src/gromacs/mdlib/settle.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013,2014,2015,2016,2017,2018, 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 /*! \internal \file
  38  * \brief Defines SETTLE code.
  39  *
  40  * \author Berk Hess <hess@kth.se>
  41  * \author Mark Abraham <mark.j.abraham@gmail.com>
  42  * \ingroup module_mdlib
  43  */
  44 #include "gmxpre.h"
  45
  46 #include "settle.h"
  47
  48 #include <cassert>
  49 #include <cmath>
  50 #include <cstdio>
  51
  52 #include <algorithm>
  53
  54 #include "gromacs/math/functions.h"
  55 #include "gromacs/math/invertmatrix.h"
  56 #include "gromacs/math/vec.h"
  57 #include "gromacs/mdlib/constr.h"
  58 #include "gromacs/mdtypes/mdatom.h"
  59 #include "gromacs/pbcutil/ishift.h"
  60 #include "gromacs/pbcutil/pbc.h"
  61 #include "gromacs/pbcutil/pbc-simd.h"
  62 #include "gromacs/simd/simd.h"
  63 #include "gromacs/simd/simd_math.h"
  64 #include "gromacs/topology/idef.h"
  65 #include "gromacs/topology/ifunc.h"
  66 #include "gromacs/topology/mtop_util.h"
  67 #include "gromacs/utility/fatalerror.h"
  68 #include "gromacs/utility/gmxassert.h"
  69 #include "gromacs/utility/smalloc.h"
  70
  71 namespace gmx
  72 {
  73
  74 struct settleparam_t
  75 {
  76     real   mO;
  77     real   mH;
  78     real   wh;
  79     real   dOH;
  80     real   dHH;
  81     real   ra;
  82     real   rb;
  83     real   rc;
  84     real   irc2;
  85     /* For projection */
  86     real   imO;
  87     real   imH;
  88     real   invdOH;
  89     real   invdHH;
  90     matrix invmat;
  91 };
  92
  93 struct settledata
  94 {
  95     settleparam_t massw;    /* Parameters for SETTLE for coordinates */
  96     settleparam_t mass1;    /* Parameters with all masses 1, for forces */
  97
  98     int           nsettle;  /* The number of settles on our rank */
  99     int          *ow1;      /* Index to OW1 atoms, size nsettle + SIMD padding */
 100     int          *hw2;      /* Index to HW2 atoms, size nsettle + SIMD padding */
 101     int          *hw3;      /* Index to HW3 atoms, size nsettle + SIMD padding */
 102     real         *virfac;   /* Virial factor 0 or 1, size nsettle + SIMD pad. */
 103     int           nalloc;   /* Allocation size of ow1, hw2, hw3, virfac */
 104
 105     bool          bUseSimd; /* Use SIMD intrinsics code, if possible */
 106 };
 107
 108
 109 //! Initializes a projection matrix.
 110 static void init_proj_matrix(real invmO, real invmH, real dOH, real dHH,
 111                              matrix inverseCouplingMatrix)
 112 {
 113     /* We normalize the inverse masses with invmO for the matrix inversion.
 114      * so we can keep using masses of almost zero for frozen particles,
 115      * without running out of the float range in invertMatrix.
 116      */
 117     double invmORelative = 1.0;
 118     double invmHRelative = invmH/static_cast<double>(invmO);
 119     double distanceRatio = dHH/static_cast<double>(dOH);
 120
 121     /* Construct the constraint coupling matrix */
 122     matrix mat;
 123     mat[0][0] = invmORelative + invmHRelative;
 124     mat[0][1] = invmORelative*(1.0 - 0.5*gmx::square(distanceRatio));
 125     mat[0][2] = invmHRelative*0.5*distanceRatio;
 126     mat[1][1] = mat[0][0];
 127     mat[1][2] = mat[0][2];
 128     mat[2][2] = invmHRelative + invmHRelative;
 129     mat[1][0] = mat[0][1];
 130     mat[2][0] = mat[0][2];
 131     mat[2][1] = mat[1][2];
 132
 133     invertMatrix(mat, inverseCouplingMatrix);
 134
 135     msmul(inverseCouplingMatrix, 1/invmO, inverseCouplingMatrix);
 136 }
 137
 138 //! Initializes settle parameters.
 139 static void settleparam_init(settleparam_t *p,
 140                              real mO, real mH, real invmO, real invmH,
 141                              real dOH, real dHH)
 142 {
 143     /* We calculate parameters in double precision to minimize errors.
 144      * The velocity correction applied during SETTLE coordinate constraining
 145      * introduces a systematic error of approximately 1 bit per atom,
 146      * depending on what the compiler does with the code.
