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, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwjidx0A,vdwjidx0B;
  87     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  94     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  95     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     int              nvdwtype;
 104     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 105     int              *vdwtype;
 106     real             *vdwparam;
 107     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 108     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 109     __m128d          dummy_mask,cutoff_mask;
 110     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 111     __m128d          one     = _mm_set1_pd(1.0);
 112     __m128d          two     = _mm_set1_pd(2.0);
 113     x                = xx[0];
 114     f                = ff[0];
 115
 116     nri              = nlist->nri;
 117     iinr             = nlist->iinr;
 118     jindex           = nlist->jindex;
 119     jjnr             = nlist->jjnr;
 120     shiftidx         = nlist->shift;
 121     gid              = nlist->gid;
 122     shiftvec         = fr->shift_vec[0];
 123     fshift           = fr->fshift[0];
 124     facel            = _mm_set1_pd(fr->epsfac);
 125     charge           = mdatoms->chargeA;
 126     krf              = _mm_set1_pd(fr->ic->k_rf);
 127     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 128     crf              = _mm_set1_pd(fr->ic->c_rf);
 129     nvdwtype         = fr->ntype;
 130     vdwparam         = fr->nbfp;
 131     vdwtype          = mdatoms->typeA;
 132
 133     /* Setup water-specific parameters */
 134     inr              = nlist->iinr[0];
 135     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 136     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 137     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 138     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 139
 140     jq0              = _mm_set1_pd(charge[inr+0]);
 141     jq1              = _mm_set1_pd(charge[inr+1]);
 142     jq2              = _mm_set1_pd(charge[inr+2]);
 143     vdwjidx0A        = 2*vdwtype[inr+0];
 144     qq00             = _mm_mul_pd(iq0,jq0);
 145     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 146     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 147     qq01             = _mm_mul_pd(iq0,jq1);
 148     qq02             = _mm_mul_pd(iq0,jq2);
 149     qq10             = _mm_mul_pd(iq1,jq0);
 150     qq11             = _mm_mul_pd(iq1,jq1);
 151     qq12             = _mm_mul_pd(iq1,jq2);
 152     qq20             = _mm_mul_pd(iq2,jq0);
 153     qq21             = _mm_mul_pd(iq2,jq1);
 154     qq22             = _mm_mul_pd(iq2,jq2);
 155
 156     /* Avoid stupid compiler warnings */
 157     jnrA = jnrB = 0;
 158     j_coord_offsetA = 0;
 159     j_coord_offsetB = 0;
 160
 161     outeriter        = 0;
 162     inneriter        = 0;
 163
 164     /* Start outer loop over neighborlists */
 165     for(iidx=0; iidx<nri; iidx++)
 166     {
 167         /* Load shift vector for this list */
 168         i_shift_offset   = DIM*shiftidx[iidx];
 169
 170         /* Load limits for loop over neighbors */
 171         j_index_start    = jindex[iidx];
 172         j_index_end      = jindex[iidx+1];
 173
 174         /* Get outer coordinate index */
 175         inr              = iinr[iidx];
 176         i_coord_offset   = DIM*inr;
 177
 178         /* Load i particle coords and add shift vector */
 179         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 180                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 181
 182         fix0             = _mm_setzero_pd();
 183         fiy0             = _mm_setzero_pd();
 184         fiz0             = _mm_setzero_pd();
 185         fix1             = _mm_setzero_pd();
 186         fiy1             = _mm_setzero_pd();
 187         fiz1             = _mm_setzero_pd();
 188         fix2             = _mm_setzero_pd();
 189         fiy2             = _mm_setzero_pd();
 190         fiz2             = _mm_setzero_pd();
 191
 192         /* Reset potential sums */
 193         velecsum         = _mm_setzero_pd();
 194         vvdwsum          = _mm_setzero_pd();
 195
 196         /* Start inner kernel loop */
 197         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 198         {
 199
 200             /* Get j neighbor index, and coordinate index */
 201             jnrA             = jjnr[jidx];
 202             jnrB             = jjnr[jidx+1];
 203             j_coord_offsetA  = DIM*jnrA;
 204             j_coord_offsetB  = DIM*jnrB;
 205
 206             /* load j atom coordinates */
 207             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 208                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 209
 210             /* Calculate displacement vector */
 211             dx00             = _mm_sub_pd(ix0,jx0);
 212             dy00             = _mm_sub_pd(iy0,jy0);
 213             dz00             = _mm_sub_pd(iz0,jz0);
 214             dx01             = _mm_sub_pd(ix0,jx1);
 215             dy01             = _mm_sub_pd(iy0,jy1);
 216             dz01             = _mm_sub_pd(iz0,jz1);
 217             dx02             = _mm_sub_pd(ix0,jx2);
 218             dy02             = _mm_sub_pd(iy0,jy2);
 219             dz02             = _mm_sub_pd(iz0,jz2);
 220             dx10             = _mm_sub_pd(ix1,jx0);
 221             dy10             = _mm_sub_pd(iy1,jy0);
 222             dz10             = _mm_sub_pd(iz1,jz0);
 223             dx11             = _mm_sub_pd(ix1,jx1);
 224             dy11             = _mm_sub_pd(iy1,jy1);
 225             dz11             = _mm_sub_pd(iz1,jz1);
 226             dx12             = _mm_sub_pd(ix1,jx2);
 227             dy12             = _mm_sub_pd(iy1,jy2);
 228             dz12             = _mm_sub_pd(iz1,jz2);
 229             dx20             = _mm_sub_pd(ix2,jx0);
 230             dy20             = _mm_sub_pd(iy2,jy0);
 231             dz20             = _mm_sub_pd(iz2,jz0);
 232             dx21             = _mm_sub_pd(ix2,jx1);
 233             dy21             = _mm_sub_pd(iy2,jy1);
 234             dz21             = _mm_sub_pd(iz2,jz1);
 235             dx22             = _mm_sub_pd(ix2,jx2);
 236             dy22             = _mm_sub_pd(iy2,jy2);
 237             dz22             = _mm_sub_pd(iz2,jz2);
 238
 239             /* Calculate squared distance and things based on it */
 240             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 241             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 242             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 243             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 244             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 245             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 246             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 247             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 248             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 249
 250             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 251             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 252             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 253             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 254             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 255             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 256             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 257             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 258             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 259
 260             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 261             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 262             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 263             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 264             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 265             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 266             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 267             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 268             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 269
 270             fjx0             = _mm_setzero_pd();
 271             fjy0             = _mm_setzero_pd();
 272             fjz0             = _mm_setzero_pd();
 273             fjx1             = _mm_setzero_pd();
