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, 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.
48 #include "gromacs/domdec/domdec_struct.h"
49 #include "gromacs/fileio/confio.h"
50 #include "gromacs/gmxlib/network.h"
51 #include "gromacs/gmxlib/nrnb.h"
52 #include "gromacs/listed_forces/disre.h"
53 #include "gromacs/listed_forces/orires.h"
54 #include "gromacs/math/functions.h"
55 #include "gromacs/math/invertmatrix.h"
56 #include "gromacs/math/paddedvector.h"
57 #include "gromacs/math/units.h"
58 #include "gromacs/math/vec.h"
59 #include "gromacs/math/vecdump.h"
60 #include "gromacs/mdlib/boxdeformation.h"
61 #include "gromacs/mdlib/constr.h"
62 #include "gromacs/mdlib/gmx_omp_nthreads.h"
63 #include "gromacs/mdlib/mdatoms.h"
64 #include "gromacs/mdlib/stat.h"
65 #include "gromacs/mdlib/tgroup.h"
66 #include "gromacs/mdtypes/commrec.h"
67 #include "gromacs/mdtypes/fcdata.h"
68 #include "gromacs/mdtypes/group.h"
69 #include "gromacs/mdtypes/inputrec.h"
70 #include "gromacs/mdtypes/md_enums.h"
71 #include "gromacs/mdtypes/mdatom.h"
72 #include "gromacs/mdtypes/state.h"
73 #include "gromacs/pbcutil/boxutilities.h"
74 #include "gromacs/pbcutil/pbc.h"
75 #include "gromacs/pulling/pull.h"
76 #include "gromacs/random/tabulatednormaldistribution.h"
77 #include "gromacs/random/threefry.h"
78 #include "gromacs/simd/simd.h"
79 #include "gromacs/timing/wallcycle.h"
80 #include "gromacs/topology/atoms.h"
81 #include "gromacs/utility/exceptions.h"
82 #include "gromacs/utility/fatalerror.h"
83 #include "gromacs/utility/futil.h"
84 #include "gromacs/utility/gmxassert.h"
85 #include "gromacs/utility/gmxomp.h"
86 #include "gromacs/utility/smalloc.h"
88 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
103 std::vector<real> bd_rf;
105 std::vector<gmx_sd_const_t> sdc;
106 std::vector<gmx_sd_sigma_t> sdsig;
107 /* andersen temperature control stuff */
108 std::vector<bool> randomize_group;
109 std::vector<real> boltzfac;
111 explicit gmx_stochd_t(const t_inputrec& inputRecord);
114 //! pImpled implementation for Update
119 Impl(const t_inputrec& inputRecord, BoxDeformation* boxDeformation);
123 void update_coords(const t_inputrec& inputRecord,
127 const gmx::ArrayRefWithPadding<const gmx::RVec>& f,
128 const t_fcdata& fcdata,
129 const gmx_ekindata_t* ekind,
133 bool haveConstraints);
135 void finish_update(const t_inputrec& inputRecord,
138 gmx_wallcycle_t wcycle,
139 bool haveConstraints);
141 void update_sd_second_half(const t_inputrec& inputRecord,
148 gmx_wallcycle_t wcycle,
149 gmx::Constraints* constr,
153 void update_temperature_constants(const t_inputrec& inputRecord);
155 const std::vector<bool>& getAndersenRandomizeGroup() const { return sd_.randomize_group; }
157 const std::vector<real>& getBoltzmanFactor() const { return sd_.boltzfac; }
159 PaddedVector<RVec>* xp() { return &xp_; }
161 BoxDeformation* deform() const { return deform_; }
164 //! stochastic dynamics struct
166 //! xprime for constraint algorithms
167 PaddedVector<RVec> xp_;
168 //! Box deformation handler (or nullptr if inactive).
169 BoxDeformation* deform_ = nullptr;
172 Update::Update(const t_inputrec& inputRecord, BoxDeformation* boxDeformation) :
173 impl_(new Impl(inputRecord, boxDeformation)){};
177 const std::vector<bool>& Update::getAndersenRandomizeGroup() const
179 return impl_->getAndersenRandomizeGroup();
182 const std::vector<real>& Update::getBoltzmanFactor() const
184 return impl_->getBoltzmanFactor();
187 PaddedVector<RVec>* Update::xp()
192 BoxDeformation* Update::deform() const
194 return impl_->deform();
197 void Update::update_coords(const t_inputrec& inputRecord,
201 const gmx::ArrayRefWithPadding<const gmx::RVec>& f,
202 const t_fcdata& fcdata,
203 const gmx_ekindata_t* ekind,
207 const bool haveConstraints)
209 return impl_->update_coords(inputRecord, step, md, state, f, fcdata, ekind, M, updatePart, cr,
213 void Update::finish_update(const t_inputrec& inputRecord,
216 gmx_wallcycle_t wcycle,
217 const bool haveConstraints)
219 return impl_->finish_update(inputRecord, md, state, wcycle, haveConstraints);
222 void Update::update_sd_second_half(const t_inputrec& inputRecord,
229 gmx_wallcycle_t wcycle,
230 gmx::Constraints* constr,
234 return impl_->update_sd_second_half(inputRecord, step, dvdlambda, md, state, cr, nrnb, wcycle,
235 constr, do_log, do_ene);
238 void Update::update_temperature_constants(const t_inputrec& inputRecord)
240 return impl_->update_temperature_constants(inputRecord);
243 /*! \brief Sets the velocities of virtual sites to zero */
244 static void clearVsiteVelocities(int start, int nrend, const unsigned short* particleType, rvec* gmx_restrict v)
246 for (int a = start; a < nrend; a++)
248 if (particleType[a] == eptVSite)
255 /*! \brief Sets the number of different temperature coupling values */
256 enum class NumTempScaleValues
258 single, //!< Single T-scaling value (either one group or all values =1)
259 multiple //!< Multiple T-scaling values, need to use T-group indices
262 /*! \brief Sets if to apply no or diagonal Parrinello-Rahman pressure scaling
264 * Note that this enum is only used in updateMDLeapfrogSimple(), which does
265 * not handle fully anistropic Parrinello-Rahman scaling, so we only have
266 * options \p no and \p diagonal here and no anistropic option.
