vdw_swF3 + d * vdw_swF4));
635 fScalV[i] = potSwitchScalarForceMod(
636 fScalV[i], vVdw[i], sw, rV, rVdw, dsw, zero);
637 vVdw[i] = potSwitchPotentialMod(vVdw[i], sw, rV, rVdw, zero);
641 /* fScalC (and fScalV) now contain: dV/drC * rC
642 * Now we multiply by rC^-p, so it will be: dV/drC * rC^1-p
643 * Further down we first multiply by r^p-2 and then by
644 * the vector r, which in total gives: dV/drC * (r/rC)^1-p
649 } // end for (int i = 0; i < NSTATES; i++)
651 /* Assemble A and B states */
652 for (int i = 0; i < NSTATES; i++)
654 vCTot += LFC[i] * vCoul[i];
655 vVTot += LFV[i] * vVdw[i];
657 fScal += LFC[i] * fScalC[i] * rpm2;
658 fScal += LFV[i] * fScalV[i] * rpm2;
662 dvdlCoul += vCoul[i] * DLF[i]
663 + LFC[i] * alphaCoulEff * dlFacCoul[i] * fScalC[i] * sigma6[i];
664 dvdlVdw += vVdw[i] * DLF[i]
665 + LFV[i] * alphaVdwEff * dlFacVdw[i] * fScalV[i] * sigma6[i];
669 dvdlCoul += vCoul[i] * DLF[i];
670 dvdlVdw += vVdw[i] * DLF[i];
673 } // end if (bPairIncluded)
674 else if (icoul == NbkernelElecType::ReactionField)
676 /* For excluded pairs, which are only in this pair list when
677 * using the Verlet scheme, we don't use soft-core.
678 * As there is no singularity, there is no need for soft-core.
680 const real FF = -two * krf;
681 RealType VV = krf * rSq - crf;
688 for (int i = 0; i < NSTATES; i++)
690 vCTot += LFC[i] * qq[i] * VV;
691 fScal += LFC[i] * qq[i] * FF;
692 dvdlCoul += DLF[i] * qq[i] * VV;
696 if (elecInteractionTypeIsEwald && (r < rCoulomb || !bPairIncluded))
698 /* See comment in the preamble. When using Ewald interactions
699 * (unless we use a switch modifier) we subtract the reciprocal-space
700 * Ewald component here which made it possible to apply the free
701 * energy interaction to 1/r (vanilla coulomb short-range part)
702 * above. This gets us closer to the ideal case of applying
703 * the softcore to the entire electrostatic interaction,
704 * including the reciprocal-space component.
708 const RealType ewrt = r * coulombTableScale;
709 IntType ewitab = static_cast<IntType>(ewrt);
710 const RealType eweps = ewrt - ewitab;
712 f_lr = ewtab[ewitab] + eweps * ewtab[ewitab + 1];
713 v_lr = (ewtab[ewitab + 2] - coulombTableScaleInvHalf * eweps * (ewtab[ewitab] + f_lr));
716 /* Note that any possible Ewald shift has already been applied in
717 * the normal interaction part above.
722 /* If we get here, the i particle (ii) has itself (jnr)
723 * in its neighborlist. This can only happen with the Verlet
724 * scheme, and corresponds to a self-interaction that will
725 * occur twice. Scale it down by 50% to only include it once.
730 for (int i = 0; i < NSTATES; i++)
732 vCTot -= LFC[i] * qq[i] * v_lr;
733 fScal -= LFC[i] * qq[i] * f_lr;
734 dvdlCoul -= (DLF[i] * qq[i]) * v_lr;
738 if (vdwInteractionTypeIsEwald && (r < rVdw || !bPairIncluded))
740 /* See comment in the preamble. When using LJ-Ewald interactions
741 * (unless we use a switch modifier) we subtract the reciprocal-space
742 * Ewald component here which made it possible to apply the free
743 * energy interaction to r^-6 (vanilla LJ6 short-range part)
744 * above. This gets us closer to the ideal case of applying
745 * the softcore to the entire VdW interaction,
746 * including the reciprocal-space component.
748 /* We could also use the analytical form here
749 * iso a table, but that can cause issues for
750 * r close to 0 for non-interacting pairs.
753 const RealType rs = rSq * rInv * vdwTableScale;
754 const IntType ri = static_cast<IntType>(rs);
755 const RealType frac = rs - ri;
756 const RealType f_lr = (1 - frac) * tab_ewald_F_lj[ri] + frac * tab_ewald_F_lj[ri + 1];
757 /* TODO: Currently the Ewald LJ table does not contain
758 * the factor 1/6, we should add this.
760 const RealType FF = f_lr * rInv / six;
762 (tab_ewald_V_lj[ri] - vdwTableScaleInvHalf * frac * (tab_ewald_F_lj[ri] + f_lr))
767 /* If we get here, the i particle (ii) has itself (jnr)
768 * in its neighborlist. This can only happen with the Verlet
769 * scheme, and corresponds to a self-interaction that will
770 * occur twice. Scale it down by 50% to only include it once.
