src/gromacs/mdlib/csettle.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, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include <assert.h>
  40 #include <math.h>
  41 #include <stdio.h>
  42
  43 #include <algorithm>
  44
  45 #include "gromacs/math/functions.h"
  46 #include "gromacs/math/invertmatrix.h"
  47 #include "gromacs/math/vec.h"
  48 #include "gromacs/mdlib/constr.h"
  49 #include "gromacs/mdtypes/mdatom.h"
  50 #include "gromacs/pbcutil/ishift.h"
  51 #include "gromacs/pbcutil/pbc.h"
  52 #include "gromacs/pbcutil/pbc-simd.h"
  53 #include "gromacs/simd/simd.h"
  54 #include "gromacs/simd/simd_math.h"
  55 #include "gromacs/topology/mtop_util.h"
  56 #include "gromacs/utility/fatalerror.h"
  57 #include "gromacs/utility/gmxassert.h"
  58 #include "gromacs/utility/smalloc.h"
  59
  60 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
  61
  62 typedef struct
  63 {
  64     real   mO;
  65     real   mH;
  66     real   wh;
  67     real   dOH;
  68     real   dHH;
  69     real   ra;
  70     real   rb;
  71     real   rc;
  72     real   irc2;
  73     /* For projection */
  74     real   imO;
  75     real   imH;
  76     real   invdOH;
  77     real   invdHH;
  78     matrix invmat;
  79 } settleparam_t;
  80
  81 typedef struct gmx_settledata
  82 {
  83     settleparam_t massw;    /* Parameters for SETTLE for coordinates */
  84     settleparam_t mass1;    /* Parameters with all masses 1, for forces */
  85
  86     int           nsettle;  /* The number of settles on our rank */
  87     int          *ow1;      /* Index to OW1 atoms, size nsettle + SIMD padding */
  88     int          *hw2;      /* Index to HW2 atoms, size nsettle + SIMD padding */
  89     int          *hw3;      /* Index to HW3 atoms, size nsettle + SIMD padding */
  90     real         *virfac;   /* Virial factor 0 or 1, size nsettle + SIMD pad. */
  91     int           nalloc;   /* Allocation size of ow1, hw2, hw3, virfac */
  92
  93     bool          bUseSimd; /* Use SIMD intrinsics code, if possible */
  94 } t_gmx_settledata;
  95
  96
  97 static void init_proj_matrix(real invmO, real invmH, real dOH, real dHH,
  98                              matrix inverseCouplingMatrix)
  99 {
 100     /* We normalize the inverse masses with invmO for the matrix inversion.
 101      * so we can keep using masses of almost zero for frozen particles,
 102      * without running out of the float range in invertMatrix.
 103      */
 104     double invmORelative = 1.0;
 105     double invmHRelative = invmH/static_cast<double>(invmO);
 106     double distanceRatio = dHH/static_cast<double>(dOH);
 107
 108     /* Construct the constraint coupling matrix */
 109     matrix mat;
 110     mat[0][0] = invmORelative + invmHRelative;
 111     mat[0][1] = invmORelative*(1.0 - 0.5*gmx::square(distanceRatio));
 112     mat[0][2] = invmHRelative*0.5*distanceRatio;
 113     mat[1][1] = mat[0][0];
 114     mat[1][2] = mat[0][2];
 115     mat[2][2] = invmHRelative + invmHRelative;
 116     mat[1][0] = mat[0][1];
 117     mat[2][0] = mat[0][2];
 118     mat[2][1] = mat[1][2];
 119
 120     invertMatrix(mat, inverseCouplingMatrix);
 121
 122     msmul(inverseCouplingMatrix, 1/invmO, inverseCouplingMatrix);
 123 }
 124
 125 static void settleparam_init(settleparam_t *p,
 126                              real mO, real mH, real invmO, real invmH,
 127                              real dOH, real dHH)
 128 {
 129     /* We calculate parameters in double precision to minimize errors.
 130      * The velocity correction applied during SETTLE coordinate constraining
 131      * introduces a systematic error of approximately 1 bit per atom,
 132      * depending on what the compiler does with the code.
