src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRF_VdwLJ_GeomW3P1_sse4_1_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse4_1_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse4_1_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sse4_1_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sse4_1_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwjidx0A,vdwjidx0B;
  86     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  89     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  90     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  97     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  98     __m128d          dummy_mask,cutoff_mask;
  99     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 100     __m128d          one     = _mm_set1_pd(1.0);
 101     __m128d          two     = _mm_set1_pd(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm_set1_pd(fr->ic->epsfac);
 114     charge           = mdatoms->chargeA;
 115     krf              = _mm_set1_pd(fr->ic->k_rf);
 116     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 117     crf              = _mm_set1_pd(fr->ic->c_rf);
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     /* Setup water-specific parameters */
 123     inr              = nlist->iinr[0];
 124     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 125     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 126     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 127     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133
 134     outeriter        = 0;
 135     inneriter        = 0;
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 153                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 154
 155         fix0             = _mm_setzero_pd();
 156         fiy0             = _mm_setzero_pd();
 157         fiz0             = _mm_setzero_pd();
 158         fix1             = _mm_setzero_pd();
 159         fiy1             = _mm_setzero_pd();
 160         fiz1             = _mm_setzero_pd();
 161         fix2             = _mm_setzero_pd();
 162         fiy2             = _mm_setzero_pd();
 163         fiz2             = _mm_setzero_pd();
 164
 165         /* Reset potential sums */
 166         velecsum         = _mm_setzero_pd();
 167         vvdwsum          = _mm_setzero_pd();
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 171         {
 172
 173             /* Get j neighbor index, and coordinate index */
 174             jnrA             = jjnr[jidx];
 175             jnrB             = jjnr[jidx+1];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178
 179             /* load j atom coordinates */
 180             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 181                                               &jx0,&jy0,&jz0);
 182
 183             /* Calculate displacement vector */
 184             dx00             = _mm_sub_pd(ix0,jx0);
 185             dy00             = _mm_sub_pd(iy0,jy0);
 186             dz00             = _mm_sub_pd(iz0,jz0);
 187             dx10             = _mm_sub_pd(ix1,jx0);
 188             dy10             = _mm_sub_pd(iy1,jy0);
 189             dz10             = _mm_sub_pd(iz1,jz0);
 190             dx20             = _mm_sub_pd(ix2,jx0);
 191             dy20             = _mm_sub_pd(iy2,jy0);
 192             dz20             = _mm_sub_pd(iz2,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 196             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 197             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 198
 199             rinv00           = sse41_invsqrt_d(rsq00);
 200             rinv10           = sse41_invsqrt_d(rsq10);
 201             rinv20           = sse41_invsqrt_d(rsq20);
 202
 203             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 204             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 205             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 206
 207             /* Load parameters for j particles */
 208             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 209             vdwjidx0A        = 2*vdwtype[jnrA+0];
 210             vdwjidx0B        = 2*vdwtype[jnrB+0];
 211
 212             fjx0             = _mm_setzero_pd();
 213             fjy0             = _mm_setzero_pd();
 214             fjz0             = _mm_setzero_pd();
 215
 216             /**************************
 217              * CALCULATE INTERACTIONS *
 218              **************************/
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             qq00             = _mm_mul_pd(iq0,jq0);
 222             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 223                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 224
 225             /* REACTION-FIELD ELECTROSTATICS */
 226             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 227             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 228