 274             fjy1             = _mm_setzero_pd();
 275             fjz1             = _mm_setzero_pd();
 276             fjx2             = _mm_setzero_pd();
 277             fjy2             = _mm_setzero_pd();
 278             fjz2             = _mm_setzero_pd();
 279
 280             /**************************
 281              * CALCULATE INTERACTIONS *
 282              **************************/
 283
 284             /* REACTION-FIELD ELECTROSTATICS */
 285             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 286             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 287
 288             /* LENNARD-JONES DISPERSION/REPULSION */
 289
 290             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 291             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 292             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 293             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 294             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 295
 296             /* Update potential sum for this i atom from the interaction with this j atom. */
 297             velecsum         = _mm_add_pd(velecsum,velec);
 298             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 299
 300             fscal            = _mm_add_pd(felec,fvdw);
 301
 302             /* Update vectorial force */
 303             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 304             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 305             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 306
 307             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 308             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 309             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             /* REACTION-FIELD ELECTROSTATICS */
 316             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 317             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velecsum         = _mm_add_pd(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324             /* Update vectorial force */
 325             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 326             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 327             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 328
 329             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 330             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 331             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             /* REACTION-FIELD ELECTROSTATICS */
 338             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 339             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 340
 341             /* Update potential sum for this i atom from the interaction with this j atom. */
 342             velecsum         = _mm_add_pd(velecsum,velec);
 343
 344             fscal            = felec;
 345
 346             /* Update vectorial force */
 347             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 348             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 349             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 350
 351             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 352             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 353             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             /* REACTION-FIELD ELECTROSTATICS */
 360             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 361             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velecsum         = _mm_add_pd(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368             /* Update vectorial force */
 369             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 370             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 371             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 372
 373             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 374             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 375             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 376
 377             /**************************
 378              * CALCULATE INTERACTIONS *
 379              **************************/
 380
 381             /* REACTION-FIELD ELECTROSTATICS */
 382             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 383             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velecsum         = _mm_add_pd(velecsum,velec);
 387
 388             fscal            = felec;
 389
 390             /* Update vectorial force */
 391             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 392             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 393             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 394
 395             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 396             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 397             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             /* REACTION-FIELD ELECTROSTATICS */
 404             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 405             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velecsum         = _mm_add_pd(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412             /* Update vectorial force */
 413             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 414             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 415             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 416
 417             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 418             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 419             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             /* REACTION-FIELD ELECTROSTATICS */
 426             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 427             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 428
 429             /* Update potential sum for this i atom from the interaction with this j atom. */
 430             velecsum         = _mm_add_pd(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434             /* Update vectorial force */
 435             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 436             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 437             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 438
 439             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 440             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 441             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             /* REACTION-FIELD ELECTROSTATICS */
 448             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 449             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 450
 451             /* Update potential sum for this i atom from the interaction with this j atom. */
 452             velecsum         = _mm_add_pd(velecsum,velec);
 453
 454             fscal            = felec;
 455
 456             /* Update vectorial force */
 457             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 458             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 459             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 460
 461             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 462             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 463             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             /* REACTION-FIELD ELECTROSTATICS */
 470             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 471             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 472
 473             /* Update potential sum for this i atom from the interaction with this j atom. */
 474             velecsum         = _mm_add_pd(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478             /* Update vectorial force */
 479             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 480             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 481             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 482
 483             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 484             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 485             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 486
 487             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 488
 489             /* Inner loop uses 327 flops */
 490         }
 491
 492         if(jidx<j_index_end)