775 for (int i = 0; i < NSTATES; i++)
777 const real c6grid = nbfp_grid[tj[i]];
778 vVTot += LFV[i] * c6grid * VV;
779 fScal += LFV[i] * c6grid * FF;
780 dvdlVdw += (DLF[i] * c6grid) * VV;
786 const real tX = fScal * dX;
787 const real tY = fScal * dY;
788 const real tZ = fScal * dZ;
792 /* OpenMP atomics are expensive, but this kernels is also
793 * expensive, so we can take this hit, instead of using
794 * thread-local output buffers and extra reduction.
796 * All the OpenMP regions in this file are trivial and should
797 * not throw, so no need for try/catch.
806 } // end for (int k = nj0; k < nj1; k++)
808 /* The atomics below are expensive with many OpenMP threads.
809 * Here unperturbed i-particles will usually only have a few
810 * (perturbed) j-particles in the list. Thus with a buffered list
811 * we can skip a significant number of i-reductions with a check.
813 if (npair_within_cutoff > 0)
829 fshift[is3 + 1] += fIY;
831 fshift[is3 + 2] += fIZ;
842 } // end for (int n = 0; n < nri; n++)
845 dvdl[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)] += dvdlCoul;
847 dvdl[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)] += dvdlVdw;
849 /* Estimate flops, average for free energy stuff:
850 * 12 flops per outer iteration
851 * 150 flops per inner iteration
853 atomicNrnbIncrement(nrnb, eNR_NBKERNEL_FREE_ENERGY, nlist->nri * 12 + nlist->jindex[nri] * 150);
855 if (numExcludedPairsBeyondRlist > 0)
858 "There are %d perturbed non-bonded pair interactions beyond the pair-list cutoff "
859 "of %g nm, which is not supported. This can happen because the system is "
860 "unstable or because intra-molecular interactions at long distances are "
862 "latter is the case, you can try to increase nstlist or rlist to avoid this."
863 "The error is likely triggered by the use of couple-intramol=no "
864 "and the maximal distance in the decoupled molecule exceeding rlist.",
865 numExcludedPairsBeyondRlist,
870 typedef void (*KernelFunction)(const t_nblist* gmx_restrict nlist,
871 rvec* gmx_restrict xx,
872 gmx::ForceWithShiftForces* forceWithShiftForces,
873 const t_forcerec* gmx_restrict fr,
874 const t_mdatoms* gmx_restrict mdatoms,
875 nb_kernel_data_t* gmx_restrict kernel_data,
876 t_nrnb* gmx_restrict nrnb);
878 template<bool useSoftCore, bool scLambdasOrAlphasDiffer, bool vdwInteractionTypeIsEwald, bool elecInteractionTypeIsEwald, bool vdwModifierIsPotSwitch>
879 static KernelFunction dispatchKernelOnUseSimd(const bool useSimd)
883 #if GMX_SIMD_HAVE_REAL && GMX_SIMD_HAVE_INT32_ARITHMETICS && GMX_USE_SIMD_KERNELS
884 /* FIXME: Here SimdDataTypes should be used to enable SIMD. So far, the code in nb_free_energy_kernel is not adapted to SIMD */
885 return (nb_free_energy_kernel<ScalarDataTypes, useSoftCore, scLambdasOrAlphasDiffer, vdwInteractionTypeIsEwald, elecInteractionTypeIsEwald, vdwModifierIsPotSwitch>);
887 return (nb_free_energy_kernel<ScalarDataTypes, useSoftCore, scLambdasOrAlphasDiffer, vdwInteractionTypeIsEwald, elecInteractionTypeIsEwald, vdwModifierIsPotSwitch>);
892 return (nb_free_energy_kernel<ScalarDataTypes, useSoftCore, scLambdasOrAlphasDiffer, vdwInteractionTypeIsEwald, elecInteractionTypeIsEwald, vdwModifierIsPotSwitch>);
896 template<bool useSoftCore, bool scLambdasOrAlphasDiffer, bool vdwInteractionTypeIsEwald, bool elecInteractionTypeIsEwald>
897 static KernelFunction dispatchKernelOnVdwModifier(const bool vdwModifierIsPotSwitch, const bool useSimd)
899 if (vdwModifierIsPotSwitch)
901 return (dispatchKernelOnUseSimd<useSoftCore, scLambdasOrAlphasDiffer, vdwInteractionTypeIsEwald, elecInteractionTypeIsEwald, true>(
906 return (dispatchKernelOnUseSimd<useSoftCore, scLambdasOrAlphasDiffer, vdwInteractionTypeIsEwald, elecInteractionTypeIsEwald, false>(
911 template<bool useSoftCore, bool scLambdasOrAlphasDiffer, bool vdwInteractionTypeIsEwald>