 147      */
 148     double wohh;
 149
 150     p->mO     = mO;
 151     p->mH     = mH;
 152     wohh      = mO + 2.0*mH;
 153     p->wh     = mH/wohh;
 154     p->dOH    = dOH;
 155     p->dHH    = dHH;
 156     double rc = dHH/2.0;
 157     double ra = 2.0*mH*std::sqrt(dOH*dOH - rc*rc)/wohh;
 158     p->rb     = std::sqrt(dOH*dOH - rc*rc) - ra;
 159     p->rc     = rc;
 160     p->ra     = ra;
 161     p->irc2   = 1.0/dHH;
 162
 163     /* For projection: inverse masses and coupling matrix inversion */
 164     p->imO    = invmO;
 165     p->imH    = invmH;
 166
 167     p->invdOH = 1.0/dOH;
 168     p->invdHH = 1.0/dHH;
 169
 170     init_proj_matrix(invmO, invmH, dOH, dHH, p->invmat);
 171
 172     if (debug)
 173     {
 174         fprintf(debug, "wh =%g, rc = %g, ra = %g\n",
 175                 p->wh, p->rc, p->ra);
 176         fprintf(debug, "rb = %g, irc2 = %g, dHH = %g, dOH = %g\n",
 177                 p->rb, p->irc2, p->dHH, p->dOH);
 178     }
 179 }
 180
 181 settledata *settle_init(const gmx_mtop_t &mtop)
 182 {
 183     /* Check that we have only one settle type */
 184     int                    settle_type = -1;
 185     gmx_mtop_ilistloop_t   iloop       = gmx_mtop_ilistloop_init(mtop);
 186     int                    nmol;
 187     const int              nral1       = 1 + NRAL(F_SETTLE);
 188     while (const InteractionLists *ilists = gmx_mtop_ilistloop_next(iloop, &nmol))
 189     {
 190         const InteractionList &ilist = (*ilists)[F_SETTLE];
 191         for (int i = 0; i < ilist.size(); i += nral1)
 192         {
 193             if (settle_type == -1)
 194             {
 195                 settle_type = ilist.iatoms[i];
 196             }
 197             else if (ilist.iatoms[i] != settle_type)
 198             {
 199                 gmx_fatal(FARGS,
 200                           "The [molecules] section of your topology specifies more than one block of\n"
 201                           "a [moleculetype] with a [settles] block. Only one such is allowed.\n"
 202                           "If you are trying to partition your solvent into different *groups*\n"
 203                           "(e.g. for freezing, T-coupling, etc.), you are using the wrong approach. Index\n"
 204                           "files specify groups. Otherwise, you may wish to change the least-used\n"
 205                           "block of molecules with SETTLE constraints into 3 normal constraints.");
 206             }
 207         }
 208     }
 209     GMX_RELEASE_ASSERT(settle_type >= 0, "settle_init called without settles");
 210
 211     settledata *settled;
 212
 213     snew(settled, 1);
 214
 215     /* We will not initialize the normal SETTLE parameters here yet,
 216      * since the atom (inv)masses can depend on the integrator and
 217      * free-energy perturbation. We set mO=-1 to trigger later initialization.
 218      */
 219     settled->massw.mO = -1;
 220
 221     real dOH = mtop.ffparams.iparams[settle_type].settle.doh;
 222     real dHH = mtop.ffparams.iparams[settle_type].settle.dhh;
 223     settleparam_init(&settled->mass1, 1.0, 1.0, 1.0, 1.0, dOH, dHH);
 224
 225     settled->ow1    = nullptr;
 226     settled->hw2    = nullptr;
 227     settled->hw3    = nullptr;
 228     settled->virfac = nullptr;
 229     settled->nalloc = 0;
 230
 231     /* Without SIMD configured, this bool is not used */
 232     settled->bUseSimd = (getenv("GMX_DISABLE_SIMD_KERNELS") == nullptr);
 233
 234     return settled;
 235 }
 236
 237 void settle_free(settledata *settled)
 238 {
 239     sfree_aligned(settled->ow1);
 240     sfree_aligned(settled->hw2);
 241     sfree_aligned(settled->hw3);
 242     sfree_aligned(settled->virfac);
 243     sfree(settled);
 244 }
 245
 246 void settle_set_constraints(settledata       *settled,
 247                             const t_ilist    *il_settle,
 248                             const t_mdatoms  &mdatoms)
 249 {
 250 #if GMX_SIMD_HAVE_REAL
 251     const int pack_size = GMX_SIMD_REAL_WIDTH;
 252 #else
 253     const int pack_size = 1;
 254 #endif
 255
 256     const int nral1     = 1 + NRAL(F_SETTLE);
 257     int       nsettle   = il_settle->nr/nral1;
 258     settled->nsettle = nsettle;
 259
 260     if (nsettle > 0)
 261     {
 262         const t_iatom *iatoms = il_settle->iatoms;
 263
 264         /* Here we initialize the normal SETTLE parameters */
 265         if (settled->massw.mO < 0)
 266         {
 267             int firstO = iatoms[1];
 268             int firstH = iatoms[2];
 269             settleparam_init(&settled->massw,
 270                              mdatoms.massT[firstO],
 271                              mdatoms.massT[firstH],
 272                              mdatoms.invmass[firstO],
 273                              mdatoms.invmass[firstH],
 274                              settled->mass1.dOH,
 275                              settled->mass1.dHH);
 276         }
 277
 278         if (nsettle + pack_size > settled->nalloc)
 279         {
 280             settled->nalloc = over_alloc_dd(nsettle + pack_size);
 281             sfree_aligned(settled->ow1);
 282             sfree_aligned(settled->hw2);
 283             sfree_aligned(settled->hw3);
 284             sfree_aligned(settled->virfac);
 285             snew_aligned(settled->ow1, settled->nalloc, 64);
 286             snew_aligned(settled->hw2, settled->nalloc, 64);
 287             snew_aligned(settled->hw3, settled->nalloc, 64);
 288             snew_aligned(settled->virfac, settled->nalloc, 64);
 289         }
 290
 291         for (int i = 0; i < nsettle; i++)
 292         {
 293             settled->ow1[i]    = iatoms[i*nral1 + 1];
 294             settled->hw2[i]    = iatoms[i*nral1 + 2];
 295             settled->hw3[i]    = iatoms[i*nral1 + 3];
 296             /* We should avoid double counting of virial contributions for
 297              * SETTLEs that appear in multiple DD domains, so we only count
 298              * the contribution on the home range of the oxygen atom.