 493         {
 494
 495             jnrA             = jjnr[jidx];
 496             j_coord_offsetA  = DIM*jnrA;
 497
 498             /* load j atom coordinates */
 499             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 500                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 501
 502             /* Calculate displacement vector */
 503             dx00             = _mm_sub_pd(ix0,jx0);
 504             dy00             = _mm_sub_pd(iy0,jy0);
 505             dz00             = _mm_sub_pd(iz0,jz0);
 506             dx01             = _mm_sub_pd(ix0,jx1);
 507             dy01             = _mm_sub_pd(iy0,jy1);
 508             dz01             = _mm_sub_pd(iz0,jz1);
 509             dx02             = _mm_sub_pd(ix0,jx2);
 510             dy02             = _mm_sub_pd(iy0,jy2);
 511             dz02             = _mm_sub_pd(iz0,jz2);
 512             dx10             = _mm_sub_pd(ix1,jx0);
 513             dy10             = _mm_sub_pd(iy1,jy0);
 514             dz10             = _mm_sub_pd(iz1,jz0);
 515             dx11             = _mm_sub_pd(ix1,jx1);
 516             dy11             = _mm_sub_pd(iy1,jy1);
 517             dz11             = _mm_sub_pd(iz1,jz1);
 518             dx12             = _mm_sub_pd(ix1,jx2);
 519             dy12             = _mm_sub_pd(iy1,jy2);
 520             dz12             = _mm_sub_pd(iz1,jz2);
 521             dx20             = _mm_sub_pd(ix2,jx0);
 522             dy20             = _mm_sub_pd(iy2,jy0);
 523             dz20             = _mm_sub_pd(iz2,jz0);
 524             dx21             = _mm_sub_pd(ix2,jx1);
 525             dy21             = _mm_sub_pd(iy2,jy1);
 526             dz21             = _mm_sub_pd(iz2,jz1);
 527             dx22             = _mm_sub_pd(ix2,jx2);
 528             dy22             = _mm_sub_pd(iy2,jy2);
 529             dz22             = _mm_sub_pd(iz2,jz2);
 530
 531             /* Calculate squared distance and things based on it */
 532             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 533             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 534             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 535             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 536             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 537             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 538             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 539             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 540             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 541
 542             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 543             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 544             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 545             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 546             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 547             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 548             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 549             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 550             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 551
 552             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 553             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 554             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 555             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 556             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 557             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 558             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 559             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 560             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 561
 562             fjx0             = _mm_setzero_pd();
 563             fjy0             = _mm_setzero_pd();
 564             fjz0             = _mm_setzero_pd();
 565             fjx1             = _mm_setzero_pd();
 566             fjy1             = _mm_setzero_pd();
 567             fjz1             = _mm_setzero_pd();
 568             fjx2             = _mm_setzero_pd();
 569             fjy2             = _mm_setzero_pd();
 570             fjz2             = _mm_setzero_pd();
 571
 572             /**************************
 573              * CALCULATE INTERACTIONS *
 574              **************************/
 575
 576             /* REACTION-FIELD ELECTROSTATICS */
 577             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 578             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 579
 580             /* LENNARD-JONES DISPERSION/REPULSION */
 581
 582             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 583             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 584             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 585             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 586             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 587
 588             /* Update potential sum for this i atom from the interaction with this j atom. */
 589             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 590             velecsum         = _mm_add_pd(velecsum,velec);
 591             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 592             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 593
 594             fscal            = _mm_add_pd(felec,fvdw);
 595
 596             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 597
 598             /* Update vectorial force */
 599             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 600             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 601             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 602
 603             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 604             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 605             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             /* REACTION-FIELD ELECTROSTATICS */
 612             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 613             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 614
 615             /* Update potential sum for this i atom from the interaction with this j atom. */
 616             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 617             velecsum         = _mm_add_pd(velecsum,velec);
 618
 619             fscal            = felec;
 620
 621             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 622
 623             /* Update vectorial force */
 624             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 625             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 626             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 627
 628             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 629             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 630             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 631
 632             /**************************
 633              * CALCULATE INTERACTIONS *
 634              **************************/
 635
 636             /* REACTION-FIELD ELECTROSTATICS */
 637             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 638             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 639
 640             /* Update potential sum for this i atom from the interaction with this j atom. */
 641             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 642             velecsum         = _mm_add_pd(velecsum,velec);
 643
 644             fscal            = felec;
 645
 646             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 647
 648             /* Update vectorial force */
 649             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 650             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 651             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 652
 653             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 654             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 655             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 656
 657             /**************************
 658              * CALCULATE INTERACTIONS *
 659              **************************/
 660
 661             /* REACTION-FIELD ELECTROSTATICS */
 662             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 663             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 664
