src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRF_VdwLJ_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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 avx_128_fma_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_avx_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_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     int              vdwjidx1A,vdwjidx1B;
  88     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     int              nvdwtype;
 103     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 104     int              *vdwtype;
 105     real             *vdwparam;
 106     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 107     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 108     __m128d          dummy_mask,cutoff_mask;
 109     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 110     __m128d          one     = _mm_set1_pd(1.0);
 111     __m128d          two     = _mm_set1_pd(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm_set1_pd(fr->ic->epsfac);
 124     charge           = mdatoms->chargeA;
 125     krf              = _mm_set1_pd(fr->ic->k_rf);
 126     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 127     crf              = _mm_set1_pd(fr->ic->c_rf);
 128     nvdwtype         = fr->ntype;
 129     vdwparam         = fr->nbfp;
 130     vdwtype          = mdatoms->typeA;
 131
 132     /* Setup water-specific parameters */
 133     inr              = nlist->iinr[0];
 134     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 135     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 136     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 137     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 138
 139     jq0              = _mm_set1_pd(charge[inr+0]);
 140     jq1              = _mm_set1_pd(charge[inr+1]);
 141     jq2              = _mm_set1_pd(charge[inr+2]);
 142     vdwjidx0A        = 2*vdwtype[inr+0];
 143     qq00             = _mm_mul_pd(iq0,jq0);
 144     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 145     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 146     qq01             = _mm_mul_pd(iq0,jq1);
 147     qq02             = _mm_mul_pd(iq0,jq2);
 148     qq10             = _mm_mul_pd(iq1,jq0);
 149     qq11             = _mm_mul_pd(iq1,jq1);
 150     qq12             = _mm_mul_pd(iq1,jq2);
 151     qq20             = _mm_mul_pd(iq2,jq0);
 152     qq21             = _mm_mul_pd(iq2,jq1);
 153     qq22             = _mm_mul_pd(iq2,jq2);
 154
 155     /* Avoid stupid compiler warnings */
 156     jnrA = jnrB = 0;
 157     j_coord_offsetA = 0;
 158     j_coord_offsetB = 0;
 159
 160     outeriter        = 0;
 161     inneriter        = 0;
 162
 163     /* Start outer loop over neighborlists */
 164     for(iidx=0; iidx<nri; iidx++)
 165     {
 166         /* Load shift vector for this list */
 167         i_shift_offset   = DIM*shiftidx[iidx];
 168
 169         /* Load limits for loop over neighbors */
 170         j_index_start    = jindex[iidx];
 171         j_index_end      = jindex[iidx+1];
 172
 173         /* Get outer coordinate index */
 174         inr              = iinr[iidx];
 175         i_coord_offset   = DIM*inr;
 176
 177         /* Load i particle coords and add shift vector */
 178         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 179                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 180
 181         fix0             = _mm_setzero_pd();
 182         fiy0             = _mm_setzero_pd();
 183         fiz0             = _mm_setzero_pd();
 184         fix1             = _mm_setzero_pd();
 185         fiy1             = _mm_setzero_pd();
 186         fiz1             = _mm_setzero_pd();
 187         fix2             = _mm_setzero_pd();
 188         fiy2             = _mm_setzero_pd();
 189         fiz2             = _mm_setzero_pd();
 190
 191         /* Reset potential sums */
 192         velecsum         = _mm_setzero_pd();
 193         vvdwsum          = _mm_setzero_pd();
 194
 195         /* Start inner kernel loop */
 196         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 197         {
 198
 199             /* Get j neighbor index, and coordinate index */
 200             jnrA             = jjnr[jidx];
 201             jnrB             = jjnr[jidx+1];
 202             j_coord_offsetA  = DIM*jnrA;
 203             j_coord_offsetB  = DIM*jnrB;
 204
 205             /* load j atom coordinates */
 206             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 207                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 208
 209             /* Calculate displacement vector */
 210             dx00             = _mm_sub_pd(ix0,jx0);
 211             dy00             = _mm_sub_pd(iy0,jy0);
 212             dz00             = _mm_sub_pd(iz0,jz0);
 213             dx01             = _mm_sub_pd(ix0,jx1);
 214             dy01             = _mm_sub_pd(iy0,jy1);
 215             dz01             = _mm_sub_pd(iz0,jz1);
 216             dx02             = _mm_sub_pd(ix0,jx2);
 217             dy02             = _mm_sub_pd(iy0,jy2);
 218             dz02             = _mm_sub_pd(iz0,jz2);
 219             dx10             = _mm_sub_pd(ix1,jx0);
 220             dy10             = _mm_sub_pd(iy1,jy0);
 221             dz10             = _mm_sub_pd(iz1,jz0);
 222             dx11             = _mm_sub_pd(ix1,jx1);
 223             dy11             = _mm_sub_pd(iy1,jy1);
 224             dz11             = _mm_sub_pd(iz1,jz1);
 225             dx12             = _mm_sub_pd(ix1,jx2);
 226             dy12             = _mm_sub_pd(iy1,jy2);
 227             dz12             = _mm_sub_pd(iz1,jz2);
 228             dx20             = _mm_sub_pd(ix2,jx0);
 229             dy20             = _mm_sub_pd(iy2,jy0);
 230             dz20             = _mm_sub_pd(iz2,jz0);
 231             dx21             = _mm_sub_pd(ix2,jx1);
 232             dy21             = _mm_sub_pd(iy2,jy1);
 233             dz21             = _mm_sub_pd(iz2,jz1);
 234             dx22             = _mm_sub_pd(ix2,jx2);
 235             dy22             = _mm_sub_pd(iy2,jy2);
 236             dz22             = _mm_sub_pd(iz2,jz2);
 237
 238             /* Calculate squared distance and things based on it */
 239             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 240             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 241             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 242             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 243             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 244             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 245             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 246             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 247             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 248
 249             rinv00           = avx128fma_invsqrt_d(rsq00);
 250             rinv01           = avx128fma_invsqrt_d(rsq01);
 251             rinv02           = avx128fma_invsqrt_d(rsq02);
 252             rinv10           = avx128fma_invsqrt_d(rsq10);
 253             rinv11           = avx128fma_invsqrt_d(rsq11);
 254             rinv12           = avx128fma_invsqrt_d(rsq12);
 255             rinv20           = avx128fma_invsqrt_d(rsq20);
 256             rinv21           = avx128fma_invsqrt_d(rsq21);
 257             rinv22           = avx128fma_invsqrt_d(rsq22);
 258
 259             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 260             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 261             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 262             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 263             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 264             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 265             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 266             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 267             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 268
 269             fjx0             = _mm_setzero_pd();
 270             fjy0             = _mm_setzero_pd();
 271             fjz0             = _mm_setzero_pd();
 272             fjx1             = _mm_setzero_pd();
 273             fjy1             = _mm_setzero_pd();
 274             fjz1             = _mm_setzero_pd();
 275             fjx2             = _mm_setzero_pd();
 276             fjy2             = _mm_setzero_pd();
 277             fjz2             = _mm_setzero_pd();
 278
 279             /**************************
 280              * CALCULATE INTERACTIONS *