912 static KernelFunction dispatchKernelOnElecInteractionType(const bool elecInteractionTypeIsEwald,
913 const bool vdwModifierIsPotSwitch,
916 if (elecInteractionTypeIsEwald)
918 return (dispatchKernelOnVdwModifier<useSoftCore, scLambdasOrAlphasDiffer, vdwInteractionTypeIsEwald, true>(
919 vdwModifierIsPotSwitch, useSimd));
923 return (dispatchKernelOnVdwModifier<useSoftCore, scLambdasOrAlphasDiffer, vdwInteractionTypeIsEwald, false>(
924 vdwModifierIsPotSwitch, useSimd));
928 template<bool useSoftCore, bool scLambdasOrAlphasDiffer>
929 static KernelFunction dispatchKernelOnVdwInteractionType(const bool vdwInteractionTypeIsEwald,
930 const bool elecInteractionTypeIsEwald,
931 const bool vdwModifierIsPotSwitch,
934 if (vdwInteractionTypeIsEwald)
936 return (dispatchKernelOnElecInteractionType<useSoftCore, scLambdasOrAlphasDiffer, true>(
937 elecInteractionTypeIsEwald, vdwModifierIsPotSwitch, useSimd));
941 return (dispatchKernelOnElecInteractionType<useSoftCore, scLambdasOrAlphasDiffer, false>(
942 elecInteractionTypeIsEwald, vdwModifierIsPotSwitch, useSimd));
946 template<bool useSoftCore>
947 static KernelFunction dispatchKernelOnScLambdasOrAlphasDifference(const bool scLambdasOrAlphasDiffer,
948 const bool vdwInteractionTypeIsEwald,
949 const bool elecInteractionTypeIsEwald,
950 const bool vdwModifierIsPotSwitch,
953 if (scLambdasOrAlphasDiffer)
955 return (dispatchKernelOnVdwInteractionType<useSoftCore, true>(
956 vdwInteractionTypeIsEwald, elecInteractionTypeIsEwald, vdwModifierIsPotSwitch, useSimd));
960 return (dispatchKernelOnVdwInteractionType<useSoftCore, false>(
961 vdwInteractionTypeIsEwald, elecInteractionTypeIsEwald, vdwModifierIsPotSwitch, useSimd));
965 static KernelFunction dispatchKernel(const bool scLambdasOrAlphasDiffer,
966 const bool vdwInteractionTypeIsEwald,
967 const bool elecInteractionTypeIsEwald,
968 const bool vdwModifierIsPotSwitch,
970 const interaction_const_t& ic)
972 if (ic.softCoreParameters->alphaCoulomb == 0 && ic.softCoreParameters->alphaVdw == 0)
974 return (dispatchKernelOnScLambdasOrAlphasDifference<false>(scLambdasOrAlphasDiffer,
975 vdwInteractionTypeIsEwald,
976 elecInteractionTypeIsEwald,
977 vdwModifierIsPotSwitch,
982 return (dispatchKernelOnScLambdasOrAlphasDifference<true>(scLambdasOrAlphasDiffer,
983 vdwInteractionTypeIsEwald,
984 elecInteractionTypeIsEwald,
985 vdwModifierIsPotSwitch,
991 void gmx_nb_free_energy_kernel(const t_nblist* nlist,
993 gmx::ForceWithShiftForces* ff,
994 const t_forcerec* fr,
995 const t_mdatoms* mdatoms,
996 nb_kernel_data_t* kernel_data,
999 const interaction_const_t& ic = *fr->ic;
1000 GMX_ASSERT(EEL_PME_EWALD(ic.eeltype) || ic.eeltype == CoulombInteractionType::Cut || EEL_RF(ic.eeltype),
1001 "Unsupported eeltype with free energy");
1002 GMX_ASSERT(ic.softCoreParameters, "We need soft-core parameters");
1004 const auto& scParams = *ic.softCoreParameters;
1005 const bool vdwInteractionTypeIsEwald = (EVDW_PME(ic.vdwtype));
1006 const bool elecInteractionTypeIsEwald = (EEL_PME_EWALD(ic.eeltype));
1007 const bool vdwModifierIsPotSwitch = (ic.vdw_modifier == InteractionModifiers::PotSwitch);
1008 bool scLambdasOrAlphasDiffer = true;
1009 const bool useSimd = fr->use_simd_kernels;
1011 if (scParams.alphaCoulomb == 0 && scParams.alphaVdw == 0)
1013 scLambdasOrAlphasDiffer = false;
1017 if (kernel_data->lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Coul)]
1018 == kernel_data->lambda[static_cast<int>(FreeEnergyPerturbationCouplingType::Vdw)]
1019 && scParams.alphaCoulomb == scParams.alphaVdw)
1021 scLambdasOrAlphasDiffer = false;
1025 KernelFunction kernelFunc;
1026 kernelFunc = dispatchKernel(scLambdasOrAlphasDiffer,
1027 vdwInteractionTypeIsEwald,
1028 elecInteractionTypeIsEwald,
1029 vdwModifierIsPotSwitch,
1032 kernelFunc(nlist, xx, ff, fr, mdatoms, kernel_data, nrnb);