268 enum class ApplyParrinelloRahmanVScaling
270 no, //!< Do not apply velocity scaling (not a PR-coupling run or step)
271 diagonal //!< Apply velocity scaling using a diagonal matrix
274 /*! \brief Integrate using leap-frog with T-scaling and optionally diagonal Parrinello-Rahman p-coupling
276 * \tparam numTempScaleValues The number of different T-couple values
277 * \tparam applyPRVScaling Apply Parrinello-Rahman velocity scaling
278 * \param[in] start Index of first atom to update
279 * \param[in] nrend Last atom to update: \p nrend - 1
280 * \param[in] dt The time step
281 * \param[in] dtPressureCouple Time step for pressure coupling
282 * \param[in] invMassPerDim 1/mass per atom and dimension
283 * \param[in] tcstat Temperature coupling information
284 * \param[in] cTC T-coupling group index per atom
285 * \param[in] pRVScaleMatrixDiagonal Parrinello-Rahman v-scale matrix diagonal
286 * \param[in] x Input coordinates
287 * \param[out] xprime Updated coordinates
288 * \param[inout] v Velocities
289 * \param[in] f Forces
291 * We expect this template to get good SIMD acceleration by most compilers,
292 * unlike the more complex general template.
293 * Note that we might get even better SIMD acceleration when we introduce
294 * aligned (and padded) memory, possibly with some hints for the compilers.
296 template<NumTempScaleValues numTempScaleValues, ApplyParrinelloRahmanVScaling applyPRVScaling>
297 static void updateMDLeapfrogSimple(int start,
300 real dtPressureCouple,
301 const rvec* gmx_restrict invMassPerDim,
302 gmx::ArrayRef<const t_grp_tcstat> tcstat,
303 const unsigned short* cTC,
304 const rvec pRVScaleMatrixDiagonal,
305 const rvec* gmx_restrict x,
306 rvec* gmx_restrict xprime,
307 rvec* gmx_restrict v,
308 const rvec* gmx_restrict f)
312 if (numTempScaleValues == NumTempScaleValues::single)
314 lambdaGroup = tcstat[0].lambda;
317 for (int a = start; a < nrend; a++)
319 if (numTempScaleValues == NumTempScaleValues::multiple)
321 lambdaGroup = tcstat[cTC[a]].lambda;
324 for (int d = 0; d < DIM; d++)
326 /* Note that using rvec invMassPerDim results in more efficient
327 * SIMD code, but this increases the cache pressure.
328 * For large systems with PME on the CPU this slows down the
329 * (then already slow) update by 20%. If all data remains in cache,
330 * using rvec is much faster.
332 real vNew = lambdaGroup * v[a][d] + f[a][d] * invMassPerDim[a][d] * dt;
334 if (applyPRVScaling == ApplyParrinelloRahmanVScaling::diagonal)
336 vNew -= dtPressureCouple * pRVScaleMatrixDiagonal[d] * v[a][d];
339 xprime[a][d] = x[a][d] + vNew * dt;
344 #if GMX_SIMD && GMX_SIMD_HAVE_REAL
345 # define GMX_HAVE_SIMD_UPDATE 1
347 # define GMX_HAVE_SIMD_UPDATE 0
350 #if GMX_HAVE_SIMD_UPDATE
352 /*! \brief Load (aligned) the contents of GMX_SIMD_REAL_WIDTH rvec elements sequentially into 3 SIMD registers
354 * The loaded output is:
355 * \p r0: { r[index][XX], r[index][YY], ... }
357 * \p r2: { ..., r[index+GMX_SIMD_REAL_WIDTH-1][YY], r[index+GMX_SIMD_REAL_WIDTH-1][ZZ] }
359 * \param[in] r Real to an rvec array, has to be aligned to SIMD register width
360 * \param[in] index Index of the first rvec triplet of reals to load
361 * \param[out] r0 Pointer to first SIMD register
362 * \param[out] r1 Pointer to second SIMD register
363 * \param[out] r2 Pointer to third SIMD register
365 static inline void simdLoadRvecs(const rvec* r, int index, SimdReal* r0, SimdReal* r1, SimdReal* r2)
367 const real* realPtr = r[index];
369 GMX_ASSERT(isSimdAligned(realPtr), "Pointer should be SIMD aligned");
371 *r0 = simdLoad(realPtr + 0 * GMX_SIMD_REAL_WIDTH);
372 *r1 = simdLoad(realPtr + 1 * GMX_SIMD_REAL_WIDTH);
373 *r2 = simdLoad(realPtr + 2 * GMX_SIMD_REAL_WIDTH);
376 /*! \brief Store (aligned) 3 SIMD registers sequentially to GMX_SIMD_REAL_WIDTH rvec elements
378 * The stored output is:
379 * \p r[index] = { { r0[0], r0[1], ... }
381 * \p r[index+GMX_SIMD_REAL_WIDTH-1] = { ... , r2[GMX_SIMD_REAL_WIDTH-2], r2[GMX_SIMD_REAL_WIDTH-1] }
383 * \param[out] r Pointer to an rvec array, has to be aligned to SIMD register width
384 * \param[in] index Index of the first rvec triplet of reals to store to
385 * \param[in] r0 First SIMD register
386 * \param[in] r1 Second SIMD register
387 * \param[in] r2 Third SIMD register
389 static inline void simdStoreRvecs(rvec* r, int index, SimdReal r0, SimdReal r1, SimdReal r2)
391 real* realPtr = r[index];
393 GMX_ASSERT(isSimdAligned(realPtr), "Pointer should be SIMD aligned");
395 store(realPtr + 0 * GMX_SIMD_REAL_WIDTH, r0);
396 store(realPtr + 1 * GMX_SIMD_REAL_WIDTH, r1);
397 store(realPtr + 2 * GMX_SIMD_REAL_WIDTH, r2);