 133      */
 134     double wohh;
 135
 136     p->mO     = mO;
 137     p->mH     = mH;
 138     wohh      = mO + 2.0*mH;
 139     p->wh     = mH/wohh;
 140     p->dOH    = dOH;
 141     p->dHH    = dHH;
 142     double rc = dHH/2.0;
 143     double ra = 2.0*mH*std::sqrt(dOH*dOH - rc*rc)/wohh;
 144     p->rb     = std::sqrt(dOH*dOH - rc*rc) - ra;
 145     p->rc     = rc;
 146     p->ra     = ra;
 147     p->irc2   = 1.0/dHH;
 148
 149     /* For projection: inverse masses and coupling matrix inversion */
 150     p->imO    = invmO;
 151     p->imH    = invmH;
 152
 153     p->invdOH = 1.0/dOH;
 154     p->invdHH = 1.0/dHH;
 155
 156     init_proj_matrix(invmO, invmH, dOH, dHH, p->invmat);
 157
 158     if (debug)
 159     {
 160         fprintf(debug, "wh =%g, rc = %g, ra = %g\n",
 161                 p->wh, p->rc, p->ra);
 162         fprintf(debug, "rb = %g, irc2 = %g, dHH = %g, dOH = %g\n",
 163                 p->rb, p->irc2, p->dHH, p->dOH);
 164     }
 165 }
 166
 167 gmx_settledata_t settle_init(const gmx_mtop_t *mtop)
 168 {
 169     /* Check that we have only one settle type */
 170     int                   settle_type = -1;
 171     gmx_mtop_ilistloop_t  iloop       = gmx_mtop_ilistloop_init(mtop);
 172     t_ilist              *ilist;
 173     int                   nmol;
 174     const int             nral1       = 1 + NRAL(F_SETTLE);
 175     while (gmx_mtop_ilistloop_next(iloop, &ilist, &nmol))
 176     {
 177         for (int i = 0; i < ilist[F_SETTLE].nr; i += nral1)
 178         {
 179             if (settle_type == -1)
 180             {
 181                 settle_type = ilist[F_SETTLE].iatoms[i];
 182             }
 183             else if (ilist[F_SETTLE].iatoms[i] != settle_type)
 184             {
 185                 gmx_fatal(FARGS,
 186                           "The [molecules] section of your topology specifies more than one block of\n"
 187                           "a [moleculetype] with a [settles] block. Only one such is allowed.\n"
 188                           "If you are trying to partition your solvent into different *groups*\n"
 189                           "(e.g. for freezing, T-coupling, etc.), you are using the wrong approach. Index\n"
 190                           "files specify groups. Otherwise, you may wish to change the least-used\n"
 191                           "block of molecules with SETTLE constraints into 3 normal constraints.");
 192             }
 193         }
 194     }
 195     GMX_RELEASE_ASSERT(settle_type >= 0, "settle_init called without settles");
 196
 197     gmx_settledata_t settled;
 198
 199     snew(settled, 1);
 200
 201     /* We will not initialize the normal SETTLE parameters here yet,
 202      * since the atom (inv)masses can depend on the integrator and
 203      * free-energy perturbation. We set mO=-1 to trigger later initialization.
 204      */
 205     settled->massw.mO = -1;
 206
 207     real dOH = mtop->ffparams.iparams[settle_type].settle.doh;
 208     real dHH = mtop->ffparams.iparams[settle_type].settle.dhh;
 209     settleparam_init(&settled->mass1, 1.0, 1.0, 1.0, 1.0, dOH, dHH);
 210
 211     settled->ow1    = NULL;
 212     settled->hw2    = NULL;
 213     settled->hw3    = NULL;
 214     settled->virfac = NULL;
 215     settled->nalloc = 0;
 216
 217     /* Without SIMD configured, this bool is not used */
 218     settled->bUseSimd = (getenv("GMX_DISABLE_SIMD_KERNELS") == NULL);
 219
 220     return settled;
 221 }
 222
 223 void settle_free(gmx_settledata_t settled)
 224 {
 225     sfree_aligned(settled->ow1);
 226     sfree_aligned(settled->hw2);
 227     sfree_aligned(settled->hw3);
 228     sfree_aligned(settled->virfac);
 229     sfree(settled);
 230 }
 231
 232 void settle_set_constraints(gmx_settledata_t  settled,
 233                             const t_ilist    *il_settle,
 234                             const t_mdatoms  *mdatoms)
 235 {
 236 #if GMX_SIMD_HAVE_REAL
 237     const int pack_size = GMX_SIMD_REAL_WIDTH;
 238 #else
 239     const int pack_size = 1;
 240 #endif
 241
 242     const int nral1     = 1 + NRAL(F_SETTLE);
 243     int       nsettle   = il_settle->nr/nral1;
 244     settled->nsettle = nsettle;
 245
 246     if (nsettle > 0)
 247     {
 248         const t_iatom *iatoms = il_settle->iatoms;
 249
 250         /* Here we initialize the normal SETTLE parameters */
 251         if (settled->massw.mO < 0)
 252         {
 253             int firstO = iatoms[1];
 254             int firstH = iatoms[2];
 255             settleparam_init(&settled->massw,
 256                              mdatoms->massT[firstO],
 257                              mdatoms->massT[firstH],
 258                              mdatoms->invmass[firstO],
 259                              mdatoms->invmass[firstH],
 260                              settled->mass1.dOH,
 261                              settled->mass1.dHH);
 262         }
 263
 264         if (nsettle + pack_size > settled->nalloc)
 265         {
 266             settled->nalloc = over_alloc_dd(nsettle + pack_size);
 267             sfree_aligned(settled->ow1);
 268             sfree_aligned(settled->hw2);
 269             sfree_aligned(settled->hw3);
 270             sfree_aligned(settled->virfac);
 271             snew_aligned(settled->ow1, settled->nalloc, 64);
 272             snew_aligned(settled->hw2, settled->nalloc, 64);
 273             snew_aligned(settled->hw3, settled->nalloc, 64);
 274             snew_aligned(settled->virfac, settled->nalloc, 64);
 275         }
 276
 277         for (int i = 0; i < nsettle; i++)
 278         {
 279             settled->ow1[i]    = iatoms[i*nral1 + 1];
 280             settled->hw2[i]    = iatoms[i*nral1 + 2];
 281             settled->hw3[i]    = iatoms[i*nral1 + 3];
 282             /* We should avoid double counting of virial contributions for
 283              * SETTLEs that appear in multiple DD domains, so we only count
 284              * the contribution on the home range of the oxygen atom.
 285              */
 286             settled->virfac[i] = (iatoms[i*nral1 + 1] < mdatoms->homenr ? 1 : 0);
 287         }
 288
 289         /* Pack the index array to the full SIMD width with copies from
 290          * the last normal entry, but with no virial contribution.
 291          */
 292         int end_packed = ((nsettle + pack_size - 1)/pack_size)*pack_size;
 293         for (int i = nsettle; i < end_packed; i++)
 294         {
 295             settled->ow1[i]    = settled->ow1[nsettle - 1];
 296             settled->hw2[i]    = settled->hw2[nsettle - 1];
 297             settled->hw3[i]    = settled->hw3[nsettle - 1];
 298             settled->virfac[i] = 0;
 299         }
 300     }
 301 }
 302
 303 void settle_proj(gmx_settledata_t settled, int econq,
 304                  int nsettle, t_iatom iatoms[],
 305                  const t_pbc *pbc,
 306                  rvec x[],
 307                  rvec *der, rvec *derp,
 308                  int calcvir_atom_end, tensor vir_r_m_dder)
 309 {
 310     /* Settle for projection out constraint components
 311      * of derivatives of the coordinates.
 312      * Berk Hess 2008-1-10
 313      */
 314
 315     settleparam_t *p;
 316     real           imO, imH, dOH, dHH, invdOH, invdHH;
 317     matrix         invmat;
 318     int            i, m, m2, ow1, hw2, hw3;
 319     rvec           roh2, roh3, rhh, dc, fc;
 320
 321     calcvir_atom_end *= DIM;
 322
 323     if (econq == econqForce)
 324     {
 325         p = &settled->mass1;
 326     }
 327     else
 328     {
 329         p = &settled->massw;
 330     }
 331     imO    = p->imO;
 332     imH    = p->imH;
 333     copy_mat(p->invmat, invmat);
 334     dOH    = p->dOH;
 335     dHH    = p->dHH;
 336     invdOH = p->invdOH;
 337     invdHH = p->invdHH;
 338
 339 #ifdef PRAGMAS
 340 #pragma ivdep
 341 #endif
 342
 343     const int nral1 = 1 + NRAL(F_SETTLE);
 344
 345     for (i = 0; i < nsettle; i++)
 346     {
 347         ow1 = iatoms[i*nral1 + 1];
 348         hw2 = iatoms[i*nral1 + 2];
 349         hw3 = iatoms[i*nral1 + 3];
 350
 351         if (pbc == NULL)
 352         {
 353             rvec_sub(x[ow1], x[hw2], roh2);
 354             rvec_sub(x[ow1], x[hw3], roh3);
 355             rvec_sub(x[hw2], x[hw3], rhh);
 356         }
 357         else
 358         {
 359             pbc_dx_aiuc(pbc, x[ow1], x[hw2], roh2);
 360             pbc_dx_aiuc(pbc, x[ow1], x[hw3], roh3);
 361             pbc_dx_aiuc(pbc, x[hw2], x[hw3], rhh);
 362         }
 363         svmul(invdOH, roh2, roh2);
 364         svmul(invdOH, roh3, roh3);
 365         svmul(invdHH, rhh, rhh);
 366         /* 18 flops */
 367
 368         /* Determine the projections of der on the bonds */
 369         clear_rvec(dc);
 370         for (m = 0; m < DIM; m++)
 371         {
 372             dc[0] += (der[ow1][m] - der[hw2][m])*roh2[m];
 373             dc[1] += (der[ow1][m] - der[hw3][m])*roh3[m];
 374             dc[2] += (der[hw2][m] - der[hw3][m])*rhh [m];
 375         }
 376         /* 27 flops */
 377
 378         /* Determine the correction for the three bonds */
 379         mvmul(invmat, dc, fc);
 380         /* 15 flops */
 381
 382         /* Subtract the corrections from derp */
 383         for (m = 0; m < DIM; m++)
 384         {
 385             derp[ow1][m] -= imO*( fc[0]*roh2[m] + fc[1]*roh3[m]);
 386             derp[hw2][m] -= imH*(-fc[0]*roh2[m] + fc[2]*rhh [m]);
 387             derp[hw3][m] -= imH*(-fc[1]*roh3[m] - fc[2]*rhh [m]);
 388         }
 389
 390         /* 45 flops */
 391
 392         if (ow1 < calcvir_atom_end)
 393         {
 394             /* Determining r \dot m der is easy,
 395              * since fc contains the mass weighted corrections for der.
 396              */
 397
 398             for (m = 0; m < DIM; m++)
 399             {
 400                 for (m2 = 0; m2 < DIM; m2++)
 401                 {
 402                     vir_r_m_dder[m][m2] +=
 403                         dOH*roh2[m]*roh2[m2]*fc[0] +
 404                         dOH*roh3[m]*roh3[m2]*fc[1] +
 405                         dHH*rhh [m]*rhh [m2]*fc[2];
 406                 }
 407             }
 408         }
 409     }
 410 }
 411
 412
 413 /* The actual settle code, templated for real/SimdReal and for optimization */
 414 template<typename T, typename TypeBool, int packSize,
 415          typename TypePbc,
 416          bool bCorrectVelocity,
 417          bool bCalcVirial>
 418 static void settleTemplate(const gmx_settledata_t settled,
 419                            int settleStart, int settleEnd,
 420                            const TypePbc pbc,
 421                            const real *x, real *xprime,
 422                            real invdt, real * gmx_restrict v,
 423                            tensor vir_r_m_dr,
 424                            bool *bErrorHasOccurred)
 425 {
 426     /* ******************************************************************* */
 427     /*                                                                  ** */
 428     /*    Original code by Shuichi Miyamoto, last update Oct. 1, 1992   ** */
 429     /*                                                                  ** */
 430     /*    Algorithm changes by Berk Hess:                               ** */
 431     /*    2004-07-15 Convert COM to double precision to avoid drift     ** */
 432     /*    2006-10-16 Changed velocity update to use differences         ** */
 433     /*    2012-09-24 Use oxygen as reference instead of COM             ** */
 434     /*    2016-02    Complete rewrite of the code for SIMD              ** */
 435     /*                                                                  ** */
 436     /*    Reference for the SETTLE algorithm                            ** */
 437     /*           S. Miyamoto et al., J. Comp. Chem., 13, 952 (1992).    ** */
 438     /*                                                                  ** */
 439     /* ******************************************************************* */
 440
 441     assert(settleStart % packSize == 0);
 442     assert(settleEnd   % packSize == 0);
 443
 444     TypeBool       bError = TypeBool(false);
 445
 446     settleparam_t *p    = &settled->massw;
 447     T              wh   = T(p->wh);
 448     T              rc   = T(p->rc);
 449     T              ra   = T(p->ra);
 450     T              rb   = T(p->rb);
 451     T              irc2 = T(p->irc2);
 452     T              mO   = T(p->mO);
 453     T              mH   = T(p->mH);
 454
 455     T              almost_zero = T(1e-12);
 456
 457     T              sum_r_m_dr[DIM][DIM];
 458
 459     if (bCalcVirial)
 460     {
 461         for (int d2 = 0; d2 < DIM; d2++)
 462         {
 463             for (int d = 0; d < DIM; d++)
 464             {
 465                 sum_r_m_dr[d2][d] = T(0);
 466             }
 467         }
 468     }
 469
 470     for (int i = settleStart; i < settleEnd; i += packSize)
 471     {
 472         /* Here we pad up to packSize with copies from the last valid entry.
 473          * This gives correct results, since we store (not increment) all
 474          * output, so we store the same output multiple times.
 475          */
 476         const int *ow1 = settled->ow1 + i;
 477         const int *hw2 = settled->hw2 + i;
 478         const int *hw3 = settled->hw3 + i;
 479
 480         T          x_ow1[DIM], x_hw2[DIM], x_hw3[DIM];
 481
 482         gatherLoadUTranspose<3>(x, ow1, &x_ow1[XX], &x_ow1[YY], &x_ow1[ZZ]);
 483         gatherLoadUTranspose<3>(x, hw2, &x_hw2[XX], &x_hw2[YY], &x_hw2[ZZ]);
 484         gatherLoadUTranspose<3>(x, hw3, &x_hw3[XX], &x_hw3[YY], &x_hw3[ZZ]);
 485
 486         T xprime_ow1[DIM], xprime_hw2[DIM], xprime_hw3[DIM];
 487
 488         gatherLoadUTranspose<3>(xprime, ow1, &xprime_ow1[XX], &xprime_ow1[YY], &xprime_ow1[ZZ]);
 489         gatherLoadUTranspose<3>(xprime, hw2, &xprime_hw2[XX], &xprime_hw2[YY], &xprime_hw2[ZZ]);
 490         gatherLoadUTranspose<3>(xprime, hw3, &xprime_hw3[XX], &xprime_hw3[YY], &xprime_hw3[ZZ]);
 491
 492         T dist21[DIM], dist31[DIM];
 493         T doh2[DIM], doh3[DIM];
 494         T sh_hw2[DIM], sh_hw3[DIM];
 495
 496         pbc_dx_aiuc(pbc, x_hw2, x_ow1, dist21);
 497
 498         pbc_dx_aiuc(pbc, x_hw3, x_ow1, dist31);
 499
 500         /* Tedious way of doing pbc */
 501         pbc_dx_aiuc(pbc, xprime_hw2, xprime_ow1, doh2);
 502         for (int d = 0; d < DIM; d++)
 503         {
 504             sh_hw2[d]     = xprime_hw2[d] - (xprime_ow1[d] + doh2[d]);
 505             xprime_hw2[d] = xprime_hw2[d] - sh_hw2[d];
 506         }
 507         pbc_dx_aiuc(pbc, xprime_hw3, xprime_ow1, doh3);
 508         for (int d = 0; d < DIM; d++)
 509         {
 510             sh_hw3[d]     = xprime_hw3[d] - (xprime_ow1[d] + doh3[d]);
 511             xprime_hw3[d] = xprime_hw3[d] - sh_hw3[d];
 512         }
 513
 514         /* Not calculating the center of mass using the oxygen position
 515          * and the O-H distances, as done below, will make SETTLE
 516          * the largest source of energy drift for simulations of water,
 517          * as then the oxygen coordinate is multiplied by 0.89 at every step,
 518          * which can then transfer a systematic rounding to the oxygen velocity.
 519          */
 520         T a1[DIM], com[DIM];
 521         for (int d = 0; d < DIM; d++)
 522         {
 523             a1[d]  = -(doh2[d] + doh3[d]) * wh;
 524             com[d] = xprime_ow1[d] - a1[d];
 525         }
 526         T b1[DIM];
 527         for (int d = 0; d < DIM; d++)
 528         {
 529             b1[d]  = xprime_hw2[d] - com[d];
 530         }
 531         T c1[DIM];
 532         for (int d = 0; d < DIM; d++)
 533         {
 534             c1[d]  = xprime_hw3[d] - com[d];
 535         }
 536         /* 15 flops */
 537
 538         T xakszd = dist21[YY] * dist31[ZZ] - dist21[ZZ] * dist31[YY];
 539         T yakszd = dist21[ZZ] * dist31[XX] - dist21[XX] * dist31[ZZ];
 540         T zakszd = dist21[XX] * dist31[YY] - dist21[YY] * dist31[XX];
 541         T xaksxd = a1[YY] * zakszd - a1[ZZ] * yakszd;
 542         T yaksxd = a1[ZZ] * xakszd - a1[XX] * zakszd;
 543         T zaksxd = a1[XX] * yakszd - a1[YY] * xakszd;
 544         T xaksyd = yakszd * zaksxd - zakszd * yaksxd;
 545         T yaksyd = zakszd * xaksxd - xakszd * zaksxd;
 546         T zaksyd = xakszd * yaksxd - yakszd * xaksxd;
 547         /* 27 flops */
 548
 549         T axlng = gmx::invsqrt(xaksxd * xaksxd + yaksxd * yaksxd + zaksxd * zaksxd);
 550         T aylng = gmx::invsqrt(xaksyd * xaksyd + yaksyd * yaksyd + zaksyd * zaksyd);
 551         T azlng = gmx::invsqrt(xakszd * xakszd + yakszd * yakszd + zakszd * zakszd);
 552
 553         T trns1[DIM], trns2[DIM], trns3[DIM];
 554
 555         trns1[XX] = xaksxd * axlng;
 556         trns2[XX] = yaksxd * axlng;
 557         trns3[XX] = zaksxd * axlng;
 558         trns1[YY] = xaksyd * aylng;
 559         trns2[YY] = yaksyd * aylng;
 560         trns3[YY] = zaksyd * aylng;
 561         trns1[ZZ] = xakszd * azlng;
 562         trns2[ZZ] = yakszd * azlng;
 563         trns3[ZZ] = zakszd * azlng;
 564         /* 24 flops */
 565
 566         T b0d[2], c0d[2];
 567
 568         for (int d = 0; d < 2; d++)
 569         {
 570             b0d[d] = trns1[d] * dist21[XX] + trns2[d] * dist21[YY] + trns3[d] * dist21[ZZ];
 571             c0d[d] = trns1[d] * dist31[XX] + trns2[d] * dist31[YY] + trns3[d] * dist31[ZZ];
 572         }
 573
 574         T a1d_z, b1d[DIM], c1d[DIM];
 575
 576         a1d_z = trns1[ZZ] * a1[XX] + trns2[ZZ] * a1[YY] + trns3[ZZ] * a1[ZZ];
 577         for (int d = 0; d < DIM; d++)
 578         {
 579             b1d[d] = trns1[d] * b1[XX] + trns2[d] * b1[YY] + trns3[d] * b1[ZZ];
 580             c1d[d] = trns1[d] * c1[XX] + trns2[d] * c1[YY] + trns3[d] * c1[ZZ];
 581         }
 582         /* 65 flops */
 583
 584         T tmp, tmp2;
 585
 586         T sinphi = a1d_z * gmx::invsqrt(ra*ra);
 587         tmp2     = 1.0 - sinphi * sinphi;
 588
 589         /* If tmp2 gets close to or beyond zero we have severly distorted
 590          * water molecules and we should terminate the simulation.
 591          * Below we take the max with almost_zero to continue the loop.
 592          */
 593         bError   = bError || (tmp2 <= almost_zero);
 594
 595         tmp2     = max(tmp2, almost_zero);
 596         tmp      = gmx::invsqrt(tmp2);
 597         T cosphi = tmp2*tmp;
 598         T sinpsi = (b1d[ZZ] - c1d[ZZ]) * irc2 * tmp;
 599         tmp2     = 1.0 - sinpsi * sinpsi;
 600
 601         T cospsi = tmp2*gmx::invsqrt(tmp2);
 602         /* 46 flops */
 603
 604         T a2d_y  =  ra * cosphi;
 605         T b2d_x  = -rc * cospsi;
 606         T t1     = -rb * cosphi;
 607         T t2     =  rc * sinpsi * sinphi;
 608         T b2d_y  =  t1 - t2;
 609         T c2d_y  =  t1 + t2;
 610         /* 7 flops */
 611
 612         /*     --- Step3  al,be,ga            --- */
 613         T alpha  = b2d_x * (b0d[XX] - c0d[XX]) + b0d[YY] * b2d_y + c0d[YY] * c2d_y;
 614         T beta   = b2d_x * (c0d[YY] - b0d[YY]) + b0d[XX] * b2d_y + c0d[XX] * c2d_y;
 615         T gamma  = b0d[XX] * b1d[YY] - b1d[XX] * b0d[YY] + c0d[XX] * c1d[YY] - c1d[XX] * c0d[YY];
 616         T al2be2 = alpha * alpha + beta * beta;
 617         tmp2     = (al2be2 - gamma * gamma);
 618         T sinthe = (alpha * gamma - beta * tmp2*gmx::invsqrt(tmp2)) * gmx::invsqrt(al2be2*al2be2);
 619         /* 47 flops */
 620
 621         /*  --- Step4  A3' --- */
 622         tmp2     = 1.0 - sinthe * sinthe;
 623         T costhe = tmp2*gmx::invsqrt(tmp2);
 624
 625         T a3d[DIM], b3d[DIM], c3d[DIM];
 626
 627         a3d[XX]  = -a2d_y * sinthe;
 628         a3d[YY]  = a2d_y * costhe;
 629         a3d[ZZ]  = a1d_z;
 630         b3d[XX]  = b2d_x * costhe - b2d_y * sinthe;
 631         b3d[YY]  = b2d_x * sinthe + b2d_y * costhe;
 632         b3d[ZZ]  = b1d[ZZ];
 633         c3d[XX]  = -b2d_x * costhe - c2d_y * sinthe;
 634         c3d[YY]  = -b2d_x * sinthe + c2d_y * costhe;
 635         c3d[ZZ]  = c1d[ZZ];
 636         /* 26 flops */
 637
 638         /*    --- Step5  A3 --- */
 639         T a3[DIM], b3[DIM], c3[DIM];
 640
 641         a3[XX] = trns1[XX]*a3d[XX] + trns1[YY]*a3d[YY] + trns1[ZZ]*a3d[ZZ];
 642         a3[YY] = trns2[XX]*a3d[XX] + trns2[YY]*a3d[YY] + trns2[ZZ]*a3d[ZZ];
 643         a3[ZZ] = trns3[XX]*a3d[XX] + trns3[YY]*a3d[YY] + trns3[ZZ]*a3d[ZZ];
 644         b3[XX] = trns1[XX]*b3d[XX] + trns1[YY]*b3d[YY] + trns1[ZZ]*b3d[ZZ];
 645         b3[YY] = trns2[XX]*b3d[XX] + trns2[YY]*b3d[YY] + trns2[ZZ]*b3d[ZZ];
 646         b3[ZZ] = trns3[XX]*b3d[XX] + trns3[YY]*b3d[YY] + trns3[ZZ]*b3d[ZZ];
 647         c3[XX] = trns1[XX]*c3d[XX] + trns1[YY]*c3d[YY] + trns1[ZZ]*c3d[ZZ];
 648         c3[YY] = trns2[XX]*c3d[XX] + trns2[YY]*c3d[YY] + trns2[ZZ]*c3d[ZZ];
 649         c3[ZZ] = trns3[XX]*c3d[XX] + trns3[YY]*c3d[YY] + trns3[ZZ]*c3d[ZZ];
 650         /* 45 flops */
 651
 652         /* Compute and store the corrected new coordinate */
 653         for (int d = 0; d < DIM; d++)
 654         {
 655             xprime_ow1[d] = com[d] + a3[d];
 656         }
 657         for (int d = 0; d < DIM; d++)
 658         {
 659             xprime_hw2[d] = com[d] + b3[d] + sh_hw2[d];;
 660         }
 661         for (int d = 0; d < DIM; d++)
 662         {
 663             xprime_hw3[d] = com[d] + c3[d] + sh_hw3[d];
 664         }
 665         /* 9 flops + 6 pbc flops */
 666
 667         transposeScatterStoreU<3>(xprime, ow1, xprime_ow1[XX], xprime_ow1[YY], xprime_ow1[ZZ]);
 668         transposeScatterStoreU<3>(xprime, hw2, xprime_hw2[XX], xprime_hw2[YY], xprime_hw2[ZZ]);
 669         transposeScatterStoreU<3>(xprime, hw3, xprime_hw3[XX], xprime_hw3[YY], xprime_hw3[ZZ]);
 670
 671         // cppcheck-suppress duplicateExpression
 672         if (bCorrectVelocity || bCalcVirial)
 673         {
 674             T da[DIM], db[DIM], dc[DIM];
 675             for (int d = 0; d < DIM; d++)
 676             {
 677                 da[d] = a3[d] - a1[d];
 678             }
 679             for (int d = 0; d < DIM; d++)
 680             {
 681                 db[d] = b3[d] - b1[d];
 682             }
 683             for (int d = 0; d < DIM; d++)
 684             {
 685                 dc[d] = c3[d] - c1[d];
 686             }
 687             /* 9 flops */
 688
 689             if (bCorrectVelocity)
 690             {
 691                 T v_ow1[DIM], v_hw2[DIM], v_hw3[DIM];
 692
 693                 gatherLoadUTranspose<3>(v, ow1, &v_ow1[XX], &v_ow1[YY], &v_ow1[ZZ]);
 694                 gatherLoadUTranspose<3>(v, hw2, &v_hw2[XX], &v_hw2[YY], &v_hw2[ZZ]);
 695                 gatherLoadUTranspose<3>(v, hw3, &v_hw3[XX], &v_hw3[YY], &v_hw3[ZZ]);
 696
 697                 /* Add the position correction divided by dt to the velocity */
 698                 for (int d = 0; d < DIM; d++)
 699                 {
 700                     v_ow1[d] = gmx::fma(da[d], invdt, v_ow1[d]);
 701                 }
 702                 for (int d = 0; d < DIM; d++)
 703                 {
 704                     v_hw2[d] = gmx::fma(db[d], invdt, v_hw2[d]);
 705                 }
 706                 for (int d = 0; d < DIM; d++)
 707                 {
 708                     v_hw3[d] = gmx::fma(dc[d], invdt, v_hw3[d]);
 709                 }
 710                 /* 3*6 flops */
 711
 712                 transposeScatterStoreU<3>(v, ow1, v_ow1[XX], v_ow1[YY], v_ow1[ZZ]);
 713                 transposeScatterStoreU<3>(v, hw2, v_hw2[XX], v_hw2[YY], v_hw2[ZZ]);
 714                 transposeScatterStoreU<3>(v, hw3, v_hw3[XX], v_hw3[YY], v_hw3[ZZ]);
 715             }
 716
 717             if (bCalcVirial)
 718             {
 719                 /* Filter out the non-local settles */
 720                 T filter = load(settled->virfac + i);
 721                 T mOf    = filter*mO;
 722                 T mHf    = filter*mH;
 723
 724                 T mdo[DIM], mdb[DIM], mdc[DIM];
 725
 726                 for (int d = 0; d < DIM; d++)
 727                 {
 728                     mdb[d] = mHf*db[d];
 729                     mdc[d] = mHf*dc[d];
 730                     mdo[d] = mOf*da[d] + mdb[d] + mdc[d];
 731                 }
 732
 733                 for (int d2 = 0; d2 < DIM; d2++)
 734                 {
 735                     for (int d = 0; d < DIM; d++)
 736                     {
 737                         sum_r_m_dr[d2][d] = sum_r_m_dr[d2][d] -
 738                             (x_ow1[d2]*mdo[d] +
 739                              dist21[d2]*mdb[d] +
 740                              dist31[d2]*mdc[d]);
 741                     }
 742                 }
 743                 /* 71 flops */
 744             }
 745         }
 746     }
 747
 748     if (bCalcVirial)
 749     {
 750         for (int d2 = 0; d2 < DIM; d2++)
 751         {
 752             for (int d = 0; d < DIM; d++)
 753             {
 754                 vir_r_m_dr[d2][d] += reduce(sum_r_m_dr[d2][d]);
 755             }
 756         }
 757     }
 758
 759     *bErrorHasOccurred = anyTrue(bError);
 760 }
 761
 762 /* Wrapper template function that divides the settles over threads
 763  * and instantiates the core template with instantiated booleans.
 764  */
 765 template<typename T, typename TypeBool, int packSize, typename TypePbc>
 766 static void settleTemplateWrapper(gmx_settledata_t settled,
 767                                   int nthread, int thread,
 768                                   TypePbc pbc,
 769                                   const real x[], real xprime[],
 770                                   real invdt, real *v,
 771                                   bool bCalcVirial, tensor vir_r_m_dr,
 772                                   bool *bErrorHasOccurred)
 773 {
 774     /* We need to assign settles to threads in groups of pack_size */
 775     int numSettlePacks = (settled->nsettle + packSize - 1)/packSize;
 776     /* Round the end value up to give thread 0 more work */
 777     int settleStart    = ((numSettlePacks* thread      + nthread - 1)/nthread)*packSize;
 778     int settleEnd      = ((numSettlePacks*(thread + 1) + nthread - 1)/nthread)*packSize;
 779
 780     if (v != NULL)
 781     {
 782         if (!bCalcVirial)
 783         {
 784             settleTemplate<T, TypeBool, packSize,
 785                            TypePbc,
 786                            true,
 787                            false>
 788                 (settled, settleStart, settleEnd,
 789                 pbc,
 790                 x, xprime,
 791                 invdt, v,
 792                 NULL,
 793                 bErrorHasOccurred);
 794         }
 795         else
 796         {
 797             settleTemplate<T, TypeBool, packSize,
 798                            TypePbc,
 799                            true,
 800                            true>
 801                 (settled, settleStart, settleEnd,
 802                 pbc,
 803                 x, xprime,
 804                 invdt, v,
 805                 vir_r_m_dr,
 806                 bErrorHasOccurred);
 807         }
 808     }
 809     else
 810     {
 811         if (!bCalcVirial)
 812         {
 813             settleTemplate<T, TypeBool, packSize,
 814                            TypePbc,
 815                            false,
 816                            false>
 817                 (settled, settleStart, settleEnd,
 818                 pbc,
 819                 x, xprime,
 820                 invdt, v,
 821                 NULL,
 822                 bErrorHasOccurred);
 823         }
 824         else
 825         {
 826             settleTemplate<T, TypeBool, packSize,
 827                            TypePbc,
 828                            false,
 829                            true>
 830                 (settled, settleStart, settleEnd,
 831                 pbc,
 832                 x, xprime,
 833                 invdt, v,
 834                 vir_r_m_dr,
 835                 bErrorHasOccurred);
 836         }
 837     }
 838 }
 839
 840 void csettle(gmx_settledata_t settled,
 841              int nthread, int thread,
 842              const t_pbc *pbc,
 843              const real x[], real xprime[],
 844              real invdt, real *v,
 845              bool bCalcVirial, tensor vir_r_m_dr,
 846              bool *bErrorHasOccurred)
 847 {
 848 #if GMX_SIMD_HAVE_REAL
 849     if (settled->bUseSimd)
 850     {
 851         /* Convert the pbc struct for SIMD */
 852         GMX_ALIGNED(real, GMX_SIMD_REAL_WIDTH) pbcSimd[9*GMX_SIMD_REAL_WIDTH];
 853         set_pbc_simd(pbc, pbcSimd);
 854
 855         settleTemplateWrapper<SimdReal, SimdBool, GMX_SIMD_REAL_WIDTH,
 856                               const real *>(settled,
 857                                             nthread, thread,
 858                                             pbcSimd,
 859                                             x, xprime,
 860                                             invdt,
 861                                             v,
 862                                             bCalcVirial, vir_r_m_dr,
 863                                             bErrorHasOccurred);
 864     }
 865     else
 866 #endif
 867     {
 868         /* This construct is needed because pbc_dx_aiuc doesn't accept pbc=NULL */
 869         t_pbc        pbcNo;
 870         const t_pbc *pbcNonNull;
 871
 872         if (pbc != NULL)
 873         {
 874             pbcNonNull = pbc;
 875         }
 876         else
 877         {
 878             set_pbc(&pbcNo, epbcNONE, NULL);
 879             pbcNonNull = &pbcNo;
 880         }
 881
 882         settleTemplateWrapper<real, bool, 1,
 883                               const t_pbc *>(settled,
 884                                              nthread, thread,
 885                                              pbcNonNull,
 886                                              x, xprime,
 887                                              invdt,
 888                                              v,
 889                                              bCalcVirial, vir_r_m_dr,
 890                                              bErrorHasOccurred);
 891     }
 892 }