 281              **************************/
 282
 283             /* REACTION-FIELD ELECTROSTATICS */
 284             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 285             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 286
 287             /* LENNARD-JONES DISPERSION/REPULSION */
 288
 289             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 290             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 291             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 292             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 293             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 294
 295             /* Update potential sum for this i atom from the interaction with this j atom. */
 296             velecsum         = _mm_add_pd(velecsum,velec);
 297             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 298
 299             fscal            = _mm_add_pd(felec,fvdw);
 300
 301             /* Update vectorial force */
 302             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 303             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 304             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 305
 306             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 307             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 308             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 309
 310             /**************************
 311              * CALCULATE INTERACTIONS *
 312              **************************/
 313
 314             /* REACTION-FIELD ELECTROSTATICS */
 315             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 316             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velecsum         = _mm_add_pd(velecsum,velec);
 320
 321             fscal            = felec;
 322
 323             /* Update vectorial force */
 324             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 325             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 326             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 327
 328             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 329             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 330             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 331
 332             /**************************
 333              * CALCULATE INTERACTIONS *
 334              **************************/
 335
 336             /* REACTION-FIELD ELECTROSTATICS */
 337             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 338             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 339
 340             /* Update potential sum for this i atom from the interaction with this j atom. */
 341             velecsum         = _mm_add_pd(velecsum,velec);
 342
 343             fscal            = felec;
 344
 345             /* Update vectorial force */
 346             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 347             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 348             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 349
 350             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 351             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 352             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             /* REACTION-FIELD ELECTROSTATICS */
 359             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 360             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velecsum         = _mm_add_pd(velecsum,velec);
 364
 365             fscal            = felec;
 366
 367             /* Update vectorial force */
 368             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 369             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 370             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 371
 372             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 373             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 374             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 375
 376             /**************************
 377              * CALCULATE INTERACTIONS *
 378              **************************/
 379
 380             /* REACTION-FIELD ELECTROSTATICS */
 381             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 382             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 383
 384             /* Update potential sum for this i atom from the interaction with this j atom. */
 385             velecsum         = _mm_add_pd(velecsum,velec);
 386
 387             fscal            = felec;
 388
 389             /* Update vectorial force */
 390             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 391             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 392             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 393
 394             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 395             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 396             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             /* REACTION-FIELD ELECTROSTATICS */
 403             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 404             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velecsum         = _mm_add_pd(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411             /* Update vectorial force */
 412             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 413             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 414             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 415
 416             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 417             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 418             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             /* REACTION-FIELD ELECTROSTATICS */
 425             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 426             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 427
 428             /* Update potential sum for this i atom from the interaction with this j atom. */
 429             velecsum         = _mm_add_pd(velecsum,velec);
 430
 431             fscal            = felec;
 432
 433             /* Update vectorial force */
 434             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 435             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 436             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 437
 438             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 439             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 440             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 441
 442             /**************************
 443              * CALCULATE INTERACTIONS *
 444              **************************/
 445
 446             /* REACTION-FIELD ELECTROSTATICS */
 447             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 448             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 449
 450             /* Update potential sum for this i atom from the interaction with this j atom. */
 451             velecsum         = _mm_add_pd(velecsum,velec);
 452
 453             fscal            = felec;
 454
 455             /* Update vectorial force */
 456             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 457             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 458             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 459
 460             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 461             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 462             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 470             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velecsum         = _mm_add_pd(velecsum,velec);
 474
 475             fscal            = felec;
 476
 477             /* Update vectorial force */
 478             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 479             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 480             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 481
 482             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 483             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 484             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 485
 486             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 487
 488             /* Inner loop uses 327 flops */
 489         }
 490
 491         if(jidx<j_index_end)
 492         {
 493
 494             jnrA             = jjnr[jidx];