 665             /* Update potential sum for this i atom from the interaction with this j atom. */
 666             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 667             velecsum         = _mm_add_pd(velecsum,velec);
 668
 669             fscal            = felec;
 670
 671             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 672
 673             /* Update vectorial force */
 674             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 675             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 676             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 677
 678             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 679             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 680             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             /* REACTION-FIELD ELECTROSTATICS */
 687             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 688             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 689
 690             /* Update potential sum for this i atom from the interaction with this j atom. */
 691             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 692             velecsum         = _mm_add_pd(velecsum,velec);
 693
 694             fscal            = felec;
 695
 696             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 697
 698             /* Update vectorial force */
 699             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 700             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 701             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 702
 703             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 704             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 705             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 706
 707             /**************************
 708              * CALCULATE INTERACTIONS *
 709              **************************/
 710
 711             /* REACTION-FIELD ELECTROSTATICS */
 712             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 713             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 714
 715             /* Update potential sum for this i atom from the interaction with this j atom. */
 716             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 717             velecsum         = _mm_add_pd(velecsum,velec);
 718
 719             fscal            = felec;
 720
 721             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 722
 723             /* Update vectorial force */
 724             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 725             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 726             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 727
 728             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 729             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 730             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 731
 732             /**************************
 733              * CALCULATE INTERACTIONS *
 734              **************************/
 735
 736             /* REACTION-FIELD ELECTROSTATICS */
 737             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 738             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 739
 740             /* Update potential sum for this i atom from the interaction with this j atom. */
 741             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 742             velecsum         = _mm_add_pd(velecsum,velec);
 743
 744             fscal            = felec;
 745
 746             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 747
 748             /* Update vectorial force */
 749             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 750             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 751             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 752
 753             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 754             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 755             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 756
 757             /**************************
 758              * CALCULATE INTERACTIONS *
 759              **************************/
 760
 761             /* REACTION-FIELD ELECTROSTATICS */
 762             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 763             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 764
 765             /* Update potential sum for this i atom from the interaction with this j atom. */
 766             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 767             velecsum         = _mm_add_pd(velecsum,velec);
 768
 769             fscal            = felec;
 770
 771             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 772
 773             /* Update vectorial force */
 774             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 775             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 776             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 777
 778             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 779             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 780             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 781
 782             /**************************
 783              * CALCULATE INTERACTIONS *
 784              **************************/
 785
 786             /* REACTION-FIELD ELECTROSTATICS */
 787             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 788             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 789
 790             /* Update potential sum for this i atom from the interaction with this j atom. */
 791             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 792             velecsum         = _mm_add_pd(velecsum,velec);
 793
 794             fscal            = felec;
 795
 796             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 797
 798             /* Update vectorial force */
 799             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 800             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 801             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 802
 803             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 804             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 805             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 806
 807             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 808
 809             /* Inner loop uses 327 flops */
 810         }
 811
 812         /* End of innermost loop */
 813
 814         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 815                                               f+i_coord_offset,fshift+i_shift_offset);
 816
 817         ggid                        = gid[iidx];
 818         /* Update potential energies */
 819         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 820         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 821
 822         /* Increment number of inner iterations */
 823         inneriter                  += j_index_end - j_index_start;
 824
 825         /* Outer loop uses 20 flops */
 826     }
 827
 828     /* Increment number of outer iterations */
 829     outeriter        += nri;
 830
 831     /* Update outer/inner flops */
 832
 833     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*327);
 834 }
 835 /*
 836  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_double
 837  * Electrostatics interaction: ReactionField
 838  * VdW interaction:            LennardJones
 839  * Geometry:                   Water3-Water3
 840  * Calculate force/pot:        Force
 841  */
 842 void
 843 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_double
 844                     (t_nblist                    * gmx_restrict       nlist,
 845                      rvec                        * gmx_restrict          xx,
 846                      rvec                        * gmx_restrict          ff,
 847                      t_forcerec                  * gmx_restrict          fr,
 848                      t_mdatoms                   * gmx_restrict     mdatoms,
 849                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 850                      t_nrnb                      * gmx_restrict        nrnb)
 851 {
 852     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 853      * just 0 for non-waters.