 229             /* LENNARD-JONES DISPERSION/REPULSION */
 230
 231             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 232             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 233             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 234             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 235             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 236
 237             /* Update potential sum for this i atom from the interaction with this j atom. */
 238             velecsum         = _mm_add_pd(velecsum,velec);
 239             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 240
 241             fscal            = _mm_add_pd(felec,fvdw);
 242
 243             /* Calculate temporary vectorial force */
 244             tx               = _mm_mul_pd(fscal,dx00);
 245             ty               = _mm_mul_pd(fscal,dy00);
 246             tz               = _mm_mul_pd(fscal,dz00);
 247
 248             /* Update vectorial force */
 249             fix0             = _mm_add_pd(fix0,tx);
 250             fiy0             = _mm_add_pd(fiy0,ty);
 251             fiz0             = _mm_add_pd(fiz0,tz);
 252
 253             fjx0             = _mm_add_pd(fjx0,tx);
 254             fjy0             = _mm_add_pd(fjy0,ty);
 255             fjz0             = _mm_add_pd(fjz0,tz);
 256
 257             /**************************
 258              * CALCULATE INTERACTIONS *
 259              **************************/
 260
 261             /* Compute parameters for interactions between i and j atoms */
 262             qq10             = _mm_mul_pd(iq1,jq0);
 263
 264             /* REACTION-FIELD ELECTROSTATICS */
 265             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 266             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 267
 268             /* Update potential sum for this i atom from the interaction with this j atom. */
 269             velecsum         = _mm_add_pd(velecsum,velec);
 270
 271             fscal            = felec;
 272
 273             /* Calculate temporary vectorial force */
 274             tx               = _mm_mul_pd(fscal,dx10);
 275             ty               = _mm_mul_pd(fscal,dy10);
 276             tz               = _mm_mul_pd(fscal,dz10);
 277
 278             /* Update vectorial force */
 279             fix1             = _mm_add_pd(fix1,tx);
 280             fiy1             = _mm_add_pd(fiy1,ty);
 281             fiz1             = _mm_add_pd(fiz1,tz);
 282
 283             fjx0             = _mm_add_pd(fjx0,tx);
 284             fjy0             = _mm_add_pd(fjy0,ty);
 285             fjz0             = _mm_add_pd(fjz0,tz);
 286
 287             /**************************
 288              * CALCULATE INTERACTIONS *
 289              **************************/
 290
 291             /* Compute parameters for interactions between i and j atoms */
 292             qq20             = _mm_mul_pd(iq2,jq0);
 293
 294             /* REACTION-FIELD ELECTROSTATICS */
 295             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 296             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 297
 298             /* Update potential sum for this i atom from the interaction with this j atom. */
 299             velecsum         = _mm_add_pd(velecsum,velec);
 300
 301             fscal            = felec;
 302
 303             /* Calculate temporary vectorial force */
 304             tx               = _mm_mul_pd(fscal,dx20);
 305             ty               = _mm_mul_pd(fscal,dy20);
 306             tz               = _mm_mul_pd(fscal,dz20);
 307
 308             /* Update vectorial force */
 309             fix2             = _mm_add_pd(fix2,tx);
 310             fiy2             = _mm_add_pd(fiy2,ty);
 311             fiz2             = _mm_add_pd(fiz2,tz);
 312
 313             fjx0             = _mm_add_pd(fjx0,tx);
 314             fjy0             = _mm_add_pd(fjy0,ty);
 315             fjz0             = _mm_add_pd(fjz0,tz);
 316
 317             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 318
 319             /* Inner loop uses 111 flops */
 320         }
 321
 322         if(jidx<j_index_end)
 323         {
 324
 325             jnrA             = jjnr[jidx];
 326             j_coord_offsetA  = DIM*jnrA;
 327
 328             /* load j atom coordinates */
 329             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 330                                               &jx0,&jy0,&jz0);
 331
 332             /* Calculate displacement vector */
 333             dx00             = _mm_sub_pd(ix0,jx0);
 334             dy00             = _mm_sub_pd(iy0,jy0);
 335             dz00             = _mm_sub_pd(iz0,jz0);
 336             dx10             = _mm_sub_pd(ix1,jx0);
 337             dy10             = _mm_sub_pd(iy1,jy0);
 338             dz10             = _mm_sub_pd(iz1,jz0);
 339             dx20             = _mm_sub_pd(ix2,jx0);
 340             dy20             = _mm_sub_pd(iy2,jy0);
 341             dz20             = _mm_sub_pd(iz2,jz0);
 342
 343             /* Calculate squared distance and things based on it */
 344             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 345             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 346             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 347
 348             rinv00           = sse41_invsqrt_d(rsq00);
 349             rinv10           = sse41_invsqrt_d(rsq10);
 350             rinv20           = sse41_invsqrt_d(rsq20);
 351
 352             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 353             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 354             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 355
 356             /* Load parameters for j particles */
 357             jq0              = _mm_load_sd(charge+jnrA+0);
 358             vdwjidx0A        = 2*vdwtype[jnrA+0];
 359
 360             fjx0             = _mm_setzero_pd();
 361             fjy0             = _mm_setzero_pd();
 362             fjz0             = _mm_setzero_pd();
 363
 364             /**************************
 365              * CALCULATE INTERACTIONS *
 366              **************************/
 367
 368             /* Compute parameters for interactions between i and j atoms */
 369             qq00             = _mm_mul_pd(iq0,jq0);
 370             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 371
 372             /* REACTION-FIELD ELECTROSTATICS */
 373             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 374             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 375
 376             /* LENNARD-JONES DISPERSION/REPULSION */
 377
 378             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 379             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 380             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 381             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 382             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 383
 384             /* Update potential sum for this i atom from the interaction with this j atom. */
 385             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 386             velecsum         = _mm_add_pd(velecsum,velec);
 387             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 388             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 389
 390             fscal            = _mm_add_pd(felec,fvdw);
 391
 392             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 393
 394             /* Calculate temporary vectorial force */
 395             tx               = _mm_mul_pd(fscal,dx00);
 396             ty               = _mm_mul_pd(fscal,dy00);
 397             tz               = _mm_mul_pd(fscal,dz00);
 398
 399             /* Update vectorial force */
 400             fix0             = _mm_add_pd(fix0,tx);
 401             fiy0             = _mm_add_pd(fiy0,ty);
 402             fiz0             = _mm_add_pd(fiz0,tz);
 403
 404             fjx0             = _mm_add_pd(fjx0,tx);
 405             fjy0             = _mm_add_pd(fjy0,ty);
 406             fjz0             = _mm_add_pd(fjz0,tz);
 407
 408             /**************************
 409              * CALCULATE INTERACTIONS *
 410              **************************/
 411
 412             /* Compute parameters for interactions between i and j atoms */
 413             qq10             = _mm_mul_pd(iq1,jq0);
 414
 415             /* REACTION-FIELD ELECTROSTATICS */
 416             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 417             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 421             velecsum         = _mm_add_pd(velecsum,velec);
 422
 423             fscal            = felec;
 424
 425             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 426
 427             /* Calculate temporary vectorial force */
 428             tx               = _mm_mul_pd(fscal,dx10);
 429             ty               = _mm_mul_pd(fscal,dy10);
 430             tz               = _mm_mul_pd(fscal,dz10);
 431
 432             /* Update vectorial force */
 433             fix1             = _mm_add_pd(fix1,tx);
 434             fiy1             = _mm_add_pd(fiy1,ty);
 435             fiz1             = _mm_add_pd(fiz1,tz);
 436
 437             fjx0             = _mm_add_pd(fjx0,tx);
 438             fjy0             = _mm_add_pd(fjy0,ty);
 439             fjz0             = _mm_add_pd(fjz0,tz);
 440
 441             /**************************
 442              * CALCULATE INTERACTIONS *
 443              **************************/
 444
 445             /* Compute parameters for interactions between i and j atoms */
 446             qq20             = _mm_mul_pd(iq2,jq0);
 447
 448             /* REACTION-FIELD ELECTROSTATICS */
 449             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 450             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 451
 452             /* Update potential sum for this i atom from the interaction with this j atom. */
 453             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 454             velecsum         = _mm_add_pd(velecsum,velec);
 455
 456             fscal            = felec;
 457
 458             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 459
 460             /* Calculate temporary vectorial force */
 461             tx               = _mm_mul_pd(fscal,dx20);
 462             ty               = _mm_mul_pd(fscal,dy20);
 463             tz               = _mm_mul_pd(fscal,dz20);
 464
 465             /* Update vectorial force */
 466             fix2             = _mm_add_pd(fix2,tx);
 467             fiy2             = _mm_add_pd(fiy2,ty);
 468             fiz2             = _mm_add_pd(fiz2,tz);
 469
 470             fjx0             = _mm_add_pd(fjx0,tx);
 471             fjy0             = _mm_add_pd(fjy0,ty);
 472             fjz0             = _mm_add_pd(fjz0,tz);
 473
 474             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 475
 476             /* Inner loop uses 111 flops */
 477         }
 478
 479         /* End of innermost loop */
 480
 481         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 482                                               f+i_coord_offset,fshift+i_shift_offset);
 483
 484         ggid                        = gid[iidx];
 485         /* Update potential energies */
 486         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 487         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 488
 489         /* Increment number of inner iterations */
 490         inneriter                  += j_index_end - j_index_start;
 491
 492         /* Outer loop uses 20 flops */
 493     }
 494
 495     /* Increment number of outer iterations */
 496     outeriter        += nri;
 497
 498     /* Update outer/inner flops */
 499
 500     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*111);
 501 }
 502 /*
 503  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sse4_1_double
 504  * Electrostatics interaction: ReactionField
 505  * VdW interaction:            LennardJones
 506  * Geometry:                   Water3-Particle
 507  * Calculate force/pot:        Force
 508  */
 509 void
 510 nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sse4_1_double
 511                     (t_nblist                    * gmx_restrict       nlist,
 512                      rvec                        * gmx_restrict          xx,
 513                      rvec                        * gmx_restrict          ff,
 514                      struct t_forcerec           * gmx_restrict          fr,
 515                      t_mdatoms                   * gmx_restrict     mdatoms,
 516                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 517                      t_nrnb                      * gmx_restrict        nrnb)
 518 {
 519     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 520      * just 0 for non-waters.
 521      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 522      * jnr indices corresponding to data put in the four positions in the SIMD register.
 523      */
 524     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 525     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 526     int              jnrA,jnrB;
 527     int              j_coord_offsetA,j_coord_offsetB;
 528     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 529     real             rcutoff_scalar;
 530     real             *shiftvec,*fshift,*x,*f;
 531     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 532     int              vdwioffset0;
 533     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 534     int              vdwioffset1;
 535     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 536     int              vdwioffset2;
 537     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 538     int              vdwjidx0A,vdwjidx0B;
 539     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 540     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 541     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 542     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 543     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 544     real             *charge;
 545     int              nvdwtype;
 546     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 547     int              *vdwtype;
 548     real             *vdwparam;
 549     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 550     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 551     __m128d          dummy_mask,cutoff_mask;
 552     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 553     __m128d          one     = _mm_set1_pd(1.0);
 554     __m128d          two     = _mm_set1_pd(2.0);
 555     x                = xx[0];
 556     f                = ff[0];
 557
 558     nri              = nlist->nri;
 559     iinr             = nlist->iinr;
 560     jindex           = nlist->jindex;
 561     jjnr             = nlist->jjnr;
 562     shiftidx         = nlist->shift;
 563     gid              = nlist->gid;
 564     shiftvec         = fr->shift_vec[0];
 565     fshift           = fr->fshift[0];
 566     facel            = _mm_set1_pd(fr->ic->epsfac);
 567     charge           = mdatoms->chargeA;
 568     krf              = _mm_set1_pd(fr->ic->k_rf);
 569     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 570     crf              = _mm_set1_pd(fr->ic->c_rf);
 571     nvdwtype         = fr->ntype;
 572     vdwparam         = fr->nbfp;
 573     vdwtype          = mdatoms->typeA;
 574
 575     /* Setup water-specific parameters */
 576     inr              = nlist->iinr[0];
 577     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 578     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 579     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 580     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 581
 582     /* Avoid stupid compiler warnings */
 583     jnrA = jnrB = 0;
 584     j_coord_offsetA = 0;
 585     j_coord_offsetB = 0;
 586
 587     outeriter        = 0;
 588     inneriter        = 0;
 589
 590     /* Start outer loop over neighborlists */
 591     for(iidx=0; iidx<nri; iidx++)
 592     {
 593         /* Load shift vector for this list */
 594         i_shift_offset   = DIM*shiftidx[iidx];
 595
 596         /* Load limits for loop over neighbors */
 597         j_index_start    = jindex[iidx];
 598         j_index_end      = jindex[iidx+1];
 599
 600         /* Get outer coordinate index */
 601         inr              = iinr[iidx];
 602         i_coord_offset   = DIM*inr;
 603
 604         /* Load i particle coords and add shift vector */
 605         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 606                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 607
 608         fix0             = _mm_setzero_pd();
 609         fiy0             = _mm_setzero_pd();
 610         fiz0             = _mm_setzero_pd();
 611         fix1             = _mm_setzero_pd();
 612         fiy1             = _mm_setzero_pd();
 613         fiz1             = _mm_setzero_pd();
 614         fix2             = _mm_setzero_pd();
 615         fiy2             = _mm_setzero_pd();
 616         fiz2             = _mm_setzero_pd();
 617
 618         /* Start inner kernel loop */
 619         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 620         {
 621
 622             /* Get j neighbor index, and coordinate index */
 623             jnrA             = jjnr[jidx];
 624             jnrB             = jjnr[jidx+1];
 625             j_coord_offsetA  = DIM*jnrA;
 626             j_coord_offsetB  = DIM*jnrB;
 627
 628             /* load j atom coordinates */
 629             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 630                                               &jx0,&jy0,&jz0);
 631
 632             /* Calculate displacement vector */
 633             dx00             = _mm_sub_pd(ix0,jx0);
 634             dy00             = _mm_sub_pd(iy0,jy0);
 635             dz00             = _mm_sub_pd(iz0,jz0);
 636             dx10             = _mm_sub_pd(ix1,jx0);
 637             dy10             = _mm_sub_pd(iy1,jy0);
 638             dz10             = _mm_sub_pd(iz1,jz0);
 639             dx20             = _mm_sub_pd(ix2,jx0);
 640             dy20             = _mm_sub_pd(iy2,jy0);
 641             dz20             = _mm_sub_pd(iz2,jz0);
 642
 643             /* Calculate squared distance and things based on it */
 644             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 645             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 646             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 647
 648             rinv00           = sse41_invsqrt_d(rsq00);
 649             rinv10           = sse41_invsqrt_d(rsq10);
 650             rinv20           = sse41_invsqrt_d(rsq20);
 651
 652             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 653             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 654             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 655
 656             /* Load parameters for j particles */
 657             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 658             vdwjidx0A        = 2*vdwtype[jnrA+0];
 659             vdwjidx0B        = 2*vdwtype[jnrB+0];
 660
 661             fjx0             = _mm_setzero_pd();
 662             fjy0             = _mm_setzero_pd();
 663             fjz0             = _mm_setzero_pd();
 664
 665             /**************************
 666              * CALCULATE INTERACTIONS *
 667              **************************/
 668
 669             /* Compute parameters for interactions between i and j atoms */
 670             qq00             = _mm_mul_pd(iq0,jq0);
 671             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 672                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 673
 674             /* REACTION-FIELD ELECTROSTATICS */
 675             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 676
 677             /* LENNARD-JONES DISPERSION/REPULSION */
 678
 679             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 680             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 681
 682             fscal            = _mm_add_pd(felec,fvdw);
 683
 684             /* Calculate temporary vectorial force */
 685             tx               = _mm_mul_pd(fscal,dx00);
 686             ty               = _mm_mul_pd(fscal,dy00);
 687             tz               = _mm_mul_pd(fscal,dz00);
 688
 689             /* Update vectorial force */
 690             fix0             = _mm_add_pd(fix0,tx);
 691             fiy0             = _mm_add_pd(fiy0,ty);
 692             fiz0             = _mm_add_pd(fiz0,tz);
 693
 694             fjx0             = _mm_add_pd(fjx0,tx);
 695             fjy0             = _mm_add_pd(fjy0,ty);
 696             fjz0             = _mm_add_pd(fjz0,tz);
 697
 698             /**************************
 699              * CALCULATE INTERACTIONS *
 700              **************************/
 701
 702             /* Compute parameters for interactions between i and j atoms */
 703             qq10             = _mm_mul_pd(iq1,jq0);
 704
 705             /* REACTION-FIELD ELECTROSTATICS */
 706             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 707
 708             fscal            = felec;
 709
 710             /* Calculate temporary vectorial force */
 711             tx               = _mm_mul_pd(fscal,dx10);
 712             ty               = _mm_mul_pd(fscal,dy10);
 713             tz               = _mm_mul_pd(fscal,dz10);
 714
 715             /* Update vectorial force */
 716             fix1             = _mm_add_pd(fix1,tx);
 717             fiy1             = _mm_add_pd(fiy1,ty);
 718             fiz1             = _mm_add_pd(fiz1,tz);
 719
 720             fjx0             = _mm_add_pd(fjx0,tx);
 721             fjy0             = _mm_add_pd(fjy0,ty);
 722             fjz0             = _mm_add_pd(fjz0,tz);
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             /* Compute parameters for interactions between i and j atoms */
 729             qq20             = _mm_mul_pd(iq2,jq0);
 730
 731             /* REACTION-FIELD ELECTROSTATICS */
 732             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 733
 734             fscal            = felec;
 735
 736             /* Calculate temporary vectorial force */
 737             tx               = _mm_mul_pd(fscal,dx20);
 738             ty               = _mm_mul_pd(fscal,dy20);
 739             tz               = _mm_mul_pd(fscal,dz20);
 740
 741             /* Update vectorial force */
 742             fix2             = _mm_add_pd(fix2,tx);
 743             fiy2             = _mm_add_pd(fiy2,ty);
 744             fiz2             = _mm_add_pd(fiz2,tz);
 745
 746             fjx0             = _mm_add_pd(fjx0,tx);
 747             fjy0             = _mm_add_pd(fjy0,ty);
 748             fjz0             = _mm_add_pd(fjz0,tz);
 749
 750             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 751
 752             /* Inner loop uses 91 flops */
 753         }
 754
 755         if(jidx<j_index_end)
 756         {
 757
 758             jnrA             = jjnr[jidx];
 759             j_coord_offsetA  = DIM*jnrA;
 760
 761             /* load j atom coordinates */
 762             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 763                                               &jx0,&jy0,&jz0);
 764
 765             /* Calculate displacement vector */
 766             dx00             = _mm_sub_pd(ix0,jx0);
 767             dy00             = _mm_sub_pd(iy0,jy0);
 768             dz00             = _mm_sub_pd(iz0,jz0);
 769             dx10             = _mm_sub_pd(ix1,jx0);
 770             dy10             = _mm_sub_pd(iy1,jy0);
 771             dz10             = _mm_sub_pd(iz1,jz0);
 772             dx20             = _mm_sub_pd(ix2,jx0);
 773             dy20             = _mm_sub_pd(iy2,jy0);
 774             dz20             = _mm_sub_pd(iz2,jz0);
 775
 776             /* Calculate squared distance and things based on it */
 777             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 778             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 779             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 780
 781             rinv00           = sse41_invsqrt_d(rsq00);
 782             rinv10           = sse41_invsqrt_d(rsq10);
 783             rinv20           = sse41_invsqrt_d(rsq20);
 784
 785             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 786             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 787             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 788
 789             /* Load parameters for j particles */
 790             jq0              = _mm_load_sd(charge+jnrA+0);
 791             vdwjidx0A        = 2*vdwtype[jnrA+0];
 792
 793             fjx0             = _mm_setzero_pd();
 794             fjy0             = _mm_setzero_pd();
 795             fjz0             = _mm_setzero_pd();
 796
 797             /**************************
 798              * CALCULATE INTERACTIONS *
 799              **************************/
 800
 801             /* Compute parameters for interactions between i and j atoms */
 802             qq00             = _mm_mul_pd(iq0,jq0);
 803             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 804
 805             /* REACTION-FIELD ELECTROSTATICS */
 806             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 807
 808             /* LENNARD-JONES DISPERSION/REPULSION */
 809
 810             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 811             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 812
 813             fscal            = _mm_add_pd(felec,fvdw);
 814
 815             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 816
 817             /* Calculate temporary vectorial force */
 818             tx               = _mm_mul_pd(fscal,dx00);
 819             ty               = _mm_mul_pd(fscal,dy00);
 820             tz               = _mm_mul_pd(fscal,dz00);
 821
 822             /* Update vectorial force */
 823             fix0             = _mm_add_pd(fix0,tx);
 824             fiy0             = _mm_add_pd(fiy0,ty);
 825             fiz0             = _mm_add_pd(fiz0,tz);
 826
 827             fjx0             = _mm_add_pd(fjx0,tx);
 828             fjy0             = _mm_add_pd(fjy0,ty);
 829             fjz0             = _mm_add_pd(fjz0,tz);
 830
 831             /**************************
 832              * CALCULATE INTERACTIONS *
 833              **************************/
 834
 835             /* Compute parameters for interactions between i and j atoms */
 836             qq10             = _mm_mul_pd(iq1,jq0);
 837
 838             /* REACTION-FIELD ELECTROSTATICS */
 839             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 840
 841             fscal            = felec;
 842
 843             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 844
 845             /* Calculate temporary vectorial force */
 846             tx               = _mm_mul_pd(fscal,dx10);
 847             ty               = _mm_mul_pd(fscal,dy10);
 848             tz               = _mm_mul_pd(fscal,dz10);
 849
 850             /* Update vectorial force */
 851             fix1             = _mm_add_pd(fix1,tx);
 852             fiy1             = _mm_add_pd(fiy1,ty);
 853             fiz1             = _mm_add_pd(fiz1,tz);
 854
 855             fjx0             = _mm_add_pd(fjx0,tx);
 856             fjy0             = _mm_add_pd(fjy0,ty);
 857             fjz0             = _mm_add_pd(fjz0,tz);
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             /* Compute parameters for interactions between i and j atoms */
 864             qq20             = _mm_mul_pd(iq2,jq0);
 865
 866             /* REACTION-FIELD ELECTROSTATICS */
 867             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 868
 869             fscal            = felec;
 870
 871             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 872
 873             /* Calculate temporary vectorial force */
 874             tx               = _mm_mul_pd(fscal,dx20);
 875             ty               = _mm_mul_pd(fscal,dy20);
 876             tz               = _mm_mul_pd(fscal,dz20);
 877
 878             /* Update vectorial force */
 879             fix2             = _mm_add_pd(fix2,tx);
 880             fiy2             = _mm_add_pd(fiy2,ty);
 881             fiz2             = _mm_add_pd(fiz2,tz);
 882
 883             fjx0             = _mm_add_pd(fjx0,tx);
 884             fjy0             = _mm_add_pd(fjy0,ty);
 885             fjz0             = _mm_add_pd(fjz0,tz);
 886
 887             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 888
 889             /* Inner loop uses 91 flops */
 890         }
 891
 892         /* End of innermost loop */
 893
 894         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 895                                               f+i_coord_offset,fshift+i_shift_offset);
 896
 897         /* Increment number of inner iterations */
 898         inneriter                  += j_index_end - j_index_start;
 899
 900         /* Outer loop uses 18 flops */
 901     }
 902
 903     /* Increment number of outer iterations */
 904     outeriter        += nri;
 905
 906     /* Update outer/inner flops */
 907
 908     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*91);
 909 }