400 /*! \brief Integrate using leap-frog with single group T-scaling and SIMD
402 * \param[in] start Index of first atom to update
403 * \param[in] nrend Last atom to update: \p nrend - 1
404 * \param[in] dt The time step
405 * \param[in] invMass 1/mass per atom
406 * \param[in] tcstat Temperature coupling information
407 * \param[in] x Input coordinates
408 * \param[out] xprime Updated coordinates
409 * \param[inout] v Velocities
410 * \param[in] f Forces
412 static void updateMDLeapfrogSimpleSimd(int start,
415 const real* gmx_restrict invMass,
416 gmx::ArrayRef<const t_grp_tcstat> tcstat,
417 const rvec* gmx_restrict x,
418 rvec* gmx_restrict xprime,
419 rvec* gmx_restrict v,
420 const rvec* gmx_restrict f)
422 SimdReal timestep(dt);
423 SimdReal lambdaSystem(tcstat[0].lambda);
425 /* We declare variables here, since code is often slower when declaring them inside the loop */
427 /* Note: We should implement a proper PaddedVector, so we don't need this check */
428 GMX_ASSERT(isSimdAligned(invMass), "invMass should be aligned");
430 for (int a = start; a < nrend; a += GMX_SIMD_REAL_WIDTH)
432 SimdReal invMass0, invMass1, invMass2;
433 expandScalarsToTriplets(simdLoad(invMass + a), &invMass0, &invMass1, &invMass2);
437 simdLoadRvecs(v, a, &v0, &v1, &v2);
438 simdLoadRvecs(f, a, &f0, &f1, &f2);
440 v0 = fma(f0 * invMass0, timestep, lambdaSystem * v0);
441 v1 = fma(f1 * invMass1, timestep, lambdaSystem * v1);
442 v2 = fma(f2 * invMass2, timestep, lambdaSystem * v2);
444 simdStoreRvecs(v, a, v0, v1, v2);
447 simdLoadRvecs(x, a, &x0, &x1, &x2);
449 SimdReal xprime0 = fma(v0, timestep, x0);
450 SimdReal xprime1 = fma(v1, timestep, x1);
451 SimdReal xprime2 = fma(v2, timestep, x2);
453 simdStoreRvecs(xprime, a, xprime0, xprime1, xprime2);
457 #endif // GMX_HAVE_SIMD_UPDATE
459 /*! \brief Sets the NEMD acceleration type */
460 enum class AccelerationType
467 /*! \brief Integrate using leap-frog with support for everything.
469 * \tparam accelerationType Type of NEMD acceleration.
470 * \param[in] start Index of first atom to update.
471 * \param[in] nrend Last atom to update: \p nrend - 1.
472 * \param[in] doNoseHoover If to apply Nose-Hoover T-coupling.
473 * \param[in] dt The time step.
474 * \param[in] dtPressureCouple Time step for pressure coupling, is 0 when no pressure
475 * coupling should be applied at this step.
476 * \param[in] accel Acceleration per group.
477 * \param[in] md Atom properties.
478 * \param[in] ekind Kinetic energy data.
479 * \param[in] box The box dimensions.
480 * \param[in] x Input coordinates.
481 * \param[out] xprime Updated coordinates.
482 * \param[inout] v Velocities.
483 * \param[in] f Forces.
484 * \param[in] nh_vxi Nose-Hoover velocity scaling factors.
485 * \param[in] nsttcouple Frequency of the temperature coupling steps.
486 * \param[in] M Parrinello-Rahman scaling matrix.
488 template<AccelerationType accelerationType>
489 static void updateMDLeapfrogGeneral(int start,
493 real dtPressureCouple,
496 const gmx_ekindata_t* ekind,
498 const rvec* gmx_restrict x,
499 rvec* gmx_restrict xprime,
500 rvec* gmx_restrict v,
501 const rvec* gmx_restrict f,
502 const double* gmx_restrict nh_vxi,
503 const int nsttcouple,
506 /* This is a version of the leap-frog integrator that supports
507 * all combinations of T-coupling, P-coupling and NEMD.
508 * Nose-Hoover uses the reversible leap-frog integrator from
509 * Holian et al. Phys Rev E 52(3) : 2338, 1995
512 gmx::ArrayRef<const t_grp_tcstat> tcstat = ekind->tcstat;
513 gmx::ArrayRef<const t_grp_acc> grpstat = ekind->grpstat;
514 const unsigned short* cTC = md->cTC;
515 const unsigned short* cACC = md->cACC;
517 const rvec* gmx_restrict invMassPerDim = md->invMassPerDim;
519 /* Initialize group values, changed later when multiple groups are used */
523 real omega_Z = 2 * static_cast<real>(M_PI) / box[ZZ][ZZ];
525 for (int n = start; n < nrend; n++)
531 real lg = tcstat[gt].lambda;
534 real cosineZ, vCosine;
535 #ifdef __INTEL_COMPILER
536 # pragma warning(disable : 280)
538 switch (accelerationType)
540 case AccelerationType::none: copy_rvec(v[n], vRel); break;
541 case AccelerationType::group:
546 /* Avoid scaling the group velocity */
547 rvec_sub(v[n], grpstat[ga].u, vRel);
549 case AccelerationType::cosine:
550 cosineZ = std::cos(x[n][ZZ] * omega_Z);
551 vCosine = cosineZ * ekind->cosacc.vcos;
552 /* Avoid scaling the cosine profile velocity */
553 copy_rvec(v[n], vRel);
561 /* Here we account for multiple time stepping, by increasing
562 * the Nose-Hoover correction by nsttcouple
563 * TODO: This can be pre-computed.