 495             j_coord_offsetA  = DIM*jnrA;
 496
 497             /* load j atom coordinates */
 498             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 499                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 500
 501             /* Calculate displacement vector */
 502             dx00             = _mm_sub_pd(ix0,jx0);
 503             dy00             = _mm_sub_pd(iy0,jy0);
 504             dz00             = _mm_sub_pd(iz0,jz0);
 505             dx01             = _mm_sub_pd(ix0,jx1);
 506             dy01             = _mm_sub_pd(iy0,jy1);
 507             dz01             = _mm_sub_pd(iz0,jz1);
 508             dx02             = _mm_sub_pd(ix0,jx2);
 509             dy02             = _mm_sub_pd(iy0,jy2);
 510             dz02             = _mm_sub_pd(iz0,jz2);
 511             dx10             = _mm_sub_pd(ix1,jx0);
 512             dy10             = _mm_sub_pd(iy1,jy0);
 513             dz10             = _mm_sub_pd(iz1,jz0);
 514             dx11             = _mm_sub_pd(ix1,jx1);
 515             dy11             = _mm_sub_pd(iy1,jy1);
 516             dz11             = _mm_sub_pd(iz1,jz1);
 517             dx12             = _mm_sub_pd(ix1,jx2);
 518             dy12             = _mm_sub_pd(iy1,jy2);
 519             dz12             = _mm_sub_pd(iz1,jz2);
 520             dx20             = _mm_sub_pd(ix2,jx0);
 521             dy20             = _mm_sub_pd(iy2,jy0);
 522             dz20             = _mm_sub_pd(iz2,jz0);
 523             dx21             = _mm_sub_pd(ix2,jx1);
 524             dy21             = _mm_sub_pd(iy2,jy1);
 525             dz21             = _mm_sub_pd(iz2,jz1);
 526             dx22             = _mm_sub_pd(ix2,jx2);
 527             dy22             = _mm_sub_pd(iy2,jy2);
 528             dz22             = _mm_sub_pd(iz2,jz2);
 529
 530             /* Calculate squared distance and things based on it */
 531             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 532             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 533             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 534             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 535             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 536             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 537             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 538             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 539             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 540
 541             rinv00           = avx128fma_invsqrt_d(rsq00);
 542             rinv01           = avx128fma_invsqrt_d(rsq01);
 543             rinv02           = avx128fma_invsqrt_d(rsq02);
 544             rinv10           = avx128fma_invsqrt_d(rsq10);
 545             rinv11           = avx128fma_invsqrt_d(rsq11);
 546             rinv12           = avx128fma_invsqrt_d(rsq12);
 547             rinv20           = avx128fma_invsqrt_d(rsq20);
 548             rinv21           = avx128fma_invsqrt_d(rsq21);
 549             rinv22           = avx128fma_invsqrt_d(rsq22);
 550
 551             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 552             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 553             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 554             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 555             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 556             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 557             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 558             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 559             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 560
 561             fjx0             = _mm_setzero_pd();
 562             fjy0             = _mm_setzero_pd();
 563             fjz0             = _mm_setzero_pd();
 564             fjx1             = _mm_setzero_pd();
 565             fjy1             = _mm_setzero_pd();
 566             fjz1             = _mm_setzero_pd();
 567             fjx2             = _mm_setzero_pd();
 568             fjy2             = _mm_setzero_pd();
 569             fjz2             = _mm_setzero_pd();
 570
 571             /**************************
 572              * CALCULATE INTERACTIONS *
 573              **************************/
 574
 575             /* REACTION-FIELD ELECTROSTATICS */
 576             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 577             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 578
 579             /* LENNARD-JONES DISPERSION/REPULSION */
 580
 581             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 582             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 583             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 584             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 585             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 586
 587             /* Update potential sum for this i atom from the interaction with this j atom. */
 588             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 589             velecsum         = _mm_add_pd(velecsum,velec);
 590             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 591             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 592
 593             fscal            = _mm_add_pd(felec,fvdw);
 594
 595             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 596
 597             /* Update vectorial force */
 598             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 599             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 600             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 601
 602             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 603             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 604             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 605
 606             /**************************
 607              * CALCULATE INTERACTIONS *
 608              **************************/
 609
 610             /* REACTION-FIELD ELECTROSTATICS */
 611             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 612             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 613
 614             /* Update potential sum for this i atom from the interaction with this j atom. */
 615             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 616             velecsum         = _mm_add_pd(velecsum,velec);
 617
 618             fscal            = felec;
 619
 620             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 621
 622             /* Update vectorial force */
 623             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 624             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 625             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 626
 627             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 628             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 629             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 630
 631             /**************************
 632              * CALCULATE INTERACTIONS *
 633              **************************/
 634
 635             /* REACTION-FIELD ELECTROSTATICS */
 636             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 637             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 638
 639             /* Update potential sum for this i atom from the interaction with this j atom. */
 640             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 641             velecsum         = _mm_add_pd(velecsum,velec);
 642
 643             fscal            = felec;
 644
 645             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 646
 647             /* Update vectorial force */
 648             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 649             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 650             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 651
 652             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 653             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 654             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 655
 656             /**************************
 657              * CALCULATE INTERACTIONS *
 658              **************************/
 659
 660             /* REACTION-FIELD ELECTROSTATICS */
 661             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 662             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 663