 854      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 855      * jnr indices corresponding to data put in the four positions in the SIMD register.
 856      */
 857     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 858     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 859     int              jnrA,jnrB;
 860     int              j_coord_offsetA,j_coord_offsetB;
 861     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 862     real             rcutoff_scalar;
 863     real             *shiftvec,*fshift,*x,*f;
 864     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 865     int              vdwioffset0;
 866     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 867     int              vdwioffset1;
 868     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 869     int              vdwioffset2;
 870     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 871     int              vdwjidx0A,vdwjidx0B;
 872     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 873     int              vdwjidx1A,vdwjidx1B;
 874     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 875     int              vdwjidx2A,vdwjidx2B;
 876     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 877     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 878     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 879     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 880     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 881     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 882     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 883     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 884     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 885     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 886     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 887     real             *charge;
 888     int              nvdwtype;
 889     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 890     int              *vdwtype;
 891     real             *vdwparam;
 892     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 893     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 894     __m128d          dummy_mask,cutoff_mask;
 895     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 896     __m128d          one     = _mm_set1_pd(1.0);
 897     __m128d          two     = _mm_set1_pd(2.0);
 898     x                = xx[0];
 899     f                = ff[0];
 900
 901     nri              = nlist->nri;
 902     iinr             = nlist->iinr;
 903     jindex           = nlist->jindex;
 904     jjnr             = nlist->jjnr;
 905     shiftidx         = nlist->shift;
 906     gid              = nlist->gid;
 907     shiftvec         = fr->shift_vec[0];
 908     fshift           = fr->fshift[0];
 909     facel            = _mm_set1_pd(fr->epsfac);
 910     charge           = mdatoms->chargeA;
 911     krf              = _mm_set1_pd(fr->ic->k_rf);
 912     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 913     crf              = _mm_set1_pd(fr->ic->c_rf);
 914     nvdwtype         = fr->ntype;
 915     vdwparam         = fr->nbfp;
 916     vdwtype          = mdatoms->typeA;
 917
 918     /* Setup water-specific parameters */
 919     inr              = nlist->iinr[0];
 920     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 921     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 922     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 923     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 924
 925     jq0              = _mm_set1_pd(charge[inr+0]);
 926     jq1              = _mm_set1_pd(charge[inr+1]);
 927     jq2              = _mm_set1_pd(charge[inr+2]);
 928     vdwjidx0A        = 2*vdwtype[inr+0];
 929     qq00             = _mm_mul_pd(iq0,jq0);
 930     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 931     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 932     qq01             = _mm_mul_pd(iq0,jq1);
 933     qq02             = _mm_mul_pd(iq0,jq2);
 934     qq10             = _mm_mul_pd(iq1,jq0);
 935     qq11             = _mm_mul_pd(iq1,jq1);
 936     qq12             = _mm_mul_pd(iq1,jq2);
 937     qq20             = _mm_mul_pd(iq2,jq0);
 938     qq21             = _mm_mul_pd(iq2,jq1);
 939     qq22             = _mm_mul_pd(iq2,jq2);
 940
 941     /* Avoid stupid compiler warnings */
 942     jnrA = jnrB = 0;
 943     j_coord_offsetA = 0;
 944     j_coord_offsetB = 0;
 945
 946     outeriter        = 0;
 947     inneriter        = 0;
 948
 949     /* Start outer loop over neighborlists */
 950     for(iidx=0; iidx<nri; iidx++)
 951     {
 952         /* Load shift vector for this list */
 953         i_shift_offset   = DIM*shiftidx[iidx];
 954
 955         /* Load limits for loop over neighbors */
 956         j_index_start    = jindex[iidx];
 957         j_index_end      = jindex[iidx+1];
 958
 959         /* Get outer coordinate index */
 960         inr              = iinr[iidx];
 961         i_coord_offset   = DIM*inr;
 962
 963         /* Load i particle coords and add shift vector */
 964         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 965                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 966
 967         fix0             = _mm_setzero_pd();
 968         fiy0             = _mm_setzero_pd();
 969         fiz0             = _mm_setzero_pd();
 970         fix1             = _mm_setzero_pd();
 971         fiy1             = _mm_setzero_pd();
 972         fiz1             = _mm_setzero_pd();
 973         fix2             = _mm_setzero_pd();
 974         fiy2             = _mm_setzero_pd();
 975         fiz2             = _mm_setzero_pd();
 976
 977         /* Start inner kernel loop */
 978         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 979         {
 980
 981             /* Get j neighbor index, and coordinate index */
 982             jnrA             = jjnr[jidx];
 983             jnrB             = jjnr[jidx+1];
 984             j_coord_offsetA  = DIM*jnrA;
 985             j_coord_offsetB  = DIM*jnrB;
 986
 987             /* load j atom coordinates */
 988             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 989                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 990
 991             /* Calculate displacement vector */
 992             dx00             = _mm_sub_pd(ix0,jx0);
 993             dy00             = _mm_sub_pd(iy0,jy0);
 994             dz00             = _mm_sub_pd(iz0,jz0);
 995             dx01             = _mm_sub_pd(ix0,jx1);
 996             dy01             = _mm_sub_pd(iy0,jy1);
 997             dz01             = _mm_sub_pd(iz0,jz1);
 998             dx02             = _mm_sub_pd(ix0,jx2);
 999             dy02             = _mm_sub_pd(iy0,jy2);
1000             dz02             = _mm_sub_pd(iz0,jz2);
1001             dx10             = _mm_sub_pd(ix1,jx0);
1002             dy10             = _mm_sub_pd(iy1,jy0);
1003             dz10             = _mm_sub_pd(iz1,jz0);
1004             dx11             = _mm_sub_pd(ix1,jx1);
1005             dy11             = _mm_sub_pd(iy1,jy1);
1006             dz11             = _mm_sub_pd(iz1,jz1);
1007             dx12             = _mm_sub_pd(ix1,jx2);
1008             dy12             = _mm_sub_pd(iy1,jy2);
1009             dz12             = _mm_sub_pd(iz1,jz2);
1010             dx20             = _mm_sub_pd(ix2,jx0);
1011             dy20             = _mm_sub_pd(iy2,jy0);
1012             dz20             = _mm_sub_pd(iz2,jz0);
1013             dx21             = _mm_sub_pd(ix2,jx1);
1014             dy21             = _mm_sub_pd(iy2,jy1);
1015             dz21             = _mm_sub_pd(iz2,jz1);