565 factorNH = 0.5 * nsttcouple * dt * nh_vxi[gt];
568 for (int d = 0; d < DIM; d++)
570 real vNew = (lg * vRel[d]
571 + (f[n][d] * invMassPerDim[n][d] * dt - factorNH * vRel[d]
572 - dtPressureCouple * iprod(M[d], vRel)))
574 switch (accelerationType)
576 case AccelerationType::none: break;
577 case AccelerationType::group:
578 /* Add back the mean velocity and apply acceleration */
579 vNew += grpstat[ga].u[d] + accel[ga][d] * dt;
581 case AccelerationType::cosine:
584 /* Add back the mean velocity and apply acceleration */
585 vNew += vCosine + cosineZ * ekind->cosacc.cos_accel * dt;
590 xprime[n][d] = x[n][d] + vNew * dt;
595 /*! \brief Handles the Leap-frog MD x and v integration */
596 static void do_update_md(int start,
600 const rvec* gmx_restrict x,
601 rvec* gmx_restrict xprime,
602 rvec* gmx_restrict v,
603 const rvec* gmx_restrict f,
604 const rvec* gmx_restrict accel,
607 const int nsttcouple,
608 const int nstpcouple,
610 const gmx_ekindata_t* ekind,
612 const double* gmx_restrict nh_vxi,
615 GMX_ASSERT(nrend == start || xprime != x,
616 "For SIMD optimization certain compilers need to have xprime != x");
618 /* Note: Berendsen pressure scaling is handled after do_update_md() */
619 bool doTempCouple = (etc != etcNO && do_per_step(step + nsttcouple - 1, nsttcouple));
620 bool doNoseHoover = (etc == etcNOSEHOOVER && doTempCouple);
621 bool doParrinelloRahman =
622 (epc == epcPARRINELLORAHMAN && do_per_step(step + nstpcouple - 1, nstpcouple));
623 bool doPROffDiagonal = (doParrinelloRahman && (M[YY][XX] != 0 || M[ZZ][XX] != 0 || M[ZZ][YY] != 0));
625 real dtPressureCouple = (doParrinelloRahman ? nstpcouple * dt : 0);
627 /* NEMD (also cosine) acceleration is applied in updateMDLeapFrogGeneral */
628 bool doAcceleration = (ekind->bNEMD || ekind->cosacc.cos_accel != 0);
630 if (doNoseHoover || doPROffDiagonal || doAcceleration)
633 if (!doParrinelloRahman)
635 /* We should not apply PR scaling at this step */
645 updateMDLeapfrogGeneral<AccelerationType::none>(start, nrend, doNoseHoover, dt,
646 dtPressureCouple, accel, md, ekind, box, x,
647 xprime, v, f, nh_vxi, nsttcouple, stepM);
649 else if (ekind->bNEMD)
651 updateMDLeapfrogGeneral<AccelerationType::group>(
652 start, nrend, doNoseHoover, dt, dtPressureCouple, accel, md, ekind, box, x,
653 xprime, v, f, nh_vxi, nsttcouple, stepM);
657 updateMDLeapfrogGeneral<AccelerationType::cosine>(
658 start, nrend, doNoseHoover, dt, dtPressureCouple, accel, md, ekind, box, x,
659 xprime, v, f, nh_vxi, nsttcouple, stepM);
664 /* Use a simple and thus more efficient integration loop. */
665 /* The simple loop does not check for particle type (so it can
666 * be vectorized), which means we need to clear the velocities
667 * of virtual sites in advance, when present. Note that vsite
668 * velocities are computed after update and constraints from
669 * their displacement.
673 /* Note: The overhead of this loop is completely neligible */
674 clearVsiteVelocities(start, nrend, md->ptype, v);
677 /* We determine if we have a single T-coupling lambda value for all
678 * atoms. That allows for better SIMD acceleration in the template.
679 * If we do not do temperature coupling (in the run or this step),
680 * all scaling values are 1, so we effectively have a single value.
682 bool haveSingleTempScaleValue = (!doTempCouple || ekind->ngtc == 1);
684 /* Extract some pointers needed by all cases */
685 const unsigned short* cTC = md->cTC;
686 gmx::ArrayRef<const t_grp_tcstat> tcstat = ekind->tcstat;
687 const rvec* invMassPerDim = md->invMassPerDim;
689 if (doParrinelloRahman)
691 GMX_ASSERT(!doPROffDiagonal,
692 "updateMDLeapfrogSimple only support diagonal Parrinello-Rahman scaling "
696 for (int d = 0; d < DIM; d++)
701 if (haveSingleTempScaleValue)
703 updateMDLeapfrogSimple<NumTempScaleValues::single, ApplyParrinelloRahmanVScaling::diagonal>(
704 start, nrend, dt, dtPressureCouple, invMassPerDim, tcstat, cTC, diagM, x,
709 updateMDLeapfrogSimple<NumTempScaleValues::multiple, ApplyParrinelloRahmanVScaling::diagonal>(
710 start, nrend, dt, dtPressureCouple, invMassPerDim, tcstat, cTC, diagM, x,
716 if (haveSingleTempScaleValue)
718 /* Note that modern compilers are pretty good at vectorizing
719 * updateMDLeapfrogSimple(). But the SIMD version will still
720 * be faster because invMass lowers the cache pressure
721 * compared to invMassPerDim.