 299              */
 300             settled->virfac[i] = (iatoms[i*nral1 + 1] < mdatoms.homenr ? 1 : 0);
 301         }
 302
 303         /* Pack the index array to the full SIMD width with copies from
 304          * the last normal entry, but with no virial contribution.
 305          */
 306         int end_packed = ((nsettle + pack_size - 1)/pack_size)*pack_size;
 307         for (int i = nsettle; i < end_packed; i++)
 308         {
 309             settled->ow1[i]    = settled->ow1[nsettle - 1];
 310             settled->hw2[i]    = settled->hw2[nsettle - 1];
 311             settled->hw3[i]    = settled->hw3[nsettle - 1];
 312             settled->virfac[i] = 0;
 313         }
 314     }
 315 }
 316
 317 void settle_proj(settledata *settled, ConstraintVariable econq,
 318                  int nsettle, const t_iatom iatoms[],
 319                  const t_pbc *pbc,
 320                  const rvec x[],
 321                  rvec *der, rvec *derp,
 322                  int calcvir_atom_end, tensor vir_r_m_dder)
 323 {
 324     /* Settle for projection out constraint components
 325      * of derivatives of the coordinates.
 326      * Berk Hess 2008-1-10
 327      */
 328
 329     settleparam_t *p;
 330     real           imO, imH, dOH, dHH, invdOH, invdHH;
 331     matrix         invmat;
 332     int            i, m, m2, ow1, hw2, hw3;
 333     rvec           roh2, roh3, rhh, dc, fc;
 334
 335     calcvir_atom_end *= DIM;
 336
 337     if (econq == ConstraintVariable::Force)
 338     {
 339         p = &settled->mass1;
 340     }
 341     else
 342     {
 343         p = &settled->massw;
 344     }
 345     imO    = p->imO;
 346     imH    = p->imH;
 347     copy_mat(p->invmat, invmat);
 348     dOH    = p->dOH;
 349     dHH    = p->dHH;
 350     invdOH = p->invdOH;
 351     invdHH = p->invdHH;
 352
 353     const int nral1 = 1 + NRAL(F_SETTLE);
 354
 355     for (i = 0; i < nsettle; i++)
 356     {
 357         ow1 = iatoms[i*nral1 + 1];
 358         hw2 = iatoms[i*nral1 + 2];
 359         hw3 = iatoms[i*nral1 + 3];
 360
 361         if (pbc == nullptr)
 362         {
 363             rvec_sub(x[ow1], x[hw2], roh2);
 364             rvec_sub(x[ow1], x[hw3], roh3);
 365             rvec_sub(x[hw2], x[hw3], rhh);
 366         }
 367         else
 368         {
 369             pbc_dx_aiuc(pbc, x[ow1], x[hw2], roh2);
 370             pbc_dx_aiuc(pbc, x[ow1], x[hw3], roh3);
 371             pbc_dx_aiuc(pbc, x[hw2], x[hw3], rhh);
 372         }
 373         svmul(invdOH, roh2, roh2);
 374         svmul(invdOH, roh3, roh3);
 375         svmul(invdHH, rhh, rhh);
 376         /* 18 flops */
 377
 378         /* Determine the projections of der on the bonds */
 379         clear_rvec(dc);
 380         for (m = 0; m < DIM; m++)
 381         {
 382             dc[0] += (der[ow1][m] - der[hw2][m])*roh2[m];
 383             dc[1] += (der[ow1][m] - der[hw3][m])*roh3[m];
 384             dc[2] += (der[hw2][m] - der[hw3][m])*rhh [m];
 385         }
 386         /* 27 flops */
 387
 388         /* Determine the correction for the three bonds */
 389         mvmul(invmat, dc, fc);
 390         /* 15 flops */
 391
 392         /* Subtract the corrections from derp */
 393         for (m = 0; m < DIM; m++)
 394         {
 395             derp[ow1][m] -= imO*( fc[0]*roh2[m] + fc[1]*roh3[m]);
 396             derp[hw2][m] -= imH*(-fc[0]*roh2[m] + fc[2]*rhh [m]);
 397             derp[hw3][m] -= imH*(-fc[1]*roh3[m] - fc[2]*rhh [m]);
 398         }
 399
 400         /* 45 flops */
 401
 402         if (ow1 < calcvir_atom_end)
 403         {
 404             /* Determining r \dot m der is easy,
 405              * since fc contains the mass weighted corrections for der.