1016             dx22             = _mm_sub_pd(ix2,jx2);
1017             dy22             = _mm_sub_pd(iy2,jy2);
1018             dz22             = _mm_sub_pd(iz2,jz2);
1019
1020             /* Calculate squared distance and things based on it */
1021             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1022             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1023             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1024             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1025             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1026             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1027             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1028             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1029             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1030
1031             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1032             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1033             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1034             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1035             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1036             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1037             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1038             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1039             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1040
1041             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1042             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1043             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1044             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1045             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1046             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1047             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1048             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1049             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1050
1051             fjx0             = _mm_setzero_pd();
1052             fjy0             = _mm_setzero_pd();
1053             fjz0             = _mm_setzero_pd();
1054             fjx1             = _mm_setzero_pd();
1055             fjy1             = _mm_setzero_pd();
1056             fjz1             = _mm_setzero_pd();
1057             fjx2             = _mm_setzero_pd();
1058             fjy2             = _mm_setzero_pd();
1059             fjz2             = _mm_setzero_pd();
1060
1061             /**************************
1062              * CALCULATE INTERACTIONS *
1063              **************************/
1064
1065             /* REACTION-FIELD ELECTROSTATICS */
1066             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1067
1068             /* LENNARD-JONES DISPERSION/REPULSION */
1069
1070             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1071             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1072
1073             fscal            = _mm_add_pd(felec,fvdw);
1074
1075             /* Update vectorial force */
1076             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1077             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1078             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1079
1080             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1081             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1082             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1083
1084             /**************************
1085              * CALCULATE INTERACTIONS *
1086              **************************/
1087
1088             /* REACTION-FIELD ELECTROSTATICS */
1089             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1090
1091             fscal            = felec;
1092
1093             /* Update vectorial force */
1094             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1095             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1096             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1097
1098             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1099             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1100             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1101
1102             /**************************
1103              * CALCULATE INTERACTIONS *
1104              **************************/
1105
1106             /* REACTION-FIELD ELECTROSTATICS */
1107             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1108
1109             fscal            = felec;
1110
1111             /* Update vectorial force */
1112             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1113             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1114             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1115
1116             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1117             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1118             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1119
1120             /**************************
1121              * CALCULATE INTERACTIONS *
1122              **************************/
1123
1124             /* REACTION-FIELD ELECTROSTATICS */
1125             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1126
1127             fscal            = felec;
1128
1129             /* Update vectorial force */
1130             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1131             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1132             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1133
1134             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1135             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1136             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1137
1138             /**************************
1139              * CALCULATE INTERACTIONS *
1140              **************************/
1141
1142             /* REACTION-FIELD ELECTROSTATICS */
1143             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1144
1145             fscal            = felec;
1146
1147             /* Update vectorial force */
1148             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1149             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1150             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1151
1152             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1153             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1154             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1155
1156             /**************************
1157              * CALCULATE INTERACTIONS *
1158              **************************/
1159
1160             /* REACTION-FIELD ELECTROSTATICS */
1161             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1162
1163             fscal            = felec;
1164
1165             /* Update vectorial force */
1166             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1167             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1168             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1169
1170             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1171             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1172             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1173
1174             /**************************
1175              * CALCULATE INTERACTIONS *
1176              **************************/
1177
1178             /* REACTION-FIELD ELECTROSTATICS */
1179             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1180
1181             fscal            = felec;
1182
1183             /* Update vectorial force */
1184             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1185             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1186             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1187
1188             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1189             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1190             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1191
1192             /**************************
1193              * CALCULATE INTERACTIONS *
1194              **************************/
1195
1196             /* REACTION-FIELD ELECTROSTATICS */
1197             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1198
1199             fscal            = felec;
1200
1201             /* Update vectorial force */
1202             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1203             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1204             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1205
1206             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1207             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1208             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1209
1210             /**************************
1211              * CALCULATE INTERACTIONS *
1212              **************************/
1213
1214             /* REACTION-FIELD ELECTROSTATICS */
1215             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1216
1217             fscal            = felec;
1218
1219             /* Update vectorial force */
1220             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1221             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1222             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1223
1224             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1225             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1226             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1227
1228             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1229
1230             /* Inner loop uses 277 flops */
1231         }
1232
1233         if(jidx<j_index_end)
1234         {
1235
1236             jnrA             = jjnr[jidx];
1237             j_coord_offsetA  = DIM*jnrA;
1238
1239             /* load j atom coordinates */
1240             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1241                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1242
1243             /* Calculate displacement vector */
1244             dx00             = _mm_sub_pd(ix0,jx0);
1245             dy00             = _mm_sub_pd(iy0,jy0);
1246             dz00             = _mm_sub_pd(iz0,jz0);
1247             dx01             = _mm_sub_pd(ix0,jx1);
1248             dy01             = _mm_sub_pd(iy0,jy1);
1249             dz01             = _mm_sub_pd(iz0,jz1);
1250             dx02             = _mm_sub_pd(ix0,jx2);
1251             dy02             = _mm_sub_pd(iy0,jy2);
1252             dz02             = _mm_sub_pd(iz0,jz2);
1253             dx10             = _mm_sub_pd(ix1,jx0);
1254             dy10             = _mm_sub_pd(iy1,jy0);
1255             dz10             = _mm_sub_pd(iz1,jz0);
1256             dx11             = _mm_sub_pd(ix1,jx1);
1257             dy11             = _mm_sub_pd(iy1,jy1);
1258             dz11             = _mm_sub_pd(iz1,jz1);
1259             dx12             = _mm_sub_pd(ix1,jx2);
1260             dy12             = _mm_sub_pd(iy1,jy2);
1261             dz12             = _mm_sub_pd(iz1,jz2);
1262             dx20             = _mm_sub_pd(ix2,jx0);
1263             dy20             = _mm_sub_pd(iy2,jy0);
1264             dz20             = _mm_sub_pd(iz2,jz0);
1265             dx21             = _mm_sub_pd(ix2,jx1);
1266             dy21             = _mm_sub_pd(iy2,jy1);
1267             dz21             = _mm_sub_pd(iz2,jz1);
1268             dx22             = _mm_sub_pd(ix2,jx2);
1269             dy22             = _mm_sub_pd(iy2,jy2);
1270             dz22             = _mm_sub_pd(iz2,jz2);
1271
1272             /* Calculate squared distance and things based on it */
1273             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1274             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1275             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1276             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1277             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1278             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1279             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1280             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1281             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1282
1283             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1284             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1285             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1286             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1287             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1288             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1289             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1290             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1291             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1292
1293             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1294             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1295             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1296             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1297             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1298             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1299             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1300             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1301             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1302
1303             fjx0             = _mm_setzero_pd();
1304             fjy0             = _mm_setzero_pd();
1305             fjz0             = _mm_setzero_pd();
1306             fjx1             = _mm_setzero_pd();
1307             fjy1             = _mm_setzero_pd();
1308             fjz1             = _mm_setzero_pd();
1309             fjx2             = _mm_setzero_pd();
1310             fjy2             = _mm_setzero_pd();
1311             fjz2             = _mm_setzero_pd();
1312
1313             /**************************
1314              * CALCULATE INTERACTIONS *
1315              **************************/
1316
1317             /* REACTION-FIELD ELECTROSTATICS */
1318             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1319
1320             /* LENNARD-JONES DISPERSION/REPULSION */
1321
1322             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1323             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1324
1325             fscal            = _mm_add_pd(felec,fvdw);
1326
1327             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1328
1329             /* Update vectorial force */
1330             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1331             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1332             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1333