723 #if GMX_HAVE_SIMD_UPDATE
724 /* Check if we can use invmass instead of invMassPerDim */
725 if (!md->havePartiallyFrozenAtoms)
727 updateMDLeapfrogSimpleSimd(start, nrend, dt, md->invmass, tcstat, x, xprime, v, f);
732 updateMDLeapfrogSimple<NumTempScaleValues::single, ApplyParrinelloRahmanVScaling::no>(
733 start, nrend, dt, dtPressureCouple, invMassPerDim, tcstat, cTC, nullptr,
739 updateMDLeapfrogSimple<NumTempScaleValues::multiple, ApplyParrinelloRahmanVScaling::no>(
740 start, nrend, dt, dtPressureCouple, invMassPerDim, tcstat, cTC, nullptr, x,
747 static void do_update_vv_vel(int start,
751 const ivec nFreeze[],
752 const real invmass[],
753 const unsigned short ptype[],
754 const unsigned short cFREEZE[],
755 const unsigned short cACC[],
768 g = 0.25 * dt * veta * alpha;
770 mv2 = gmx::series_sinhx(g);
777 for (n = start; n < nrend; n++)
779 real w_dt = invmass[n] * dt;
789 for (d = 0; d < DIM; d++)
791 if ((ptype[n] != eptVSite) && (ptype[n] != eptShell) && !nFreeze[gf][d])
793 v[n][d] = mv1 * (mv1 * v[n][d] + 0.5 * (w_dt * mv2 * f[n][d])) + 0.5 * accel[ga][d] * dt;
801 } /* do_update_vv_vel */
803 static void do_update_vv_pos(int start,
806 const ivec nFreeze[],
807 const unsigned short ptype[],
808 const unsigned short cFREEZE[],
819 /* Would it make more sense if Parrinello-Rahman was put here? */
824 mr2 = gmx::series_sinhx(g);
832 for (n = start; n < nrend; n++)
840 for (d = 0; d < DIM; d++)
842 if ((ptype[n] != eptVSite) && (ptype[n] != eptShell) && !nFreeze[gf][d])
844 xprime[n][d] = mr1 * (mr1 * x[n][d] + mr2 * dt * v[n][d]);
848 xprime[n][d] = x[n][d];
852 } /* do_update_vv_pos */
854 gmx_stochd_t::gmx_stochd_t(const t_inputrec& inputRecord)
856 const t_grpopts* opts = &inputRecord.opts;
857 const int ngtc = opts->ngtc;
859 if (inputRecord.eI == eiBD)
863 else if (EI_SD(inputRecord.eI))
868 for (int gt = 0; gt < ngtc; gt++)
870 if (opts->tau_t[gt] > 0)
872 sdc[gt].em = std::exp(-inputRecord.delta_t / opts->tau_t[gt]);
876 /* No friction and noise on this group */
881 else if (ETC_ANDERSEN(inputRecord.etc))
883 randomize_group.resize(ngtc);
884 boltzfac.resize(ngtc);
886 /* for now, assume that all groups, if randomized, are randomized at the same rate, i.e. tau_t is the same. */
887 /* since constraint groups don't necessarily match up with temperature groups! This is checked in readir.c */
889 for (int gt = 0; gt < ngtc; gt++)
891 real reft = std::max<real>(0, opts->ref_t[gt]);
892 if ((opts->tau_t[gt] > 0)
893 && (reft > 0)) /* tau_t or ref_t = 0 means that no randomization is done */
895 randomize_group[gt] = true;
896 boltzfac[gt] = BOLTZ * opts->ref_t[gt];
900 randomize_group[gt] = false;
906 void Update::Impl::update_temperature_constants(const t_inputrec& inputRecord)
908 if (inputRecord.eI == eiBD)
910 if (inputRecord.bd_fric != 0)
912 for (int gt = 0; gt < inputRecord.opts.ngtc; gt++)
914 sd_.bd_rf[gt] = std::sqrt(2.0 * BOLTZ * inputRecord.opts.ref_t[gt]
915 / (inputRecord.bd_fric * inputRecord.delta_t));
920 for (int gt = 0; gt < inputRecord.opts.ngtc; gt++)
922 sd_.bd_rf[gt] = std::sqrt(2.0 * BOLTZ * inputRecord.opts.ref_t[gt]);
926 if (inputRecord.eI == eiSD1)
928 for (int gt = 0; gt < inputRecord.opts.ngtc; gt++)
930 real kT = BOLTZ * inputRecord.opts.ref_t[gt];
931 /* The mass is accounted for later, since this differs per atom */
932 sd_.sdsig[gt].V = std::sqrt(kT * (1 - sd_.sdc[gt].em * sd_.sdc[gt].em));
937 Update::Impl::Impl(const t_inputrec& inputRecord, BoxDeformation* boxDeformation) :
939 deform_(boxDeformation)
941 update_temperature_constants(inputRecord);
942 xp_.resizeWithPadding(0);
945 void Update::setNumAtoms(int numAtoms)
948 impl_->xp()->resizeWithPadding(numAtoms);
951 /*! \brief Sets the SD update type */
952 enum class SDUpdate : int
955 FrictionAndNoiseOnly,
959 /*! \brief SD integrator update
961 * Two phases are required in the general case of a constrained
962 * update, the first phase from the contribution of forces, before
963 * applying constraints, and then a second phase applying the friction
964 * and noise, and then further constraining. For details, see
967 * Without constraints, the two phases can be combined, for
970 * Thus three instantiations of this templated function will be made,
971 * two with only one contribution, and one with both contributions. */
972 template<SDUpdate updateType>
973 static void doSDUpdateGeneral(const gmx_stochd_t& sd,
978 const ivec nFreeze[],
979 const real invmass[],
980 const unsigned short ptype[],
981 const unsigned short cFREEZE[],
982 const unsigned short cACC[],
983 const unsigned short cTC[],
992 // cTC, cACC and cFREEZE can be nullptr any time, but various
993 // instantiations do not make sense with particular pointer
995 if (updateType == SDUpdate::ForcesOnly)
997 GMX_ASSERT(f != nullptr, "SD update with only forces requires forces");
998 GMX_ASSERT(cTC == nullptr, "SD update with only forces cannot handle temperature groups");
1000 if (updateType == SDUpdate::FrictionAndNoiseOnly)
1002 GMX_ASSERT(f == nullptr, "SD update with only noise cannot handle forces");
1003 GMX_ASSERT(cACC == nullptr, "SD update with only noise cannot handle acceleration groups");
1005 if (updateType == SDUpdate::Combined)
1007 GMX_ASSERT(f != nullptr, "SD update with forces and noise requires forces");
1010 // Even 0 bits internal counter gives 2x64 ints (more than enough for three table lookups)
1011 gmx::ThreeFry2x64<0> rng(seed, gmx::RandomDomain::UpdateCoordinates);
1012 gmx::TabulatedNormalDistribution<real, 14> dist;
1014 for (int n = start; n < nrend; n++)
1016 int globalAtomIndex = gatindex ? gatindex[n] : n;
1017 rng.restart(step, globalAtomIndex);
1020 real inverseMass = invmass[n];
1021 real invsqrtMass = std::sqrt(inverseMass);
1023 int freezeGroup = cFREEZE ? cFREEZE[n] : 0;
1024 int accelerationGroup = cACC ? cACC[n] : 0;
1025 int temperatureGroup = cTC ? cTC[n] : 0;
1027 for (int d = 0; d < DIM; d++)
1029 if ((ptype[n] != eptVSite) && (ptype[n] != eptShell) && !nFreeze[freezeGroup][d])
1031 if (updateType == SDUpdate::ForcesOnly)
1033 real vn = v[n][d] + (inverseMass * f[n][d] + accel[accelerationGroup][d]) * dt;
1035 // Simple position update.