 664             /* Update potential sum for this i atom from the interaction with this j atom. */
 665             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 666             velecsum         = _mm_add_pd(velecsum,velec);
 667
 668             fscal            = felec;
 669
 670             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 671
 672             /* Update vectorial force */
 673             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 674             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 675             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 676
 677             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 678             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 679             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 680
 681             /**************************
 682              * CALCULATE INTERACTIONS *
 683              **************************/
 684
 685             /* REACTION-FIELD ELECTROSTATICS */
 686             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 687             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 688
 689             /* Update potential sum for this i atom from the interaction with this j atom. */
 690             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 691             velecsum         = _mm_add_pd(velecsum,velec);
 692
 693             fscal            = felec;
 694
 695             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 696
 697             /* Update vectorial force */
 698             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 699             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 700             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 701
 702             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 703             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 704             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 705
 706             /**************************
 707              * CALCULATE INTERACTIONS *
 708              **************************/
 709
 710             /* REACTION-FIELD ELECTROSTATICS */
 711             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 712             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 713
 714             /* Update potential sum for this i atom from the interaction with this j atom. */
 715             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 716             velecsum         = _mm_add_pd(velecsum,velec);
 717
 718             fscal            = felec;
 719
 720             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 721
 722             /* Update vectorial force */
 723             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 724             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 725             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 726
 727             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 728             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 729             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 730
 731             /**************************
 732              * CALCULATE INTERACTIONS *
 733              **************************/
 734
 735             /* REACTION-FIELD ELECTROSTATICS */
 736             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 737             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 738
 739             /* Update potential sum for this i atom from the interaction with this j atom. */
 740             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 741             velecsum         = _mm_add_pd(velecsum,velec);
 742
 743             fscal            = felec;
 744
 745             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 746
 747             /* Update vectorial force */
 748             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 749             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 750             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 751
 752             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 753             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 754             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 755
 756             /**************************
 757              * CALCULATE INTERACTIONS *
 758              **************************/
 759
 760             /* REACTION-FIELD ELECTROSTATICS */
 761             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 762             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 763
 764             /* Update potential sum for this i atom from the interaction with this j atom. */
 765             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 766             velecsum         = _mm_add_pd(velecsum,velec);
 767
 768             fscal            = felec;
 769
 770             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 771
 772             /* Update vectorial force */
 773             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 774             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 775             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 776
 777             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 778             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 779             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 780
 781             /**************************
 782              * CALCULATE INTERACTIONS *
 783              **************************/
 784
 785             /* REACTION-FIELD ELECTROSTATICS */
 786             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 787             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 788
 789             /* Update potential sum for this i atom from the interaction with this j atom. */
 790             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 791             velecsum         = _mm_add_pd(velecsum,velec);
 792
 793             fscal            = felec;
 794
 795             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 796
 797             /* Update vectorial force */
 798             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 799             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 800             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 801
 802             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 803             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 804             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 805
 806             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 807
 808             /* Inner loop uses 327 flops */
 809         }
 810
 811         /* End of innermost loop */
 812
 813         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 814                                               f+i_coord_offset,fshift+i_shift_offset);
 815
 816         ggid                        = gid[iidx];
 817         /* Update potential energies */
 818         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 819         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 820
 821         /* Increment number of inner iterations */
 822         inneriter                  += j_index_end - j_index_start;
 823
 824         /* Outer loop uses 20 flops */
 825     }
 826
 827     /* Increment number of outer iterations */
 828     outeriter        += nri;
 829
 830     /* Update outer/inner flops */
 831
 832     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*327);
 833 }
 834 /*
 835  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_double
 836  * Electrostatics interaction: ReactionField
 837  * VdW interaction:            LennardJones
 838  * Geometry:                   Water3-Water3
 839  * Calculate force/pot:        Force
 840  */
 841 void
 842 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_double
 843                     (t_nblist                    * gmx_restrict       nlist,
 844                      rvec                        * gmx_restrict          xx,
 845                      rvec                        * gmx_restrict          ff,
 846                      struct t_forcerec           * gmx_restrict          fr,
 847                      t_mdatoms                   * gmx_restrict     mdatoms,
 848                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 849                      t_nrnb                      * gmx_restrict        nrnb)
 850 {
 851     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 852      * just 0 for non-waters.