 406              */
 407
 408             for (m = 0; m < DIM; m++)
 409             {
 410                 for (m2 = 0; m2 < DIM; m2++)
 411                 {
 412                     vir_r_m_dder[m][m2] +=
 413                         dOH*roh2[m]*roh2[m2]*fc[0] +
 414                         dOH*roh3[m]*roh3[m2]*fc[1] +
 415                         dHH*rhh [m]*rhh [m2]*fc[2];
 416                 }
 417             }
 418         }
 419     }
 420 }
 421
 422
 423 /*! \brief The actual settle code, templated for real/SimdReal and for optimization */
 424 template<typename T, typename TypeBool, int packSize,
 425          typename TypePbc,
 426          bool bCorrectVelocity,
 427          bool bCalcVirial>
 428 static void settleTemplate(const settledata *settled,
 429                            int settleStart, int settleEnd,
 430                            const TypePbc pbc,
 431                            const real *x, real *xprime,
 432                            real invdt, real * gmx_restrict v,
 433                            tensor vir_r_m_dr,
 434                            bool *bErrorHasOccurred)
 435 {
 436     /* ******************************************************************* */
 437     /*                                                                  ** */
 438     /*    Original code by Shuichi Miyamoto, last update Oct. 1, 1992   ** */
 439     /*                                                                  ** */
 440     /*    Algorithm changes by Berk Hess:                               ** */
 441     /*    2004-07-15 Convert COM to double precision to avoid drift     ** */
 442     /*    2006-10-16 Changed velocity update to use differences         ** */
 443     /*    2012-09-24 Use oxygen as reference instead of COM             ** */
 444     /*    2016-02    Complete rewrite of the code for SIMD              ** */
 445     /*                                                                  ** */
 446     /*    Reference for the SETTLE algorithm                            ** */
 447     /*           S. Miyamoto et al., J. Comp. Chem., 13, 952 (1992).    ** */
 448     /*                                                                  ** */
 449     /* ******************************************************************* */
 450
 451     assert(settleStart % packSize == 0);
 452     assert(settleEnd   % packSize == 0);
 453
 454     TypeBool             bError = TypeBool(false);
 455
 456     const settleparam_t *p    = &settled->massw;
 457     T                    wh   = T(p->wh);
 458     T                    rc   = T(p->rc);
 459     T                    ra   = T(p->ra);
 460     T                    rb   = T(p->rb);
 461     T                    irc2 = T(p->irc2);
 462     T                    mO   = T(p->mO);
 463     T                    mH   = T(p->mH);
 464
 465     T                    almost_zero = T(1e-12);
 466
 467     T                    sum_r_m_dr[DIM][DIM];
 468
 469     if (bCalcVirial)
 470     {
 471         for (int d2 = 0; d2 < DIM; d2++)
 472         {
 473             for (int d = 0; d < DIM; d++)
 474             {
 475                 sum_r_m_dr[d2][d] = T(0);
 476             }
 477         }
 478     }
 479
 480     for (int i = settleStart; i < settleEnd; i += packSize)
 481     {
 482         /* Here we pad up to packSize with copies from the last valid entry.
 483          * This gives correct results, since we store (not increment) all
 484          * output, so we store the same output multiple times.
 485          */
 486         const int *ow1 = settled->ow1 + i;
 487         const int *hw2 = settled->hw2 + i;
 488         const int *hw3 = settled->hw3 + i;
 489
 490         T          x_ow1[DIM], x_hw2[DIM], x_hw3[DIM];
 491
 492         gatherLoadUTranspose<3>(x, ow1, &x_ow1[XX], &x_ow1[YY], &x_ow1[ZZ]);
 493         gatherLoadUTranspose<3>(x, hw2, &x_hw2[XX], &x_hw2[YY], &x_hw2[ZZ]);
 494         gatherLoadUTranspose<3>(x, hw3, &x_hw3[XX], &x_hw3[YY], &x_hw3[ZZ]);
 495
 496         T xprime_ow1[DIM], xprime_hw2[DIM], xprime_hw3[DIM];
 497
 498         gatherLoadUTranspose<3>(xprime, ow1, &xprime_ow1[XX], &xprime_ow1[YY], &xprime_ow1[ZZ]);
 499         gatherLoadUTranspose<3>(xprime, hw2, &xprime_hw2[XX], &xprime_hw2[YY], &xprime_hw2[ZZ]);
 500         gatherLoadUTranspose<3>(xprime, hw3, &xprime_hw3[XX], &xprime_hw3[YY], &xprime_hw3[ZZ]);
 501
 502         T dist21[DIM], dist31[DIM];
 503         T doh2[DIM], doh3[DIM];
 504         T sh_hw2[DIM], sh_hw3[DIM];
 505
 506         pbc_dx_aiuc(pbc, x_hw2, x_ow1, dist21);
 507
 508         pbc_dx_aiuc(pbc, x_hw3, x_ow1, dist31);
 509
 510         /* Tedious way of doing pbc */
 511         pbc_dx_aiuc(pbc, xprime_hw2, xprime_ow1, doh2);
 512         for (int d = 0; d < DIM; d++)
 513         {
 514             sh_hw2[d]     = xprime_hw2[d] - (xprime_ow1[d] + doh2[d]);
 515             xprime_hw2[d] = xprime_hw2[d] - sh_hw2[d];
 516         }
 517         pbc_dx_aiuc(pbc, xprime_hw3, xprime_ow1, doh3);
 518         for (int d = 0; d < DIM; d++)
 519         {
 520             sh_hw3[d]     = xprime_hw3[d] - (xprime_ow1[d] + doh3[d]);
 521             xprime_hw3[d] = xprime_hw3[d] - sh_hw3[d];
 522         }
 523
 524         /* Not calculating the center of mass using the oxygen position
 525          * and the O-H distances, as done below, will make SETTLE
 526          * the largest source of energy drift for simulations of water,
 527          * as then the oxygen coordinate is multiplied by 0.89 at every step,
 528          * which can then transfer a systematic rounding to the oxygen velocity.
 529          */
 530         T a1[DIM], com[DIM];
 531         for (int d = 0; d < DIM; d++)
 532         {
 533             a1[d]  = -(doh2[d] + doh3[d]) * wh;
 534             com[d] = xprime_ow1[d] - a1[d];
 535         }
 536         T b1[DIM];
 537         for (int d = 0; d < DIM; d++)
 538         {
 539             b1[d]  = xprime_hw2[d] - com[d];
 540         }
 541         T c1[DIM];
 542         for (int d = 0; d < DIM; d++)
 543         {
 544             c1[d]  = xprime_hw3[d] - com[d];
 545         }
 546         /* 15 flops */
 547
 548         T xakszd = dist21[YY] * dist31[ZZ] - dist21[ZZ] * dist31[YY];
 549         T yakszd = dist21[ZZ] * dist31[XX] - dist21[XX] * dist31[ZZ];
 550         T zakszd = dist21[XX] * dist31[YY] - dist21[YY] * dist31[XX];
 551         T xaksxd = a1[YY] * zakszd - a1[ZZ] * yakszd;
 552         T yaksxd = a1[ZZ] * xakszd - a1[XX] * zakszd;
 553         T zaksxd = a1[XX] * yakszd - a1[YY] * xakszd;
 554         T xaksyd = yakszd * zaksxd - zakszd * yaksxd;
 555         T yaksyd = zakszd * xaksxd - xakszd * zaksxd;
 556         T zaksyd = xakszd * yaksxd - yakszd * xaksxd;
 557         /* 27 flops */
 558
 559         T axlng = gmx::invsqrt(xaksxd * xaksxd + yaksxd * yaksxd + zaksxd * zaksxd);
 560         T aylng = gmx::invsqrt(xaksyd * xaksyd + yaksyd * yaksyd + zaksyd * zaksyd);
 561         T azlng = gmx::invsqrt(xakszd * xakszd + yakszd * yakszd + zakszd * zakszd);
 562
 563         T trns1[DIM], trns2[DIM], trns3[DIM];
 564
 565         trns1[XX] = xaksxd * axlng;
 566         trns2[XX] = yaksxd * axlng;
 567         trns3[XX] = zaksxd * axlng;
 568         trns1[YY] = xaksyd * aylng;
 569         trns2[YY] = yaksyd * aylng;
 570         trns3[YY] = zaksyd * aylng;
 571         trns1[ZZ] = xakszd * azlng;
 572         trns2[ZZ] = yakszd * azlng;
 573         trns3[ZZ] = zakszd * azlng;
 574         /* 24 flops */
 575
 576         T b0d[2], c0d[2];
 577
 578         for (int d = 0; d < 2; d++)
 579         {
 580             b0d[d] = trns1[d] * dist21[XX] + trns2[d] * dist21[YY] + trns3[d] * dist21[ZZ];
 581             c0d[d] = trns1[d] * dist31[XX] + trns2[d] * dist31[YY] + trns3[d] * dist31[ZZ];
 582         }
 583
 584         T a1d_z, b1d[DIM], c1d[DIM];
 585
 586         a1d_z = trns1[ZZ] * a1[XX] + trns2[ZZ] * a1[YY] + trns3[ZZ] * a1[ZZ];
 587         for (int d = 0; d < DIM; d++)
 588         {
 589             b1d[d] = trns1[d] * b1[XX] + trns2[d] * b1[YY] + trns3[d] * b1[ZZ];
 590             c1d[d] = trns1[d] * c1[XX] + trns2[d] * c1[YY] + trns3[d] * c1[ZZ];
 591         }
 592         /* 65 flops */
 593
 594         T tmp, tmp2;
 595
 596         T sinphi = a1d_z * gmx::invsqrt(ra*ra);
 597         tmp2     = 1.0 - sinphi * sinphi;
 598
 599         /* If tmp2 gets close to or beyond zero we have severly distorted
 600          * water molecules and we should terminate the simulation.
 601          * Below we take the max with almost_zero to continue the loop.
 602          */
 603         bError   = bError || (tmp2 <= almost_zero);
 604
 605         tmp2     = max(tmp2, almost_zero);
 606         tmp      = gmx::invsqrt(tmp2);
 607         T cosphi = tmp2*tmp;
 608         T sinpsi = (b1d[ZZ] - c1d[ZZ]) * irc2 * tmp;
 609         tmp2     = 1.0 - sinpsi * sinpsi;
 610
 611         T cospsi = tmp2*gmx::invsqrt(tmp2);
 612         /* 46 flops */
 613
 614         T a2d_y  =  ra * cosphi;
 615         T b2d_x  = -rc * cospsi;
 616         T t1     = -rb * cosphi;
 617         T t2     =  rc * sinpsi * sinphi;
 618         T b2d_y  =  t1 - t2;
 619         T c2d_y  =  t1 + t2;
 620         /* 7 flops */
 621
 622         /*     --- Step3  al,be,ga            --- */
 623         T alpha  = b2d_x * (b0d[XX] - c0d[XX]) + b0d[YY] * b2d_y + c0d[YY] * c2d_y;
 624         T beta   = b2d_x * (c0d[YY] - b0d[YY]) + b0d[XX] * b2d_y + c0d[XX] * c2d_y;
 625         T gamma  = b0d[XX] * b1d[YY] - b1d[XX] * b0d[YY] + c0d[XX] * c1d[YY] - c1d[XX] * c0d[YY];
 626         T al2be2 = alpha * alpha + beta * beta;
 627         tmp2     = (al2be2 - gamma * gamma);
 628         T sinthe = (alpha * gamma - beta * tmp2*gmx::invsqrt(tmp2)) * gmx::invsqrt(al2be2*al2be2);
 629         /* 47 flops */
 630
 631         /*  --- Step4  A3' --- */
 632         tmp2     = 1.0 - sinthe * sinthe;
 633         T costhe = tmp2*gmx::invsqrt(tmp2);
 634
 635         T a3d[DIM], b3d[DIM], c3d[DIM];
 636
 637         a3d[XX]  = -a2d_y * sinthe;
 638         a3d[YY]  = a2d_y * costhe;
 639         a3d[ZZ]  = a1d_z;
 640         b3d[XX]  = b2d_x * costhe - b2d_y * sinthe;
 641         b3d[YY]  = b2d_x * sinthe + b2d_y * costhe;
 642         b3d[ZZ]  = b1d[ZZ];
 643         c3d[XX]  = -b2d_x * costhe - c2d_y * sinthe;
 644         c3d[YY]  = -b2d_x * sinthe + c2d_y * costhe;
 645         c3d[ZZ]  = c1d[ZZ];
 646         /* 26 flops */
 647
 648         /*    --- Step5  A3 --- */
 649         T a3[DIM], b3[DIM], c3[DIM];
 650
 651         a3[XX] = trns1[XX]*a3d[XX] + trns1[YY]*a3d[YY] + trns1[ZZ]*a3d[ZZ];
 652         a3[YY] = trns2[XX]*a3d[XX] + trns2[YY]*a3d[YY] + trns2[ZZ]*a3d[ZZ];
 653         a3[ZZ] = trns3[XX]*a3d[XX] + trns3[YY]*a3d[YY] + trns3[ZZ]*a3d[ZZ];
 654         b3[XX] = trns1[XX]*b3d[XX] + trns1[YY]*b3d[YY] + trns1[ZZ]*b3d[ZZ];
 655         b3[YY] = trns2[XX]*b3d[XX] + trns2[YY]*b3d[YY] + trns2[ZZ]*b3d[ZZ];
 656         b3[ZZ] = trns3[XX]*b3d[XX] + trns3[YY]*b3d[YY] + trns3[ZZ]*b3d[ZZ];
 657         c3[XX] = trns1[XX]*c3d[XX] + trns1[YY]*c3d[YY] + trns1[ZZ]*c3d[ZZ];
 658         c3[YY] = trns2[XX]*c3d[XX] + trns2[YY]*c3d[YY] + trns2[ZZ]*c3d[ZZ];
 659         c3[ZZ] = trns3[XX]*c3d[XX] + trns3[YY]*c3d[YY] + trns3[ZZ]*c3d[ZZ];
 660         /* 45 flops */
 661
 662         /* Compute and store the corrected new coordinate */
 663         for (int d = 0; d < DIM; d++)
 664         {
 665             xprime_ow1[d] = com[d] + a3[d];
 666         }
 667         for (int d = 0; d < DIM; d++)
 668         {
 669             xprime_hw2[d] = com[d] + b3[d] + sh_hw2[d];;
 670         }
 671         for (int d = 0; d < DIM; d++)
 672         {
 673             xprime_hw3[d] = com[d] + c3[d] + sh_hw3[d];
 674         }
 675         /* 9 flops + 6 pbc flops */
 676
 677         transposeScatterStoreU<3>(xprime, ow1, xprime_ow1[XX], xprime_ow1[YY], xprime_ow1[ZZ]);
 678         transposeScatterStoreU<3>(xprime, hw2, xprime_hw2[XX], xprime_hw2[YY], xprime_hw2[ZZ]);
 679         transposeScatterStoreU<3>(xprime, hw3, xprime_hw3[XX], xprime_hw3[YY], xprime_hw3[ZZ]);
 680
 681         if (bCorrectVelocity || bCalcVirial)
 682         {
 683             T da[DIM], db[DIM], dc[DIM];
 684             for (int d = 0; d < DIM; d++)
 685             {
 686                 da[d] = a3[d] - a1[d];
 687             }
 688             for (int d = 0; d < DIM; d++)
 689             {
 690                 db[d] = b3[d] - b1[d];
 691             }
 692             for (int d = 0; d < DIM; d++)
 693             {
 694                 dc[d] = c3[d] - c1[d];
 695             }
 696             /* 9 flops */
 697
 698             if (bCorrectVelocity)
 699             {
 700                 T v_ow1[DIM], v_hw2[DIM], v_hw3[DIM];
 701
 702                 gatherLoadUTranspose<3>(v, ow1, &v_ow1[XX], &v_ow1[YY], &v_ow1[ZZ]);
 703                 gatherLoadUTranspose<3>(v, hw2, &v_hw2[XX], &v_hw2[YY], &v_hw2[ZZ]);
 704                 gatherLoadUTranspose<3>(v, hw3, &v_hw3[XX], &v_hw3[YY], &v_hw3[ZZ]);
 705
 706                 /* Add the position correction divided by dt to the velocity */
 707                 for (int d = 0; d < DIM; d++)
 708                 {
 709                     v_ow1[d] = gmx::fma(da[d], invdt, v_ow1[d]);
 710                 }
 711                 for (int d = 0; d < DIM; d++)
 712                 {
 713                     v_hw2[d] = gmx::fma(db[d], invdt, v_hw2[d]);
 714                 }
 715                 for (int d = 0; d < DIM; d++)
 716                 {
 717                     v_hw3[d] = gmx::fma(dc[d], invdt, v_hw3[d]);
 718                 }
 719                 /* 3*6 flops */
 720
 721                 transposeScatterStoreU<3>(v, ow1, v_ow1[XX], v_ow1[YY], v_ow1[ZZ]);
 722                 transposeScatterStoreU<3>(v, hw2, v_hw2[XX], v_hw2[YY], v_hw2[ZZ]);
 723                 transposeScatterStoreU<3>(v, hw3, v_hw3[XX], v_hw3[YY], v_hw3[ZZ]);
 724             }
 725
 726             if (bCalcVirial)
 727             {
 728                 /* Filter out the non-local settles */
 729                 T filter = load<T>(settled->virfac + i);
 730                 T mOf    = filter*mO;
 731                 T mHf    = filter*mH;
 732
 733                 T mdo[DIM], mdb[DIM], mdc[DIM];
 734
 735                 for (int d = 0; d < DIM; d++)
 736                 {
 737                     mdb[d] = mHf*db[d];
 738                     mdc[d] = mHf*dc[d];
 739                     mdo[d] = mOf*da[d] + mdb[d] + mdc[d];
 740                 }
 741
 742                 for (int d2 = 0; d2 < DIM; d2++)
 743                 {
 744                     for (int d = 0; d < DIM; d++)
 745                     {
 746                         sum_r_m_dr[d2][d] = sum_r_m_dr[d2][d] -
 747                             (x_ow1[d2]*mdo[d] +
 748                              dist21[d2]*mdb[d] +
 749                              dist31[d2]*mdc[d]);
 750                     }
 751                 }
 752                 /* 71 flops */
 753             }
 754         }
 755     }
 756
 757     if (bCalcVirial)
 758     {
 759         for (int d2 = 0; d2 < DIM; d2++)
 760         {
 761             for (int d = 0; d < DIM; d++)
 762             {
 763                 vir_r_m_dr[d2][d] += reduce(sum_r_m_dr[d2][d]);
 764             }
 765         }
 766     }
 767
 768     *bErrorHasOccurred = anyTrue(bError);
 769 }
 770
 771 /*! \brief Wrapper template function that divides the settles over threads
 772  * and instantiates the core template with instantiated booleans.
 773  */
 774 template<typename T, typename TypeBool, int packSize, typename TypePbc>
 775 static void settleTemplateWrapper(settledata *settled,
 776                                   int nthread, int thread,
 777                                   TypePbc pbc,
 778                                   const real x[], real xprime[],
 779                                   real invdt, real *v,
 780                                   bool bCalcVirial, tensor vir_r_m_dr,
 781                                   bool *bErrorHasOccurred)
 782 {
 783     /* We need to assign settles to threads in groups of pack_size */
 784     int numSettlePacks = (settled->nsettle + packSize - 1)/packSize;
 785     /* Round the end value up to give thread 0 more work */
 786     int settleStart    = ((numSettlePacks* thread      + nthread - 1)/nthread)*packSize;
 787     int settleEnd      = ((numSettlePacks*(thread + 1) + nthread - 1)/nthread)*packSize;
 788
 789     if (v != nullptr)
 790     {
 791         if (!bCalcVirial)
 792         {
 793             settleTemplate<T, TypeBool, packSize,
 794                            TypePbc,
 795                            true,
 796                            false>
 797                 (settled, settleStart, settleEnd,
 798                 pbc,
 799                 x, xprime,
 800                 invdt, v,
 801                 nullptr,
 802                 bErrorHasOccurred);
 803         }
 804         else
 805         {
 806             settleTemplate<T, TypeBool, packSize,
 807                            TypePbc,
 808                            true,
 809                            true>
 810                 (settled, settleStart, settleEnd,
 811                 pbc,
 812                 x, xprime,
 813                 invdt, v,
 814                 vir_r_m_dr,
 815                 bErrorHasOccurred);
 816         }
 817     }
 818     else
 819     {
 820         if (!bCalcVirial)
 821         {
 822             settleTemplate<T, TypeBool, packSize,
 823                            TypePbc,
 824                            false,
 825                            false>
 826                 (settled, settleStart, settleEnd,
 827                 pbc,
 828                 x, xprime,
 829                 invdt, v,
 830                 nullptr,
 831                 bErrorHasOccurred);
 832         }
 833         else
 834         {
 835             settleTemplate<T, TypeBool, packSize,
 836                            TypePbc,
 837                            false,
 838                            true>
 839                 (settled, settleStart, settleEnd,
 840                 pbc,
 841                 x, xprime,
 842                 invdt, v,
 843                 vir_r_m_dr,
 844                 bErrorHasOccurred);
 845         }
 846     }
 847 }
 848
 849 void csettle(settledata *settled,
 850              int nthread, int thread,
 851              const t_pbc *pbc,
 852              const real x[], real xprime[],
 853              real invdt, real *v,
 854              bool bCalcVirial, tensor vir_r_m_dr,
 855              bool *bErrorHasOccurred)
 856 {
 857 #if GMX_SIMD_HAVE_REAL
 858     if (settled->bUseSimd)
 859     {
 860         /* Convert the pbc struct for SIMD */
 861         alignas(GMX_SIMD_ALIGNMENT) real    pbcSimd[9*GMX_SIMD_REAL_WIDTH];
 862         set_pbc_simd(pbc, pbcSimd);
 863
 864         settleTemplateWrapper<SimdReal, SimdBool, GMX_SIMD_REAL_WIDTH,
 865                               const real *>(settled,
 866                                             nthread, thread,
 867                                             pbcSimd,
 868                                             x, xprime,
 869                                             invdt,
 870                                             v,
 871                                             bCalcVirial, vir_r_m_dr,
 872                                             bErrorHasOccurred);
 873     }
 874     else
 875 #endif
 876     {
 877         /* This construct is needed because pbc_dx_aiuc doesn't accept pbc=NULL */
 878         t_pbc        pbcNo;
 879         const t_pbc *pbcNonNull;
 880
 881         if (pbc != nullptr)
 882         {
 883             pbcNonNull = pbc;
 884         }
 885         else
 886         {
 887             set_pbc(&pbcNo, epbcNONE, nullptr);
 888             pbcNonNull = &pbcNo;
 889         }
 890
 891         settleTemplateWrapper<real, bool, 1,
 892                               const t_pbc *>(settled,
 893                                              nthread, thread,
 894                                              pbcNonNull,
 895                                              x, xprime,
 896                                              invdt,
 897                                              v,
 898                                              bCalcVirial, vir_r_m_dr,
 899                                              bErrorHasOccurred);
 900     }
 901 }
 902
 903 }  // namespace gmx