1036 xprime[n][d] = x[n][d] + v[n][d] * dt;
1038 else if (updateType == SDUpdate::FrictionAndNoiseOnly)
1041 v[n][d] = (vn * sd.sdc[temperatureGroup].em
1042 + invsqrtMass * sd.sdsig[temperatureGroup].V * dist(rng));
1043 // The previous phase already updated the
1044 // positions with a full v*dt term that must
1045 // now be half removed.
1046 xprime[n][d] = xprime[n][d] + 0.5 * (v[n][d] - vn) * dt;
1050 real vn = v[n][d] + (inverseMass * f[n][d] + accel[accelerationGroup][d]) * dt;
1051 v[n][d] = (vn * sd.sdc[temperatureGroup].em
1052 + invsqrtMass * sd.sdsig[temperatureGroup].V * dist(rng));
1053 // Here we include half of the friction+noise
1054 // update of v into the position update.
1055 xprime[n][d] = x[n][d] + 0.5 * (vn + v[n][d]) * dt;
1060 // When using constraints, the update is split into
1061 // two phases, but we only need to zero the update of
1062 // virtual, shell or frozen particles in at most one
1064 if (updateType != SDUpdate::FrictionAndNoiseOnly)
1067 xprime[n][d] = x[n][d];
1074 static void do_update_sd(int start,
1078 const rvec* gmx_restrict x,
1079 rvec* gmx_restrict xprime,
1080 rvec* gmx_restrict v,
1081 const rvec* gmx_restrict f,
1083 const ivec nFreeze[],
1084 const real invmass[],
1085 const unsigned short ptype[],
1086 const unsigned short cFREEZE[],
1087 const unsigned short cACC[],
1088 const unsigned short cTC[],
1090 const t_commrec* cr,
1091 const gmx_stochd_t& sd,
1092 bool haveConstraints)
1094 if (haveConstraints)
1096 // With constraints, the SD update is done in 2 parts
1097 doSDUpdateGeneral<SDUpdate::ForcesOnly>(sd, start, nrend, dt, accel, nFreeze, invmass,
1098 ptype, cFREEZE, cACC, nullptr, x, xprime, v, f,
1099 step, seed, nullptr);
1103 doSDUpdateGeneral<SDUpdate::Combined>(
1104 sd, start, nrend, dt, accel, nFreeze, invmass, ptype, cFREEZE, cACC, cTC, x, xprime,
1105 v, f, step, seed, DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr);
1109 static void do_update_bd(int start,
1113 const rvec* gmx_restrict x,
1114 rvec* gmx_restrict xprime,
1115 rvec* gmx_restrict v,
1116 const rvec* gmx_restrict f,
1117 const ivec nFreeze[],
1118 const real invmass[],
1119 const unsigned short ptype[],
1120 const unsigned short cFREEZE[],
1121 const unsigned short cTC[],
1122 real friction_coefficient,
1125 const int* gatindex)
1127 /* note -- these appear to be full step velocities . . . */
1132 // Use 1 bit of internal counters to give us 2*2 64-bits values per stream
1133 // Each 64-bit value is enough for 4 normal distribution table numbers.
1134 gmx::ThreeFry2x64<0> rng(seed, gmx::RandomDomain::UpdateCoordinates);
1135 gmx::TabulatedNormalDistribution<real, 14> dist;
1137 if (friction_coefficient != 0)
1139 invfr = 1.0 / friction_coefficient;
1142 for (n = start; (n < nrend); n++)
1144 int ng = gatindex ? gatindex[n] : n;
1146 rng.restart(step, ng);
1157 for (d = 0; (d < DIM); d++)
1159 if ((ptype[n] != eptVSite) && (ptype[n] != eptShell) && !nFreeze[gf][d])
1161 if (friction_coefficient != 0)
1163 vn = invfr * f[n][d] + rf[gt] * dist(rng);
1167 /* NOTE: invmass = 2/(mass*friction_constant*dt) */
1168 vn = 0.5 * invmass[n] * f[n][d] * dt
1169 + std::sqrt(0.5 * invmass[n]) * rf[gt] * dist(rng);
1173 xprime[n][d] = x[n][d] + vn * dt;
1178 xprime[n][d] = x[n][d];
1184 extern void init_ekinstate(ekinstate_t* ekinstate, const t_inputrec* ir)
1186 ekinstate->ekin_n = ir->opts.ngtc;
1187 snew(ekinstate->ekinh, ekinstate->ekin_n);
1188 snew(ekinstate->ekinf, ekinstate->ekin_n);
1189 snew(ekinstate->ekinh_old, ekinstate->ekin_n);
1190 ekinstate->ekinscalef_nhc.resize(ekinstate->ekin_n);
1191 ekinstate->ekinscaleh_nhc.resize(ekinstate->ekin_n);
1192 ekinstate->vscale_nhc.resize(ekinstate->ekin_n);
1193 ekinstate->dekindl = 0;
1194 ekinstate->mvcos = 0;
1195 ekinstate->hasReadEkinState = false;
1198 void update_ekinstate(ekinstate_t* ekinstate, const gmx_ekindata_t* ekind)
1202 for (i = 0; i < ekinstate->ekin_n; i++)
1204 copy_mat(ekind->tcstat[i].ekinh, ekinstate->ekinh[i]);
1205 copy_mat(ekind->tcstat[i].ekinf, ekinstate->ekinf[i]);
1206 copy_mat(ekind->tcstat[i].ekinh_old, ekinstate->ekinh_old[i]);
1207 ekinstate->ekinscalef_nhc[i] = ekind->tcstat[i].ekinscalef_nhc;
1208 ekinstate->ekinscaleh_nhc[i] = ekind->tcstat[i].ekinscaleh_nhc;
1209 ekinstate->vscale_nhc[i] = ekind->tcstat[i].vscale_nhc;
1212 copy_mat(ekind->ekin, ekinstate->ekin_total);
1213 ekinstate->dekindl = ekind->dekindl;
1214 ekinstate->mvcos = ekind->cosacc.mvcos;