 853      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 854      * jnr indices corresponding to data put in the four positions in the SIMD register.
 855      */
 856     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 857     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 858     int              jnrA,jnrB;
 859     int              j_coord_offsetA,j_coord_offsetB;
 860     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 861     real             rcutoff_scalar;
 862     real             *shiftvec,*fshift,*x,*f;
 863     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 864     int              vdwioffset0;
 865     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 866     int              vdwioffset1;
 867     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 868     int              vdwioffset2;
 869     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 870     int              vdwjidx0A,vdwjidx0B;
 871     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 872     int              vdwjidx1A,vdwjidx1B;
 873     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 874     int              vdwjidx2A,vdwjidx2B;
 875     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 876     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 877     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 878     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 879     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 880     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 881     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 882     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 883     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 884     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 885     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 886     real             *charge;
 887     int              nvdwtype;
 888     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 889     int              *vdwtype;
 890     real             *vdwparam;
 891     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 892     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 893     __m128d          dummy_mask,cutoff_mask;
 894     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 895     __m128d          one     = _mm_set1_pd(1.0);
 896     __m128d          two     = _mm_set1_pd(2.0);
 897     x                = xx[0];
 898     f                = ff[0];
 899
 900     nri              = nlist->nri;
 901     iinr             = nlist->iinr;
 902     jindex           = nlist->jindex;
 903     jjnr             = nlist->jjnr;
 904     shiftidx         = nlist->shift;
 905     gid              = nlist->gid;
 906     shiftvec         = fr->shift_vec[0];
 907     fshift           = fr->fshift[0];
 908     facel            = _mm_set1_pd(fr->ic->epsfac);
 909     charge           = mdatoms->chargeA;
 910     krf              = _mm_set1_pd(fr->ic->k_rf);
 911     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 912     crf              = _mm_set1_pd(fr->ic->c_rf);
 913     nvdwtype         = fr->ntype;
 914     vdwparam         = fr->nbfp;
 915     vdwtype          = mdatoms->typeA;
 916
 917     /* Setup water-specific parameters */
 918     inr              = nlist->iinr[0];
 919     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 920     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 921     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 922     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 923
 924     jq0              = _mm_set1_pd(charge[inr+0]);
 925     jq1              = _mm_set1_pd(charge[inr+1]);
 926     jq2              = _mm_set1_pd(charge[inr+2]);
 927     vdwjidx0A        = 2*vdwtype[inr+0];
 928     qq00             = _mm_mul_pd(iq0,jq0);
 929     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 930     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 931     qq01             = _mm_mul_pd(iq0,jq1);
 932     qq02             = _mm_mul_pd(iq0,jq2);
 933     qq10             = _mm_mul_pd(iq1,jq0);
 934     qq11             = _mm_mul_pd(iq1,jq1);
 935     qq12             = _mm_mul_pd(iq1,jq2);
 936     qq20             = _mm_mul_pd(iq2,jq0);
 937     qq21             = _mm_mul_pd(iq2,jq1);
 938     qq22             = _mm_mul_pd(iq2,jq2);
 939
 940     /* Avoid stupid compiler warnings */
 941     jnrA = jnrB = 0;
 942     j_coord_offsetA = 0;
 943     j_coord_offsetB = 0;
 944
 945     outeriter        = 0;
 946     inneriter        = 0;
 947
 948     /* Start outer loop over neighborlists */
 949     for(iidx=0; iidx<nri; iidx++)
 950     {
 951         /* Load shift vector for this list */
 952         i_shift_offset   = DIM*shiftidx[iidx];
 953
 954         /* Load limits for loop over neighbors */
 955         j_index_start    = jindex[iidx];
 956         j_index_end      = jindex[iidx+1];
 957
 958         /* Get outer coordinate index */
 959         inr              = iinr[iidx];
 960         i_coord_offset   = DIM*inr;
 961
 962         /* Load i particle coords and add shift vector */
 963         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 964                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 965
 966         fix0             = _mm_setzero_pd();
 967         fiy0             = _mm_setzero_pd();
 968         fiz0             = _mm_setzero_pd();
 969         fix1             = _mm_setzero_pd();
 970         fiy1             = _mm_setzero_pd();
 971         fiz1             = _mm_setzero_pd();
 972         fix2             = _mm_setzero_pd();
 973         fiy2             = _mm_setzero_pd();
 974         fiz2             = _mm_setzero_pd();
 975
 976         /* Start inner kernel loop */
 977         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 978         {
 979
 980             /* Get j neighbor index, and coordinate index */
 981             jnrA             = jjnr[jidx];
 982             jnrB             = jjnr[jidx+1];
 983             j_coord_offsetA  = DIM*jnrA;
 984             j_coord_offsetB  = DIM*jnrB;
 985
 986             /* load j atom coordinates */
 987             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 988                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 989
 990             /* Calculate displacement vector */
 991             dx00             = _mm_sub_pd(ix0,jx0);
 992             dy00             = _mm_sub_pd(iy0,jy0);
 993             dz00             = _mm_sub_pd(iz0,jz0);
 994             dx01             = _mm_sub_pd(ix0,jx1);
 995             dy01             = _mm_sub_pd(iy0,jy1);
 996             dz01             = _mm_sub_pd(iz0,jz1);
 997             dx02             = _mm_sub_pd(ix0,jx2);
 998             dy02             = _mm_sub_pd(iy0,jy2);
 999             dz02             = _mm_sub_pd(iz0,jz2);
1000             dx10             = _mm_sub_pd(ix1,jx0);
1001             dy10             = _mm_sub_pd(iy1,jy0);
1002             dz10             = _mm_sub_pd(iz1,jz0);
1003             dx11             = _mm_sub_pd(ix1,jx1);
1004             dy11             = _mm_sub_pd(iy1,jy1);
1005             dz11             = _mm_sub_pd(iz1,jz1);
1006             dx12             = _mm_sub_pd(ix1,jx2);
1007             dy12             = _mm_sub_pd(iy1,jy2);
1008             dz12             = _mm_sub_pd(iz1,jz2);
1009             dx20             = _mm_sub_pd(ix2,jx0);
1010             dy20             = _mm_sub_pd(iy2,jy0);
1011             dz20             = _mm_sub_pd(iz2,jz0);
1012             dx21             = _mm_sub_pd(ix2,jx1);
1013             dy21             = _mm_sub_pd(iy2,jy1);
1014             dz21             = _mm_sub_pd(iz2,jz1);
1015             dx22             = _mm_sub_pd(ix2,jx2);
1016             dy22             = _mm_sub_pd(iy2,jy2);
1017             dz22             = _mm_sub_pd(iz2,jz2);
1018
1019             /* Calculate squared distance and things based on it */
1020             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1021             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1022             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1023             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1024             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1025             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1026             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1027             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1028             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1029
1030             rinv00           = avx128fma_invsqrt_d(rsq00);
1031             rinv01           = avx128fma_invsqrt_d(rsq01);
1032             rinv02           = avx128fma_invsqrt_d(rsq02);
1033             rinv10           = avx128fma_invsqrt_d(rsq10);
1034             rinv11           = avx128fma_invsqrt_d(rsq11);
1035             rinv12           = avx128fma_invsqrt_d(rsq12);
1036             rinv20           = avx128fma_invsqrt_d(rsq20);
1037             rinv21           = avx128fma_invsqrt_d(rsq21);
1038             rinv22           = avx128fma_invsqrt_d(rsq22);
1039
1040             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1041             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1042             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1043             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1044             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1045             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1046             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1047             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1048             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1049
1050             fjx0             = _mm_setzero_pd();
1051             fjy0             = _mm_setzero_pd();
1052             fjz0             = _mm_setzero_pd();
1053             fjx1             = _mm_setzero_pd();
1054             fjy1             = _mm_setzero_pd();
1055             fjz1             = _mm_setzero_pd();
1056             fjx2             = _mm_setzero_pd();
1057             fjy2             = _mm_setzero_pd();
1058             fjz2             = _mm_setzero_pd();
1059
1060             /**************************
1061              * CALCULATE INTERACTIONS *
1062              **************************/
1063
1064             /* REACTION-FIELD ELECTROSTATICS */
1065             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1066
1067             /* LENNARD-JONES DISPERSION/REPULSION */
1068
1069             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1070             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1071
1072             fscal            = _mm_add_pd(felec,fvdw);
1073
1074             /* Update vectorial force */
1075             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1076             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1077             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1078
1079             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1080             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1081             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1082
1083             /**************************
1084              * CALCULATE INTERACTIONS *
1085              **************************/
1086
1087             /* REACTION-FIELD ELECTROSTATICS */
1088             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1089
1090             fscal            = felec;
1091
1092             /* Update vectorial force */
1093             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1094             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1095             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1096
1097             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1098             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1099             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1100
1101             /**************************
1102              * CALCULATE INTERACTIONS *
1103              **************************/
1104
1105             /* REACTION-FIELD ELECTROSTATICS */
1106             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1107
1108             fscal            = felec;
1109
1110             /* Update vectorial force */
1111             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1112             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1113             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1114
1115             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1116             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1117             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1118
1119             /**************************
1120              * CALCULATE INTERACTIONS *
1121              **************************/
1122
1123             /* REACTION-FIELD ELECTROSTATICS */
1124             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1125
1126             fscal            = felec;
1127
1128             /* Update vectorial force */
1129             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1130             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1131             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1132
1133             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1134             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1135             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1136
1137             /**************************
1138              * CALCULATE INTERACTIONS *
1139              **************************/
1140
1141             /* REACTION-FIELD ELECTROSTATICS */
1142             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1143
1144             fscal            = felec;
1145
1146             /* Update vectorial force */
1147             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1148             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1149             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1150
1151             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1152             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1153             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1154
1155             /**************************
1156              * CALCULATE INTERACTIONS *
1157              **************************/
1158
1159             /* REACTION-FIELD ELECTROSTATICS */
1160             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1161
1162             fscal            = felec;
1163
1164             /* Update vectorial force */
1165             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1166             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1167             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1168
1169             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1170             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1171             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1172
1173             /**************************
1174              * CALCULATE INTERACTIONS *
1175              **************************/
1176
1177             /* REACTION-FIELD ELECTROSTATICS */