1334             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1335             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1336             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1337
1338             /**************************
1339              * CALCULATE INTERACTIONS *
1340              **************************/
1341
1342             /* REACTION-FIELD ELECTROSTATICS */
1343             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1344
1345             fscal            = felec;
1346
1347             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1348
1349             /* Update vectorial force */
1350             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1351             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1352             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1353
1354             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1355             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1356             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1357
1358             /**************************
1359              * CALCULATE INTERACTIONS *
1360              **************************/
1361
1362             /* REACTION-FIELD ELECTROSTATICS */
1363             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1364
1365             fscal            = felec;
1366
1367             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1368
1369             /* Update vectorial force */
1370             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1371             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1372             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1373
1374             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1375             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1376             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1377
1378             /**************************
1379              * CALCULATE INTERACTIONS *
1380              **************************/
1381
1382             /* REACTION-FIELD ELECTROSTATICS */
1383             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1384
1385             fscal            = felec;
1386
1387             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1388
1389             /* Update vectorial force */
1390             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1391             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1392             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1393
1394             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1395             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1396             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1397
1398             /**************************
1399              * CALCULATE INTERACTIONS *
1400              **************************/
1401
1402             /* REACTION-FIELD ELECTROSTATICS */
1403             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1404
1405             fscal            = felec;
1406
1407             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1408
1409             /* Update vectorial force */
1410             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1411             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1412             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1413
1414             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1415             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1416             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             /* REACTION-FIELD ELECTROSTATICS */
1423             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1424
1425             fscal            = felec;
1426
1427             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1428
1429             /* Update vectorial force */
1430             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1431             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1432             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1433
1434             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1435             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1436             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             /* REACTION-FIELD ELECTROSTATICS */
1443             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1444
1445             fscal            = felec;
1446
1447             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1448
1449             /* Update vectorial force */
1450             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1451             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1452             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1453
1454             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1455             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1456             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1457
1458             /**************************
1459              * CALCULATE INTERACTIONS *
1460              **************************/
1461
1462             /* REACTION-FIELD ELECTROSTATICS */
1463             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1464
1465             fscal            = felec;
1466
1467             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1468
1469             /* Update vectorial force */
1470             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1471             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1472             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1473
1474             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1475             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1476             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1477
1478             /**************************
1479              * CALCULATE INTERACTIONS *
1480              **************************/
1481
1482             /* REACTION-FIELD ELECTROSTATICS */
1483             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1484
1485             fscal            = felec;
1486
1487             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1488
1489             /* Update vectorial force */
1490             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1491             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1492             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1493
1494             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1495             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1496             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1497
1498             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1499
1500             /* Inner loop uses 277 flops */
1501         }
1502
1503         /* End of innermost loop */
1504
1505         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1506                                               f+i_coord_offset,fshift+i_shift_offset);
1507
1508         /* Increment number of inner iterations */
1509         inneriter                  += j_index_end - j_index_start;
1510
1511         /* Outer loop uses 18 flops */
1512     }
1513
1514     /* Increment number of outer iterations */
1515     outeriter        += nri;
1516
1517     /* Update outer/inner flops */
1518
1519     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1520 }