1217 void restore_ekinstate_from_state(const t_commrec* cr, gmx_ekindata_t* ekind, const ekinstate_t* ekinstate)
1223 for (i = 0; i < ekinstate->ekin_n; i++)
1225 copy_mat(ekinstate->ekinh[i], ekind->tcstat[i].ekinh);
1226 copy_mat(ekinstate->ekinf[i], ekind->tcstat[i].ekinf);
1227 copy_mat(ekinstate->ekinh_old[i], ekind->tcstat[i].ekinh_old);
1228 ekind->tcstat[i].ekinscalef_nhc = ekinstate->ekinscalef_nhc[i];
1229 ekind->tcstat[i].ekinscaleh_nhc = ekinstate->ekinscaleh_nhc[i];
1230 ekind->tcstat[i].vscale_nhc = ekinstate->vscale_nhc[i];
1233 copy_mat(ekinstate->ekin_total, ekind->ekin);
1235 ekind->dekindl = ekinstate->dekindl;
1236 ekind->cosacc.mvcos = ekinstate->mvcos;
1237 n = ekinstate->ekin_n;
1242 gmx_bcast(sizeof(n), &n, cr->mpi_comm_mygroup);
1243 for (i = 0; i < n; i++)
1245 gmx_bcast(DIM * DIM * sizeof(ekind->tcstat[i].ekinh[0][0]), ekind->tcstat[i].ekinh[0],
1246 cr->mpi_comm_mygroup);
1247 gmx_bcast(DIM * DIM * sizeof(ekind->tcstat[i].ekinf[0][0]), ekind->tcstat[i].ekinf[0],
1248 cr->mpi_comm_mygroup);
1249 gmx_bcast(DIM * DIM * sizeof(ekind->tcstat[i].ekinh_old[0][0]),
1250 ekind->tcstat[i].ekinh_old[0], cr->mpi_comm_mygroup);
1252 gmx_bcast(sizeof(ekind->tcstat[i].ekinscalef_nhc), &(ekind->tcstat[i].ekinscalef_nhc),
1253 cr->mpi_comm_mygroup);
1254 gmx_bcast(sizeof(ekind->tcstat[i].ekinscaleh_nhc), &(ekind->tcstat[i].ekinscaleh_nhc),
1255 cr->mpi_comm_mygroup);
1256 gmx_bcast(sizeof(ekind->tcstat[i].vscale_nhc), &(ekind->tcstat[i].vscale_nhc),
1257 cr->mpi_comm_mygroup);
1259 gmx_bcast(DIM * DIM * sizeof(ekind->ekin[0][0]), ekind->ekin[0], cr->mpi_comm_mygroup);
1261 gmx_bcast(sizeof(ekind->dekindl), &ekind->dekindl, cr->mpi_comm_mygroup);
1262 gmx_bcast(sizeof(ekind->cosacc.mvcos), &ekind->cosacc.mvcos, cr->mpi_comm_mygroup);
1266 void getThreadAtomRange(int numThreads, int threadIndex, int numAtoms, int* startAtom, int* endAtom)
1268 #if GMX_HAVE_SIMD_UPDATE
1269 constexpr int blockSize = GMX_SIMD_REAL_WIDTH;
1271 constexpr int blockSize = 1;
1273 int numBlocks = (numAtoms + blockSize - 1) / blockSize;
1275 *startAtom = ((numBlocks * threadIndex) / numThreads) * blockSize;
1276 *endAtom = ((numBlocks * (threadIndex + 1)) / numThreads) * blockSize;
1277 if (threadIndex == numThreads - 1)
1279 *endAtom = numAtoms;
1283 void Update::Impl::update_sd_second_half(const t_inputrec& inputRecord,
1286 const t_mdatoms* md,
1288 const t_commrec* cr,
1290 gmx_wallcycle_t wcycle,
1291 gmx::Constraints* constr,
1299 if (inputRecord.eI == eiSD1)
1301 int homenr = md->homenr;
1303 /* Cast delta_t from double to real to make the integrators faster.
1304 * The only reason for having delta_t double is to get accurate values
1305 * for t=delta_t*step when step is larger than float precision.
1306 * For integration dt the accuracy of real suffices, since with
1307 * integral += dt*integrand the increment is nearly always (much) smaller
1308 * than the integral (and the integrand has real precision).
1310 real dt = inputRecord.delta_t;
1312 wallcycle_start(wcycle, ewcUPDATE);
1314 int nth = gmx_omp_nthreads_get(emntUpdate);
1316 #pragma omp parallel for num_threads(nth) schedule(static)
1317 for (int th = 0; th < nth; th++)
1321 int start_th, end_th;
1322 getThreadAtomRange(nth, th, homenr, &start_th, &end_th);
1324 doSDUpdateGeneral<SDUpdate::FrictionAndNoiseOnly>(
1325 sd_, start_th, end_th, dt, inputRecord.opts.acc, inputRecord.opts.nFreeze,
1326 md->invmass, md->ptype, md->cFREEZE, nullptr, md->cTC, state->x.rvec_array(),
1327 xp_.rvec_array(), state->v.rvec_array(), nullptr, step, inputRecord.ld_seed,
1328 DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr);
1330 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1332 inc_nrnb(nrnb, eNR_UPDATE, homenr);
1333 wallcycle_stop(wcycle, ewcUPDATE);
1335 /* Constrain the coordinates upd->xp for half a time step */
1336 bool computeVirial = false;
1337 constr->apply(do_log, do_ene, step, 1, 0.5, state->x.arrayRefWithPadding(),
1338 xp_.arrayRefWithPadding(), ArrayRef<RVec>(), state->box,
1339 state->lambda[efptBONDED], dvdlambda, state->v.arrayRefWithPadding(),
1340 computeVirial, nullptr, ConstraintVariable::Positions);
1344 void Update::Impl::finish_update(const t_inputrec& inputRecord,
1345 const t_mdatoms* md,
1347 gmx_wallcycle_t wcycle,
1348 const bool haveConstraints)
1350 /* NOTE: Currently we always integrate to a temporary buffer and
1351 * then copy the results back here.