1178             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1179
1180             fscal            = felec;
1181
1182             /* Update vectorial force */
1183             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1184             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1185             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1186
1187             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1188             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1189             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1190
1191             /**************************
1192              * CALCULATE INTERACTIONS *
1193              **************************/
1194
1195             /* REACTION-FIELD ELECTROSTATICS */
1196             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1197
1198             fscal            = felec;
1199
1200             /* Update vectorial force */
1201             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1202             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1203             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1204
1205             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1206             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1207             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1208
1209             /**************************
1210              * CALCULATE INTERACTIONS *
1211              **************************/
1212
1213             /* REACTION-FIELD ELECTROSTATICS */
1214             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1215
1216             fscal            = felec;
1217
1218             /* Update vectorial force */
1219             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1220             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1221             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1222
1223             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1224             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1225             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1226
1227             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1228
1229             /* Inner loop uses 277 flops */
1230         }
1231
1232         if(jidx<j_index_end)
1233         {
1234
1235             jnrA             = jjnr[jidx];
1236             j_coord_offsetA  = DIM*jnrA;
1237
1238             /* load j atom coordinates */
1239             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1240                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1241
1242             /* Calculate displacement vector */
1243             dx00             = _mm_sub_pd(ix0,jx0);
1244             dy00             = _mm_sub_pd(iy0,jy0);
1245             dz00             = _mm_sub_pd(iz0,jz0);
1246             dx01             = _mm_sub_pd(ix0,jx1);
1247             dy01             = _mm_sub_pd(iy0,jy1);
1248             dz01             = _mm_sub_pd(iz0,jz1);
1249             dx02             = _mm_sub_pd(ix0,jx2);
1250             dy02             = _mm_sub_pd(iy0,jy2);
1251             dz02             = _mm_sub_pd(iz0,jz2);
1252             dx10             = _mm_sub_pd(ix1,jx0);
1253             dy10             = _mm_sub_pd(iy1,jy0);
1254             dz10             = _mm_sub_pd(iz1,jz0);
1255             dx11             = _mm_sub_pd(ix1,jx1);
1256             dy11             = _mm_sub_pd(iy1,jy1);
1257             dz11             = _mm_sub_pd(iz1,jz1);
1258             dx12             = _mm_sub_pd(ix1,jx2);
1259             dy12             = _mm_sub_pd(iy1,jy2);
1260             dz12             = _mm_sub_pd(iz1,jz2);
1261             dx20             = _mm_sub_pd(ix2,jx0);
1262             dy20             = _mm_sub_pd(iy2,jy0);
1263             dz20             = _mm_sub_pd(iz2,jz0);
1264             dx21             = _mm_sub_pd(ix2,jx1);
1265             dy21             = _mm_sub_pd(iy2,jy1);
1266             dz21             = _mm_sub_pd(iz2,jz1);
1267             dx22             = _mm_sub_pd(ix2,jx2);
1268             dy22             = _mm_sub_pd(iy2,jy2);
1269             dz22             = _mm_sub_pd(iz2,jz2);
1270
1271             /* Calculate squared distance and things based on it */
1272             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1273             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1274             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1275             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1276             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1277             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1278             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1279             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1280             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1281
1282             rinv00           = avx128fma_invsqrt_d(rsq00);
1283             rinv01           = avx128fma_invsqrt_d(rsq01);
1284             rinv02           = avx128fma_invsqrt_d(rsq02);
1285             rinv10           = avx128fma_invsqrt_d(rsq10);
1286             rinv11           = avx128fma_invsqrt_d(rsq11);
1287             rinv12           = avx128fma_invsqrt_d(rsq12);
1288             rinv20           = avx128fma_invsqrt_d(rsq20);
1289             rinv21           = avx128fma_invsqrt_d(rsq21);
1290             rinv22           = avx128fma_invsqrt_d(rsq22);
1291
1292             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1293             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1294             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1295             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1296             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1297             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1298             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1299             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1300             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1301
1302             fjx0             = _mm_setzero_pd();
1303             fjy0             = _mm_setzero_pd();
1304             fjz0             = _mm_setzero_pd();
1305             fjx1             = _mm_setzero_pd();
1306             fjy1             = _mm_setzero_pd();
1307             fjz1             = _mm_setzero_pd();
1308             fjx2             = _mm_setzero_pd();
1309             fjy2             = _mm_setzero_pd();
1310             fjz2             = _mm_setzero_pd();
1311
1312             /**************************
1313              * CALCULATE INTERACTIONS *
1314              **************************/
1315
1316             /* REACTION-FIELD ELECTROSTATICS */
1317             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1318
1319             /* LENNARD-JONES DISPERSION/REPULSION */
1320
1321             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1322             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1323
1324             fscal            = _mm_add_pd(felec,fvdw);
1325
1326             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1327
1328             /* Update vectorial force */
1329             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1330             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1331             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1332
1333             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1334             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1335             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1336
1337             /**************************
1338              * CALCULATE INTERACTIONS *
1339              **************************/
1340
1341             /* REACTION-FIELD ELECTROSTATICS */
1342             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1343
1344             fscal            = felec;
1345
1346             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1347
1348             /* Update vectorial force */
1349             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1350             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1351             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1352
1353             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1354             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1355             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1356
1357             /**************************
1358              * CALCULATE INTERACTIONS *
1359              **************************/
1360
1361             /* REACTION-FIELD ELECTROSTATICS */
1362             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1363
1364             fscal            = felec;
1365
1366             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1367
1368             /* Update vectorial force */
1369             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1370             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1371             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1372
1373             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1374             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1375             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1376
1377             /**************************
1378              * CALCULATE INTERACTIONS *
1379              **************************/
1380
1381             /* REACTION-FIELD ELECTROSTATICS */
1382             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1383
1384             fscal            = felec;
1385
1386             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1387
1388             /* Update vectorial force */
1389             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1390             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1391             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1392
1393             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1394             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1395             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1396
1397             /**************************
1398              * CALCULATE INTERACTIONS *
1399              **************************/
1400
1401             /* REACTION-FIELD ELECTROSTATICS */
1402             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1403
1404             fscal            = felec;
1405
1406             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1407
1408             /* Update vectorial force */
1409             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1410             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1411             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1412
1413             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1414             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1415             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1416
1417             /**************************
1418              * CALCULATE INTERACTIONS *
1419              **************************/
1420
1421             /* REACTION-FIELD ELECTROSTATICS */
1422             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1423
1424             fscal            = felec;
1425
1426             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1427
1428             /* Update vectorial force */
1429             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1430             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1431             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1432
1433             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1434             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1435             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1436
1437             /**************************
1438              * CALCULATE INTERACTIONS *
1439              **************************/
1440
1441             /* REACTION-FIELD ELECTROSTATICS */
1442             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1443
1444             fscal            = felec;
1445
1446             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1447
1448             /* Update vectorial force */
1449             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1450             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1451             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1452
1453             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1454             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1455             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1456
1457             /**************************
1458              * CALCULATE INTERACTIONS *
1459              **************************/
1460
1461             /* REACTION-FIELD ELECTROSTATICS */
1462             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1463
1464             fscal            = felec;
1465
1466             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1467
1468             /* Update vectorial force */
1469             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1470             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1471             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1472
1473             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1474             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1475             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1476
1477             /**************************
1478              * CALCULATE INTERACTIONS *
1479              **************************/
1480
1481             /* REACTION-FIELD ELECTROSTATICS */
1482             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1483
1484             fscal            = felec;
1485
1486             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1487
1488             /* Update vectorial force */
1489             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1490             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1491             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1492
1493             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1494             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1495             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1496
1497             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1498
1499             /* Inner loop uses 277 flops */
1500         }
1501
1502         /* End of innermost loop */
1503
1504         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1505                                               f+i_coord_offset,fshift+i_shift_offset);
1506
1507         /* Increment number of inner iterations */
1508         inneriter                  += j_index_end - j_index_start;
1509
1510         /* Outer loop uses 18 flops */
1511     }
1512
1513     /* Increment number of outer iterations */
1514     outeriter        += nri;
1515
1516     /* Update outer/inner flops */
1517
1518     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1519 }