1354 wallcycle_start_nocount(wcycle, ewcUPDATE);
1356 const int homenr = md->homenr;
1357 auto xp = makeConstArrayRef(xp_).subArray(0, homenr);
1358 auto x = makeArrayRef(state->x).subArray(0, homenr);
1360 if (md->havePartiallyFrozenAtoms && haveConstraints)
1362 /* We have atoms that are frozen along some, but not all dimensions,
1363 * then constraints will have moved them also along the frozen dimensions.
1364 * To freeze such degrees of freedom we do not copy them back here.
1366 const ivec* nFreeze = inputRecord.opts.nFreeze;
1368 for (int i = 0; i < homenr; i++)
1370 const int g = md->cFREEZE[i];
1372 for (int d = 0; d < DIM; d++)
1374 if (nFreeze[g][d] == 0)
1383 /* We have no frozen atoms or fully frozen atoms which have not
1384 * been moved by the update, so we can simply copy all coordinates.
1386 int gmx_unused nth = gmx_omp_nthreads_get(emntUpdate);
1387 #pragma omp parallel for num_threads(nth) schedule(static)
1388 for (int i = 0; i < homenr; i++)
1390 // Trivial statement, does not throw
1395 wallcycle_stop(wcycle, ewcUPDATE);
1398 void Update::Impl::update_coords(const t_inputrec& inputRecord,
1400 const t_mdatoms* md,
1402 const gmx::ArrayRefWithPadding<const gmx::RVec>& f,
1403 const t_fcdata& fcdata,
1404 const gmx_ekindata_t* ekind,
1407 const t_commrec* cr,
1408 const bool haveConstraints)
1410 /* Running the velocity half does nothing except for velocity verlet */
1411 if ((updatePart == etrtVELOCITY1 || updatePart == etrtVELOCITY2) && !EI_VV(inputRecord.eI))
1413 gmx_incons("update_coords called for velocity without VV integrator");
1416 int homenr = md->homenr;
1418 /* Cast to real for faster code, no loss in precision (see comment above) */
1419 real dt = inputRecord.delta_t;
1421 /* We need to update the NMR restraint history when time averaging is used */
1422 if (state->flags & (1 << estDISRE_RM3TAV))
1424 update_disres_history(*fcdata.disres, &state->hist);
1426 if (state->flags & (1 << estORIRE_DTAV))
1428 update_orires_history(*fcdata.orires, &state->hist);
1431 /* ############# START The update of velocities and positions ######### */
1432 int nth = gmx_omp_nthreads_get(emntUpdate);
1434 #pragma omp parallel for num_threads(nth) schedule(static)
1435 for (int th = 0; th < nth; th++)
1439 int start_th, end_th;
1440 getThreadAtomRange(nth, th, homenr, &start_th, &end_th);
1442 const rvec* x_rvec = state->x.rvec_array();
1443 rvec* xp_rvec = xp_.rvec_array();
1444 rvec* v_rvec = state->v.rvec_array();
1445 const rvec* f_rvec = as_rvec_array(f.unpaddedConstArrayRef().data());
1447 switch (inputRecord.eI)
1450 do_update_md(start_th, end_th, dt, step, x_rvec, xp_rvec, v_rvec, f_rvec,
1451 inputRecord.opts.acc, inputRecord.etc, inputRecord.epc,
1452 inputRecord.nsttcouple, inputRecord.nstpcouple, md, ekind,
1453 state->box, state->nosehoover_vxi.data(), M);
1456 do_update_sd(start_th, end_th, dt, step, x_rvec, xp_rvec, v_rvec, f_rvec,
1457 inputRecord.opts.acc, inputRecord.opts.nFreeze, md->invmass,
1458 md->ptype, md->cFREEZE, md->cACC, md->cTC, inputRecord.ld_seed, cr,
1459 sd_, haveConstraints);
1462 do_update_bd(start_th, end_th, dt, step, x_rvec, xp_rvec, v_rvec, f_rvec,
1463 inputRecord.opts.nFreeze, md->invmass, md->ptype, md->cFREEZE,
1464 md->cTC, inputRecord.bd_fric, sd_.bd_rf.data(), inputRecord.ld_seed,
1465 DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr);
1470 gmx_bool bExtended = (inputRecord.etc == etcNOSEHOOVER || inputRecord.epc == epcPARRINELLORAHMAN
1471 || inputRecord.epc == epcMTTK);
1473 /* assuming barostat coupled to group 0 */
1474 real alpha = 1.0 + DIM / static_cast<real>(inputRecord.opts.nrdf[0]);
1479 do_update_vv_vel(start_th, end_th, dt, inputRecord.opts.acc,
1480 inputRecord.opts.nFreeze, md->invmass, md->ptype, md->cFREEZE,
1481 md->cACC, v_rvec, f_rvec, bExtended, state->veta, alpha);
1484 do_update_vv_pos(start_th, end_th, dt, inputRecord.opts.nFreeze,
1485 md->ptype, md->cFREEZE, x_rvec, xp_rvec, v_rvec,
1486 bExtended, state->veta);
1491 default: gmx_fatal(FARGS, "Don't know how to update coordinates");
1494 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR